All states were surrounded by nonstate peoples, but owing to their dispersal, we know precious little about their coming and going, their shifting relationship to states, and their political structures. When a city is burned to the ground, it is often hard to tell whether it was an accidental

fire such as plagued all ancient cities built of combustible ma- terials, a civil war or uprising, or a raid from outside.

The author states that state funded archeology often leads to bias in the work.

The author explains homo sapiens have inhabited the earth for 200,000 years in and around Africa first. The first tax collecting society was around 3,100 BCE 4 millennia after the first crop domestication society's.

Its interesting how only the last 5 percent of our history as a species on this planet agrarian society's have been around.

"latest five percent of our history" "Civilization" is fairly recent and has existed during a rapid change in development and living.

The author argues that much of early human history, especially aspects that don’t support the idea of states as “civilizing forces”, is missing from traditional historical records. Events like migration, rebellion, or resistance to state control often left few physical traces or were deliberately ignored in written sources.

By comparing the fossil fuel era to the entire span of human history, the author shows how recent industrialization really is. Even though it was a short period, it changed societies and ecosystems greatly.

Scott describes how many indigenous and nomadic peoples in North America responded to colonization and state expansion. Some groups were forcibly confined to reservations after military defeat.

Scott challenges the assumption that Human Beings naturally progressed toward agriculture, cities, and states. He argues for the opposite, for most of human history, people chose not to live in concentrated and hierarchical societies.

real price in year Y = nominal price in year X * (CPI of year Y / CPI of year X)

例如，考虑公式 one x, y : A | x->y in r 和 one x : A | one y : A | x in r ，并设 r = {(A0,A0),(A0,A1),(A1,A0)} 。前一个公式为假，因为在 r 中有三个元组。然而，后一个公式为真，因为在 r 中恰好有一个 x （ A1 ），对于它恰好有一个 y （ A0 ）。 one x,y : A | P --- 将x,y看做整体。只存在唯一的一对 (x, y) 组合，使得这个组合出现在我们的配对名单 r 上。 one x : A | one y : A | x->y in r -- 剥洋葱。这是对应x的判定，是否存在唯一一个x，使得存在唯一一个y，使这个组合出现在我们的配对名单 r 上。

请注意， disj 关键字仅适用于对同一集合进行量化的变量组。这意味着在公式 all disj x, y : A , w, z : B | P 中， w 和 z 可以取相同的值（但 x 和 y 不行）

Función Clave de la Comunicación: Su rol principal es permitir que el plan estratégico se ejecute eficientemente en todos los niveles jerárquicos, optimizando tiempo, recursos (humanos y económicos) y esfuerzo.

Necesidad de Planificación: La comunicación no debe ser un proceso improvisado. El artículo subraya que debe ser estructurada, ordenada y planificada desde la alta dirección, e integrada en el plan estratégico general de la compañía.

Impacto en el Clima Organizacional y la Productividad: Se establece una relación directa entre una comunicación eficiente y un buen clima organizacional. Una comunicación deficiente, incluso en empresas rentables, genera alta rotación de personal, desmotivación y, en última instancia, una disminución de la productividad.

La Comunicación como Pilar Estratégico: El artículo posiciona la comunicación no solo como una herramienta operativa, sino como un recurso estratégico indispensable para la gestión empresarial. Es fundamental para alinear a toda la organización con la visión, misión y objetivos establecidos.

La Comunicación como Pilar Estratégico: El artículo posiciona la comunicación no solo como una herramienta operativa, sino como un recurso estratégico indispensable para la gestión empresarial. Es fundamental para alinear a toda la organización con la visión, misión y objetivos establecidos.

considero que este subtema refleja como una comunicación interna efectiva mejora los procesos operativos al igual que fortalece el sentido de pertenencia y confianza entre los miembros de una organización.

Coincido plenamente con la observación de que la globalización y la competencia obligan a las empresas a diferenciarse mediante estrategias comunicacionales efectivas. Desde mi punto de vista, esta sección destaca un punto clave: la comunicación no solo influye en el marketing o la imagen corporativa, sino también en la eficiencia operativa interna, ya que una información mal transmitida puede generar retrasos o duplicación de esfuerzos. En este sentido, la gestión comunicacional actúa como un sistema nervioso de la empresa, permitiendo la coordinación entre sus distintas áreas.

Las conclusiones resumen con claridad la relevancia de adoptar sistemas de comunicación multidireccionales y aprovechar los canales digitales. Personalmente, creo que este punto conecta con el desafío actual de muchas empresas: mantener la cohesión humana en entornos tecnológicos. Es decir, la digitalización no debe sustituir el contacto personal, sino fortalecerlo a través de la inmediatez y la transparencia. Por tanto, la comunicación empresarial en este siglo XXI debe equilibrar lo tecnológico con lo humano para mantener su efectividad.

El argumento de que una comunicación eficiente mejora el clima laboral y promueve la confianza es totalmente válido. Considero que este punto debería verse también desde la óptica de la inteligencia emocional organizacional: cuando los líderes practican una comunicación empática, abierta y transparente, los equipos se sienten valorados y esto impacta directamente en la productividad. En otras palabras, la comunicación no solo transmite datos, sino que construye relaciones humanas dentro de la empresa.

Me parece muy acertada la relación que los autores establecen entre la deficiente comunicación y el deterioro del clima laboral. Un entorno donde los mensajes no fluyen o son ambiguos tiende a generar desmotivación, frustración y alta rotación de empleados. En mi experiencia, la comunicación organizacional no solo busca informar, sino también dar sentido al trabajo. Cuando el personal entiende el propósito de sus tareas y percibe coherencia entre lo que se dice y lo que se hace, aumenta su compromiso y satisfacción. Esto convierte la comunicación en un pilar de la cultura organizacional.

Las preguntas omitidas en un cuestiona- rio son tan importantes como las preguntas que sí se hacen (Westmarland, 2001).

Esto es muy cierto lo que no se pregunta también importa, porque deja fuera parte de la realidad y no se logra un análisis realmente crítico.

De los muchos mitos que existen respecto a México y su sociedad, quizá dos de los más perversos son que el racismo no existe y que somos un país de puertas abiertas frente a la inmigración extranjera.

Estoy de acuerdo es un mito que no existe el racismo, aunque muchas personas creen esto o están escépticas a la idea , es importante verlo de manera mas critica ya que en México sí hay racismo y discriminación.

La falta de datos sobre las mujeres migrantes deja fuera sus necesidades e ignora que el género interactúa con variables como la edad, el origen y la condición social. Hay una ausencia de preguntas en las encuestas sobre las diferencias que podrían existir entre los dos géneros. Esto solo refleja la complejidad del tema y que los programas de protección no están especializados para atender adecuadamente a las mujeres y sus distintas situaciones de vulnerabilidad.

La ceguera de género nos habla de toda la falta de consideración e información sobre migración y discriminación dependiendo del sexo, lo que provoca la invisibilización de las experiencias femeninas en estos temas. Se asume al instante al migrante como masculino; esto limita la comprensión total del fenómeno, creando un problema donde no se incluyen las variables de género ni se reconocen como algo relevante.

El texto menciona que se asume que todos los migrantes son únicamente del sexo masculino, por lo que se refuerza la ignorancia hacia las experiencias femeninas. Esta escasa perspectiva impide que se reconozcan en las investigaciones particularidades de la migración femenina como los distintos riesgos y las estrategias de supervivencia específicas de cada mujer. En consecuencia, las mujeres migrantes no reciben la suficiente protección.

La nada no sería nada, no puede ser un estado porque la nada no tiene atributos. Esa parte del argumento no tiene sentido, no merece ser abordada. No hay posibilidad de que la nada exista porque su existencia implicaría que contiene el atributo del ser.

Leibniz se da un tiro en el pie. Si Dios es necesario pero también es perfecto, el mundo como creación suya no pudo ser de otra manera y no pudo no existir porque Dios no pudo no haberlo creado ya que su decisión de crear el mundo es perfecta y no hay otra decisión posible derivada de su perfección, por lo tanto el mundo es necesario, no contingente ya que es una consecuencia necesaria de un ser necesario por lo tanto ambos existen necesariamente. Además, si no pudo haber momento ni instancia en la que el mundo no existiera porque dios no pudo haber permanecido en un estado de imperfección (ya que crear el mundo y coexistir con el es perfecto, entonces su contrario es imperfecto) el mundo tiene que existir desde siempre con Dios mismo por lo que no comenzó a existir.

Heidegger se apropia del concepto de "nada" para dar su explicación de la experiencia mental humana de imaginar la nada y sus consecuencias, derivando en un aprecio profundo por el ser. Parlotea sobre el concepto de nada deformandolo dificultando la comprensión de su idea. El lector de por sí siempre da su toque de deformación de la idea, pero elegir el parloteo por sobre la expresión explícita añade una capa inecesaria mas sobre la interpretación de la idea.

working towards a sweet-spot

achive what's above the threshold that is required

working towards a sweet-spot

achive what's above the threshold that is required

https://hyp.is/95kSDrJPEfCvcauZZHECLg/www.futuresplatform.com/blog/s-curve-analysis-foresight

working towards a sweet-spot

REALize what's just above the threshold that is required

groan zone of perseverance where progress appears to be stalled for a long long time

working towards a sweet-spot

achive what's above the threshold that is required

from

Test

DOI: 10.1038/s44385-025-00040-y

Resource: University of Florida Molecular Pathology Core (RRID:SCR_016601)

Curator: @scibot

SciCrunch record: RRID:SCR_016601

What is this?

Traduce en tu mente el siguiente extracto al español sin mirar el texto y luego comprueba si lo has hecho correctamente o si existen otras alternativas de vocabulario:

The advantages of sharing housing are extensive: the most prominent is the savings you have by paying only for a room and not for the whole house. By sharing expenses, there is the possibility of having a larger space for a lower price. As if that were not enough, following the roommate guide that you can see in the video, it is most likely that no misunderstandings will be generated between the cohabitants and you can even make plans together.

Although there are also disadvantages, the list of advantages can be longer than the list of problems. It is up to the individual to decide which lifestyle he or she prefers to lead.

Sin mirar el texto intenta traducir en tu mente esta parte al español y luego, mirando el texto, comprueba si lo has hecho correctamente:

Most apartment sharers are students or young people who have just started working. The fact that several people of the same age live together can sometimes lead to clashes of character.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review): 

This paper measures the positioning and diffusivity of RNaseE-mEos3.2 proteins in E. coli as a function of rifampicin treatment, compares RNaseE to other E. coli proteins, and measures the effect of changes in domain composition on this localization and motion. The straightforward study is thoroughly presented, including very good descriptions of the imaging parameters and the image analysis/modeling involved, which is good because the key impact of the work lies in presenting this clear methodology for determining the position and mobility of a series of proteins in living bacteria cells. 

Thank you for the nice summary and positive feedback on the descriptions and methodology. 

My key notes and concerns are listed below; the most important concerns are indicated with asterisks. 

(1) The very start of the abstract mentions that the domain composition of RNase E varies among species, which leads the reader to believe that the modifications made to E. coli RNase E would be to swap in the domains from other species, but the experiment is actually to swap in domains from other E. coli proteins. The impact of this work would be increased by examining, for instance, RNase E domains from B. subtilis and C. crescentus as mentioned in the introduction. 

Thank you for the suggestions. We agree that the sentence may convey an unintended expectation. Our original intention was to note the presence and absence of certain domains of RNase E (e.g. membrane-binding motif and CTD) vary across species, rather than the actual sequence variations. To avoid any misinterpretation, we decided to remove the sentence from the abstract. Using the domains of B. subtilis and C. crescentus RNase E in E. coli is a very interesting suggestion, but we will leave that for a future study. 

(2) Furthermore, the introduction ends by suggesting that this work will modulate the localization, diffusion, and activity of RNase E for "various applications", but no applications are discussed in the discussion or conclusion. The impact of this work would be increased by actually indicating potential reasons why one would want to modulate the activity of RNase E. 

Thank you for this suggestion. For example, an E. coli strain expressing membranebound RNase E without CTD can help stabilize mRNAs and enhance protein expression. In fact, this idea was used in a commercial BL21 cell line (Invitrogen’s One Shot BL21 Star), to increase the yield of protein expression. We also think that environmentally modulated MB% of RNase E can be useful for controlling the mRNA half-lives and protein expression levels in different conditions. We discussed these ideas at the end of the Discussion.

(3) Lines 114 - 115: "The xNorm histogram of RNase E shows two peaks corresponding to each side edge of the membrane": "side edge" is not a helpful term. I suggest instead: "...corresponding to the membrane at each side of the cell" 

Thank you. We made the suggested change.

(4) A key concern of this reviewer is that, since membrane-bound proteins diffuse more slowly than cytoplasmic proteins, some significant undercounting of the % of cytoplasmic proteins is expected due to decreased detectability of the faster-moving proteins. This would not be a problem for the LacZ imaging where essentially all proteins are cytoplasmic, but would significantly affect the reported MB% for the intermediate protein constructs. How is this undercounting considered and taken into account? One could, for instance, compare LacZ vs. LacY (or RNase E) copy numbers detected in fixed cells to those detected in living cells to estimate it.  

Thank you for raising this point and suggesting a possible way to address this. We compared the number of tracks for mEos3.2-fused proteins in live vs fixed cells and tested the undercounting effect of cytoplasmic molecules. We compared WT RNase E molecules in live and fixed cells and found that there are about 50% lower molecules detected in the fixed cells, which agrees with the expectation that fluorescent proteins lose their signal upon fixation. Similarly, cytoplasmic RNase E (RNase E ΔMTS) copy number was also ~50% less in the fixed cells compared to live cells. If cytoplasmic molecules were undercounted compared membrane-bound molecules in live cells, fixation would reduce the copy number less than 50%. The comparable ratio of 50% indicates that the undercounting issue is not significant. This control analysis is provided in Figure S1B-C, and we made corresponding textual change in the result section as below:

For this analysis, we first confirmed that proteins localized on the membrane and in the cytoplasm are detected with equal probability, despite differences in their mobilities (Fig. S1B-C). 

(5) The rifampicin treatment study is not presented well. Firstly, it is found that LacY diffuses more rapidly upon rifampicin treatment. This change is attributed to changes in crowding at the membrane due to mRNA. Several other things change in cells after adding rif, including ATP levels, and these factors should be considered. More importantly, since the change in the diffusivity of RNaseE is similar to the change in diffusivity of LacY, then it seems that most of the change in RNaseE diffusion is NOT due to RNaseE-mRNAribosome binding, but rather due to whatever crowding/viscosity effects are experienced by LacY (along these lines: the error reported for D is SEM, but really should be a confidence interval, as in Figure 1, to give the reader a better sense of how different (or similar) 1.47 and 1.25 are). 

We agree with the reviewer that upon rifampicin treatment, RNase E’s D increases to a similar extent as that of LacY. Hence, the increase likely arises from a factor common to both proteins. We have added the reviewer’s suggested interpretation as a possible explanation in the manuscript as below. 

The similar fold change in D_RNE and D_LacY upon rif treatment suggests that the change in RNE diffusion may largely be attributed to physical changes in the intracellular environment (such as reduced viscosity or macromolecular crowding[41,42]), rather than a loss of RNA-RNE interactions.

As requested by the reviewer, we have provided confidence intervals for our D values in Table S8. Because these intervals are very narrow, we chose to present the SEM as the error metric for D and have also reported the corresponding errors for the fold-change values whenever we describe the fold differences between D values. 

(6) Lines 185-189: it is surprising to me that the CTD mutants both have the same change in D (5.5x and 5.3x) relative to their full-length counterparts since D for the membranebound WT protein should be much less sensitive to protein size than D for the cytoplasmic MTS mutant. Can the authors comment? 

Perhaps the reviewer understood that these differences are the ratios between +/-CTD (e.g. WT RNE vs ΔCTD). However, the differences we mentioned were from membrane-bound vs cytoplasmic versions of RNase E with comparable sizes (e.g. WT RNase E vs RNase E ΔMTS). We modified text and added a summary sentence at the end of the paragraph to clarify the point.

We found that D_ΔMTS is ~5.5 times that of D_RNE (Fig. 3B). [...] Together, these results suggest that the membrane binding reduces RNE mobility by a factor of 5.

That being said, we also realized a similar fold difference between +/-CTD. Specifically, WT RNE vs RNE ΔCTD (both membrane-bound) show a ~4.1-fold difference and RNE ΔMTS vs RNE ΔMTS ΔCTD (both cytoplasmic) show ~3.9-fold difference. We do not currently do not have a clear explanation for this pattern. Given that these two pairs have a similar change in mass, we speculate that the relationship between D and molecular mass may be comparable for membrane-bound and free-floating RNE variants. 

(7) Lines 190-194. Again, the confidence intervals and experimental uncertainties should be considered before drawing biological conclusions. It would seem that there is "no significant change" in the rhlB and pnp mutants, and I would avoid saying "especially for ∆pnp" when the same conclusion is true for both (one shouldn't say 1.04 is "very minute" and 1.08 is just kind of small - they are pretty much the same within experiments like this). 

Thank you for raising this point, which we fully agree with. That being said, we decided to remove results related to the degradosome proteins to improve the flow of the paper. We are preparing another paper related to the RNA degradosome complex formation. 

(8) Lines 221-223 " This is remarkable because their molecular masses (and thus size) are expected to be larger than that of MTS" should be reconsidered: diffusion in a membrane does not follow the Einstein law (indeed lines 223-225 agree with me and disagree with lines 221-223). (Also the discussion paragraph starting at line 375). Rather, it is generally limited by the interactions with the transmembrane segments with the membrane. So Figure 3D does not contain the right data for a comparison, and what is surprising to me is that MTS doesn't diffuse considerably faster than LacY2. 

We agree with the reviewer’s point that diffusion in a membrane does not follow the Stokes-Einstein law. That is why we introduced Saffman’s model. However, even in this model, proteins of larger size (or mass) should be slower than smaller size (a reason why we presented Figure 3D, now 4D). In other words, both Einstein and Saffman models predict that larger particles diffuse slower, although the exact scaling relationship differs between two models. Here, we assume that mass is related to the size. Contrary to Saffman’s model for membrane proteins, LacY2 diffuses faster than MTS despite of large size. Using MD simulations, we showed that this discrepancy can be explained by different interaction energies as the reviewer mentioned. This analysis further demonstrates that the size is not the only factor to consider protein diffusion in the membrane. We edited the paragraph to clarify the expectations and our interpretations.

According to the Stokes-Einstein relation for diffusion in simple fluids[49] and the Saffman-Delbruck diffusion model for membrane proteins, D decreases as particle size increases, albeit with different scaling behaviors. […] Thus, if size (or mass) were the primary determinant of diffusion, LacY2 and LacY6 would diffuse more slowly than the smaller MTS. The observed discrepancy instead implies that D may be governed by how each motif interacts with the membrane. For example, the way that TM domains are anchored to the membrane may facilitate faster lateral diffusion with surrounding lipids. 

(9) The logical connection between the membrane-association discussion (which seems to ignore associations with other proteins in the cell) and the preceding +/- rifampicin discussion (which seeks to attribute very small changes to mRNA association) is confusing.

Thank you for raising this point. We re-arranged the second result section to present diffusion due to membrane binding first before rifampicin. Furthermore, we stated our hypothesis and expectations in the beginning of the results section. This addition will legitimate our logic flow.

(10) Separately, the manuscript should be read through again for grammar and usage. For instance, the title should be: "Single-molecule imaging reveals the *roles* of *the* membrane-binding motif and *the* C-terminal domain of RNase E in its localization and diffusion in Escherichia coli". Also, some writing is unwieldy, for instance, "RNase E's D" would be easier to read if written as D_{RNaseE}. (underscore = subscript), and there is a lot of repetition in the sentence structures. 

Thank you for catching grammar mistakes. We went through extensive proofreading to avoid these mistakes and also used simple notation suggested by the reviewer, such as D_RNE, to make it easier to read. Thank you again for your suggestions.

Reviewer #2 (Public review): 

Summary: 

Troyer and colleagues have studied the in vivo localisation and mobility of the E.coli RNaseE (a protein key for mRNA degradation in all bacteria) as well as the impact of two key protein segments (MTS and CTD) on RNase E cellular localisation and mobility. Such sequences are important to study since there is significant sequence diversity within bacteria, as well as a lack of clarity about their functional effects. Using single-molecule tracking in living bacteria, the authors confirmed that >90% of RNaseE localised on the membrane, and measured its diffusion coefficient. Via a series of mutants, they also showed that MTS leads to stronger membrane association and slower diffusion compared to a transmembrane motif (despite the latter being more embedded in the membrane), and that the CTD weakens membrane binding. The study also rationalised how the interplay of MTS and CTD modulate mRNA metabolism (and hence gene expression) in different cellular contexts. 

Strengths: 

The study uses powerful single-molecule tracking in living cells along with solid quantitative analysis, and provides direct measurements for the mobility and localisation of E.coli RNaseE, adding to information from complementary studies and other bacteria. The exploration of different membrane-binding motifs (both MTS and CTD) has novelty and provides insight on how sequence and membrane interactions can control function of protein-associated membranes and complexes. The methods and membrane-protein standards used contribute to the toolbox for molecular analysis in live bacteria. 

Thank you for the nice summary of our work and positive comments about the paper’s strengths.

Weaknesses: 

The Results sections can be structured better to present the main hypotheses to be tested. For example, since it is well known that RNase E is membrane-localised (via its MTS), one expects its mobility to be mainly controlled by the interaction with the membrane (rather than with other molecules, such as polysomes and the degradosome). The results indeed support this expectation - however, the manuscript in its current form does not lay down the dominant hypothesis early on (see second Results chapter), and instead considers the rifampicin-addition results as "surprising"; it will be best to outline the most likely hypotheses, and then discuss the results in that light. 

Thank you for this comment. We addressed this point by stating our main hypothesis from the beginning of the results section. We also agree with the reviewer that the membrane binding effect should be discussed first; hence, we re-arranged the result section. In the revised manuscript, we discuss the effect of membrane binding on diffusion first, followed by rif effects.

Similarly, the authors should first discuss the different modes of interaction for a peripheral anchor vs a transmembrane anchor, outline the state of knowledge and possibilities, and then discuss their result; in its current version, the ms considers the LacY2 and LacY6 faster diffusion compared to MTS "remarkable", but considering the very different mode of interaction, there is no clear expectation prior to the experiment. In the same section, it would be good to see how the MD simulations capture the motion of LacY6 and LacY12, since this will provide a set of results consistent with the experimental set. 

Thank you for pointing this out. In fact, there is little discussion in the literature about the different modes of interaction for a peripheral anchor vs a transmembrane anchor. To our knowledge, our work (experiments and MD simulations) is the first that directly compared the two to reveal that the peripheral anchor has higher interaction energy than the transmembrane anchor. We added a sentence “Despite the prevalence of peripheral membrane proteins, how they interact with the membrane and how this differs from TM proteins remain poorly understood”. Furthermore, we added the MD simulation result of LacY6 and LacY12 in Figure 4E-F.

The work will benefit from further exploration of the membrane-RNase E interactions; e.g., the effect of membrane composition is explored by just using two different growth media (which on its own is not a well-controlled setting), and no attempts to change the MTS itself were made. The manuscript will benefit from considering experiments that explore the diversity of RNaseE interactions in different species; for example, the authors may want to consider the possibility of using the membrane-localisation signals of functional homologs of RNaseE in different bacteria (e.g., B. subtilis). It would be good to look at the effect of CTD deletions in a similar context (i.e., in addition to the MTS substitution by LacY2 and LacY6). 

Thank you very much for this suggestion. During revision, we engineered point mutations in MTS and analyzed critical hydrophobic residues for membrane binding. We characterized MB% in both +/-CTD variants (Fig. 2 and Fig. S6) and their effect on lacZ mRNA degradation (Fig. 6). We will leave the use of membrane motif of B. subtilis RNase E for future study. 

The manuscript will benefit from further discussion of the unstructured nature of the CTD, especially since the RNase CTD is well known to form condensates in Caulobacter crescentus; it is unclear how the authors excluded any roles for RNaseE phase separation in the mobility of RNaseE in E.coli cells. 

Yes, we agree with the reviewer that the intrinsically disordered nature of the CTD might contribute to condensate formation. We explored this possibility using both epifluorescence microscopy (with a YFP fusion) and single-molecule imaging with cluster analysis (using an mEos3.2 fusion). Please see Figure S8. We did observe some weak de-clustering of RNase E upon CTD deletion. In the current study, we are unable to quantify the extent to which clustering contributes to the slow diffusion of RNase E. However, we speculate that the clustering may be linked to the low MB% of certain RNE mutants containing CTD, and we discussed this possibility in the Discussion.

[…] further supporting that the CTD decreases membrane association across RNE variants. We speculate that this effect may be related to the CTD’s role in promoting phase-separated ribonucleoprotein condensates, as observed in Caulobacter crescentus[19]. In E. coli, we also observed a modest increase in the clustering tendency of RNE compared to ΔCTD (Fig. S8). 

Some statements in the Discussion require support with example calculations or toning down substantially. Specifically, it is not clear how the authors conclude that RNaseE interacts with its substrate for a short time (and what this time may actually be); further, the speculation about the MTS "not being an efficient membrane-binding motif for diffusion" lacks adequate support as it stands. 

Thank you for these points. To elaborate our point on transient interaction between RNase E and RNA, we added a sentence “Specifically, if RNE interacts with mRNAs for ~20 ms or less, the slow-diffusing state would last shorter than the frame interval and remain undetected in our experiment.” Also, we added this sentence in the discussion.

One possible explanation is that RNA-bound RNE (and RNase Y) is short-lived compared to our frame interval (~20 ms), unlike other RNA-binding proteins related to transcription and translation, interacting with RNA for ~1 min for elongation [48].

Plus, we clarified the wording used in the second sentence that the reviewer pointed out as follows,

Lastly, the slow diffusion of the MTS in comparison to LacY2 and LacY6 suggests that MTS is less favorable for rapid lateral motion in the membrane. 

Reviewer #3 (Public review): 

Summary: 

The manuscript by Troyer et al quantitatively measured the membrane localization and diffusion of RNase E, an essential ribonuclease for mRNA turnover as well as tRNA and rRNA processing in bacteria cells. Using single-molecule tracking in live E. coli cells, the authors investigated the impact of membrane targeting sequence (MTS) and the Cterminal domain (CTD) on the membrane localization and diffusion of RNase E under various perturbations. Finally, the authors tried to correlate the membrane localization of RNase E to its function on co- and post-transcriptional mRNA decay using lacZ mRNA as a model. 

The major findings of the manuscripts include: 

(1) WT RNase E is mostly membrane localized via MTS, confirming previous results. The diffusion of RNase E is increased upon removal of MTS or CTD, and more significantly increased upon removal of both regions. 

(2) By tagging RNase E MTS and different lengths of LacY transmembrane domain (LacY2, LacY6, or LacY12) to mEos3.2, the results demonstrate that short LacY transmembrane sequence (LacY2 and LacY6) can increase the diffusion of mEos3.2 on the membrane compared to MTS, further supported by the molecular dynamics simulation. A similar trend was roughly observed in RNase E mutants with MTS switched to LacY transmembrane domains. 

(3) The removal of RNase E MTS significantly increases the co-transcriptional degradation of lacZ mRNA, but has minimal effect on the post-transcriptional degradation of lacZ mRNA. Removal of CTD of RNase E overall decreases the mRNA decay rates, suggesting the synergistic effect of CTD on RNase E activity. 

Strengths: 

(1) The manuscript is clearly written with very detailed method descriptions and analysis parameters. 

(2) The conclusions are mostly supported by the data and analysis. 

(3) Some of the main conclusions are interesting and important for understanding the cellular behavior and function of RNase E. 

Thank you for your thorough summary of our work and positive comments.

Weaknesses: 

(1) Some of the observations show inconsistent or context-dependent trends that make it hard to generalize certain conclusions. Those points are worth discussion at least. Examples include: 

(a) The authors conclude that MTS segment exhibits reduced MB% when succinate is used as a carbon source compared to glycerol, whereas LacY2 segment maintains 100% membrane localization, suggesting that MTS can lose membrane affinity in the former growth condition (Ln 341-342). However, the opposite case was observed for the WT RNase E and RNase E-LacY2-CTD, in which RNase E-LacY2-CTD showed reduced MB% in the succinate-containing M9 media compared to the WT RNase E (Ln 264-267). This opposite trend was not discussed. In the absence of CTD, would the media-dependent membrane localization be similar to the membrane localization sequence or to the fulllength RNase E? 

This is a great point. Thank you for pointing out the discrepancy in data. We think the weak membrane interaction of RNaseE-lacY2-CTD likely stems from the structure instability in the presence of the CTD. Our data shows that an RNase E variant with a cytoplasmic population under a normal growth condition exhibits a greater cytoplasmic fraction in a poor growth media. In contrast, RNaseE-MTS and RNaseE-LacY2 lacking the CTD both showed 100% MB% under both normal and poor growth conditions. These results are presented in Figure S6 and further discussed in the Discussion section.

The loss of MB% in LacY2-based RNE was observed only in the presence of the CTD (Fig. S6D), suggesting that the CTD negatively affects membrane binding of RNE, possibly by altering protein conformation. In fact, all ΔCTD RNE mutants we tested exhibited higher MB% than their CTD-containing counterparts (Fig. S6A-B). 

(b) When using mEos3.2 reporter only, LacY2 and LacY6 both increase the diffusion of mEos3.2 compared to MTS. However, when inserting the LacY transmembrane sequence into RNase E or RNase E without CTD, only the LacY2 increases the diffusion of RNase E. This should also be discussed. 

Thank you for raising this point. As the reviewer pointed out, as the membrane motifs, both LacY2 and LacY6 diffuse faster than the MTS, but when they are fused to RNE, only LacY2-based RNE diffuses faster than MTS-based RNE. We speculate that it is possibly due to a structural reason—having four (large) LacY6 in a tetrameric arrangement may cancel out the original fast-diffusing property of LacY6. We added this idea in the result section:

This result may be due to the high TM load (24 helices) created by four LacY6 anchors in the RNE tetramer. Although all constructs are tetrameric, the 24-helix load (LacY6), compared with 8 (LacY2) and 4 (MTS), likely enlarges the membrane-embedded footprint and increases drag, thereby changing the mobility advantages assessed as standalone membrane anchors.

(2) The authors interpret that in some cases the increase in the diffusion coefficient is related to the increase in the cytoplasm localization portion, such as for the LacY2 inserted RNase E with CTD, which is rational. However, the authors can directly measure the diffusion coefficient of the membrane and cytoplasm portion of RNase E by classifying the trajectories based on their localizations first, rather than just the ensemble calculation. 

Thank you for this suggestion. Currently, because of the 2D projection effect from imaging, we cannot clearly distinguish which individual tracks are from the cytoplasm or from the inner membrane based on the localization. Therefore, we are unable to assign individual tracks as membrane-bound or cytoplasmic. However, we can demonstrate that the xNorm data can be separated into two different spatial populations based on the diffusion coefficient. D. That is we can plot xNorm of slow tracks vs xNorm of fast tracks. This analysis showed that the slow tracks have LacY-like xNorm profiles while the fast tracks have LacZ-like xNorm profiles, also quantitatively supporting our MB% fitting results. We have added this analysis to Figure S2.

(3) The error bars of the diffusion coefficient and MB% are all SEM from bootstrapping, which are very small. I am wondering how much of the difference is simply due to a batch effect. Were the data mixed from multiple biological replicates? The number of biological replicates should also be reported. 

Thank you for raising this point. In the original manuscript, we reported the number of tracks analyzed and noted that all data was from at least three separate biological replicates (measurements were repeated at least three different days). Furthermore, in the revised manuscript, we have provided the number of cells imaged in Table S6. 

(4) Some figures lack p-values, such as Figures 4 and 5C-D. Also, adding p-values directly to the bar graphs will make it easier to read. 

Thank you for checking these details. We added p values in the graphs showing k_d1 and k_d2 (Table S7).

Reviewer #2 (Recommendations for the authors): 

Minor and technical points: 

(1) Clarity and flow will be improved if each section first highlights the objective for the experiments that are described (e.g., line 240). 

Thank you for the suggestion. We addressed this point by editing the beginning of each subsection in the Results. 

(2) Line 272 (and elsewhere)."1.33-times faster is wrong". The authors mean 33% faster (from 0.075 to 1, see Figure 4G), and not 133% faster. Needs fixing. 

Thanks for pointing this out. We changed this as well as other incidences where we talk about the fold difference. For example, this particular incidence was changed to:

Indeed, in the absence of the CTD, we found that the D of LacY2-based RNE was 1.33 ± 0.01 times as fast as the MTS-based RNE. 

(3) The authors need to consider the fitting of two species on their D population. e.g., how will a 93% - 7% split between diffusive species would have looked for the distribution in S4B? Note also the L1 profile in Fig S4C - while it is not hugely different from Figure S4B, the analysis gives a 41% amplitude for the fast-diffusing species. The 2-species analysis can also be used on some of the samples with much higher cytoplasmic components. Further, tracks that are in the more central region can be analysed to see whether the fast-diffusing species increase in amplitude. 

Thank you for this comment. The D histograms of L1 and RNase E show a dominant peak at around 0.015, but L1 has a residual population in the shoulder (note the difference between L1’s experimental data and D1 fit, a yellow line in now Figure S3B). This residual shoulder population is absent in the D histogram of RNase E. We also performed two-species analysis as suggested by the reviewer and provided the result in Figure S3C. The analysis shows that the two-population fit (black line) is very close to one one-population fit (yellow line). While we agree with the reviewer that subpopulation analysis is helpful for other proteins that show <90% MB% (>10% significant cytoplasmic population). we found it useful to divide xNorm histogram into two populations based on the diffusivity (rather than doing two-population fit to the D histogram, which does not have spatial information). This analysis, shown in Figure S2, supports our MB% fit results.

(4) The authors suggest that the sequestration of RNaseE to the membrane limits its interaction with cytoplasmic mRNAs, and may increase mRNA lifetime. While this is true and supported by the authors' preprint (Ref15), it will also be good to consider (and discuss) that highly-transcribed regions are in the nucleoid periphery (and thus close to the membrane) and that ribosomes/polysomes are likewise predominantly peripheral (coregulation of transcription/translation) and membrane proximal. 

This is an interesting point, which we appreciate very much. The lacZ gene, when induced, is shown to move to the nucleoid periphery (Yang et al. 2019, Nat Comm). Also, in our preprint (Ref 15), we engineered to have lacZ closer to the membrane, by translationally fusing it to lacY. However, the degradation rate of lacZ mRNA was not enhanced by the proximity to the membrane (for both k<Sub>d1</sub> and k_d2). For lacZ mRNA, we mainly see the change in k_d1 when RNE localization changes. We think it is due to the slow diffusion of the nascent mRNA (attached to the chromosome) and the slow diffusion of membrane-bound RNE, such that regardless of the location of the nascent mRNA, the degradation by the membrane-bound RNE is inefficient. Only when RNE is free diffusing in the cytoplasm, it seems to increase k_d1 (the decay of nascent mRNAs).

Reviewer #3 (Recommendations for the authors):

(1) It will increase the clarity of the manuscript if the authors can provide better nomenclatures for different constructs, such as for different membrane targeting sequences fused to mEos3.2, full-length RNase E, or CDT truncated RNaseE. 

Thank you for this suggestion. We agree that many constructions were discussed, and their naming can be confusing. To help with clarity, we have abbreviated RNase E as RNE throughout the text where appropriate. 

(2) Line 342, Figure S7D should be cited instead of S6D. 

Thank you for finding this error. We made a proper change in the revised manuscript.

Relaciona cada juego con su objetivo: 1. Gana quien consigue mayor puntuación al sumar el valor de las letras y/o palabras. 2. Gana quien saca las fichas fuera del tablero antes. 3. Gana quien lleva las fichas al centro del tablero. 4. Gana quien adivina el mayor número de palabras representadas.

Encuentra en este extracto un sinónimo de amenazar, adversario, misión, enganchar.

Preguntas de Comprensión

¿Qué experiencia tiene Alex Rawlings con el aprendizaje de idiomas?

Según el texto, ¿por qué algunas personas se desaniman al aprender un nuevo idioma?

¿Cuál es uno de los principales consejos de Rawlings para aprender un nuevo idioma?

¿Qué método recomienda Rawlings para comenzar a aprender un nuevo idioma?

¿Qué beneficios adicionales menciona el texto sobre hablar más de un idioma?

Preguntas de Reflexión ¿Por qué crees que es importante no desanimarse al empezar a aprender un nuevo idioma, sin importar la edad? ¿Qué métodos de aprendizaje de idiomas te parecen más efectivos y por qué?

Explica esta frase en tus propias palabras y di si estás de acuerdo o no y por qué.

Responde a las siguientes preguntas de vocabulario y comprensión y después contrasta tus respuestas con las de la clave que encontrarás más abajo.

1. Preguntas de vocabulario: Define las siguientes palabras según el contexto del texto: Gastronomía Platillo Cocción Mortero Marinada

2. Encuentra sinónimos en el texto para las siguientes palabras: Únicos Fama Tradicional

3. Preguntas de comprensión:

3.1 ¿Cuántos platillos de Latinoamérica están entre los más populares del mundo según TasteAtlas?

3.2 ¿Qué técnica de cocción se asocia con la barbacoa mexicana?

3.3 ¿Cuál es la base del mole mexicano?

3.4 Describe el método de preparación del churrasco brasileño.

3.5 ¿Qué ingredientes básicos se utilizan en el ceviche peruano?

3.6 Explica el origen del guacamole.

3.7 ¿Qué diferencia a las quesadillas de Ciudad de México respecto al resto del país?

3.8 ¿De qué están hechos los tamales y cómo se cocinan?

3.9 ¿En qué consisten los burritos y dónde surgieron?

3.10 ¿Qué ingredientes llevan los nachos tradicionales?

3.11 ¿Qué importancia tiene la tortilla en la gastronomía mexicana?

3.12 ¿Cómo se definen los tacos y cuál es su origen histórico según el texto?

Relaciona estas frases con los siguientes platos: quesadilla, tamal, burrito, tacos:

1. Este plato tiene su origen en las minas de plata del siglo dieciocho en México.
2. Este plato consiste en tortillas que se pueden doblar y rellenar de frijoles, guacamole y queso.
3. Este plato tradicionalmente se envuelve en hojas de maíz y se basa en una masa de maíz rellena.
4. Este plato consiste en tortillas de harina dobladas por la mutad con papas, frijoles o carne y queso fundido por dentro.

Encuentra en este extracto: un verbo de cambio y un verbo que significa ¨que tiene raices en¨

Encuentra un sinónimo de: trozos, destemplado, aderezar, adobar

Intenta responder a las siguientes preguntas de comprensión. Después de intentarlo, comprueba tus respuestas con las claves.

Encuentra sinónimos en el texto para las siguientes palabras: Impresionante Declarada Ruinas

Preguntas de comprensión:

¿Cuántos saltos tienen las Cataratas del Iguazú y cuál es su altura aproximada?

Describe la región de la Amazonia y menciona cuántos países abarca.

¿Cuál era la importancia de Chichén Itzá en la península de Yucatán?

¿Qué representan los moáis en la Isla de Pascua según la tradición?

¿Qué tipo de figuras componen las líneas de Nazca?

¿Cuántas islas forman el conjunto de las Islas Galápagos y qué tipo de especies albergan?

¿Dónde se encuentra el Salto Ángel y qué lo hace especial?

¿Qué simboliza el Cristo Redentor en Río de Janeiro y cuándo fue inaugurado?

Explica el significado de Machu Picchu y por qué es famoso.

Describe las ruinas de Teotihuacan y menciona algunas de sus estructuras más destacadas.

**Preguntas: **

1. Menciona alguna ventaja del teletrabajo para el trabajador y para la empresa explicándola con tus propias palabras.

2. Menciona alguna desventaja del teletrabajo para el trabajador y para la empresa con tus propias palabras.

3. ¿Cómo se pueden atenuar/suavizar muchas de las desventajas del teletrabajo? Menciona algunos ejemplos explicándolos en la medida de lo posible con tus propias palabras.

4. Según el texto, ¿cómo está impactando el teletrabajo en la economía de América Latina y el Caribe?. Describe algún ejemplo.

¿Cómo se pueden atenuar/suavizar muchas de las desventajas del teletrabajo? Menciona algunos ejemplos explicándolos en la medida de lo posible con tus propias palabras.

Según el texto, ¿cómo está impactando el teletrabajo en la economía de América Latina y el Caribe?. Describe algún ejemplo.

Menciona alguna desventaja del teletrabajo para el trabajador y para la empresa mencionada en el texto con tus propias palabras.

Menciona alguna ventaja del teletrabajo para el trabajador y para la empresa mencionada en el texto con tus propias palabras.

Preguntas de comprensión:

1. "Cambiar el rumbo, transformar la educación". ¿Cómo explicarías este lema en tus propias palabras?

2. ¿Verdadero o falso? Canadá es líder entre las potencias mundiales en destinar más dinero para la educación.

3. ¿Verdadero o falso? En Finlandia los padres intervienen con frecuencia en la toma de decisiones escolares.

4. ¿Verdadero o falso? En Hong Kong se pone énfasis en la memorización y en la creatividad.

¿Verdadero o falso? En Finlandia los padres intervienen con frecuencia en la toma de decisiones escolares.

¿Verdadero o falso? En Hong Kong se pone énfasis en la memorización y en la creatividad.

Preguntas de comprensión:

1. ¿En qué situación la autora cogió accidentalmente la mano de una viejecilla en Salzburgo?

2. ¿Por qué el grupo de seis personas tuvo problemas al pagar la cuenta en el Hard Rock Pekín?

3. ¿Qué decisión tomó la autora durante la excursión en bicicleta por las Salinas de Maras y cómo afectó esto al resto de la expedición?

4. ¿Qué botón apretó la autora por error en el baño de la estación de tren de Osaka y cuál fue la consecuencia?

Busca en el texto una expresión idiomática que significa lo siguiente: Acabar con una situación de indiferencia, desconfianza o tensión con otra persona, iniciando la conversación con ella y procurando crear un ambiente agradable.

Decimos que algo es un _ cuando pensamos que es muy bueno, algo genial, y decimos ¡Qué __! como reacción ante algo gracioso, muy oportuno o ingenioso.

Presta atención a los conectores que dan coherencia a este texto argumentativo (si bien, aunque, por ejemplo, por otro lado). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Presta atención a los conectores que dan coherencia a este texto argumentativo (si bien, además, en este caso). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Presta atención a los conectores que dan coherencia a este texto argumentativo (si bien, sin embargo, en primer lugar, en segundo lugar, en cambio, en resumen). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Presta atención a los conectores que dan coherencia a este texto argumentativo (aunque, ya que, si bien, por otro lado, además, dado que). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Presta atención a los conectores que dan coherencia a este texto argumentativo (por lo tanto, por lo general, por otro lado, entonces). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Presta atención a los conectores que dan coherencia a este texto argumentativo (debido a, por otro lado, ya que). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Aquí tienes diez ejemplos de breves textos argumentativos para practicar la expresión de la opinión crítica sobre cualquier tema haciendo uso de los marcadores discursivos que estamos aprendiendo. Fíjate en los ejemplos y, siguiendo los modelos, reacciona con tu opinión sobre los temas que más te interesen.

Presta atención a los conectores y adverbios que dan coherencia a este texto argumentativo (pero, de manera similar, además de, por lo general, por otro lado, en su mayoría, lamentablemente). ¿Cuál es tu opinión sobre el tema? Escribe un párrafo sobre el tema usando adverbios y conectores para estructurarlo.

Synthèse sur les Événements Traumatiques et le Secourisme en Santé Mentale

Résumé

Ce document de synthèse analyse en profondeur la nature des événements traumatiques, le trouble de stress post-traumatique (TSPT) et le rôle crucial des secouristes en santé mentale.

S'appuyant sur des expertises psychiatriques et des témoignages de terrain, il ressort que la compréhension du traumatisme repose sur la distinction fondamentale entre le stress aigu, une réaction adaptative normale, et le TSPT, un trouble chronique pouvant se manifester des mois après l'événement.

L'intervention d'un secouriste PSSM (Premiers Secours en Santé Mentale) est essentielle pour créer un climat de sécurité, d'écoute non-jugeante et de réassurance.

Les études de cas de Fabienne, assistante sociale, et de Laurine, une jeune femme aidant une amie, démontrent l'application pratique des compétences PSSM :

• la patience,
• la capacité à laisser du temps à la personne pour s'exprimer et
• l'importance de la mise en sécurité

sont des piliers de l'intervention.

Le secouriste agit comme un maillon essentiel, guidant la personne en souffrance vers des ressources professionnelles adaptées (thérapies comme l'EMDR, centres médico-psychologiques), tout en apprenant à gérer son propre stress et à reconnaître les limites de son rôle.

La formation PSSM est présentée comme une démarche citoyenne indispensable pour briser les tabous et outiller chacun à apporter un premier soutien efficace.

--------------------------------------------------------------------------------

I. Définition et Compréhension des Traumatismes Psychiques

A. L'Événement à Potentiel Traumatique

Un événement est considéré comme potentiellement traumatique lorsqu'il confronte une personne, directement ou indirectement, à la mort, à une menace de mort, à la peur de mourir, à de graves blessures ou à une menace pour son intégrité physique ou celle d'un tiers.

• Caractéristiques : Il provoque généralement une détresse intense, un sentiment d'impuissance et d'horreur.

• Exemples : Actes de violence interpersonnelle (viols, agressions), accidents, catastrophes naturelles, guerres, attentats, actes de torture.

• Statistiques clés (selon l'OMS) :

◦ 70 % des personnes dans le monde vivent un événement potentiellement traumatisant au cours de leur vie.    ◦ Environ 4 % de ces personnes sont susceptibles de développer un TSPT.

B. Le Trouble de Stress Post-Traumatique (TSPT)

Le TSPT est un trouble qui peut survenir après un événement traumatique. Il se caractérise par des réactions intenses, désagréables et dysfonctionnelles qui persistent dans le temps.

• Symptômes principaux :

◦ Reviviscence : Flashbacks, cauchemars, intrusions sensorielles.  

◦ Évitement : Efforts pour éviter les lieux, personnes ou pensées liés au trauma.  

◦ Hypervigilance : État d'alerte constant, irritabilité, sursauts.   

◦ Altérations cognitives et de l'humeur : Ruminations, changements d'humeur rapides, sentiment de culpabilité ou de honte.

• Troubles associés : Le TSPT s'accompagne fréquemment de dérégulation émotionnelle, de troubles anxiodépressifs ou de troubles liés à l'usage de substances. Un rétablissement est souvent possible avec une prise en charge adaptée.

C. Distinction entre Stress Aigu et TSPT

Selon le Dr Jean-Michel de Lille, psychiatre, il est crucial de différencier ces deux états :

Caractéristique

Stress Aigu

Trouble de Stress Post-Traumatique (TSPT)

Nature

Réaction adaptative et normale face à un danger imminent.

Trouble chronique et pathologique.

Fonction

Mécanisme de défense (le corps se prépare à fuir ou combattre : cœur qui s'accélère, muscles irrigués).

Le système de stress reste activé longtemps après la disparition du danger.

Durée

Temporaire. La pression redescend en quelques jours ou semaines une fois le danger écarté.

Dure plusieurs mois, voire s'installe durablement. Peut survenir à distance de l'événement (parfois un an après).

Facteurs de risque pour le TSPT

Vulnérabilités génétiques, antécédents de traumatismes infantiles (négligence, abus) qui réactivent un stress plus ancien.

D. Les Portes d'Entrée du TSPT

Le Dr de Lille identifie trois principales manières de développer un TSPT, initialement décrites en psychiatrie militaire :

1. Être victime : Avoir subi directement la violence ou l'événement traumatique.

2. Être témoin : Le fait d'être un "témoin impuissant" est une porte d'entrée majeure.

Les taux de TSPT chez les survivants du Bataclan sont plus élevés chez les témoins que chez les victimes directement blessées.

Ce phénomène est aussi crucial chez les enfants témoins de violences conjugales, dont l'impact neuropsychique est décisif.

3. Être auteur : Plus rare, le fait d'avoir commis des actes de violence peut également générer des retours traumatiques.

II. Manifestations et Symptômes du TSPT

A. La Reviviscence Traumatique (Flashbacks et Cauchemars)

La reviviscence, ou "intrusion", est l'un des symptômes les plus douloureux du TSPT. La personne revit la scène traumatique de manière involontaire et intense.

• Flashbacks : La scène est revécue y compris physiquement (cœur qui s'accélère, panique) en l'absence de danger objectif.

Ils peuvent être déclenchés par des éléments sensoriels anodins (ex: croiser un homme avec des lunettes si l'agresseur portait des lunettes).

• Cauchemars : Le sommeil est un moment où l'évitement est impossible, les cauchemars ramènent la scène traumatique.

L'exemple est donné d'un jeune militaire hanté des années plus tard par les odeurs d'un charnier qu'il avait dû déterrer, le menant à abuser de morphine et d'alcool pour obtenir un "sommeil anesthétique".

B. Les Stratégies d'Évitement et l'Hypervigilance

En réaction à la douleur des intrusions, la personne développe des stratégies d'évitement.

• Comportements : Ne plus prendre les transports en commun, éviter le quartier de l'agression, éviter certaines dates (anniversaires).

• Conséquences : Dans les cas sévères, cela peut conduire à un isolement social complet, où toute interaction est perçue comme une menace potentielle.

• Hypervigilance anxieuse : La personne est constamment sur ses gardes, méfiante, a l'impression que le danger va resurgir, ce qui rend la vie "absolument invivable".

C. L'Impact Émotionnel et Comportemental

Le TSPT affecte profondément la sphère émotionnelle et les relations sociales.

• Culpabilité et Honte : Des sentiments de culpabilité ("pourquoi ai-je survécu et pas d'autres ?") ou de honte (particulièrement dans des métiers où l'expression des émotions est taboue) sont fréquents.

• Irritabilité et Isolement : Comme observé dans le cas de l'agent de la route, la personne peut devenir irritable, s'isoler de ses collègues et de sa famille.

III. Le Rôle du Secouriste en Santé Mentale (PSSM)

A. Principes Fondamentaux de l'Intervention

L'intervention d'un secouriste PSSM repose sur des principes clés pour aider une personne suite à un événement traumatique.

1. Créer la Sécurité : L'élément "absolument décisif" est de recréer un climat de sécurité et d'apaisement, de garantir que l'échange se déroule dans un espace sûr.

2. Adopter une Posture Non-Jugeante : Se positionner sur un terrain de parité, d'empathie et de soin, sans se présenter comme un expert.

3. Accueillir et Laisser le Temps : Laisser le temps à la personne d'exprimer ou de ne pas exprimer ses émotions, sans la presser.

La formation PSSM insiste sur cette "notion de temps" et de pauses.

B. Actions Clés face à une Personne en Crise

Face à une personne qui revit un événement, le secouriste peut :

• Rassurer sur la culpabilité : Aider la personne à comprendre qu'elle n'est pas coupable de ce qui s'est passé et qu'elle est une personne respectable.

• Écouter activement : Permettre à la personne de verbaliser ce qu'elle a vu, ressenti et comment elle se sent aujourd'hui.

• Réorienter en douceur vers le présent : Lors d'un flashback, introduire gentiment le doute pour aider la personne à faire la part des choses entre la réalité passée (l'agression) et la réalité présente (la sécurité actuelle).

Exemple : "Le monsieur avec des lunettes dans le tram n'est pas celui que vous avez vu il y a des années."

• Informer sur les ressources : Même si la personne n'est pas prête à parler, lui fournir des informations sur les professionnels et les lieux où elle peut trouver de l'aide (plateformes téléphoniques, médecin, etc.).

C. Connaître ses Limites et Orienter vers les Professionnels

Le secouriste PSSM n'est pas un professionnel de santé. Il est crucial de reconnaître les limites de son intervention.

Le rôle est d'être un pont, d'accompagner la personne vers une prise en charge adaptée et personnalisée par des professionnels qualifiés.

IV. Études de Cas : Applications Pratiques du Secourisme

A. Cas de Fabienne : L'Agent de la Route Traumatisé

Fabienne, assistante sociale et secouriste PSSM, est intervenue auprès d'un agent de la route mutique après avoir été confronté à une victime décédée.

Aspect

Description

Contexte

Un agent de la route est en retrait et isolé après avoir été témoin d'un accident mortel.

Le public est majoritairement masculin, peu habitué à parler des émotions, avec un sentiment de "honte" à exprimer sa fragilité.

Défi

L'agent reste silencieux lors des deux premières tentatives d'entretien. Il refuse de se livrer.

Approche PSSM

Fabienne fait preuve d'une grande patience. Elle laisse passer près d'un mois entre la première et la troisième rencontre.

Elle lui fournit des informations sur les ressources dès le deuxième entretien, anticipant un besoin urgent. Lors du troisième entretien (initié par l'agent), elle applique les principes PSSM :

elle prend le temps, accueille les émotions avec une "qualité d'écoute différente", et laisse des temps de pause importants.

Résultat

L'agent parvient enfin à s'exprimer, livrant son mal-être (cauchemars, irritabilité).

Fabienne l'oriente précisément vers un centre médico-psychologique spécialisé. Plus tard, l'agent la remercie pour son aide et sa confiance.

B. Cas de Laurine : L'Amie en Pleine Reviviscence

Laurine, secouriste PSSM, a aidé une amie proche, diagnostiquée TSPT suite à une enfance violente, lors d'un flashback.

Aspect

Description

Contexte

L'amie a une enfance marquée par une mère alcoolique et un père violent.

Elle se sent coupable de la mort de ses animaux de compagnie, survenue dans des circonstances troubles.

L'alcool est souvent un "facilitateur d'anxiété" pour elle.

Déclencheur

En sortie de boîte de nuit, la vue d'un chien déclenche une reviviscence intense.

Elle s'effondre en appelant le nom de son chien décédé, persuadée de l'avoir retrouvé.

Intervention

1. Mise en sécurité : Laurine l'isole d'un groupe d'hommes alcoolisés.

1. Écoute et validation : Elle ne la contredit pas frontalement ("je veux bien te croire qu'il ressemble beaucoup") mais tente de la ramener à la réalité.

2. Fermeté bienveillante : Face à l'insistance de son amie, elle reste ferme pour la ramener en lieu sûr ("moi en tant qu'amie je peux pas te laisser y retourner").

Réflexion et Résultat

Avec le recul, Laurine pense qu'elle aurait dû être "plus dans l'écoute" et moins dans la volonté de "raisonner".

Le lendemain, elle encourage son amie à en parler à sa psychologue.

L'amie aborde ce trauma en thérapie EMDR et comprend qu'elle a fait un transfert : son désir de sauver le chien était le reflet de son propre désir d'avoir été sauvée enfant.

V. Prise en Charge Professionnelle et Ressources

A. Thérapies Spécifiques pour le TSPT

Le Dr de Lille souligne des avancées majeures dans les thérapies du TSPT, notamment les thérapies d'exposition prolongée.

L'objectif est de permettre à la personne d'évoquer le souvenir traumatique sans subir les symptômes de panique, afin de "désamorcer le couplage" entre la mémoire et la souffrance physique.

• Méthodes psychothérapeutiques :

◦ EMDR (MDR en français) : Thérapie brève utilisant des mouvements oculaires répétitifs pour "digérer" le souvenir traumatique.    ◦ ICV (Intégration du Cycle de la Vie)

• Approches pharmacologiques :

◦ Méthode Brunet : Utilisation de bêta-bloquants (ex: Vlocardie) pour diminuer la réaction physique lors de l'évocation.   

◦ Expérimentations : Des recherches sont en cours avec de nouvelles molécules comme la MDMA.

B. Ressources et Soutien Disponibles

Le podcast met en avant plusieurs ressources pour les personnes concernées et les aidants :

• Numéros d'urgence : 112, 15, 18 et le 3114 (numéro national de prévention du suicide).

• PSSM France :

◦ Le site internet pour se former aux premiers secours en santé mentale.   

◦ Le "Carnet du secouriste en santé mentale : mieux aider un adulte suite à un événement traumatique" (téléchargement gratuit).

• Centre National de Ressources et de résilience (CN2R) : Chaîne YouTube et témoignages pour améliorer la prise en charge du psychotraumatisme.

• Podcast "Émotion" par Louis Média : L'épisode "Stress post-traumatique : Comment s'inscrit-il dans notre corps ?".

VI. Apports et Valeur de la Formation PSSM France

A. Un Cadre Sécurisant pour l'Aidant

La formation PSSM est décrite comme un outil essentiel qui fournit un "cadre sécurisant" à l'aidant.

Elle permet de savoir quoi faire, mais surtout "ce qu'il ne faut pas faire" et "ne pas dire".

• Elle incite à prendre de la hauteur et à requestionner ses propres pratiques et réflexes.

• Elle aide à gérer son propre stress et à se sentir moins affecté personnellement, ce qui est crucial pour pouvoir aider efficacement sans s'épuiser.

B. Témoignages sur l'Impact de la Formation

• Pour Fabienne : La formation a enrichi son approche, notamment sur "cette notion du temps" et l'importance de "laisser à l'autre la possibilité ou pas de dire".

• Pour Laurine : La formation lui a permis de trouver un moyen d'avoir du recul, de moins culpabiliser et de prendre conscience du réseau de ressources existant (elle ne connaissait pas PSSM avant).

C. Un Enjeu Citoyen pour Briser les Tabous

Le podcast conclut en soulignant que, tout comme les premiers secours physiques, les premiers secours en santé mentale sont une démarche citoyenne.

La formation permet d'acquérir des "clés toutes simples" qui peuvent aider de nombreuses personnes.

Elle encourage chacun à jouer un rôle pour briser les tabous autour des troubles psychiques et à "apprendre à aider".

Ve el vídeo y lee el texto sobre el papel de Samuel Melinao como coordinador de un centro de salud mapuche. ¿Puedes resumir los puntos principales que se mencionan en torno a la medicina mapuche y los problemas a los que se ha enfrentado?

Ve el vídeo y lee el texto sobre el papel de Samuel Melinao como coordinador de un centro de salud mapuche. ¿Puedes resumir los puntos principales que se mencionan en torno a la medicina mapuche y los problemas a los que se ha enfrentado?

Resume el contenido de este párrafo con tus propias palabras.

¿Por qué crees que Javier es muy criticado por su gente? ¿Puedes resumir los puntos principales que se mencionan sobre su caso en el texto y en el vídeo?*

¿Por qué existe este conflicto entre el pueblo mapuche y el Estado chileno?

Lee y escucha sobre el caso de Ana Millaleo y resume los puntos principales que se mencionan en torno a su faceta como cantante de hip hop su faceta investigadora.

Preguntas de Comprensión

¿Qué etiquetas asigna el artista Martin Vargic a algunos países de América Latina en su ilustración?

¿Cómo se perciben los argentinos según los estereotipos mencionados en el estudio de la Agencia Española de Cooperación Internacional?

¿Qué características positivas se mencionan sobre los brasileños en el estudio de 2006?

¿Cómo se describen los estereotipos sobre los bolivianos en el estudio mencionado?

¿Qué problemas menciona Elier Chara García sobre la falta de información en la propia región?

Preguntas de Reflexión ¿Qué impacto crees que tienen los estereotipos en la percepción de las personas de una región específica? ¿Cómo podríamos combatir la formación y perpetuación de estereotipos negativos entre países?

¿Estás de acuerdo con esta afirmación? ¿Por qué?

Explica el significado de la frase "no todos pueden ser puestos en el mismo saco" dentro del contexto y contrasta el ejemplo de Ismael con otro ejemplo propio.

¿Puedes explicar esta frase con tus propias palabras y añadir tu opinión al respecto?

¿Puedes explicar la ironía contenida en esta frase?

Author response:

The following is the authors’ response to the previous reviews.

To the Senior Editor and the Reviewing Editor:

We sincerely appreciate the valuable comments provided by the reviewers, the reviewing editor, and the senior editor. Based on our last response and revision, we are confused by the two limitations noted in the eLife assessment. 

(1) benchmarking against comparable methods is limited.

In our last revision, we added the comparison experiments with TNDM, as the reviewers requested. Additionally, it is crucial to emphasize that our evaluation of decoding capabilities of behaviorally relevant signals has been benchmarked against the performance of the ANN on raw signals, which, as Reviewer #1 previously noted, nearly represents the upper limit of performance. Consequently, we believe that our benchmarking methods are sufficiently strong.

(2) some observations may be a byproduct of their method, and may not constitute new scientific observations.

We believe that our experimental results are sufficient to demonstrate that our conclusions are not byproducts of d-VAE based on three reasons:

(1) The d-VAE, as a latent variable model, adheres to the population doctrine, which posits that latent variables are responsible for generating the activities of individual neurons. The goal of such models is to maximize the explanation of the raw signals. At the signal level, the only criterion we can rely on is neural reconstruction performance, in which we have achieved unparalleled results. Thus, it is inappropriate to focus on the mixing process during the model's inference stage while overlooking the crucial de-mixing process during the generation stage and dismissing the significance of our neural reconstruction results. For more details, please refer to the first point in our response to Q4 from Reviewer #4.

(2) The criterion that irrelevant signals should contain minimal information can effectively demonstrate that our conclusions are not by-products of d-VAE. Unfortunately, the reviewers seem to have overlooked this criterion. For more details, please refer to the third point in our response to Q4 from Reviewer #4

(3) Our synthetic experimental results also substantiate that our conclusions are not byproducts of d-VAE. However, it appears the reviewers did not give these results adequate consideration. For more details, please refer to the fourth point in our response to Q4 from Reviewer #4.

Furthermore, our work presents not just "a useful method" but a comprehensive framework. Our study proposes, for the first time, a framework for defining, extracting, and validating behaviorally relevant signals. In our current revision, to clearly distinguish between d-VAE and other methods, we have formalized the extraction of behaviorally relevant signals into a mathematical optimization problem. To our knowledge, current methods have not explicitly proposed extracting behaviorally relevant signals, nor have they identified and addressed the key challenges of extracting relevant signals. Similarly, existing research has not yet defined and validated behaviorally relevant signals. For more details, please refer to our response to Q1 from Reviewer #4.

Based on these considerations, we respectfully request that you reconsider the eLife assessment of our work. We greatly appreciate your time and attention to this matter.

The main revisions made to the manuscript are as follows:

(1) We have formalized the extraction of behaviorally relevant signals into a mathematical optimization problem, enabling a clearer distinction between d-VAE and other models.

(2) We have moderated the assertion about linear readout to highlight its conjectural nature and have broadened the discussion regarding this conclusion. 

(3) We have elaborated on the model details of d-VAE and have removed the identifiability claim.

To Reviewer #1

Q1: “As reviewer 3 also points out, I would, however, caution to interpret this as evidence for linear read-out of the motor system - your model performs a non-linear transformation, and while this is indeed linearly decodable, the motor system would need to do something similar first to achieve the same. In fact to me it seems to show the opposite, that behaviour-related information may not be generally accessible to linear decoders (including to down-stream brain areas).”

Thank you for your comments. It's important to note that the conclusions we draw are speculative and not definitive. We use terms like "suggest" to reflect this uncertainty. To further emphasize the conjectural nature of our conclusions, we have deliberately moderated our tone.

The question of whether behaviorally-relevant signals can be accessed by linear decoders or downstream brain regions hinges on the debate over whether the brain employs a strategy of filtering before decoding. If the brain employs such a strategy, the brain can probably access these signals. In our opinion, it is likely that the brain utilizes this strategy.

Given the existence of behaviorally relevant signals, it is reasonable to assume that the brain has intrinsic mechanisms to differentiate between relevant and irrelevant signals. There is growing evidence suggesting that the brain utilizes various mechanisms, such as attention and specialized filtering, to suppress irrelevant signals and enhance relevant signals [1-3]. Therefore, it is plausible that the brain filters before decoding, thereby effectively accessing behaviorally relevant signals.

Thank you for your valuable feedback.

(1) Sreenivasan, Sameet, and Ila Fiete. "Grid cells generate an analog error-correcting code for singularly precise neural computation." Nature neuroscience 14.10 (2011): 1330-1337.

(2) Schneider, David M., Janani Sundararajan, and Richard Mooney. "A cortical filter that learns to suppress the acoustic consequences of movement." Nature 561.7723 (2018): 391-395.

(3) Nakajima, Miho, L. Ian Schmitt, and Michael M. Halassa. "Prefrontal cortex regulates sensory filtering through a basal ganglia-to-thalamus pathway." Neuron 103.3 (2019): 445-458.

Q2: “As in my initial review, I would also caution against making strong claims about identifiability although this work and TNDM seem to show that in practise such methods work quite well. CEBRA, in contrast, offers some theoretical guarantees, but it is not a generative model, so would not allow the type of analysis done in this paper. In your model there is a para,eter \alpha to balance between neural and behaviour reconstruction. This seems very similar to TNDM and has to be optimised - if this is correct, then there is manual intervention required to identify a good model.”

Thank you for your comments. 

Considering your concerns about our identifiability claims and the fact that identifiability is not directly relevant to the core of our paper, we have removed content related to identifiability.

Firstly, our model is based on the pi-VAE, which also has theoretical guarantees. However, it is important to note that all such theoretical guarantees (including pi-VAE and CEBRA) are based on certain assumptions that cannot be validated as the true distribution of latent variables remains unknown.

Secondly, it is important to clarify that the identifiability of latent variables does not impact the conclusions of this paper, nor does this paper make specific conclusions about the model's latent variables. Identifiability means that distinct latent variables correspond to distinct observations. If multiple latent variables can generate the same observation, it becomes impossible to determine which one is correct given the observation, which leads to the issue of nonidentifiability. Notably, our analysis focuses on the generated signals, not the latent variables themselves, and thus the identifiability of these variables does not affect our findings. 

Our approach, dedicated to extracting these signals, distinctly differs from methods such as TNDM, which focuses on extracting behaviorally relevant latent dynamics. To clearly set apart d-VAE from other models, we have framed the extraction of behaviorally relevant signals as the following mathematical optimization problem:

where 𝑥# denotes generated behaviorally-relevant signals, 𝑥 denotes raw noisy signals, 𝐸(⋅,⋅) demotes reconstruction loss, and 𝑅(⋅) denotes regularization loss. It is important to note that while both d-VAE and TNDM employ reconstruction loss, relying solely on this term is insufficient for determining the optimal degree of similarity between the generated and raw noisy signals. The key to accurately extracting behaviorally relevant signals lies in leveraging prior knowledge about these signals to determine the optimal similarity degree, encapsulated by 𝑅(𝒙𝒓).  Other studies have not explicitly proposed extracting behaviorally-relevant signals, nor have they identified and addressed the key challenges involved in extracting relevant signals. Consequently, our approach is distinct from other methods.

Thank you for your valuable feedback.

Q3: “Somewhat related, I also found that the now comprehensive comparison with related models shows that the using decoding performance (R2) as a metric for model comparison may be problematic: the R2 values reported in Figure 2 (e.g. the MC_RTT dataset) should be compared to the values reported in the neural latent benchmark, which represent well-tuned models (e.g. AutoLFADS). The numbers (difficult to see, a table with numbers in the appendix would be useful, see: https://eval.ai/web/challenges/challenge-page/1256/leaderboard) seem lower than what can be obtained with models without latent space disentanglement. While this does not necessarily invalidate the conclusions drawn here, it shows that decoding performance can depend on a variety of model choices, and may not be ideal to discriminate between models. I'm also surprised by the low neural R2 for LFADS I assume this is condition-averaged) - LFADS tends to perform very well on this metric.”

Thank you for your comments. The dataset we utilized is not from the same day as the neural latent benchmark dataset. Notably, there is considerable variation in the length of trials within the RTT paradigm, and the dataset lacks explicit trial information, rendering trial-averaging unsuitable. Furthermore, behaviorally relevant signals are not static averages devoid of variability; even behavioral data exhibits variability. We computed the neural R2 using individual trials rather than condition-averaged responses. 

Thank you for your valuable feedback.

Q4: “One statement I still cannot follow is how the prior of the variational distribution is modelled. You say you depart from the usual Gaussian prior, but equation 7 seems to suggest there is a normal prior. Are the parameters of this distribution learned? As I pointed out earlier, I however suspect this may not matter much as you give the prior a very low weight. I also still am not sure how you generate a sample from the variational distribution, do you just draw one for each pass?”

Thank you for your questions.

The conditional distribution of prior latent variables 𝑝%(𝒛|𝒚) is a Gaussian distribution, but the distribution of prior latent variables 𝑝(𝒛) is a mixture Gaussian distribution. The distribution of prior latent variables 𝑝(𝒛) is:

where denotes the empirical distribution of behavioral variables

𝒚, and 𝑁 denotes the number of samples, 𝒚(𝒊) denotes the 𝒊th sample, δ(⋅) denotes the Dirac delta function, and 𝑝%(𝒛|𝒚) denotes the conditional distribution of prior latent variables given the behavioral variables parameterized by network 𝑚. Based on the above equation, we can see that 𝑝(𝒛) is not a Gaussian distribution, it is a Gaussian mixture model with 𝑁 components, which is theoretically a universal approximator of continuous probability densities.

Learning this prior is important, as illustrated by our latent variable visualizations, which are not a Gaussian distribution. Upon conducting hypothesis testing for both latent variables and behavioral variables, neither conforms to Gaussian distribution (Lilliefors test and Kolmogorov-Smirnov test). Consequently, imposing a constraint on the latent variables towards N(0,1) is expected to affect performance adversely.

Regarding sampling, during training process, we draw only one sample from the approximate posterior distribution . It is worth noting that drawing multiple samples or one sample for each pass does not affect the experimental results. After training, we can generate a sample from the prior by providing input behavioral data 𝒚(𝒊) and then generating corresponding samples via and . To extract behaviorally-relevant signals from raw signals, we use and .

Thank you for your valuable feedback.

Q5: “(1) I found the figures good and useful, but the text is, in places, not easy to follow. I think the manuscript could be shortened somewhat, and in some places more concise focussed explanations would improve readability.

(2) I would not call the encoding "complex non-linear" - non-linear is a clear term, but complex can mean many things (e.g. is a quadratic function complex?) ”

Thank you for your recommendation. We have revised the manuscript for enhanced clarity.  We call the encoding “complex nonlinear” because neurons encode information with varying degrees of nonlinearity, as illustrated in Fig. 3b, f, and Fig. S3b.

Thank you for your valuable feedback.

To Reviewer #2

Q1: “I still remain unconvinced that the core findings of the paper are "unexpected". In the response to my previous Specific Comment #1, they say "We use the term 'unexpected' due to the disparity between our findings and the prior understanding concerning neural encoding and decoding." However, they provide no citations or grounding for why they make those claims. What prior understanding makes it unexpected that encoding is more complex than decoding given the entropy, sparseness, and high dimensionality of neural signals (the "encoding") compared to the smoothness and low dimensionality of typical behavioural signals (the "decoding")?” 

Thank you for your comments. We believe that both the complexity of neural encoding and the simplicity of neural decoding in motor cortex are unexpected.

The Complexity of Neural Encoding: As noted in the Introduction, neurons with small R2 values were traditionally considered noise and consequently disregarded, as detailed in references [1-3]. However, after filtering out irrelevant signals, we discovered that these neurons actually contain substantial amounts of behavioral information, previously unrecognized. Similarly, in population-level analyses, neural signals composed of small principal components (PCs) are often dismissed as noise, with analyses typically utilizing only between 6 and 18 PCs [4-10]. Yet, the discarded PC signals nonlinearly encode significant amounts of information, with practically useful dimensions found to range between 30 and 40—far exceeding the usual number analyzed. These findings underscore the complexity of neural encoding and are unexpected.

The Simplicity of Neural Decoding: In the motor cortex, nonlinear decoding of raw signals has been shown to significantly outperform linear decoding, as evidenced in references [11,12]. Interestingly, after separating behaviorally relevant and irrelevant signals, we observed that the linear decoding performance of behaviorally relevant signals is nearly equivalent to that of nonlinear decoding—a phenomenon previously undocumented in the motor cortex. This discovery is also unexpected.

Thank you for your valuable feedback.

(1) Georgopoulos, Apostolos P., Andrew B. Schwartz, and Ronald E. Kettner. "Neuronal population coding of movement direction." Science 233.4771 (1986): 1416-1419.

(2) Hochberg, Leigh R., et al. "Reach and grasp by people with tetraplegia using a neurally controlled robotic arm." Nature 485.7398 (2012): 372-375. 

(3) Inoue, Yoh, et al. "Decoding arm speed during reaching." Nature communications 9.1 (2018): 5243.

(4) Churchland, Mark M., et al. "Neural population dynamics during reaching." Nature 487.7405 (2012): 51-56.

(5) Kaufman, Matthew T., et al. "Cortical activity in the null space: permitting preparation without movement." Nature neuroscience 17.3 (2014): 440-448.

(6) Elsayed, Gamaleldin F., et al. "Reorganization between preparatory and movement population responses in motor cortex." Nature communications 7.1 (2016): 13239.

(7) Sadtler, Patrick T., et al. "Neural constraints on learning." Nature 512.7515 (2014): 423426.

(8) Golub, Matthew D., et al. "Learning by neural reassociation." Nature neuroscience 21.4 (2018): 607-616.

(9) Gallego, Juan A., et al. "Cortical population activity within a preserved neural manifold underlies multiple motor behaviors." Nature communications 9.1 (2018): 4233.

(10) Gallego, Juan A., et al. "Long-term stability of cortical population dynamics underlying consistent behavior." Nature neuroscience 23.2 (2020): 260-270.

(11) Glaser, Joshua I., et al. "Machine learning for neural decoding." Eneuro 7.4 (2020).

(12) Willsey, Matthew S., et al. "Real-time brain-machine interface in non-human primates achieves high-velocity prosthetic finger movements using a shallow feedforward neural network decoder." Nature Communications 13.1 (2022): 6899.

Q2: “I still take issue with the premise that signals in the brain are "irrelevant" simply because they do not correlate with a fixed temporal lag with a particular behavioural feature handchosen by the experimenter. In the response to my previous review, the authors say "we employ terms like 'behaviorally-relevant' and 'behaviorally-irrelevant' only regarding behavioral variables of interest measured within a given task, such as arm kinematics during a motor control task.". This is just a restatement of their definition, not a response to my concern, and does not address my concern that the method requires a fixed temporal lag and continual decoding/encoding. My example of reward signals remains. There is a huge body of literature dating back to the 70s on the linear relationships between neural and activity and arm kinematics; in a sense, the authors have chosen the "variable of interest" that proves their point. This all ties back to the previous comment: this is mostly expected, not unexpected, when relating apparently-stochastic, discrete action potential events to smoothly varying limb kinematics.”

Thank you for your comments. 

Regarding the experimenter's specification of behavioral variables of interest, we followed common practice in existing studies [1, 2]. Regarding the use of fixed temporal lags, we followed the same practice as papers related to the dataset we use, which assume fixed temporal lags [3-5]. Furthermore, many studies in the motor cortex similarly use fixed temporal lags [68].

Concerning the issue of rewards, in the paper you mentioned [9], the impact of rewards occurs after the reaching phase. It's important to note that in our experiments, we analyze only the reaching phase, without any post-movement phase. 

If the impact of rewards can be stably reflected in the signals in the reaching phase of the subsequent trial, and if the reward-induced signals do not interfere with decoding—since these signals are harmless for decoding and beneficial for reconstruction—our model is likely to capture these signals. If the signals induced by rewards during the reaching phase are randomly unstable, our model will likely be unable to capture them.

If the goal is to extract post-movement neural activity from both rewarded and unrewarded trials, and if the neural patterns differ between these conditions, one could replace the d-VAE's regression loss, used for continuous kinematics decoding, with a classification loss tailored to distinguish between rewarded and unrewarded conditions.

To clarify the definition, we have revised it in the manuscript. Specifically, before a specific definition, we briefly introduce the relevant signals and irrelevant signals. Behaviorally irrelevant signals refer to those not directly associated with the behavioral variables of interest and may include noise or signals from variables of no interest. In contrast, behaviorally relevant signals refer to those directly related to the behavioral variables of interest. For instance, rewards in the post-movement phase are not directly related to behavioral variables (kinematics) in the reaching movement phase.

It is important to note that our definition of behaviorally relevant signals not only includes decoding capabilities but also specific requirement at the signal level, based on two key requirements:

(1) they should closely resemble raw signals to preserve the underlying neuronal properties without becoming so similar that they include irrelevant signals. (encoding requirement), and  (2) they should contain behavioral information as much as possible (decoding requirement). Signals that meet both requirements are considered effective behaviorally relevant signals. In our study, we assume raw signals are additively composed of behaviorally-relevant and irrelevant signals. We define irrelevant signals as those remaining after subtracting relevant signals from raw signals. Therefore, we believe our definition is clearly articulated. 

Thank you for your valuable feedback.

(1) Sani, Omid G., et al. "Modeling behaviorally relevant neural dynamics enabled by preferential subspace identification." Nature Neuroscience 24.1 (2021): 140-149.

(2) Buetfering, Christina, et al. "Behaviorally relevant decision coding in primary somatosensory cortex neurons." Nature neuroscience 25.9 (2022): 1225-1236.

(3) Wang, Fang, et al. "Quantized attention-gated kernel reinforcement learning for brain– machine interface decoding." IEEE transactions on neural networks and learning systems 28.4 (2015): 873-886.

(4) Dyer, Eva L., et al. "A cryptography-based approach for movement decoding." Nature biomedical engineering 1.12 (2017): 967-976.

(5) Ahmadi, Nur, Timothy G. Constandinou, and Christos-Savvas Bouganis. "Robust and accurate decoding of hand kinematics from entire spiking activity using deep learning." Journal of Neural Engineering 18.2 (2021): 026011.

(6) Churchland, Mark M., et al. "Neural population dynamics during reaching." Nature 487.7405 (2012): 51-56.

(7) Kaufman, Matthew T., et al. "Cortical activity in the null space: permitting preparation without movement." Nature neuroscience 17.3 (2014): 440-448.

(8) Elsayed, Gamaleldin F., et al. "Reorganization between preparatory and movement population responses in motor cortex." Nature communications 7.1 (2016): 13239.

(9) Ramkumar, Pavan, et al. "Premotor and motor cortices encode reward." PloS one 11.8 (2016): e0160851.

Q3: “The authors seem to have missed the spirit of my critique: to say "linear readout is performed in motor cortex" is an over-interpretation of what their model can show.”

Thank you for your comments. It's important to note that the conclusions we draw are speculative and not definitive. We use terms like "suggest" to reflect this uncertainty. To further emphasize the conjectural nature of our conclusions, we have deliberately moderated our tone.

The question of whether behaviorally-relevant signals can be accessed by downstream brain regions hinges on the debate over whether the brain employs a strategy of filtering before decoding. If the brain employs such a strategy, the brain can probably access these signals. In our view, it is likely that the brain utilizes this strategy.

Given the existence of behaviorally relevant signals, it is reasonable to assume that the brain has intrinsic mechanisms to differentiate between relevant and irrelevant signals. There is growing evidence suggesting that the brain utilizes various mechanisms, such as attention and specialized filtering, to suppress irrelevant signals and enhance relevant signals [1-3]. Therefore, it is plausible that the brain filters before decoding, thereby effectively accessing behaviorally relevant signals.

Regarding the question of whether the brain employs linear readout, given the limitations of current observational methods and our incomplete understanding of brain mechanisms, it is challenging to ascertain whether the brain employs a linear readout. In many cortical areas, linear decoders have proven to be sufficiently accurate. Consequently, numerous studies [4, 5, 6], including the one you referenced [4], directly employ linear decoders to extract information and formulate conclusions based on the decoding results. Contrary to these approaches, our research has compared the performance of linear and nonlinear decoders on behaviorally relevant signals and found their decoding performance is comparable. Considering both the decoding accuracy and model complexity, our results suggest that the motor cortex may utilize linear readout to decode information from relevant signals. Given the current technological limitations, we consider it reasonable to analyze collected data to speculate on the potential workings of the brain, an approach that many studies have also embraced [7-10]. For instance, a study [7] deduces strategies the brain might employ to overcome noise by analyzing the structure of recorded data and decoding outcomes for new stimuli.

Thank you for your valuable feedback.

(1) Sreenivasan, Sameet, and Ila Fiete. "Grid cells generate an analog error-correcting code for singularly precise neural computation." Nature neuroscience 14.10 (2011): 1330-1337.

(2) Schneider, David M., Janani Sundararajan, and Richard Mooney. "A cortical filter that learns to suppress the acoustic consequences of movement." Nature 561.7723 (2018): 391-395.

(3) Nakajima, Miho, L. Ian Schmitt, and Michael M. Halassa. "Prefrontal cortex regulates sensory filtering through a basal ganglia-to-thalamus pathway." Neuron 103.3 (2019): 445-458.

(4) Jurewicz, Katarzyna, et al. "Irrational choices via a curvilinear representational geometry for value." bioRxiv (2022): 2022-03.

(5) Hong, Ha, et al. "Explicit information for category-orthogonal object properties increases along the ventral stream." Nature neuroscience 19.4 (2016): 613-622.

(6) Chang, Le, and Doris Y. Tsao. "The code for facial identity in the primate brain." Cell 169.6 (2017): 1013-1028.

(7) Ganmor, Elad, Ronen Segev, and Elad Schneidman. "A thesaurus for a neural population code." Elife 4 (2015): e06134.

(8) Churchland, Mark M., et al. "Neural population dynamics during reaching." Nature 487.7405 (2012): 51-56.

(9) Gallego, Juan A., et al. "Cortical population activity within a preserved neural manifold underlies multiple motor behaviors." Nature communications 9.1 (2018): 4233.

(10) Gallego, Juan A., et al. "Long-term stability of cortical population dynamics underlying consistent behavior." Nature neuroscience 23.2 (2020): 260-270.

Q4: “Agreeing with my critique is not sufficient; please provide the data or simulations that provides the context for the reference in the fano factor. I believe my critique is still valid.”

Thank you for your comments. As we previously replied, Churchland's research examines the variability of neural signals across different stages, including the preparation and execution phases, as well as before and after the target appears. Our study, however, focuses exclusively on the movement execution phase. Consequently, we are unable to produce comparative displays similar to those in his research. Intuitively, one might expect that the variability of behaviorally relevant signals would be lower; however, since no prior studies have accurately extracted such signals, the specific FF values of behaviorally relevant signals remain unknown. Therefore, presenting these values is meaningful, and can provide a reference for future research. While we cannot compare FF across different stages, we can numerically compare the values to the Poisson count process. An FF of 1 indicates a Poisson firing process, and our experimental data reveals that most neurons have an FF less than 1, indicating that the variance in firing counts is below the mean.  Thank you for your valuable feedback.

To Reviewer #4

Q1: “Overall, studying neural computations that are behaviorally relevant or not is an important problem, which several previous studies have explored (for example PSID in (Sani et al. 2021), TNDM in (Hurwitz et al. 2021), TAME-GP in (Balzani et al. 2023), pi-VAE in (Zhou and Wei 2020), and dPCA in (Kobak et al. 2016), etc). However, this manuscript does not properly put their work in the context of such prior works. For example, the abstract states "One solution is to accurately separate behaviorally-relevant and irrelevant signals, but this approach remains elusive", which is not the case given that these prior works have done that. The same is true for various claims in the main text, for example "Furthermore, we found that the dimensionality of primary subspace of raw signals (26, 64, and 45 for datasets A, B, and C) is significantly higher than that of behaviorally-relevant signals (7, 13, and 9), indicating that using raw signals to estimate the neural dimensionality of behaviors leads to an overestimation" (line 321). This finding was presented in (Sani et al. 2021) and (Hurwitz et al. 2021), which is not clarified here. This issue of putting the work in context has been brought up by other reviewers previously but seems to remain largely unaddressed. The introduction is inaccurate also in that it mixes up methods that were designed for separation of behaviorally relevant information with those that are unsupervised and do not aim to do so (e.g., LFADS). The introduction should be significantly revised to explicitly discuss prior models/works that specifically formulated this behavior separation and what these prior studies found, and how this study differs.”  

Thank you for your comments. Our statement about “One solution is to accurately separate behaviorally-relevant and irrelevant signals, but this approach remains elusive” is accurate. To our best knowledge, there is no prior works to do this work--- separating accurate behaviorally relevant neural signals at both single-neuron and single-trial resolution. The works you mentioned have not explicitly proposed extracting behaviorally relevant signals, nor have they identified and addressed the key challenges of extracting relevant signals, namely determining the optimal degree of similarity between the generated relevant signals and raw signals. Those works focus on the latent neural dynamics, rather than signal level.

To clearly set apart d-VAE from other models, we have framed the extraction of behaviorally relevant signals as the following mathematical optimization problem:

where 𝒙𝒓 denotes generated behaviorally-relevant signals, 𝒙 denotes raw noisy signals, 𝐸(⋅,⋅) demotes reconstruction loss, and 𝑅(⋅) denotes regularization loss. It is important to note that while both d-VAE and TNDM employ reconstruction loss, relying solely on this term is insufficient for determining the optimal degree of similarity between the generated and raw noisy signals. The key to accurately extracting behaviorally relevant signals lies in leveraging prior knowledge about these signals to determine the optimal similarity degree, encapsulated by 𝑅(𝒙𝒓). All the works you mentioned did not have the key part 𝑅(𝒙𝒓).

Regarding the dimensionality estimation, the dimensionality of neural manifolds quantifies the degrees of freedom required to describe population activity without significant information loss.

There are two differences between our work and PSID and TNDM. 

First, the dimensions they refer to are fundamentally different from ours. The dimensionality we describe pertains to a linear subspace, where a neural dimension or neural mode or principal component basis, , with N representing the number of neurons. However, the vector length of a neural mode of PSID and our approach differs; PSID requires concatenating multiple time steps T, essentially making , TNDM, on the other hand, involves nonlinear dimensionality reduction, which is different from linear dimensionality reduction.

Second, we estimate neural dimensionality by explaining the variance of neural signals, whereas PSID and TNDM determine dimensionality through decoding performance saturation. It is important to note that the dimensionality at which decoding performance saturates may not accurately reflect the true dimensionality of neural manifolds, as some dimensions may contain redundant information that does not enhance decoding performance.

We acknowledge that while LFADS can generate signals that contain some behavioral information, it was not specifically designed to do so. Following your suggestion, we have removed this reference from the Introduction.

Thank you for your valuable feedback.

Q2: “Claims about linearity of "motor cortex" readout are not supported by results yet stated even in the abstract. Instead, what the results support is that for decoding behavior from the output of the dVAE model -- that is trained specifically to have a linear behavior readout from its embedding -- a nonlinear readout does not help. This result can be biased by the very construction of the dVAE's loss that encourages a linear readout/decoding from embeddings, and thus does not imply a finding about motor cortex.”

Thank you for your comments. We respectfully disagree with the notion that the ability of relevant signals to be linearly decoded is due to constraints that allow embedding to be linearly decoded. Embedding involves reorganizing or transforming the structure of original signals, and they can be linearly decoded does not mean the corresponding signals can be decoded linearly.

Let's clarify this with three intuitive examples:

Example 1: Image denoising is a well-established field. Whether employing supervised or blind denoising methods [1, 2], both can effectively recover the original image. This denoising process closely resembles the extraction of behaviorally relevant signals from raw signals. Consider if noisy images are not amenable to linear decoding (classification); would removing the noise enable linear decoding? The answer is no. Typically, the noise in images captured under normal conditions is minimal, yet even the clear images remain challenging to decode linearly.

Example 2: Consider the task of face recognition, where face images are set against various backgrounds, in this context, the pixels representing the face corresponds to relevant signals, while the background pixels are considered irrelevant. Suppose a network is capable of extracting the face pixels and the resulting embedding can be linearly decoded. Can the face pixels themselves be linearly decoded? The answer is no. If linear decoding of face pixels were feasible, the challenging task of face recognition could be easily resolved by merely extracting the face from the background and training a linear classifier.

Example 3: In the MNIST dataset, the background is uniformly black, and its impact is minimal. However, linear SVM classifiers used directly on the original pixels significantly underperform compared to non-linear SVMs.

In summary, embedding involves reorganizing the structure of the original signals through a feature transformation function. However, the reconstruction process can recover the structure of the original signals from the embedding. The fact that the structure of the embedding can be linearly decoded does not imply that the structure of the original signals can be linearly decoded in the same way. It is inappropriate to focus on the compression process without equally considering the reconstruction process.

Thank you for your valuable feedback.

(1) Mao, Xiao-Jiao, Chunhua Shen, and Yu-Bin Yang. "Image restoration using convolutional auto-encoders with symmetric skip connections." arXiv preprint arXiv:1606.08921 (2016).

(2) Lehtinen, Jaakko, et al. "Noise2Noise: Learning image restoration without clean data." International Conference on Machine Learning. International Machine Learning Society, 2018.

Q3: “Related to the above, it is unclear what the manuscript means by readout from motor cortex. A clearer definition of "readout" (a mapping from what to what?) in general is needed. The mapping that the linearity/nonlinearity claims refer to is from the *inferred* behaviorally relevant neural signals, which themselves are inferred nonlinearly using the VAE. This should be explicitly clarified in all claims, i.e., that only the mapping from distilled signals to behavior is linear, not the whole mapping from neural data to behavior. Again, to say the readout from motor cortex is linear is not supported, including in the abstract.” 

Thank you for your comments. We have revised the manuscript to make it more clearly. Thank you for your valuable feedback.

Q4: “Claims about individual neurons are also confounded. The d-VAE distilling processing is a population level embedding so the individual distilled neurons are not obtainable on their own without using the population data. This population level approach also raises the possibility that information can leak from one neuron to another during distillation, which is indeed what the authors hope would recover true information about individual neurons that wasn't there in the recording (the pixel denoising example). The authors acknowledge the possibility that information could leak to a neuron that didn't truly have that information and try to rule it out to some extent with some simulations and by comparing the distilled behaviorally relevant signals to the original neural signals. But ultimately, the distilled signals are different enough from the original signals to substantially improve decoding of low information neurons, and one cannot be sure if all of the information in distilled signals from any individual neuron truly belongs to that neuron. It is still quite likely that some of the improved behavior prediction of the distilled version of low-information neurons is due to leakage of behaviorally relevant information from other neurons, not the former's inherent behavioral information. This should be explicitly acknowledged in the manuscript.”

Thank you for your comments. We value your insights regarding the mixing process. However, we are confident in the robustness of our conclusions. We respectfully disagree with the notion that the small R2 values containing significant information are primarily due to leakage, and we base our disagreement on four key reasons.

(1) Neural reconstruction performance is a reliable and valid criterion.

The purpose of latent variable models is to explain neuronal activity as much as possible. Given the fact that the ground truth of behaviorally-relevant signals, the latent variables, and the generative model is unknow, it becomes evident that the only reliable reference at the signal level is the raw signals. A crucial criterion for evaluating the reliability of latent variable models (including latent variables and generated relevant signals) is their capability to effectively explain the raw signals [1]. Consequently, we firmly maintain the belief that if the generated signals closely resemble the raw signals to the greatest extent possible, in accordance with an equivalence principle, we can claim that these obtained signals faithfully retain the inherent properties of single neurons. 

Reviewer #4 appears to focus on the compression (mixing) process without giving equal consideration to the reconstruction (de-mixing) process. Numerous studies have demonstrated that deep autoencoders can reconstruct the original signal very effectively. For example, in the field of image denoising, autoencoders are capable of accurately restoring the original image [2, 3]. If one persistently focuses on the fact of mixing and ignores the reconstruction （demix） process, even if the only criterion that we can rely on at the signal level is high, one still won't acknowledge it. If this were the case, many problems would become unsolvable. For instance, a fundamental criterion for latent variable models is their ability to explain the original data. If the ground truth of the latent variables remains unknown and the reconstruction criterion is disregarded, how can we validate the effectiveness of the model, the validity of the latent variables, or ensure that findings related to latent variables are not merely by-products of the model? Therefore, we disagree with the aforementioned notion. We believe that as long as the reconstruction performance is satisfactory, the extracted signals have successfully retained the characteristics of individual neurons.

In our paper, we have shown in various ways that our generated signals sufficiently resemble the raw signals, including visualizing neuronal activity (Fig. 2m, Fig. 3i, and Fig. S5), achieving the highest performance among competitors (Fig. 2d, h, l), and conducting control analyses. Therefore, we believe our results are reliable. 

(1) Cunningham, J.P. and Yu, B.M., 2014. Dimensionality reduction for large-scale neural recordings. Nature neuroscience, 17(11), pp.1500-1509.

(2) Mao, Xiao-Jiao, Chunhua Shen, and Yu-Bin Yang. "Image restoration using convolutional auto-encoders with symmetric skip connections." arXiv preprint arXiv:1606.08921 (2016).

(3) Lehtinen, Jaakko, et al. "Noise2Noise: Learning image restoration without clean data." International Conference on Machine Learning. International Machine Learning Society, 2018.

(2) There is no reason for d-VAE to add signals that do not exist in the original signals.

(1) Adding signals that does not exist in the small R2 neurons would decrease the reconstruction performance. This is because if the added signals contain significant information, they will not resemble the irrelevant signals which contain no information, and thus, the generated signals will not resemble the raw signals. The model optimizes towards reducing the reconstruction loss, and this scenario deviates from the model's optimization direction. It is worth mentioning that when the model only has reconstruction loss without the interference of decoding loss, we believe that information leakage does not happen. Because the model can only be optimized in a direction that is similar to the raw signals; adding non-existent signals to the generated signals would increase the reconstruction loss, which is contrary to the objective of optimization. 

(2) Information carried by these additional signals is redundant for larger R2 neurons, thus they do not introduce new information that can enhance the decoding performance of the neural population, which does not benefit the decoding loss.

Based on these two points, we believe the model would not perform such counterproductive and harmful operations.

(3) The criterion that irrelevant signals should contain minimal information can effectively rule out the leakage scenario.

The criterion that irrelevant signals should contain minimal information is very important, but it seems that reviewer #4 has continuously overlooked their significance. If the model's reconstruction is insufficient, or if additional information is added (which we do not believe will happen), the residuals would decode a large amount of information, and this criterion would exclude selecting such signals. To clarify, if we assume that x, y, and z denote the raw, relevant, and irrelevant signals of smaller R2 neurons, with x=y+z, and the extracted relevant signals become y+m, the irrelevant signals become z-m in this case. Consequently, the irrelevant signals contain a significant amount of information.

We presented the decoding R2 for irrelevant signals in real datasets under three distillation scenarios: a bias towards reconstruction (alpha=0, an extreme case where the model only has reconstruction loss without decoding loss), a balanced trade-off, and a bias towards decoding (alpha=0.9), as detailed in Table 1. If significant information from small R2 neurons leaks from large R2 neurons, the irrelevant signals should contain a large amount of information. However, our results indicate that the irrelevant signals contain only minimal information, and their performance closely resembles that of the model training solely with reconstruction loss, showing no significant differences (P > 0.05, Wilcoxon rank-sum test). When the model leans towards decoding, some useful information will be left in the residuals, and irrelevant signals will contain a substantial amount of information, as observed in Table 1, alpha=0.9. Therefore, we will not choose these signals for analysis.

In conclusion, the criterion that irrelevant signals should contain minimal information is a very effective measure to exclude undesirable signals.

Author response table 1.

Decoding R2 of irrelevant signals

(4) Synthetic experiments can effectively rule out the leakage scenario.

In the absence of ground truth data, synthetic experiments serve as an effective method for validating models and are commonly employed [1-3]. 

Our experimental results demonstrate that d-VAE can effectively extract neural signals that more closely resemble actual behaviorally relevant signals (Fig. S2g).  If there were information leakage, it would decrease the similarity to the ground truth signals, hence we have ruled out this possibility. Moreover, in synthetic experiments with small R2 neurons (Fig. S10), results also demonstrate that our model could make these neurons more closely resemble ground truth relevant signals and recover their information. 

In summary, synthetic experiments strongly demonstrate that our model can recover obscured neuronal information, rather than adding signals that do not exist.

(1) Pnevmatikakis, Eftychios A., et al. "Simultaneous denoising, deconvolution, and demixing of calcium imaging data." Neuron 89.2 (2016): 285-299.

(2) Schneider, Steffen, Jin Hwa Lee, and Mackenzie Weygandt Mathis. "Learnable latent embeddings for joint behavioural and neural analysis." Nature 617.7960 (2023): 360-368.

(3) Zhou, Ding, and Xue-Xin Wei. "Learning identifiable and interpretable latent models of high-dimensional neural activity using pi-VAE." Advances in Neural Information Processing Systems 33 (2020): 7234-7247.

Based on these four points, we are confident in the reliability of our results. If Reviewer #4 considers these points insufficient, we would highly appreciate it if specific concerns regarding any of these aspects could be detailed.

Thank you for your valuable feedback.

Q5: “Given the nuances involved in appropriate comparisons across methods and since two of the datasets are public, the authors should provide their complete code (not just the dVAE method code), including the code for data loading, data preprocessing, model fitting and model evaluation for all methods and public datasets. This will alleviate concerns and allow readers to confirm conclusions (e.g., figure 2) for themselves down the line.”

Thanks for your suggestion.

Our codes are now available on GitHub at https://github.com/eric0li/d-VAE. Thank you for your valuable feedback.

Q6: “Related to 1) above, the authors should explore the results if the affine network h(.) (from embedding to behavior) was replaced with a nonlinear ANN. Perhaps linear decoders would no longer be as close to nonlinear decoders. Regardless, the claim of linearity should be revised as described in 1) and 2) above, and all caveats should be discussed.”

Thank you for your suggestion. We appreciate your feasible proposal that can be empirically tested. Following your suggestion, we have replaced the decoding of the latent variable z to behavior y with a nonlinear neural network, specifically a neural network with a single hidden layer. The modified model is termed d-VAE2. We applied the d-VAE2 to the real data, and selected the optimal alpha through the validation set. As shown in Table 1, results demonstrate that the performance of KF and ANN remains comparable. Therefore, the capacity to linearly decode behaviorally relevant signals does not stem from the linear decoding of embeddings.

Author response table 2.

Decoding R2 of behaviorally relevant signals obtained by d-VAE2

Additionally, it is worth noting that this approach is uncommon and is considered somewhat inappropriate according to the Information Bottleneck theory [1]. According to the Information Bottleneck theory, information is progressively compressed in multilayer neural networks, discarding what is irrelevant to the output and retaining what is relevant. This means that as the number of layers increases, the mutual information between each layer's embedding and the model input gradually decreases, while the mutual information between each layer's embedding and the model output gradually increases. For the decoding part, if the embeddings that is not closest to the output (behaviors) is used, then these embeddings might contain behaviorally irrelevant signals. Using these embeddings to generate behaviorally relevant signals could lead to the inclusion of irrelevant signals in the behaviorally relevant signals.

To demonstrate the above statement, we conducted experiments on the synthetic data. As shown in Table 2, we present the performance (neural R2 between the generated signals and the ground truth signals) of both models at several alpha values around the optimal alpha of dVAE (alpha=0.9) selected by the validation set. The experimental results show that at the same alpha value, the performance of d-VAE2 is consistently inferior to that of d-VAE, and d-VAE2 requires a higher alpha value to achieve performance comparable to d-VAE, and the best performance of d-VAE2 is inferior to that of d-VAE.

Author response table 3.

Neural R2 between generated signals and real behaviorally relevant signals

Thank you for your valuable feedback.

(1) Shwartz-Ziv, Ravid, and Naftali Tishby. "Opening the black box of deep neural networks via information." arXiv preprint arXiv:1703.00810 (2017).

Q7: “The beginning of the section on the "smaller R2 neurons" should clearly define what R2 is being discussed. Based on the response to previous reviewers, this R2 "signifies the proportion of neuronal activity variance explained by the linear encoding model, calculated using raw signals". This should be mentioned and made clear in the main text whenever this R2 is referred to.”

Thank you for your suggestion. We have made the modifications in the main text. Thank you for your valuable feedback.

Q8: “Various terms require clear definitions. The authors sometimes use vague terminology (e.g., "useless") without a clear definition. Similarly, discussions regarding dimensionality could benefit from more precise definitions. How is neural dimensionality defined? For example, how is "neural dimensionality of specific behaviors" (line 590) defined? Related to this, I agree with Reviewer 2 that a clear definition of irrelevant should be mentioned that clarifies that relevance is roughly taken as "correlated or predictive with a fixed time lag". The analyses do not explore relevance with arbitrary time lags between neural and behavior data.”

Thanks for your suggestion. We have removed the “useless” statements and have revised the statement of “the neural dimensionality of specific behaviors” in our revised manuscripts.

Regarding the use of fixed temporal lags, we followed the same practice as papers related to the dataset we use, which assume fixed temporal lags [1-3]. Furthermore, many studies in the motor cortex similarly use fixed temporal lags [4-6]. To clarify the definition, we have revised the definition in our manuscript. For details, please refer to the response to Q2 of reviewer #2 and our revised manuscript. We believe our definition is clearly articulated.

Thank you for your valuable feedback.

(1) Wang, Fang, et al. "Quantized attention-gated kernel reinforcement learning for brain– machine interface decoding." IEEE transactions on neural networks and learning systems 28.4 (2015): 873-886.

(2) Dyer, Eva L., et al. "A cryptography-based approach for movement decoding." Nature biomedical engineering 1.12 (2017): 967-976.

(3) Ahmadi, Nur, Timothy G. Constandinou, and Christos-Savvas Bouganis. "Robust and accurate decoding of hand kinematics from entire spiking activity using deep learning." Journal of Neural Engineering 18.2 (2021): 026011.

(4) Churchland, Mark M., et al. "Neural population dynamics during reaching." Nature 487.7405 (2012): 51-56.

(5) Kaufman, Matthew T., et al. "Cortical activity in the null space: permitting preparation without movement." Nature neuroscience 17.3 (2014): 440-448.

(6) Elsayed, Gamaleldin F., et al. "Reorganization between preparatory and movement population responses in motor cortex." Nature communications 7.1 (2016): 13239. 

Q9: “CEBRA itself doesn't provide a neural reconstruction from its embeddings, but one could obtain one via a regression from extracted CEBRA embeddings to neural data. In addition to decoding results of CEBRA (figure S3), the neural reconstruction of CEBRA should be computed and CEBRA should be added to Figure 2 to see how the behaviorally relevant and irrelevant signals from CEBRA compare to other methods.”

Thank you for your question. Modifying CEBRA is beyond the scope of our work. As CEBRA is not a generative model, it cannot obtain behaviorally relevant and irrelevant signals, and therefore it lacks the results presented in Fig. 2. To avoid the same confusion encountered by reviewers #3 and #4 among our readers, we have opted to exclude the comparison with CEBRA. It is crucial to note, as previously stated, that our assessment of decoding capabilities has been benchmarked against the performance of the ANN on raw signals, which almost represents the upper limit of performance. Consequently, omitting CEBRA does not affect our conclusions.

Thank you for your valuable feedback.

Q10: “Line 923: "The optimal hyperparameter is selected based on the lowest averaged loss of five-fold training data." => why is this explained specifically under CEBRA? Isn't the same criteria used for hyperparameters of other methods? If so, clarify.”

Thank you for your question. The hyperparameter selection for CEBRA follows the practice of the original CEBRA paper. The hyperparameter selection for generative models is detailed in the Section “The strategy for selecting effective behaviorally-relevant signals”.  Thank you for your valuable feedback.

Author response:

The following is the authors’ response to the previous reviews.

Reviewer #1 (Public Review):

In this paper, the authors evaluate the utility of brain age derived metrics for predicting cognitive decline by performing a 'commonality' analysis in a downstream regression that enables the different contribution of different predictors to be assessed. The main conclusion is that brain age derived metrics do not explain much additional variation in cognition over and above what is already explained by age. The authors propose to use a regression model trained to predict cognition ('brain cognition') as an alternative suited to applications of cognitive decline. While this is less accurate overall than brain age, it explains more unique variance in the downstream regression.  

Importantly, in this revision, we clarified that we did not intend to use Brain Cognition as an alternative approach. This is because, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Here we made this point more explicit and further stated that the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. By examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. 

REVISED VERSION: while the authors have partially addressed my concerns, I do not feel they have addressed them all. I do not feel they have addressed the weight instability and concerns about the stacked regression models satisfactorily.

Please see our responses to Reviewer #1 Public Review #3 below

I also must say that I agree with Reviewer 3 about the limitations of the brain age and brain cognition methods conceptually. In particular that the regression model used to predict fluid cognition will by construction explain more variance in cognition than a brain age model that is trained to predict age. This suffers from the same problem the authors raise with brain age and would indeed disappear if the authors had a separate measure of cognition against which to validate and were then to regress this out as they do for age correction. I am aware that these conceptual problems are more widespread than this paper alone (in fact throughout the brain age literature), so I do not believe the authors should be penalised for that. However, I do think they can make these concerns more explicit and further tone down the comments they make about the utility of brain cognition. I have indicated the main considerations about these points in the recommendations section below. 

Thank you so much for raising this point. We now have the following statement in the introduction and discussion to address this concern (see below). 

Briefly, we made it explicit that, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. That is, the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. More importantly, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And this is the third goal of this present study. 

From Introduction:

“Third and finally, certain variation in fluid cognition is related to brain MRI, but to what extent does Brain Age not capture this variation? To estimate the variation in fluid cognition that is related to the brain MRI, we could build prediction models that directly predict fluid cognition (i.e., as opposed to chronological age) from brain MRI data. Previous studies found reasonable predictive performances of these cognition-prediction models, built from certain MRI modalities (Dubois et al., 2018; Pat, Wang, Anney, et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). Analogous to Brain Age, we called the predicted values from these cognition-prediction models, Brain Cognition. The strength of an out-of-sample relationship between Brain Cognition and fluid cognition reflects variation in fluid cognition that is related to the brain MRI and, therefore, indicates the upper limit of Brain Age’s capability in capturing fluid cognition. This is, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Consequently, if we included Brain Cognition, Brain Age and chronological age in the same model to explain fluid cognition, we would be able to examine the unique effects of Brain Cognition that explain fluid cognition beyond Brain Age and chronological age. These unique effects of Brain Cognition, in turn, would indicate the amount of co-variation between brain MRI and fluid cognition that is missed by Brain Age.”

From Discussion:

“Third, by introducing Brain Cognition,  we showed the extent to which Brain Age indices were not able to capture the variation in fluid cognition that is related to brain MRI. More specifically, using Brain Cognition allowed us to gauge the variation in fluid cognition that is related to the brain MRI, and thereby, to estimate the upper limit of what Brain Age can do. Moreover, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age.

From our results, Brain Cognition, especially from certain cognition-prediction models such as the stacked models, has relatively good predictive performance, consistent with previous studies (Dubois et al., 2018; Pat, Wang, Anney, et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). We then examined Brain Cognition using commonality analyses (Nimon et al., 2008) in multiple regression models having a Brain Age index, chronological age and Brain Cognition as regressors to explain fluid cognition. Similar to Brain Age indices, Brain Cognition exhibited large common effects with chronological age. But more importantly, unlike Brain Age indices, Brain Cognition showed large unique effects, up to around 11%. As explained above, the unique effects of Brain Cognition indicated the amount of co-variation between brain MRI and fluid cognition that was missed by a Brain Age index and chronological age. This missing amount was relatively high, considering that Brain Age and chronological age together explained around 32% of the total variation in fluid cognition. Accordingly, if a Brain Age index was used as a biomarker along with chronological age, we would have missed an opportunity to improve the performance of the model by around one-third of the variation explained.” 

This is a reasonably good paper and the use of a commonality analysis is a nice contribution to understanding variance partitioning across different covariates. I have some comments that I believe the authors ought to address, which mostly relate to clarity and interpretation 

Reviewer #1 Public Review #1

First, from a conceptual point of view, the authors focus exclusively on cognition as a downstream outcome. I would suggest the authors nuance their discussion to provide broader considerations of the utility of their method and on the limits of interpretation of brain age models more generally. 

Thank you for your comments on this issue. 

We now discussed the broader consideration in detail:

(1) the consistency between our findings on fluid cognition and other recent works on brain disorders, 

(2) the difference between studies investigating the utility of Brain Age in explaining cognitive functioning, including ours and others (e.g., Butler et al., 2021; Cole, 2020, 2020; Jirsaraie, Kaufmann, et al., 2023) and those explaining neurological/psychological disorders (e.g., Bashyam et al., 2020; Rokicki et al., 2021)

and 

(3) suggested solutions we and others made to optimise the utility of Brain Age for both cognitive functioning and brain disorders.

From Discussion:

“This discrepancy between the predictive performance of age-prediction models and the utility of Brain Age indices as a biomarker is consistent with recent findings (for review, see Jirsaraie, Gorelik, et al., 2023), both in the context of cognitive functioning (Jirsaraie, Kaufmann, et al., 2023) and neurological/psychological disorders (Bashyam et al., 2020; Rokicki et al., 2021). For instance,  combining different MRI modalities into the prediction models, similar to our stacked models, ocen leads to the highest performance of age prediction models, but does not likely explain the highest variance across different phenotypes, including cognitive functioning and beyond (Jirsaraie, Gorelik, et al., 2023).”

“There is a notable difference between studies investigating the utility of Brain Age in explaining cognitive functioning, including ours and others (e.g., Butler et al., 2021; Cole, 2020, 2020; Jirsaraie, Kaufmann, et al., 2023) and those explaining neurological/psychological disorders (e.g., Bashyam et al., 2020; Rokicki et al., 2021). We consider the former as a normative type of study and the lader as a case-control type of study (Insel et al., 2010; Marquand et al., 2016). Those case-control Brain Age studies focusing on neurological/psychological disorders often build age-prediction models from MRI data of largely healthy participants (e.g., controls in a case-control design or large samples in a population-based design), apply the built age-prediction models to participants without vs. with neurological/psychological disorders and compare Brain Age indices between the two groups. On the one hand, this means that case-control studies treat Brain Age as a method to detect anomalies in the neurological/psychological group (Hahn et al., 2021). On the other hand, this also means that case-control studies have to ignore underfided models when applied prediction models built from largely healthy participants to participants with neurological/psychological disorders (i.e., Brain Age may predict chronological age well for the controls, but not for those with a disorder). On the contrary, our study and other normative studies focusing on cognitive functioning often build age prediction models from MRI data of largely healthy participants and apply the built age prediction models to participants who are also largely healthy. Accordingly, the age prediction models for explaining cognitive functioning in normative studies, while not allowing us to detect group-level anomalies, do not suffer from being under-fided. This unfortunately might limit the generalisability of our study into just the normative type of study. Future work is still needed to test the utility of brain age in the case-control case.”

“Next, researchers should not select age-prediction models based solely on age-prediction performance. Instead, researchers could select age-prediction models that explained phenotypes of interest the best. Here we selected age-prediction models based on a set of features (i.e., modalities) of brain MRI. This strategy was found effective not only for fluid cognition as we demonstrated here, but also for neurological and psychological disorders as shown elsewhere (Jirsaraie, Gorelik, et al., 2023; Rokicki et al., 2021). Rokicki and colleagues (2021), for instance, found that, while integrating across MRI modalities led to age prediction models with the highest age-prediction performance, using only T1 structural MRI gave age-prediction models that were better at classifying Alzheimer’s disease. Similarly, using only cerebral blood flow gave age-prediction models that were better at classifying mild/subjective cognitive impairment, schizophrenia and bipolar disorder. 

As opposed to selecting age-prediction models based on a set of features, researchers could also select age-prediction models based on modelling methods. For instance, Jirsaraie and colleagues (2023) compared gradient tree boosting (GTB) and deep-learning brain network (DBN) algorithms in building age-prediction models. They found GTB to have higher age prediction performance but DBN to have better utility in explaining cognitive functioning. In this case, an algorithm with better utility (e.g., DBN) should be used for explaining a phenotype of interest. Similarly, Bashyam and colleagues (2020) built different DBN-based age-prediction models, varying in age-prediction performance. The DBN models with a higher number of epochs corresponded to higher age-prediction performance. However, DBN-based age-prediction models with a moderate (as opposed to higher or lower) number of epochs were better at classifying Alzheimer’s disease, mild cognitive impairment and schizophrenia. In this case, a model from the same algorithm with better utility (e.g., those DBN with a moderate epoch number) should be used for explaining a phenotype of interest.

Accordingly, this calls for a change in research practice, as recently pointed out by Jirasarie and colleagues (2023, p7), “Despite mounting evidence, there is a persisting assumption across several studies that the most accurate brain age models will have the most potential for detecting differences in a given phenotype of interest”. Future neuroimaging research should aim to build age-prediction models that are not necessarily good at predicting age, but at capturing phenotypes of interest.”

Reviewer #1 Public Review #2

Second, from a methods perspective, there is not a sufficient explanation of the methodological procedures in the current manuscript to fully understand how the stacked regression models were constructed. I would request that the authors provide more information to enable the reader to beUer understand the stacked regression models used to ensure that these models are not overfit. 

Thank you for allowing us an opportunity to clarify our stacked model. We made additional clarification to make this clearer (see below). We wanted to confirm that we did not use test sets to build a stacked model in both lower and higher levels of the Elastic Net models. Test sets were there just for testing the performance of the models.  

From Methods:

“We used nested cross-validation (CV) to build these prediction models (see Figure 7). We first split the data into five outer folds, leaving each outer fold with around 100 participants. This number of participants in each fold is to ensure the stability of the test performance across folds. In each outer-fold CV loop, one of the outer folds was treated as an outer-fold test set, and the rest was treated as an outer-fold training set. Ultimately, looping through the nested CV resulted in a) prediction models from each of the 18 sets of features as well as b) prediction models that drew information across different combinations of the 18 separate sets, known as “stacked models.” We specified eight stacked models: “All” (i.e., including all 18 sets of features),  “All excluding Task FC”, “All excluding Task Contrast”, “Non-Task” (i.e., including only Rest FC and sMRI), “Resting and Task FC”, “Task Contrast and FC”, “Task Contrast” and “Task FC”. Accordingly, there were 26 prediction models in total for both Brain Age and Brain Cognition.

To create these 26 prediction models, we applied three steps for each outer-fold loop. The first step aimed at tuning prediction models for each of 18 sets of features. This step only involved the outer-fold training set and did not involve the outer-fold test set. Here, we divided the outer-fold training set into five inner folds and applied inner-fold CV to tune hyperparameters with grid search. Specifically, in each inner-fold CV, one of the inner folds was treated as an inner-fold validation set, and the rest was treated as an inner-fold training set. Within each inner-fold CV loop, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters and applied the estimated model to the inner-fold validation set. Acer looping through the inner-fold CV, we, then, chose the prediction models that led to the highest performance, reflected by coefficient of determination (R2), on average across the inner-fold validation sets. This led to 18 tuned models, one for each of the 18 sets of features, for each outer fold.

The second step aimed at tuning stacked models. Same as the first step, the second step only involved the outer-fold training set and did not involve the outer-fold test set. Here, using the same outer-fold training set as the first step, we applied tuned models, created from the first step, one from each of the 18 sets of features, resulting in 18 predicted values for each participant. We, then, re-divided this outer-fold training set into new five inner folds. In each inner fold, we treated different combinations of the 18 predicted values from separate sets of features as features to predict the targets in separate “stacked” models. Same as the first step, in each inner-fold CV loop, we treated one out of five inner folds as an inner-fold validation set, and the rest as an inner-fold training set. Also as in the first step, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters from our grid. We tuned the hyperparameters of stacked models using grid search by selecting the models with the highest R2 on average across the inner-fold validation sets. This led to eight tuned stacked models.

The third step aimed at testing the predictive performance of the 18 tuned prediction models from each of the set of features, built from the first step, and eight tuned stacked models, built from the second step. Unlike the first two steps, here we applied the already tuned models to the outer-fold test set. We started by applying the 18 tuned prediction models from each of the sets of features to each observation in the outer-fold test set, resulting in 18 predicted values. We then applied the tuned stacked models to these predicted values from separate sets of features, resulting in eight predicted values. 

To demonstrate the predictive performance, we assessed the similarity between the observed values and the predicted values of each model across outer-fold test sets, using Pearson’s r, coefficient of determination (R2) and mean absolute error (MAE). Note that for R2, we used the sum of squares definition (i.e., R2 \= 1 – (sum of squares residuals/total sum of squares)) per a previous recommendation (Poldrack et al., 2020). We considered the predicted values from the outer-fold test sets of models predicting age or fluid cognition, as Brain Age and Brain Cognition, respectively.”

Author response image 1.

Diagram of the nested cross-validation used for creating predictions for models of each set of features as well as predictions for stacked models. 

Note some previous research, including ours (Tetereva et al., 2022), splits the observations in the outer-fold training set into layer 1 and layer 2 and applies the first and second steps to layers 1 and 2, respectively. Here we decided against this approach and used the same outer-fold training set for both first and second steps in order to avoid potential bias toward the stacked models. This is because, when the data are split into two layers, predictive models built for each separate set of features only use the data from layer 1, while the stacked models use the data from both layers 1 and 2. In practice with large enough data, these two approaches might not differ much, as we demonstrated previously (Tetereva et al., 2022).

Reviewer #1 Public Review #3

Please also provide an indication of the different regression strengths that were estimated across the different models and cross-validation splits. Also, how stable were the weights across splits? 

The focus of this article is on the predictions. Still, it is informative for readers to understand how stable the feature importance (i.e., Elastic Net coefficients) is. To demonstrate the stability of feature importance, we now examined the rank stability of feature importance using Spearman’s ρ (see Figure 4). Specifically, we correlated the feature importance between two prediction models of the same features, used in two different outer-fold test sets. Given that there were five outer-fold test sets, we computed 10 Spearman’s ρ for each prediction model of the same features.  We found Spearman’s ρ to be varied dramatically in both age-prediction (range\=.31-.94) and fluid cognition-prediction (range\=.16-.84) models. This means that some prediction models were much more stable in their feature importance than others. This is probably due to various factors such as a) the collinearity of features in the model, b) the number of features (e.g., 71,631 features in functional connectivity, which were further reduced to 75 PCAs, as compared to 19 features in subcortical volume based on the ASEG atlas), c) the penalisation of coefficients either with ‘Ridge’ or ‘Lasso’ methods, which resulted in reduction as a group of features or selection of a feature among correlated features, respectively, and d) the predictive performance of the models. Understanding the stability of feature importance is beyond the scope of the current article. As mentioned by Reviewer 1, “The predictions can be stable when the coefficients are not,” and we chose to focus on the prediction in the current article.   

Author response image 2.

Stability of feature importance (i.e., Elastic Net Coefficients) of prediction models. Each dot represents rank stability (reflected by Spearman’s ρ) in the feature importance between two prediction models of the same features, used in two different outer-fold test sets. Given that there were five outer-fold test sets, there were 10 Spearman’s ρs for each prediction model.  The numbers to the right of the plots indicate the mean of Spearman’s ρ for each prediction model.  

Reviewer #1 Public Review #4

Please provide more details about the task designs, MRI processing procedures that were employed on this sample in addition to the regression methods and bias correction methods used. For example, there are several different parameterisations of the elastic net, please provide equations to describe the method used here so that readers can easily determine how the regularisation parameters should be interpreted.  

Thank you for the opportunity for us to provide more methodical details.

First, for the task design, we included the following statements:

From Methods:

“HCP-A collected fMRI data from three tasks: Face Name (Sperling et al., 2001), Conditioned Approach Response Inhibition Task (CARIT) (Somerville et al., 2018) and VISual MOTOR (VISMOTOR) (Ances et al., 2009). 

First, the Face Name task (Sperling et al., 2001) taps into episodic memory. The task had three blocks. In the encoding block [Encoding], participants were asked to memorise the names of faces shown. These faces were then shown again in the recall block [Recall] when the participants were asked if they could remember the names of the previously shown faces. There was also the distractor block [Distractor] occurring between the encoding and recall blocks. Here participants were distracted by a Go/NoGo task. We computed six contrasts for this Face Name task: [Encode], [Recall], [Distractor], [Encode vs. Distractor], [Recall vs. Distractor] and [Encode vs. Recall].

Second, the CARIT task (Somerville et al., 2018) was adapted from the classic Go/NoGo task and taps into inhibitory control. Participants were asked to press a budon to all [Go] but not to two [NoGo] shapes. We computed three contrasts for the CARIT task: [NoGo], [Go] and [NoGo vs. Go]. 

Third, the VISMOTOR task (Ances et al., 2009) was designed to test simple activation of the motor and visual cortices. Participants saw a checkerboard with a red square either on the lec or right. They needed to press a corresponding key to indicate the location of the red square. We computed just one contrast for the VISMOTOR task: [Vismotor], which indicates the presence of the checkerboard vs. baseline.” 

Second, for MRI processing procedures, we included the following statements.

From Methods:

“HCP-A provides details of parameters for brain MRI elsewhere (Bookheimer et al., 2019; Harms et al., 2018). Here we used MRI data that were pre-processed by the HCP-A with recommended methods, including the MSMALL alignment (Glasser et al., 2016; Robinson et al., 2018) and ICA-FIX (Glasser et al., 2016) for functional MRI. We used multiple brain MRI modalities, covering task functional MRI (task fMRI), resting-state functional MRI (rsfMRI) and structural MRI (sMRI), and organised them into 19 sets of features.”

“Sets of Features 1-10: Task fMRI contrast (Task Contrast)

Task contrasts reflect fMRI activation relevant to events in each task. Bookheimer and colleagues (2019) provided detailed information about the fMRI in HCP-A. Here we focused on the pre-processed task fMRI Connectivity Informatics Technology Initiative (CIFTI) files with a suffix, “_PA_Atlas_MSMAll_hp0_clean.dtseries.nii.” These CIFTI files encompassed both the cortical mesh surface and subcortical volume (Glasser et al., 2013). Collected using the posterior-to-anterior (PA) phase, these files were aligned using MSMALL (Glasser et al., 2016; Robinson et al., 2018), linear detrended (see hdps://groups.google.com/a/humanconnectome.org/g/hcp-users/c/ZLJc092h980/m/GiihzQAUAwAJ) and cleaned from potential artifacts using ICA-FIX (Glasser et al., 2016). 

To extract Task Contrasts, we regressed the fMRI time series on the convolved task events using a double-gamma canonical hemodynamic response function via FMRIB Software Library (FSL)’s FMRI Expert Analysis Tool (FEAT) (Woolrich et al., 2001). We kept FSL’s default high pass cutoff at 200s (i.e., .005 Hz). We then parcellated the contrast ‘cope’ files, using the Glasser atlas (Gordon et al., 2016) for cortical surface regions and the Freesurfer’s automatic segmentation (aseg) (Fischl et al., 2002) for subcortical regions. This resulted in 379 regions, whose number was, in turn, the number of features for each Task Contrast set of features. “ 

“Sets of Features 11-13: Task fMRI functional connectivity (Task FC)

Task FC reflects functional connectivity (FC ) among the brain regions during each task, which is considered an important source of individual differences (Elliod et al., 2019; Fair et al., 2007; Gradon et al., 2018). We used the same CIFTI file “_PA_Atlas_MSMAll_hp0_clean.dtseries.nii.” as the task contrasts. Unlike Task Contrasts, here we treated the double-gamma, convolved task events as regressors of no interest and focused on the residuals of the regression from each task (Fair et al., 2007). We computed these regressors on FSL, and regressed them in nilearn (Abraham et al., 2014). Following previous work on task FC (Elliod et al., 2019), we applied a highpass at .008 Hz. For parcellation, we used the same atlases as Task Contrast (Fischl et al., 2002; Glasser et al., 2016). We computed Pearson’s correlations of each pair of 379 regions, resulting in a table of 71,631 non-overlapping FC indices for each task. We then applied r-to-z transformation and principal component analysis (PCA) of 75 components (Rasero et al., 2021; Sripada et al., 2019, 2020). Note to avoid data leakage, we conducted the PCA on each training set and applied its definition to the corresponding test set. Accordingly, there were three sets of 75 features for Task FC, one for each task. 

Set of Features 14: Resting-state functional MRI functional connectivity (Rest FC) Similar to Task FC, Rest FC reflects functional connectivity (FC ) among the brain regions, except that Rest FC occurred during the resting (as opposed to task-performing) period. HCPA collected Rest FC from four 6.42-min (488 frames) runs across two days, leading to 26-min long data (Harms et al., 2018). On each day, the study scanned two runs of Rest FC, starting with anterior-to-posterior (AP) and then with posterior-to-anterior (PA) phase encoding polarity. We used the “rfMRI_REST_Atlas_MSMAll_hp0_clean.dscalar.nii” file that was preprocessed and concatenated across the four runs.  We applied the same computations (i.e., highpass filter, parcellation, Pearson’s correlations, r-to-z transformation and PCA) with the Task FC. 

Sets of Features 15-18: Structural MRI (sMRI)

sMRI reflects individual differences in brain anatomy. The HCP-A used an established preprocessing pipeline for sMRI (Glasser et al., 2013). We focused on four sets of features: cortical thickness, cortical surface area, subcortical volume and total brain volume. For cortical thickness and cortical surface area, we used Destrieux’s atlas (Destrieux et al., 2010; Fischl, 2012) from FreeSurfer’s “aparc.stats” file, resulting in 148 regions for each set of features. For subcortical volume, we used the aseg atlas (Fischl et al., 2002) from FreeSurfer’s “aseg.stats” file, resulting in 19 regions. For total brain volume, we had five FreeSurfer-based features: “FS_IntraCranial_Vol” or estimated intra-cranial volume, “FS_TotCort_GM_Vol” or total cortical grey mader volume, “FS_Tot_WM_Vol” or total cortical white mader volume, “FS_SubCort_GM_Vol” or total subcortical grey mader volume and “FS_BrainSegVol_eTIV_Ratio” or ratio of brain segmentation volume to estimated total intracranial volume.”

Third, for regression methods and bias correction methods used, we included the following statements:

From Methods:

“For the machine learning algorithm, we used Elastic Net (Zou & Hastie, 2005). Elastic Net is a general form of penalised regressions (including Lasso and Ridge regression), allowing us to simultaneously draw information across different brain indices to predict one target variable. Penalised regressions are commonly used for building age-prediction models (Jirsaraie, Gorelik, et al., 2023). Previously we showed that the performance of Elastic Net in predicting cognitive abilities is on par, if not better than, many non-linear and morecomplicated algorithms (Pat, Wang, Bartonicek, et al., 2022; Tetereva et al., 2022). Moreover, Elastic Net coefficients are readily explainable, allowing us the ability to explain how our age-prediction and cognition-prediction models made the prediction from each brain feature (Molnar, 2019; Pat, Wang, Bartonicek, et al., 2022) (see below). 

Elastic Net simultaneously minimises the weighted sum of the features’ coefficients. The degree of penalty to the sum of the feature’s coefficients is determined by a shrinkage hyperparameter ‘a’: the greater the a, the more the coefficients shrink, and the more regularised the model becomes. Elastic Net also includes another hyperparameter, ‘ℓ! ratio’, which determines the degree to which the sum of either the squared (known as ‘Ridge’; ℓ! ratio=0) or absolute (known as ‘Lasso’; ℓ! ratio=1) coefficients is penalised (Zou & Hastie, 2005). The objective function of Elastic Net as implemented by sklearn (Pedregosa et al., 2011) is defined as:

where X is the features, y is the target, and b is the coefficient. In our grid search, we tuned two Elastic Net hyperparameters: a using 70 numbers in log space, ranging from .1 and 100, and ℓ!-ratio using 25 numbers in linear space, ranging from 0 and 1.

To understand how Elastic Net made a prediction based on different brain features, we examined the coefficients of the tuned model. Elastic Net coefficients can be considered as feature importance, such that more positive Elastic Net coefficients lead to more positive predicted values and, similarly, more negative Elastic Net coefficients lead to more negative predicted values (Molnar, 2019; Pat, Wang, Bartonicek, et al., 2022). While the magnitude of Elastic Net coefficients is regularised (thus making it difficult for us to interpret the magnitude itself directly), we could still indicate that a brain feature with a higher magnitude weights relatively stronger in making a prediction. Another benefit of Elastic Net as a penalised regression is that the coefficients are less susceptible to collinearity among features as they have already been regularised (Dormann et al., 2013; Pat, Wang, Bartonicek, et al., 2022).

Given that we used five-fold nested cross validation, different outer folds may have different degrees of ‘a’ and ‘ℓ! ratio’, making the final coefficients from different folds to be different. For instance, for certain sets of features, penalisation may not play a big part (i.e., higher or lower ‘a’ leads to similar predictive performance), resulting in different ‘a’ for different folds. To remedy this in the visualisation of Elastic Net feature importance, we refitted the Elastic Net model to the full dataset without spli{ng them into five folds and visualised the coefficients on brain images using Brainspace (Vos De Wael et al., 2020) and Nilern (Abraham et al., 2014) packages. Note, unlike other sets of features, Task FC and Rest FC were modelled acer data reduction via PCA. Thus, for Task FC and Rest FC, we, first, multiplied the absolute PCA scores (extracted from the ‘components_’ attribute of ‘sklearn.decomposition.PCA’) with Elastic Net coefficients and, then, summed the multiplied values across the 75 components, leaving 71,631 ROI-pair indices. “

References

Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine learning for neuroimaging with scikitlearn. Frontiers in Neuroinformatics, 8, 14. hdps://doi.org/10.3389/fninf.2014.00014

Ances, B. M., Liang, C. L., Leontiev, O., Perthen, J. E., Fleisher, A. S., Lansing, A. E., & Buxton, R. B. (2009). Effects of aging on cerebral blood flow, oxygen metabolism, and blood oxygenation level dependent responses to visual stimulation. Human Brain Mapping, 30(4), 1120–1132. hdps://doi.org/10.1002/hbm.20574

Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I. M., Truelove-Hill, M., Srinivasan, D., Mamourian, L., Pomponio, R., Fan, Y., Launer, L. J., Masters, C. L., Maruff, P., Zhuo, C., Völzke, H., Johnson, S. C., Fripp, J., Koutsouleris, N., Saderthwaite, T. D., … on behalf of the ISTAGING Consortium,  the P. A. disease C., ADNI, and CARDIA studies. (2020). MRI signatures of brain age and disease over the lifespan based on a deep brain network and 14 468 individuals worldwide. Brain, 143(7), 2312–2324. hdps://doi.org/10.1093/brain/awaa160

Bookheimer, S. Y., Salat, D. H., Terpstra, M., Ances, B. M., Barch, D. M., Buckner, R. L., Burgess, G. C., Curtiss, S. W., Diaz-Santos, M., Elam, J. S., Fischl, B., Greve, D. N., Hagy, H. A., Harms, M. P., Hatch, O. M., Hedden, T., Hodge, C., Japardi, K. C., Kuhn, T. P., … Yacoub, E. (2019). The Lifespan Human Connectome Project in Aging: An overview. NeuroImage, 185, 335–348. hdps://doi.org/10.1016/j.neuroimage.2018.10.009

Butler, E. R., Chen, A., Ramadan, R., Le, T. T., Ruparel, K., Moore, T. M., Saderthwaite, T. D., Zhang, F., Shou, H., Gur, R. C., Nichols, T. E., & Shinohara, R. T. (2021). Pi alls in brain age analyses. Human Brain Mapping, 42(13), 4092–4101. hdps://doi.org/10.1002/hbm.25533

Cole, J. H. (2020). Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical, lifestyle, and cognitive factors. Neurobiology of Aging, 92, 34–42. hdps://doi.org/10.1016/j.neurobiolaging.2020.03.014

Destrieux, C., Fischl, B., Dale, A., & Halgren, E. (2010). Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. NeuroImage, 53(1), 1–15. hdps://doi.org/10.1016/j.neuroimage.2010.06.010

Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., Marquéz, J. R. G., Gruber, B., Lafourcade, B., Leitão, P. J., Münkemüller, T., McClean, C., Osborne, P. E., Reineking, B., Schröder, B., Skidmore, A. K., Zurell, D., & Lautenbach, S. (2013). Collinearity: A review of methods to deal with it and a simulation study evaluating their performance. Ecography, 36(1), 27–46. hdps://doi.org/10.1111/j.16000587.2012.07348.x

Dubois, J., Galdi, P., Paul, L. K., & Adolphs, R. (2018). A distributed brain network predicts general intelligence from resting-state human neuroimaging data. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170284. hdps://doi.org/10.1098/rstb.2017.0284

Elliod, M. L., Knodt, A. R., Cooke, M., Kim, M. J., Melzer, T. R., Keenan, R., Ireland, D., Ramrakha, S., Poulton, R., Caspi, A., Moffid, T. E., & Hariri, A. R. (2019). General functional connectivity: Shared features of resting-state and task fMRI drive reliable and heritable individual differences in functional brain networks. NeuroImage, 189, 516–532. hdps://doi.org/10.1016/j.neuroimage.2019.01.068

Fair, D. A., Schlaggar, B. L., Cohen, A. L., Miezin, F. M., Dosenbach, N. U. F., Wenger, K. K., Fox, M. D., Snyder, A. Z., Raichle, M. E., & Petersen, S. E. (2007). A method for using blocked and event-related fMRI data to study “resting state” functional connectivity. NeuroImage, 35(1), 396–405. hdps://doi.org/10.1016/j.neuroimage.2006.11.051

Fischl, B. (2012). FreeSurfer. NeuroImage, 62(2), 774–781. hdps://doi.org/10.1016/j.neuroimage.2012.01.021

Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., van der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, S., Montillo, A., Makris, N., Rosen, B., & Dale, A. M. (2002). Whole Brain Segmentation. Neuron, 33(3), 341–355. hdps://doi.org/10.1016/S0896-6273(02)00569-X

Glasser, M. F., Smith, S. M., Marcus, D. S., Andersson, J. L. R., Auerbach, E. J., Behrens, T. E. J., Coalson, T. S., Harms, M. P., Jenkinson, M., Moeller, S., Robinson, E. C., Sotiropoulos, S. N., Xu, J., Yacoub, E., Ugurbil, K., & Van Essen, D. C. (2016). The Human Connectome Project’s neuroimaging approach. Nature Neuroscience, 19(9), 1175– 1187. hdps://doi.org/10.1038/nn.4361

Glasser, M. F., Sotiropoulos, S. N., Wilson, J. A., Coalson, T. S., Fischl, B., Andersson, J. L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J. R., Van Essen, D. C., & Jenkinson, M. (2013). The minimal preprocessing pipelines for the Human Connectome Project. NeuroImage, 80, 105–124. hdps://doi.org/10.1016/j.neuroimage.2013.04.127

Gordon, E. M., Laumann, T. O., Adeyemo, B., Huckins, J. F., Kelley, W. M., & Petersen, S. E. (2016). Generation and Evaluation of a Cortical Area Parcellation from Resting-State Correlations. Cerebral Cortex, 26(1), 288–303. hdps://doi.org/10.1093/cercor/bhu239

Gradon, C., Laumann, T. O., Nielsen, A. N., Greene, D. J., Gordon, E. M., Gilmore, A. W., Nelson, S. M., Coalson, R. S., Snyder, A. Z., Schlaggar, B. L., Dosenbach, N. U. F., & Petersen, S. E. (2018). Functional Brain Networks Are Dominated by Stable Group and Individual Factors, Not Cognitive or Daily Variation. Neuron, 98(2), 439-452.e5. hdps://doi.org/10.1016/j.neuron.2018.03.035

Hahn, T., Fisch, L., Ernsting, J., Winter, N. R., Leenings, R., Sarink, K., Emden, D., Kircher, T., Berger, K., & Dannlowski, U. (2021). From ‘loose fi{ng’ to high-performance, uncertainty-aware brain-age modelling. Brain, 144(3), e31–e31. hdps://doi.org/10.1093/brain/awaa454

Harms, M. P., Somerville, L. H., Ances, B. M., Andersson, J., Barch, D. M., Bastiani, M., Bookheimer, S. Y., Brown, T. B., Buckner, R. L., Burgess, G. C., Coalson, T. S., Chappell, M. A., Dapredo, M., Douaud, G., Fischl, B., Glasser, M. F., Greve, D. N., Hodge, C., Jamison, K. W., … Yacoub, E. (2018). Extending the Human Connectome Project across ages: Imaging protocols for the Lifespan Development and Aging projects. NeuroImage, 183, 972–984. hdps://doi.org/10.1016/j.neuroimage.2018.09.060

Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., Sanislow, C., & Wang, P. (2010). Research Domain Criteria (RDoC): Toward a New Classification Framework for Research on Mental Disorders. American Journal of Psychiatry, 167(7), 748–751. hdps://doi.org/10.1176/appi.ajp.2010.09091379

Jirsaraie, R. J., Gorelik, A. J., Gatavins, M. M., Engemann, D. A., Bogdan, R., Barch, D. M., & Sotiras, A. (2023). A systematic review of multimodal brain age studies: Uncovering a divergence between model accuracy and utility. PaUerns, 4(4), 100712. hdps://doi.org/10.1016/j.pader.2023.100712

Jirsaraie, R. J., Kaufmann, T., Bashyam, V., Erus, G., Luby, J. L., Westlye, L. T., Davatzikos, C., Barch, D. M., & Sotiras, A. (2023). Benchmarking the generalizability of brain age models: Challenges posed by scanner variance and prediction bias. Human Brain Mapping, 44(3), 1118–1128. hdps://doi.org/10.1002/hbm.26144

Marquand, A. F., Rezek, I., Buitelaar, J., & Beckmann, C. F. (2016). Understanding Heterogeneity in Clinical Cohorts Using Normative Models: Beyond Case-Control Studies. Biological Psychiatry, 80(7), 552–561. hdps://doi.org/10.1016/j.biopsych.2015.12.023

Molnar, C. (2019). Interpretable Machine Learning. A Guide for Making Black Box Models Explainable. hdps://christophm.github.io/interpretable-ml-book/

Nimon, K., Lewis, M., Kane, R., & Haynes, R. M. (2008). An R package to compute commonality coefficients in the multiple regression case: An introduction to the package and a practical example. Behavior Research Methods, 40(2), 457–466. hdps://doi.org/10.3758/BRM.40.2.457

Pat, N., Wang, Y., Anney, R., Riglin, L., Thapar, A., & Stringaris, A. (2022). Longitudinally stable, brain-based predictive models mediate the relationships between childhood cognition and socio-demographic, psychological and genetic factors. Human Brain Mapping, hbm.26027. hdps://doi.org/10.1002/hbm.26027

Pat, N., Wang, Y., Bartonicek, A., Candia, J., & Stringaris, A. (2022). Explainable machine learning approach to predict and explain the relationship between task-based fMRI and individual differences in cognition. Cerebral Cortex, bhac235. hdps://doi.org/10.1093/cercor/bhac235

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Predenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, É. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12(85), 2825–2830.

Poldrack, R. A., Huckins, G., & Varoquaux, G. (2020). Establishment of Best Practices for Evidence for Prediction: A Review. JAMA Psychiatry, 77(5), 534–540. hdps://doi.org/10.1001/jamapsychiatry.2019.3671

Rasero, J., Sentis, A. I., Yeh, F.-C., & Verstynen, T. (2021). Integrating across neuroimaging modalities boosts prediction accuracy of cognitive ability. PLOS Computational Biology, 17(3), e1008347. hdps://doi.org/10.1371/journal.pcbi.1008347

Robinson, E. C., Garcia, K., Glasser, M. F., Chen, Z., Coalson, T. S., Makropoulos, A., Bozek, J., Wright, R., Schuh, A., Webster, M., Huder, J., Price, A., Cordero Grande, L., Hughes, E., Tusor, N., Bayly, P. V., Van Essen, D. C., Smith, S. M., Edwards, A. D., … Rueckert, D. (2018). Multimodal surface matching with higher-order smoothness constraints. NeuroImage, 167, 453–465. hdps://doi.org/10.1016/j.neuroimage.2017.10.037

Rokicki, J., Wolfers, T., Nordhøy, W., Tesli, N., Quintana, D. S., Alnæs, D., Richard, G., de Lange, A.-M. G., Lund, M. J., Norbom, L., Agartz, I., Melle, I., Nærland, T., Selbæk, G., Persson, K., Nordvik, J. E., Schwarz, E., Andreassen, O. A., Kaufmann, T., & Westlye, L. T. (2021). Multimodal imaging improves brain age prediction and reveals distinct abnormalities in patients with psychiatric and neurological disorders. Human Brain Mapping, 42(6), 1714–1726. hdps://doi.org/10.1002/hbm.25323

Somerville, L. H., Bookheimer, S. Y., Buckner, R. L., Burgess, G. C., Curtiss, S. W., Dapredo, M., Elam, J. S., Gaffrey, M. S., Harms, M. P., Hodge, C., Kandala, S., Kastman, E. K., Nichols, T. E., Schlaggar, B. L., Smith, S. M., Thomas, K. M., Yacoub, E., Van Essen, D. C., & Barch, D. M. (2018). The Lifespan Human Connectome Project in Development: A large-scale study of brain connectivity development in 5–21 year olds. NeuroImage, 183, 456–468. hdps://doi.org/10.1016/j.neuroimage.2018.08.050

Sperling, R. A., Bates, J. F., Cocchiarella, A. J., Schacter, D. L., Rosen, B. R., & Albert, M. S. (2001). Encoding novel face-name associations: A functional MRI study. Human Brain Mapping, 14(3), 129–139. hdps://doi.org/10.1002/hbm.1047

Sripada, C., Angstadt, M., Rutherford, S., Kessler, D., Kim, Y., Yee, M., & Levina, E. (2019). Basic Units of Inter-Individual Variation in Resting State Connectomes. Scientific Reports, 9(1), Article 1. hdps://doi.org/10.1038/s41598-018-38406-5

Sripada, C., Angstadt, M., Rutherford, S., Taxali, A., & Shedden, K. (2020). Toward a “treadmill test” for cognition: Improved prediction of general cognitive ability from the task activated brain. Human Brain Mapping, 41(12), 3186–3197. hdps://doi.org/10.1002/hbm.25007

Tetereva, A., Li, J., Deng, J. D., Stringaris, A., & Pat, N. (2022). Capturing brain-cognition relationship: Integrating task-based fMRI across tasks markedly boosts prediction and test-retest reliability. NeuroImage, 263, 119588. hdps://doi.org/10.1016/j.neuroimage.2022.119588

Vieira, B. H., Pamplona, G. S. P., Fachinello, K., Silva, A. K., Foss, M. P., & Salmon, C. E. G. (2022). On the prediction of human intelligence from neuroimaging: A systematic review of methods and reporting. Intelligence, 93, 101654. hdps://doi.org/10.1016/j.intell.2022.101654

Vos De Wael, R., Benkarim, O., Paquola, C., Lariviere, S., Royer, J., Tavakol, S., Xu, T., Hong, S.J., Langs, G., Valk, S., Misic, B., Milham, M., Margulies, D., Smallwood, J., & Bernhardt, B. C. (2020). BrainSpace: A toolbox for the analysis of macroscale gradients in neuroimaging and connectomics datasets. Communications Biology, 3(1), 103. hdps://doi.org/10.1038/s42003-020-0794-7

Woolrich, M. W., Ripley, B. D., Brady, M., & Smith, S. M. (2001). Temporal Autocorrelation in Univariate Linear Modeling of FMRI Data. NeuroImage, 14(6), 1370–1386. hdps://doi.org/10.1006/nimg.2001.0931

Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301–320. hdps://doi.org/10.1111/j.1467-9868.2005.00503.x

Author response:

The following is the authors’ response to the original reviews.

Responses to Reviewer’s Comments:  

To Reviewer #2:

(1) The use of two m⁵C reader proteins is likely a reason for the high number of edits introduced by the DRAM-Seq method. Both ALYREF and YBX1 are ubiquitous proteins with multiple roles in RNA metabolism including splicing and mRNA export. It is reasonable to assume that both ALYREF and YBX1 bind to many mRNAs that do not contain m⁵C. 

To substantiate the author's claim that ALYREF or YBX1 binds m⁵C-modified RNAs to an extent that would allow distinguishing its binding to non-modified RNAs from binding to m⁵Cmodified RNAs, it would be recommended to provide data on the affinity of these, supposedly proven, m⁵C readers to non-modified versus m⁵C-modified RNAs. To do so, this reviewer suggests performing experiments as described in Slama et al., 2020 (doi: 10.1016/j.ymeth.2018.10.020). However, using dot blots like in so many published studies to show modification of a specific antibody or protein binding, is insufficient as an argument because no antibody, nor protein, encounters nanograms to micrograms of a specific RNA identity in a cell. This issue remains a major caveat in all studies using so-called RNA modification reader proteins as bait for detecting RNA modifications in epitranscriptomics research. It becomes a pertinent problem if used as a platform for base editing similar to the work presented in this manuscript.

The authors have tried to address the point made by this reviewer. However, rather than performing an experiment with recombinant ALYREF-fusions and m⁵C-modified to unmodified RNA oligos for testing the enrichment factor of ALYREF in vitro, the authors resorted to citing two manuscripts. One manuscript is cited by everybody when it comes to ALYREF as m⁵C reader, however none of the experiments have been repeated by another laboratory. The other manuscript is reporting on YBX1 binding to m⁵C-containing RNA and mentions PARCLiP experiments with ALYREF, the details of which are nowhere to be found in doi: 10.1038/s41556-019-0361-y.

Furthermore, the authors have added RNA pull-down assays that should substitute for the requested experiments. Interestingly, Figure S1E shows that ALYREF binds equally well to unmodified and m⁵C-modified RNA oligos, which contradicts doi:10.1038/cr.2017.55, and supports the conclusion that wild-type ALYREF is not specific m⁵C binder. The necessity of including always an overexpression of ALYREF-mut in parallel DRAM experiments, makes the developed method better controlled but not easy to handle (expression differences of the plasmid-driven proteins etc.) 

Thank you for pointing this out. First, we would like to correct our previous response: the binding ability of ALYREF to m⁵C-modified RNA was initially reported in doi: 10.1038/cr.2017.55, (and not in doi: 10.1038/s41556-019-0361-y), where it was observed through PAR-CLIP analysis that the K171 mutation weakens its binding affinity to m⁵C -modified RNA.

Our previous experimental approach was not optimal: the protein concentration in the INPUT group was too high, leading to overexposure in the experimental group. Additionally, we did not conduct a quantitative analysis of the results at that time. In response to your suggestion, we performed RNA pull-down experiments with YBX1 and ALYREF, rather than with the pan-DRAM protein, to better validate and reproduce the previously reported findings. Our quantitative analysis revealed that both ALYREF and YBX1 exhibit a stronger affinity for m⁵C -modified RNAs. Furthermore, mutating the key amino acids involved in m⁵C recognition significantly reduced the binding affinity of both readers. These results align with previous studies (doi: 10.1038/cr.2017.55 and doi: 10.1038/s41556-019-0361-y), confirming that ALYREF and YBX1 are specific readers of m⁵C -modified RNAs. However, our detection system has certain limitations. Despite mutating the critical amino acids, both readers retained a weak binding affinity for m⁵C, suggesting that while the mutation helps reduce false positives, it is still challenging to precisely map the distribution of m⁵C modifications. To address this, we plan to further investigate the protein structure and function to obtain a more accurate m⁵C sequencing of the transcriptome in future studies. Accordingly, we have updated our results and conclusions in lines 294-299 and discuss these limitations in lines 109114.

In addition, while the m⁵C assay can be performed using only the DRAM system alone, comparing it with the DRAM^mut control enhances the accuracy of m⁵C region detection. To minimize the variations in transfection efficiency across experimental groups, it is recommended to use the same batch of transfections. This approach not only ensures more consistent results but also improve the standardization of the DRAM assay, as discussed in the section added on line 308-312.

(2) Using sodium arsenite treatment of cells as a means to change the m⁵C status of transcripts through the downregulation of the two major m⁵C writer proteins NSUN2 and NSUN6 is problematic and the conclusions from these experiments are not warranted. Sodium arsenite is a chemical that poisons every protein containing thiol groups. Not only do NSUN proteins contain cysteines but also the base editor fusion proteins. Arsenite will inactivate these proteins, hence the editing frequency will drop, as observed in the experiments shown in Figure 5, which the authors explain with fewer m⁵C sites to be detected by the fusion proteins.

The authors have not addressed the point made by this reviewer. Instead the authors state that they have not addressed that possibility. They claim that they have revised the results section, but this reviewer can only see the point raised in the conclusions. An experiment would have been to purify base editors via the HA tag and then perform some kind of binding/editing assay in vitro before and after arsenite treatment of cells.

We appreciate the reviewer’s insightful comment. We fully agree with the concern raised. In the original manuscript, our intention was to use sodium arsenite treatment to downregulate NSUN mediated m⁵C levels and subsequently decrease DRAM editing efficiency, with the aim of monitoring m⁵C dynamics through the DRAM system. However, as the reviewer pointed out, sodium arsenite may inactivate both NSUN proteins and the base editor fusion proteins, and any such inactivation would likely result in a reduced DRAM editing.

This confounds the interpretation of our experimental data.

As demonstrated in Author response image 1A, western blot analysis confirmed that sodium arsenite indeed decreased the expression of fusion proteins. In addition, we attempted in vitro fusion protein purificationusing multiple fusion tags (HIS, GST, HA, MBP) for DRAM fusion protein expression, but unfortunately, we were unable to obtain purified proteins. However, using the Promega TNT T7 Rapid Coupled In Vitro Transcription/Translation Kit, we successfully purified the DRAM protein (Author response image 1B). Despite this success, subsequent in vitro deamination experiments did not yield the expected mutation results (Author response image 1C), indicating that further optimization is required. This issue is further discussed in line 314-315.

Taken together, the above evidence supports that the experiment of sodium arsenite treatment was confusing and we determined to remove the corresponding results from the main text of the revised manuscript.

Author response image 1.

(3) The authors should move high-confidence editing site data contained in Supplementary Tables 2 and 3 into one of the main Figures to substantiate what is discussed in Figure 4A. However, the data needs to be visualized in another way then excel format. Furthermore, Supplementary Table 2 does not contain a description of the columns, while Supplementary Table 3 contains a single row with letters and numbers.

The authors have not addressed the point made by this reviewer. Figure 3F shows the screening process for DRAM-seq assays and principles for screening highconfidence genes rather than the data contained in Supplementary Tables 2 and 3 of the former version of this manuscript.

Thank you for your valuable suggestion. We have visualized the data from Supplementary Tables 2 and 3 in Figure 4A as a circlize diagram (described in lines 213-216), illustrating the distribution of mutation sites detected by the DRAM system across each chromosome. Additionally, to improve the presentation and clarity of the data, we have revised Supplementary Tables 2 and 3 by adding column descriptions, merging the DRAM-ABE and DRAM-CBE sites, and including overlapping m⁵C genes from previous datasets.

Responses to Reviewer’s Comments:  

To Reviewer #3:

The authors have again tried to address the former concern by this reviewer who questioned the specificity of both m⁵C reader proteins towards modified RNA rather than unmodified RNA. The authors chose to do RNA pull down experiments which serve as a proxy for proving the specificity of ALYREF and YBX1 for m⁵C modified RNAs. Even though this reviewer asked for determining the enrichment factor of the reader-base editor fusion proteins (as wildtype or mutant for the identified m⁵C specificity motif) when presented with m⁵C-modified RNAs, the authors chose to use both reader proteins alone (without the fusion to an editor) as wildtype and as respective m⁵C-binding mutant in RNA in vitro pull-down experiments along with unmodified and m⁵C-modified RNA oligomers as binding substrates. The quantification of these pull-down experiments (n=2) have now been added, and are revealing that (according to SFigure 1 E and G) YBX1 enriches an RNA containing a single m⁵C by a factor of 1.3 over its unmodified counterpart, while ALYREF enriches by a factor of 4x. This is an acceptable approach for educated readers to question the specificity of the reader proteins, even though the quantification should be performed differently (see below).

Given that there is no specific sequence motif embedding those cytosines identified in the vicinity of the DRAM-edits (Figure 3J and K), even though it has been accepted by now that most of the m⁵C sites in mRNA are mediated by NSUN2 and NSUN6 proteins, which target tRNA like substrate structures with a particular sequence enrichment, one can conclude that DRAM-Seq is uncovering a huge number of false positives. This must be so not only because of the RNA bisulfite seq data that have been extensively studied by others, but also by the following calculations: Given that the m⁵C/C ratio in human mRNA is 0.02-0.09% (measured by mass spec) and assuming that 1/4 of the nucleotides in an average mRNA are cytosines, an mRNA of 1.000 nucleotides would contain 250 Cs. 0.02- 0.09% m⁵C/C would then translate into 0.05-0.225 methylated cytosines per 250 Cs in a 1000 nt mRNA. YBX1 would bind every C in such an mRNA since there is no m⁵C to be expected, which it could bind with 1.3 higher affinity. Even if the mRNAs would be 10.000 nt long, YBX1 would bind to half a methylated cytosine or 2.25 methylated cytosines with 1.3x higher affinity than to all the remaining cytosines (2499.5 to 2497.75 of 2.500 cytosines in 10.000 nt, respectively). These numbers indicate a 4999x to 1110x excess of cytosine over m⁵C in any substrate RNA, which the "reader" can bind as shown in the RNA pull-downs on unmodified RNAs. This reviewer spares the reader of this review the calculations for ALYREF specificity, which is slightly higher than YBX1. Hence, it is up to the capable reader of these calculations to follow the claim that this minor affinity difference allows the unambiguous detection of the few m⁵C sites in mRNA be it in the endogenous scenario of a cell or as fusion-protein with a base editor attached? 

We sincerely appreciate the reviewer’s rigorous analysis. We would like to clarify that in our RNA pulldown assays, we indeed utilized the full DRAM system (reader protein fused to the base editor) to reflect the specificity of m⁵C recognition. As previously suggested by the reviewer, to independently validate the m⁵C-binding specificity of ALYREF and YBX1, we performed separate pulldown experiments with wild-type and mutant reader proteins (without the base editor fusion) using both unmodified and m⁵C-modified RNA substrates. This approach aligns with established methodologies in the field (doi:10.1038/cr.2017.55 and doi: 10.1038/s41556-019-0361-y). We have revised the Methods section (line 230) to explicitly describe this experimental design.

Although the m⁵C/C ratios in LC/MS-assayed mRNA are relatively low (ranging from 0.02% to 0.09%), as noted by the reviewer, both our data and previous studies have demonstrated that ALYREF and YBX1 preferentially bind to m⁵C-modified RNAs over unmodified RNAs, exhibiting 4-fold and 1.3-fold enrichment, respectively (Supplementary Figure 1E–1G). Importantly, this specificity is further enhanced in the DRAM system through two key mechanisms: first, the fusion of reader proteins to the deaminase restricts editing to regions near m⁵C sites, thereby minimizing off-target effects; second, background editing observed in reader-mutant or deaminase controls (e.g., DRAM^mut-CBE in Figure 2D) is systematically corrected for during data analysis.

We agree that the theoretical challenge posed by the vast excess of unmodified cytosines. However, our approach includes stringent controls to alleviate this issue. Specifically, sites identified in NSUN2/NSUN6 knockout cells or reader-mutant controls are excluded (Figure 3F), which significantly reduces the number of false-positive detections. Additionally, we have observed deamination changes near high-confidence m⁵C methylation sites detected by RNA bisulfite sequencing, both in first-generation and high-throughput sequencing data. This observation further substantiates the validity of DRAM-Seq in accurately identifying m⁵C sites.

We fully acknowledge that residual false positives may persist due to the inherent limitations of reader protein specificity, as discussed in line 299-301 of our manuscript. To address this, we plan to optimize reader domains with enhanced m⁵C binding (e.g., through structure-guided engineering), which is also previously implemented in the discussion of the manuscript.

The reviewer supports the attempt to visualize the data. However, the usefulness of this Figure addition as a readable presentation of the data included in the supplement is up to debate.

Thank you for your kind suggestion. We understand the reviewer's concern regarding data visualization. However, due to the large volume of DRAM-seq data, it is challenging to present each mutation site and its characteristics clearly in a single figure. Therefore, we chose to categorize the data by chromosome, which not only allows for a more organized presentation of the DRAM-seq data but also facilitates comparison with other database entries. Additionally, we have updated Supplementary Tables 2 and 3 to provide comprehensive information on the mutation sites. We hope that both the reviewer and editors will understand this approach. We will, of course, continue to carefully consider the reviewer's suggestions and explore better ways to present these results in the future.

(3) A set of private Recommendations for the Authors that outline how you think the science and its presentation could be strengthened

NEW COMMENTS to TEXT:

Abstract:

"5-Methylcytosine (m⁵C) is one of the major post-transcriptional modifications in mRNA and is highly involved in the pathogenesis of various diseases."

In light of the increasing use of AI-based writing, and the proof that neither DeepSeek nor ChatGPT write truthfully statements if they collect metadata from scientific abstracts, this sentence is utterly misleading.

m⁵C is not one of the major post-transcriptional modifications in mRNA as it is only present with a m⁵C/C ratio of 0.02- 0.09% as measured by mass-spec. Also, if m⁵C is involved in the pathogenesis of various diseases, it is not through mRNA but tRNA. No single published work has shown that a single m⁵C on an mRNA has anything to do with disease. Every conclusion that is perpetuated by copying the false statements given in the many reviews on the subject is based on knock-out phenotypes of the involved writer proteins. This reviewer wishes that the authors would abstain from the common practice that is currently flooding any scientific field through relentless repetitions in the increasing volume of literature which perpetuate alternative facts.

We sincerely appreciate the reviewer’s insightful comments. While we acknowledge that m⁵C is not the most abundant post-transcriptional modification in mRNA, we believe that research into m⁵C modification holds considerable value. Numerous studies have highlighted its role in regulating gene expression and its potential contribution to disease progression. For example, recent publications have demonstrated that m⁵C modifications in mRNA can influence cancer progression, lipid metabolism, and other pathological processes (e.g., PMID: 37845385; 39013911; 39924557; 38042059; 37870216).

We fully agree with the reviewer on the importance of maintaining scientific rigor in academic writing. While m⁵C is not the most abundant RNA modification, we cannot simply draw a conclusion that the level of modification should be the sole criterion for assessing its biological significance. However, to avoid potential confusion, we have removed the word “major”.

COMMENTS ON FIGURE PRESENTATION:

Figure 2D:

The main text states: "DRAM-CBE induced C to U editing in the vicinity of the m⁵C site in AP5Z1 mRNA, with 13.6% C-to-U editing, while this effect was significantly reduced with APOBEC1 or DRAM^mut-CBE (Fig.2D)." The Figure does not fit this statement. The seq trace shows a U signal of about 1/3 of that of C (about 30%), while the quantification shows 20+ percent

Thank you for your kind suggestion. Upon visual evaluation, the sequencing trace in the figure appears to suggest a mutation rate closer to 30% rather than 22%. However, relying solely on the visual interpretation of sequencing peaks is not a rigorous approach. The trace on the left represents the visualization of Sanger sequencing results using SnapGene, while the quantification on the right is derived from EditR 1.0.10 software analysis of three independent biological replicates. The C-to-U mutation rates calculated were 22.91667%, 23.23232%, and 21.05263%, respectively. To further validate this, we have included the original EditR analysis of the Sanger sequencing results for the DRAM-CBE group used in the left panel of Figure 2D (see Author response image 2). This analysis confirms an m⁵C fraction (%) of 22/(22+74) = 22.91667, and the sequencing trace aligns well with the mutation rate we reported in Figure 2D. In conclusion, the data and conclusions presented in Figure 2D are consistent and supported by the quantitative analysis.

Author response image 2.

Figure 4B: shows now different numbers in Venn-diagrams than in the same depiction, formerly Figure 4A

We sincerely thank the reviewer for pointing out this issue, and we apologize for not clearly indicating the changes in the previous version of the manuscript. In response to the initial round of reviewer comments, we implemented a more stringent data filtering process (as described in Figure 3F and method section) : "For high-confidence filtering, we further adjusted the parameters of Find_edit_site.pl to include an edit ratio of 10%–60%, a requirement that the edit ratio in control samples be at least 2-fold higher than in NSUN2 or NSUN6knockout samples, and at least 4 editing events at a given site." As a result, we made minor adjustments to the Venn diagram data in Figure 4A, reducing the total number of DRAM-edited mRNAs from 11,977 to 10,835. These changes were consistently applied throughout the manuscript, and the modifications have been highlighted for clarity. Importantly, these adjustments do not affect any of the conclusions presented in the manuscript.

Figure 4B and D: while the overlap of the DRAM-Seq data with RNA bisulfite data might be 80% or 92%, it is obvious that the remaining data DRAM seq suggests a detection of additional sites of around 97% or 81.83%. It would be advised to mention this large number of additional sites as potential false positives, unless these data were normalized to the sites that can be allocated to NSUN2 and NSUN6 activity (NSUN mutant data sets could be substracted).

Thank you for pointing this out. The Venn diagrams presented in Figure 4B and D already reflect the exclusion of potential false-positive sites identified in methyltransferasedeficient datasets, as described in our experimental filtering process, and they represent the remaining sites after this stringent filtering. However, we acknowledge that YBX1 and ALYREF, while preferentially binding to m⁵C-modified RNA, also exhibit some affinity for unmodified RNA. Although we employed rigorous controls, including DRAM^mut and deaminase groups, to minimize false positives, the possibility of residual false positives cannot be entirely ruled out. Addressing this limitation would require even more stringent filtering methods, as discussed in lines 299–301 of the manuscript. We are committed to further optimizing the DRAM system to enhance the accuracy of transcriptome-wide m⁵C analysis in future studies.

SFigure 1: It is clear that the wild type version of both reader proteins are robustly binding to RNA that does not contain m⁵C. As for the calculations of x-fold affinity loss of RNA binding using both ALYREF -mut or YBX1 -mut, this reviewer asks the authors to determine how much less the mutated versions of the proteins bind to a m⁵C-modified RNAs. Hence, a comparison of YBX1 versus YBX1 -mut (ALYREF versus ALYREF -mut) on the same substrate RNA with the same m⁵C-modified position would allow determining the contribution of the so-called modification binding pocket in the respective proteins to their RNA binding. The way the authors chose to show the data presently is misleading because what is compared is the binding of either the wild type or the mutant protein to different RNAs.

We appreciate the reviewer’s valuable feedback and apologize for any confusion caused by the presentation of our data. We would like to clarify the rationale behind our approach. The decision to present the wild-type and mutant reader proteins in separate panels, rather than together, was made in response to comments from Reviewer 2. Below, we provide a detailed explanation of our experimental design and its justification.

First, we confirmed that YBX1 and ALYREF exhibit stronger binding affinity to m⁵Cmodified RNA compared to unmodified RNA, establishing their role as m⁵C reader proteins. Next, to validate the functional significance of the DRAM^mut group, we demonstrated that mutating key amino acids in the m⁵C-binding pocket significantly reduces the binding affinity of YBX1^mut and ALYREF^mut to m⁵C-modified RNA. This confirms that the DRAM^mut group effectively minimizes false-positive results by disrupting specific m⁵C interactions.

Crucially, in our pull-down experiments, both the wild-type and mutant proteins (YBX1/YBX1^mut and ALYREF/ALYREF^mut) were incubated with the same RNA sequences. To avoid any ambiguity, we have included the specific RNA sequence information in the Methods section (lines 463–468). This ensures a assessment of the reduced binding affinity of the mutant versions relative to the wild-type proteins, even though they are presented in separate panels.

We hope this explanation clarifies our approach and demonstrates the robustness of our findings. We sincerely appreciate the reviewer’s understanding and hope this addresses their concerns.

SFigure 2C: first two panels are duplicates of the same image.

Thank you for pointing this out. We sincerely apologize for incorrectly duplicating the images. We have now updated Supplementary Figure 2C with the correct panels and have provided the original flow cytometry data for the first two images. It is important to note that, as demonstrated by the original data analysis, the EGFP-positive quantification values (59.78% and 59.74%) remain accurate. Therefore, this correction does not affect the conclusions of our study. Thank you again for bringing this to our attention.

Author response image 3.

SFigure 4B: how would the PCR product for NSUN6 be indicative of a mutation? The used primers seem to amplify the wildtype sequence.

Thank you for your kind suggestion. In our NSUN6^-/- cell line, the NSUN6 gene is only missing a single base pair (1bp) compared to the wildtype, which results in frame shift mutation and reduction in NSUN6 protein expression. We fully agree with the reviewer that the current PCR gel electrophoresis does not provide a clear distinction of this 1bp mutation. To better illustrate our experimental design, we have included a schematic representation of the knockout sequence in SFigure 4B. Additionally, we have provided the original sequencing data, and the corresponding details have been added to lines 151-153 of the manuscript for further clarification.

Author response image 4.

SFigure 4C: the Figure legend is insufficient to understand the subfigure.

Thank you for your valuable suggestion. To improve clarity, we have revised the figure legend for SFigure 4C, as well as the corresponding text in lines 178-179. We have additionally updated the title of SFigure 4 for better clarity. The updated SFigure 4C now demonstrates that the DRAM-edited mRNAs exhibit a high degree of overlap across the three biological replicates.

SFigure 4D: the Figure legend is insufficient to understand the subfigure.

Thank you for your kind suggestion. We have revised the figure legend to provide a clearer explanation of the subfigure. Specifically, this figure illustrates the motif analysis derived from sequences spanning 10 nucleotides upstream and downstream of DRAMedited sites mediated by loci associated with NSUN2 or NSUN6. To enhance clarity, we have also rephrased the relevant results section (lines 169-175) and the corresponding discussion (lines 304-307).

SFigure 7: There is something off with all 6 panels. This reviewer can find data points in each panel that do not show up on the other two panels even though this is a pairwise comparison of three data sets (file was sent to the Editor) Available at https://elife-rp.msubmit.net/elife-rp_files/2025/01/22/00130809/02/130809_2_attach_27_15153.pdf

Response: We thank the reviewer for pointing this out. We would like to clarify the methodology behind this analysis. In this study, we conducted pairwise comparisons of the number of DRAM-edited sites per gene across three biological replicates of DRAM-ABE or DRAM-CBE, visualized as scatterplots. Each data point in the plots corresponds to a gene, and while the same gene is represented in all three panels, its position may vary vertically or horizontally across the panels. This variation arises because the number of mutation sites typically differs between replicates, making it unlikely for a data point to occupy the exact same position in all panels. A similar analytical approach has been used in previous studies on m6A (PMID: 31548708). To address the reviewer’s concern, we have annotated the corresponding positions of the questioned data points with arrows in Author response image 5.

Author response image 5.

Author Response

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public Review):

In this paper, the authors developed an image analysis pipeline to automacally idenfy individual neurons within a populaon of fluorescently tagged neurons. This applicaon is opmized to deal with mul-cell analysis and builds on a previous soware version, developed by the same team, to resolve individual neurons from whole-brain imaging stacks. Using advanced stascal approaches and several heuriscs tailored for C. elegans anatomy, the method successfully idenfies individual neurons with a fairly high accuracy. Thus, while specific to C. elegans, this method can become instrumental for a variety of research direcons such as in-vivo single-cell gene expression analysis and calcium-based neural acvity studies.

Thank you.

Reviewer #2 (Public Review):

The authors succeed in generalizing the pre-alignment procedure for their cell idenficaon method to allow it to work effecvely on data with only small subsets of cells labeled. They convincingly show that their extension accurately idenfies head angle, based on finding auto florescent ssue and looking for a symmetric l/r axis. They demonstrate method works to allow the idenficaon of a parcular subset of neurons. Their approach should be a useful one for researchers wishing to idenfy subsets of head neurons in C. elegans, and the ideas might be useful elsewhere.

The authors also assess the relave usefulness of several atlases for making identy predicons. They atempt to give some addional general insights on what makes a good atlas, but here insights seem less clear as available data does not allow for experiments that cleanly decouple: 1. the number of examples in the atlas 2. the completeness of the atlas. and 3. the match in strain and imaging modality discussed. In the presented experiments the custom atlas, besides the strain and imaging modality mismatches discussed is also the only complete atlas with more than one example. The neuroPAL atlas, is an imperfect stand in, since a significant fracon of cells could not be idenfied in these data sets, making it a 60/40 mix of Openworm and a hypothecal perfect neuroPAL comparison. This waters down general insights since it is unclear if the performance is driven by strain/imaging modality or these difficules creang a complete neuroPal atlas. The experiments do usefully explore the volume of data needed. Though generalizaon remains to be shown the insight is useful for future atlas building that for the specific (small) set of cells labeled in the experiments 5-10 examples is sufficient to build a accurate atlas.

The reviewer brings up an interesting point. As the reviewer noted, given the imperfection of the datasets (ours and others’), it is possible that artifacts from incomplete atlases can interfere with the assessment of the performances of different atlases. To address this, as the reviewer suggested, we have searched the literature and found two sets of data that give specific coordinates of identified neurons (both using NeuroPAL). We compared the performance of the atlases derived from these datasets to the strain-specific atlases, and the original conclusion stands. Details are now included in the revised manuscript (Figure 3- figure supplement 2).

Recommendaons for the authors:

Reviewer #1 (Recommendaons For The Authors):

I appreciate the new mosaic analysis (Fig. 3 -figure suppl 2). Please fix the y-axis ck label that I believe should be 0.8 (instead of 0.9).

We thank the reviewer for spotting the typo. We have fixed the error.

**Reviewer #2 (Recommendaons For The Authors):

Though I'm not familiar with the exact quality of GT labels in available neuroPAL data I know increasing volumes of published data is available. Comparison with a complete neuroPAL atlas, and a similar assessment on atlas size as made with the custom atlas would to my mind qualitavely increase the general insights on atlas construcon.

We thank the reviewer for the insightful suggestion. We have newly constructed several other NeuroPAL atlases by incorporating neuron positional data from two other published data: [Yemini E. et al. NeuroPAL: A Multicolor Atlas for Whole-Brain Neuronal Identification in C. elegans. Cell. 2021 Jan 7;184(1):272-288.e11] and [Skuhersky, M. et al. Toward a more accurate 3D atlas of C. elegans neurons. BMC Bioinformatics 23, 195 (2022)].

Interestingly, we found that the two new atlases (NP-Yemini and NP-Skuhersky) have significantly different values of PA, LR, DV, and angle relationships for certain cells compared to the OpenWorm and glr-1 atlases. For example, in both the NP atlases, SMDD is labeled as being anterior to AIB, which is the opposite of the SMDD-AIB relationship in the glr-1 atlas.

Because this relationship (and other similar cases) were missing in our original NeuroPAL atlas (NP-Chaudhary), the addition of these two NeuroPAL datasets to our NeuroPAL atlas dramatically changed the atlas. As a result, incorporating the published data sets into the NeuroPAL atlas (NP-all) actually decreased the average prediction accuracy to 44%, while the average accuracy of original NeuroPAL atlas (NP-Chaudhary) was 57%. The atlas based on the Yemini et al. data alone (NP-Yemini) had 43% accuracy, and the atlas based on the Skuhersky et al. data alone (NP-Skuhersky) had 38% accuracy.

For the rest of our analysis, we focused on comparing the NeuroPAL atlas that resulted in the highest accuracy against other atlases in figure 3 (NP-Chaudhary). Therefore, we have added Figure 3- figure supplement 2 and the following sentence in the discussion. “Several other NeuroPAL atlases from different data sources were considered, and the atlas that resulted in the highest neuron ID correspondence was selected (Figure 3- figure supplement 2).”

Author response image 1.

Figure3- figure supplement 2. Comparison of neuron ID correspondences resulng from addional atlases- atlases driven from NeuroPAL neuron posional data from mulple sources (Chaudhary et al., Yemini et al., and Skuhersky et al.) in red compared to other atlases in Figure 3. Two sample t-tests were performed for stascal analysis. The asterisk symbol denotes a significance level of p<0.05, and n.s. denotes no significance. OW: atlas driven by data from OpenWorm project, NP-source: NeuroPAL atlas driven by data from the source. NP-Chaudhary atlas corresponds to NeuroPAL atlas in Figure 3.

80% agreement among manual idenficaons seems low to me for a relavely small, (mostly) known set of cells, which seems to cast into doubt ground truth idenes based on a best 2 out of 3 vote. The authors menon 3% of cell idenes had total disagreement and were excluded, what were the fracon unanimous and 2/3? Are there any further insights about what limited human performance in the context of this parcular idenficaon task?

We closely looked into the manual annotation data. The fraction of cells in unanimous, two thirds, and no agreement are approximately 74%, 20%, and 6%, respectively. We made the corresponding change in the manuscript from 3% to 6%. Indeed, we identified certain patterns in labels that were more likely to be disagreed upon. First, cells in close proximity to each other, such as AVE and RMD, were often switched from annotator to annotator. Second, cells in the posterior part of the cluster, such as RIM, AVD, AVB, were more variable in positions, so their identities were not clear at times. Third, annotators were more likely to disagree on cells whose expressions are rare and low, and these include AIB, AVJ, and M1. These observations agree with our results in figure 4c.

Author Response

The following is the authors’ response to the previous reviews.

To the Senior Editor and the Reviewing Editor:

We sincerely appreciate the valuable comments provided by the reviewers, the reviewing editor, and the senior editor. After carefully reviewing and considering the comments, we have addressed the key concerns raised by the reviewers and made appropriate modifications to the article in the revised manuscript.

The main revisions made to the manuscript are as follows:

1) We have added comparison experiments with TNDM (see Fig. 2 and Fig. S2).

2) We conducted new synthetic experiments to demonstrate that our conclusions are not a by-product of d-VAE (see Fig. S2 and Fig. S11).

3) We have provided a detailed explanation of how our proposed criteria, especially the second criterion, can effectively exclude the selection of unsuitable signals.

4) We have included a semantic overview figure of d-VAE (Fig. S1) and a visualization plot of latent variables (Fig. S13).

5) We have elaborated on the model details of d-VAE, as well as the hyperparameter selection and experimental settings of other comparison models.

We believe these revisions have significantly improved the clarity and comprehensibility of the manuscript. Thank you for the opportunity to address these important points.

Reviewer #1

Q1: “First, the model in the paper is almost identical to an existing VAE model (TNDM) that makes use of weak supervision with behaviour in the same way [1]. This paper should at least be referenced. If the authors wish they could compare their model to TNDM, which combines a state space model with smoothing similar to LFADS. Given that TNDM achieves very good behaviour reconstructions, it may be on par with this model without the need for a Kalman filter (and hence may achieve better separation of behaviour-related and unrelated dynamics).”

Our model significantly differs from TNDM in several aspects. While TNDM also constrains latent variables to decode behavioral information, it does not impose constraints to maximize behavioral information in the generated relevant signals. The trade-off between the decoding and reconstruction capabilities of generated relevant signals is the most significant contribution of our approach, which is not reflected in TNDM. In addition, the backbone network of signal extraction and the prior distribution of the two models are also different.

It's worth noting that our method does not require a Kalman filter. Kalman filter is used for post hoc assessment of the linear decoding ability of the generated signals. Please note that extracting and evaluating relevant signals are two distinct stages.

Heeding your suggestion, we have incorporated comparison experiments involving TNDM into the revised manuscript. Detailed information on model hyperparameters and training settings can be found in the Methods section in the revised manuscripts.

Thank you for your valuable feedback.

Q2: “Second, in my opinion, the claims regarding identifiability are overstated - this matters as the results depend on this to some extent. Recent work shows that VAEs generally suffer from identifiability problems due to the Gaussian latent space [2]. This paper also hints that weak supervision may help to resolve such issues, so this model as well as TNDM and CEBRA may indeed benefit from this. In addition however, it appears that the relative weight of the KL Divergence in the VAE objective is chosen very small compared to the likelihood (0.1%), so the influence of the prior is weak and the model may essentially learn the average neural trajectories while underestimating the noise in the latent variables. This, in turn, could mean that the model will not autoencode neural activity as well as it should, note that an average R2 in this case will still be high (I could not see how this is actually computed). At the same time, the behaviour R2 will be large simply because the different movement trajectories are very distinct. Since the paper makes claims about the roles of different neurons, it would be important to understand how well their single trial activities are reconstructed, which can perhaps best be investigated by comparing the Poisson likelihood (LFADS is a good baseline model). Taken together, while it certainly makes sense that well-tuned neurons contribute more to behaviour decoding, I worry that the very interesting claim that neurons with weak tuning contain behavioural signals is not well supported.”

We don’t think our distilled signals are average neural trajectories without variability. The quality of reconstructing single trial activities can be observed in Figure 3i and Figure S4. Neural trajectories in Fig. 3i and Fig. S4 show that our distilled signals are not average neural trajectories. Furthermore, if each trial activity closely matched the average neural trajectory, the Fano Factor (FF) should theoretically approach 0. However, our distilled signals exhibit a notable departure from this expectation, as evident in Figure 3c, d, g, and f. Regarding the diminished influence of the KL Divergence: Given that the ground truth of latent variable distribution is unknown, even a learned prior distribution might not accurately reflect the true distribution. We found the pronounced impact of the KL divergence would prove detrimental to the decoding and reconstruction performance. As a result, we opt to reduce the weight of the KL divergence term. Even so, KL divergence can still effectively align the distribution of latent variables with the distribution of prior latent variables, as illustrated in Fig. S13. Notably, our goal is extracting behaviorally-relevant signals from given raw signals rather than generating diverse samples from the prior distribution. When aim to separating relevant signals, we recommend reducing the influence of KL divergence. Regarding comparing the Poisson likelihood: We compared Poisson log-likelihood among different methods (except PSID since their obtained signals have negative values), and the results show that d-VAE outperforms other methods.

Author response image 1.

Regarding how R2 is computed: , where and denote ith sample of raw signals, ith sample of distilled relevant signals, and the mean of raw signals. If the distilled signals exactly match the raw signals, the sum of squared error is zero, thus R2=1. If the distilled signals always are equal to R2=0. If the distilled signals are worse than the mean estimation, R2 is negative, negative R2 is set to zero.

Thank you for your valuable feedback.

Q3: “Third, and relating to this issue, I could not entirely follow the reasoning in the section arguing that behavioural information can be inferred from neurons with weak selectivity, but that it is not linearly decodable. It is right to test if weak supervision signals bleed into the irrelevant subspace, but I could not follow the explanations. Why, for instance, is the ANN decoder on raw data (I assume this is a decoder trained fully supervised) not equal in performance to the revenant distilled signals? Should a well-trained non-linear decoder not simply yield a performance ceiling? Next, if I understand correctly, distilled signals were obtained from the full model. How does a model perform trained only on the weakly tuned neurons? Is it possible that the subspaces obtained with the model are just not optimally aligned for decoding? This could be a result of limited identifiability or model specifics that bias reconstruction to averages (a well-known problem of VAEs). I, therefore, think this analysis should be complemented with tests that do not depend on the model.”

Regarding “Why, for instance, is the ANN decoder on raw data (I assume this is a decoder trained fully supervised) not equal in performance to the relevant distilled signals? Should a well-trained non-linear decoder not simply yield a performance ceiling?”: In fact, the decoding performance of raw signals with ANN is quite close to the ceiling. However, due to the presence of significant irrelevant signals in raw signals, decoding models like deep neural networks are more prone to overfitting when trained on noisy raw signals compared to behaviorally-relevant signals. Consequently, we anticipate that the distilled signals will demonstrate superior decoding generalization. This phenomenon is evident in Fig. 2 and Fig. S1, where the decoding performance of the distilled signals surpasses that of the raw signals, albeit not by a substantial margin.

Regarding “Next, if I understand correctly, distilled signals were obtained from the full model. How does a model perform trained only on the weakly tuned neurons? Is it possible that the subspaces obtained with the model are just not optimally aligned for decoding?”：Distilled signals (involving all neurons) are obtained by d-VAE. Subsequently, we use ANN to evaluate the performance of smaller and larger R2 neurons. Please note that separating and evaluating relevant signals are two distinct stages.

Regarding the reasoning in the section arguing that smaller R2 neurons encode rich information, we would like to provide a detailed explanation:

1) After extracting relevant signals through d-VAE, we specifically selected neurons characterized by smaller R2 values (Here, R2 signifies the proportion of neuronal activity variance explained by the linear encoding model, calculated using raw signals). Subsequently, we employed both KF and ANN to assess the decoding performance of these neurons. Remarkably, our findings revealed that smaller R2 neurons, previously believed to carry limited behavioral information, indeed encode rich information.

2) In a subsequent step, we employed d-VAE to exclusively distill the raw signals of these smaller R2 neurons (distinct from the earlier experiment where d-VAE processed signals from all neurons). We then employed KF and ANN to evaluate the distilled smaller R2 neurons. Interestingly, we observed that we could not attain the same richness of information solely through the use of these smaller R2 neurons.

3) Consequently, we put forth and tested two hypotheses: First, that larger R2 neurons introduce additional signals into the smaller R2 neurons that do not exist in the real smaller R2 neurons. Second, that larger R2 neurons aid in restoring the original appearance of impaired smaller R2 neurons. Our proposed criteria and synthetic experiments substantiate the latter scenario.

Thank you for your valuable feedback.

Q4: “Finally, a more technical issue to note is related to the choice to learn a non-parametric prior instead of using a conventional Gaussian prior. How is this implemented? Is just a single sample taken during a forward pass? I worry this may be insufficient as this would not sample the prior well, and some other strategy such as importance sampling may be required (unless the prior is not relevant as it weakly contributed to the ELBO, in which case this choice seems not very relevant). Generally, it would be useful to see visualisations of the latent variables to see how information about behaviour is represented by the model.”

Regarding "how to implement the prior?": Please refer to Equation 7 in the revised manuscript; we have added detailed descriptions in the revised manuscript.

Regarding "Generally, it would be useful to see visualizations of the latent variables to see how information about behavior is represented by the model.": Note that our focus is not on latent variables but on distilled relevant signals. Nonetheless, at your request, we have added the visualization of latent variables in the revised manuscript. Please see Fig. S13 for details.

Thank you for your valuable feedback.

Recommendations: “A minor point: the word 'distill' in the name of the model may be a little misleading - in machine learning the term refers to the construction of smaller models with the same capabilities.

It should be useful to add a schematic picture of the model to ease comparison with related approaches.”

In the context of our model's functions, it operates as a distillation process, eliminating irrelevant signals and retaining the relevant ones. Although the name of our model may be a little misleading, it faithfully reflects what our model does.

I have added a schematic picture of d-VAE in the revised manuscript. Please see Fig. S1 for details.

Thank you for your valuable feedback.

Reviewer #2

Q1: “Is the apparently increased complexity of encoding vs decoding so unexpected given the entropy, sparseness, and high dimensionality of neural signals (the "encoding") compared to the smoothness and low dimensionality of typical behavioural signals (the "decoding") recorded in neuroscience experiments? This is the title of the paper so it seems to be the main result on which the authors expect readers to focus. ”

We use the term "unexpected" due to the disparity between our findings and the prior understanding concerning neural encoding and decoding. For neural encoding, as we said in the Introduction, in previous studies, weakly-tuned neurons are considered useless, and smaller variance PCs are considered noise, but we found they encode rich behavioral information. For neural decoding, the nonlinear decoding performance of raw signals is significantly superior to linear decoding. However, after eliminating the interference of irrelevant signals, we found the linear decoding performance is comparable to nonlinear decoding. Rooted in these findings, which counter previous thought, we employ the term "unexpected" to characterize our observations.

Thank you for your valuable feedback.

Q2: “I take issue with the premise that signals in the brain are "irrelevant" simply because they do not correlate with a fixed temporal lag with a particular behavioural feature hand-chosen by the experimenter. As an example, the presence of a reward signal in motor cortex [1] after the movement is likely to be of little use from the perspective of predicting kinematics from time-bin to time-bin using a fixed model across trials (the apparent definition of "relevant" for behaviour here), but an entire sub-field of neuroscience is dedicated to understanding the impact of these reward-related signals on future behaviour. Is there method sophisticated enough to see the behavioural "relevance" of this brief, transient, post-movement signal? This may just be an issue of semantics, and perhaps I read too much into the choice of words here. Perhaps the authors truly treat "irrelevant" and "without a fixed temporal correlation" as synonymous phrases and the issue is easily resolved with a clarifying parenthetical the first time the word "irrelevant" is used. But I remain troubled by some claims in the paper which lead me to believe that they read more deeply into the "irrelevancy" of these components.”

In this paper, we employ terms like ‘behaviorally-relevant’ and ‘behaviorally-irrelevant’ only regarding behavioral variables of interest measured within a given task, such as arm kinematics during a motor control task. A similar definition can be found in the PSID[1].

Thank you for your valuable feedback.

[1] Sani, Omid G., et al. "Modeling behaviorally relevant neural dynamics enabled by preferential subspace identification." Nature Neuroscience 24.1 (2021): 140-149.

Q3: “The authors claim the "irrelevant" responses underpin an unprecedented neuronal redundancy and reveal that movement behaviors are distributed in a higher-dimensional neural space than previously thought." Perhaps I just missed the logic, but I fail to see the evidence for this. The neural space is a fixed dimensionality based on the number of neurons. A more sparse and nonlinear distribution across this set of neurons may mean that linear methods such as PCA are not effective ways to approximate the dimensionality. But ultimately the behaviourally relevant signals seem quite low-dimensional in this paper even if they show some nonlinearity may help.”

The evidence for the “useless” responses underpin an unprecedented neuronal redundancy is shown in Fig. 5a, d and Fig. S9a. Specifically, the sum of the decoding performance of smaller R2 neurons and larger R2 neurons is significantly greater than that of all neurons for relevant signals (red bar), demonstrating that movement parameters are encoded very redundantly in neuronal population. In contrast, we can not find this degree of neural redundancy in raw signals (purple bar).

The evidence for the “useless” responses reveal that movement behaviors are distributed in a higher-dimensional neural space than previously thought is shown in the left plot (involving KF decoding) of Fig. 6c, f and Fig. S9f. Specifically, the improvement of KF using secondary signals is significantly higher than using raw signals composed of the same number of dimensions as the secondary signals. These results demonstrate that these dimensions, spanning roughly from ten to thirty, encode much information, suggesting that behavioral information exists in a higher-dimensional subspace than anticipated from raw signals.

Thank you for your valuable feedback.

Q5: “there is an apparent logical fallacy that begins in the abstract and persists in the paper: "Surprisingly, when incorporating often-ignored neural dimensions, behavioral information can be decoded linearly as accurately as nonlinear decoding, suggesting linear readout is performed in motor cortex." Don't get me wrong: the equivalency of linear and nonlinear decoding approaches on this dataset is interesting, and useful for neuroscientists in a practical sense. However, the paper expends much effort trying to make fundamental scientific claims that do not feel very strongly supported. This reviewer fails to see what we can learn about a set of neurons in the brain which are presumed to "read out" from motor cortex. These neurons will not have access to the data analyzed here. That a linear model can be conceived by an experimenter does not imply that the brain must use a linear model. The claim may be true, and it may well be that a linear readout is implemented in the brain. Other work [2,3] has shown that linear readouts of nonlinear neural activity patterns can explain some behavioural features. The claim in this paper, however, is not given enough”

Due to the limitations of current observational methods and our incomplete understanding of brain mechanisms, it is indeed challenging to ascertain the specific data the brain acquires to generate behavior and whether it employs a linear readout. Conventionally, the neural data recorded in the motor cortex do encode movement behaviors and can be used to analyze neural encoding and decoding. Based on these data, we found that the linear decoder KF achieves comparable performance to that of the nonlinear decoder ANN on distilled relevant signals. This finding has undergone validation across three widely used datasets, providing substantial evidence. Furthermore, we conducted experiments on synthetic data to show that this conclusion is not a by-product of our model. In the revised manuscript, we added a more detailed description of this conclusion.

Thank you for your valuable feedback.

Q6: “Relatedly, I would like to note that the exercise of arbitrarily dividing a continuous distribution of a statistic (the "R2") based on an arbitrary threshold is a conceptually flawed exercise. The authors read too much into the fact that neurons which have a low R2 w.r.t. PDs have behavioural information w.r.t. other methods. To this reviewer, it speaks more about the irrelevance, so to speak, of the preferred direction metric than anything fundamental about the brain.”

We chose the R2 threshold in accordance with the guidelines provided in reference [1]. It's worth mentioning that this threshold does not exert any significant influence on the overall conclusions.

Thank you for your valuable feedback.

[1] Inoue, Y., Mao, H., Suway, S.B., Orellana, J. and Schwartz, A.B., 2018. Decoding arm speed during reaching. Nature communications, 9(1), p.5243.

Q7: “I am afraid I may be missing something, as I did not understand the fano factor analysis of Figure 3. In a sense the behaviourally relevant signals must have lower FF given they are in effect tied to the temporally smooth (and consistent on average across trials) behavioural covariates. The point of the original Churchland paper was to show that producing a behaviour squelches the variance; naturally these must appear in the behaviourally relevant components. A control distribution or reference of some type would possibly help here.”

We agree that including reference signals could provide more context. The Churchland paper said stimulus onset can lead to a reduction in neural variability. However, our experiment focuses specifically on the reaching process, and thus, we don't have comparative experiments involving different types of signals.

Thank you for your valuable feedback.

Q8: “The authors compare the method to LFADS. While this is a reasonable benchmark as a prominent method in the field, LFADS does not attempt to solve the same problem as d-VAE. A better and much more fair comparison would be TNDM [4], an extension of LFADS which is designed to identify behaviourally relevant dimensions.”

We have added the comparison experiments with TNDM in the revised manuscript (see Fig. 2 and Fig. S2). The details of model hyperparameters and training settings can be found in the Methods section in the revised manuscripts.

Thank you for your valuable feedback.

Reviewer #3

Q1.1: “TNDM: LFADS is not the best baseline for comparison. The authors should have compared with TNDM (Hurwitz et al. 2021), which is an extension of LFADS that (unlike LFADS) actually attempts to extract behaviorally relevant factors by adding a behavior term to the loss. The code for TNDM is also available on Github. LFADS is not even supervised by behavior and does not aim to address the problem that d-VAE aims to address, so it is not the most appropriate comparison. ”

We have added the comparison experiments with TNDM in the revised manuscript (see Fig. 2 and Fig. S2). The details of model hyperparameters and training settings can be found in the Methods section in the revised manuscripts.

Thank you for your valuable feedback.

Q1.2: “LFADS: LFADS is a sequential autoencoder that processes sections of data (e.g. trials). No explanation is given in Methods for how the data was passed to LFADS. Was the moving averaged smoothed data passed to LFADS or the raw spiking data (at what bin size)? Was a gaussian loss used or a poisson loss? What are the trial lengths used in each dataset, from which part of trials? For dataset C that has back-to-back reaches, was data chopped into segments? How long were these segments? Were the edges of segments overlapped and averaged as in (Keshtkaran et al. 2022) to avoid noisy segment edges or not? These are all critical details that are not explained. The same details would also be needed for a TNDM comparison (comment 1.1) since it has largely the same architecture as LFADS.

It is also critical to briefly discuss these fundamental differences between the inputs of methods in the main text. LFADS uses a segment of data whereas VAE methods just use one sample at a time. What does this imply in the results? I guess as long as VAEs outperform LFADS it is ok, but if LFADS outperforms VAEs in a given metric, could it be because it received more data as input (a whole segment)? Why was the factor dimension set to 50? I presume it was to match the latent dimension of the VAE methods, but is the LFADS factor dimension the correct match for that to make things comparable?

I am also surprised by the results. How do the authors justify LFADS having lower neural similarity (fig 2d) than VAE methods that operate on single time steps? LFADS is not supervised by behavior, so of course I don't expect it to necessarily outperform methods on behavior decoding. But all LFADS aims to do is to reconstruct the neural data so at least in this metric it should be able to outperform VAEs that just operate on single time steps? Is it because LFADS smooths the data too much? This is important to discuss and show examples of. These are all critical nuances that need to be discussed to validate the results and interpret them.”

Regarding “Was the moving averaged smoothed data passed to LFADS or the raw spiking data (at what bin size)? Was a gaussian loss used or a poisson loss?”: The data used by all models was applied to the same preprocessing procedure. That is, using moving averaged smoothed data with three bins, where the bin size is 100ms. For all models except PSID, we used a Poisson loss.

Regrading “What are the trial lengths used in each dataset, from which part of trials? For dataset C that has back-to-back reaches, was data chopped into segments? How long were these segments? Were the edges of segments overlapped and averaged as in (Keshtkaran et al. 2022) to avoid noisy segment edges or not?”:

For datasets A and B, a trial length of eighteen is set. Trials with lengths below the threshold are zero-padded, while trials exceeding the threshold are truncated to the threshold length from their starting point. In dataset A, there are several trials with lengths considerably longer than that of most trials. We found that padding all trials with zeros to reach the maximum length (32) led to poor performance. Consequently, we chose a trial length of eighteen, effectively encompassing the durations of most trials and leading to the removal of approximately 9% of samples. For dataset B (center-out), the trial lengths are relatively consistent with small variation, and the maximum length across all trials is eighteen. For dataset C, we set the trial length as ten because we observed the video of this paradigm and found that the time for completing a single trial was approximately one second. The segments are not overlapped.

Regarding “Why was the factor dimension set to 50? I presume it was to match the latent dimension of the VAE methods, but is the LFADS factor dimension the correct match for that to make things comparable?”: We performed a grid search for latent dimensions in {10,20,50} and found 50 is the best.

Regarding “I am also surprised by the results. How do the authors justify LFADS having lower neural similarity (fig 2d) than VAE methods that operate on single time steps? LFADS is not supervised by behavior, so of course I don't expect it to necessarily outperform methods on behavior decoding. But all LFADS aims to do is to reconstruct the neural data so at least in this metric it should be able to outperform VAEs that just operate on single time steps? Is it because LFADS smooths the data too much?”: As you pointed out, we found that LFADS tends to produce excessively smooth and consistent data, which can lead to a reduction in neural similarity.

Thank you for your valuable feedback.

Q1.3: “PSID: PSID is linear and uses past input samples to predict the next sample in the output. Again, some setup choices are not well justified, and some details are left out in the 1-line explanation given in Methods.

Why was a latent dimension of 6 chosen? Is this the behaviorally relevant latent dimension or the total latent dimension (for the use case here it would make sense to set all latent states to be behaviorally relevant)? Why was a horizon hyperparameter of 3 chosen? First, it is important to mention fundamental parameters such as latent dimension for each method in the main text (not just in methods) to make the results interpretable. Second, these hyperparameters should be chosen with a grid search in each dataset (within the training data, based on performance on the validation part of the training data), just as the authors do for their method (line 779). Given that PSID isn't a deep learning method, doing a thorough grid search in each fold should be quite feasible. It is important that high values for latent dimension and a wider range of other hyperparmeters are included in the search, because based on how well the residuals (x_i) for this method are shown predict behavior in Fig 2, the method seems to not have been used appropriately. I would expect ANN to improve decoding for PSID versus its KF decoding since PSID is fully linear, but I don't expect KF to be able to decode so well using the residuals of PSID if the method is used correctly to extract all behaviorally relevant information from neural data. The low neural reconstruction in Fid 2d could also partly be due to using too small of a latent dimension.

Again, another import nuance is the input to this method and how differs with the input to VAE methods. The learned PSID model is a filter that operates on all past samples of input to predict the output in the "next" time step. To enable a fair comparison with VAE methods, the authors should make sure that the last sample "seen" by PSID is the same as then input sample seen by VAE methods. This is absolutely critical given how large the time steps are, otherwise PSID might underperform simply because it stopped receiving input 300ms earlier than the input received by VAE methods. To fix this, I think the authors can just shift the training and testing neural time series of PSID by 1 sample into the past (relative to the behavior), so that PSID's input would include the input of VAE methods. Otherwise, VAEs outperforming PSID is confounded by PSID's input not including the time step that was provided to VAE.”

Thanks for your suggestions for letting PSID see the current neural observations. We did it per your suggestions and then performed a grid search for the hyperparameters for PSID. Specifically, we performed a grid search for the horizon hyperparameter in {2,3,4,5,6,7}. Since the relevant latent dimension should be lower than the horizon times the dimension of behavior variables (two-dimensional velocity in this paper) and increasing the dimension will reach performance saturation, we directly set the relevant latent dimensions as the maximum. The horizon number of datasets A, B, C, and synthetic datasets is 7, 6, 6 and 5, respectively.

And thus the latent dimension of datasets A, B, and C and the synthetic dataset is 14, 12, 12 and 10, respectively.

Our experiments show that KF can decode information from irrelevant signals obtained by PSID. Although PSID extracts the linear part of raw signals, KF can still use the linear part of the residuals for decoding. The low reconstruction performance of PSID may be because the relationship between latent variables and neural signals is linear, and the relationship between latent variables and behaviors is also linear; this is equivalent to the linear relationship between behaviors and neural signals, and linear models can only explain a small fraction of neural signals.

Thank you for your valuable feedback.

Q1.4: “CEBRA: results for CEBRA are incomplete. Similarity to raw signals is not shown. Decoding of behaviorally irrelevant residuals for CEBRA is not shown. Per Fig. S2, CEBRA does better or similar ANN decoding in datasets A and C, is only slightly worse in Dataset B, so it is important to show the other key metrics otherwise it is unclear whether d-VAE has some tangible advantage over CEBRA in those 2 datasets or if they are similar in every metric. Finally, it would be better if the authors show the results for CEBRA on Fig. 2, just as is done for other methods because otherwise it is hard to compare all methods.”

CEBRA is a non-generative model, this model cannot generate behaviorally-relevant signals. Therefore, we only compared the decoding performance of latent embeddings of CEBRA and signals of d-VAE.

Thank you for your valuable feedback.

Q2: “Given the fact that d-VAE infers the latent (z) based on the population activity (x), claims about properties of the inferred behaviorally relevant signals (x_r) that attribute properties to individual neurons are confounded.

The authors contrast their approach to population level approaches in that it infers behaviorally relevant signals for individual neurons. However, d-VAE is also a population method as it aggregates population information to infer the latent (z), from which behaviorally relevant part of the activity of each neuron (x_r) is inferred. The authors note this population level aggregation of information as a benefit of d-VAE, but only acknowledge it as a confound briefly in the context of one of their analyses (line 340): "The first is that the larger R2 neurons leak their information to the smaller R2 neurons, causing them contain too much behavioral information". They go on to dismiss this confounding possibility by showing that the inferred behaviorally relevant signal of each neuron is often most similar to its own raw signals (line 348-352) compared with all other neurons. They also provide another argument specific to that result section (i.e., residuals are not very behavior predictive), which is not general so I won't discuss it in depth here. These arguments however do not change the basic fact that d-VAE aggregates information from other neurons when extracting the behaviorally relevant activity of any given neuron, something that the authors note as a benefit of d-VAE in many instances. The fact that d-VAE aggregates population level info to give the inferred behaviorally relevant signal for each neuron confounds several key conclusions. For example, because information is aggregated across neurons, when trial to trial variability looks smoother after applying d-VAE (Fig 3i), or reveals better cosine tuning (Fig 3b), or when neurons that were not very predictive of behavior become more predictive of behavior (Fig 5), one cannot really attribute the new smoother single trial activity or the improved decoding to the same single neurons; rather these new signals/performances include information from other neurons. Unless the connections of the encoder network (z=f(x)) is zero for all other neurons, one cannot claim that the inferred rates for the neuron are truly solely associated with that neuron. I believe this a fundamental property of a population level VAE, and simply makes the architecture unsuitable for claims regarding inherent properties of single neurons. This confound is partly why the first claim in the abstract are not supported by data: observing that neurons that don't predict behavior very well would predict it much better after applying d-VAE does not prove that these neurons themselves "encode rich[er] behavioral information in complex nonlinear ways" (i.e., the first conclusion highlighted in the abstract) because information was also aggregated from other neurons. The other reason why this claim is not supported by data is the characterization of the encoding for smaller R2 neurons as "complex nonlinear", which the method is not well equipped to tease apart from linear mappings as I explain in my comment 3.”

We acknowledge that we cannot obtain the exact single neuronal activity that does not contain any information from other neurons. However, we believe our model can extract accurate approximation signals of the ground truth relevant signals. These signals preserve the inherent properties of single neuronal activity to some extent and can be used for analysis at the single-neuron level.

We believe d-VAE is a reasonable approach to extract effective relevant signals that preserve inherent properties of single neuronal activity for four key reasons:

1) d-VAE is a latent variable model that adheres to the neural population doctrine. The neural population doctrine posits that information is encoded within interconnected groups of neurons, with the existence of latent variables (neural modes) responsible for generating observable neuronal activity [1, 2]. If we can perfectly obtain the true generative model from latent variables to neuronal activity, then we can generate the activity of each neuron from hidden variables without containing any information from other neurons. However, without a complete understanding of the brain’s encoding strategies (or generative model), we can only get the approximation signals of the ground truth signals.

2) After the generative model is established, we need to infer the parameters of the generative model and the distribution of latent variables. During the inference process, inference algorithms such as variational inference or EM algorithms will be used. Generally, the obtained latent variables are also approximations of the real latent variables. When inferring the latent variables, it is inevitable to aggregation the information of the neural population, and latent variables are derived through weighted combinations of neuronal populations [3].

This inference process is consistent with that of d-VAE (or VAE-based models).

3) Latent variables are derived from raw neural signals and used to explain raw neural signals. Considering the unknown ground truth of latent variables and behaviorally-relevant signals, it becomes evident that the only reliable reference at the signal level is the raw signals. A crucial criterion for evaluating the reliability of latent variable models (including latent variables and generated relevant signals) is their capability to effectively explain the raw signals [3]. Consequently, we firmly maintain the belief that if the generated signals closely resemble the raw signals to the greatest extent possible, in accordance with an equivalence principle, we can claim that these obtained signals faithfully retain the inherent properties of single neurons. d-VAE explicitly constrains the generated signal to closely resemble the raw signals. These results demonstrate that d-VAE can extract effective relevant signals that preserve inherent properties of single neuronal activity.

Based on the above reasons, we hold that generating single neuronal activities with the VAE framework is a reasonable approach. The remaining question is whether our model can obtain accurate relevant signals in the absence of ground truth. To our knowledge, in cases where the ground truth of relevant signals is unknown, there are typically two approaches to verifying the reliability of extracted signals:

1) Conducting synthetic experiments where the ground truth is known.

2) Validation based on expert knowledge (Three criteria were proposed in this paper). Both our extracted signals and key conclusions have been validated using these two approaches.

Next, we will provide a detailed response to the concerns regarding our first key conclusion that smaller R2 neurons encode rich information.

We acknowledge that larger R2 neurons play a role in aiding the reconstruction of signals in smaller R2 neurons through their neural activity. However, considering that neurons are correlated rather than independent entities, we maintain the belief that larger R2 neurons assist damaged smaller R2 neurons in restoring their original appearance. Taking image denoising as an example, when restoring noisy pixels to their original appearance, relying solely on the noisy pixels themselves is often impractical. Assistance from their correlated, clean neighboring pixels becomes necessary.

The case we need to be cautious of is that the larger R2 neurons introduce additional signals (m) that contain substantial information to smaller R2 neurons, which they do not inherently possess. We believe this case does not hold for two reasons. Firstly, logically, adding extra signals decreases the reconstruction performance, and the information carried by these additional signals is redundant for larger R2 neurons, thus they do not introduce new information that can enhance the decoding performance of the neural population. Therefore, it seems unlikely and unnecessary for neural networks to engage in such counterproductive actions. Secondly, even if this occurs, our second criterion can effectively exclude the selection of these signals. To clarify, if we assume that x, y, and z denote the raw, relevant, and irrelevant signals of smaller R2 neurons, with x=y+z, and the extracted relevant signals become y+m, the irrelevant signals become z-m in this case. Consequently, the irrelevant signals contain a significant amount of information. It's essential to emphasize that this criterion holds significant importance in excluding undesirable signals.

Furthermore, we conducted a synthetic experiment to show that d-VAE can indeed restore the damaged information of smaller R2 neurons with the help of larger R2 neurons, and the restored neuronal activities are more similar to ground truth compared to damaged raw signals. Please see Fig. S11a,b for details.

Thank you for your valuable feedback.

[1] Saxena, S. and Cunningham, J.P., 2019. Towards the neural population doctrine. Current opinion in neurobiology, 55, pp.103-111.

[2] Gallego, J.A., Perich, M.G., Miller, L.E. and Solla, S.A., 2017. Neural manifolds for the control of movement. Neuron, 94(5), pp.978-984.

[3] Cunningham, J.P. and Yu, B.M., 2014. Dimensionality reduction for large-scale neural recordings. Nature neuroscience, 17(11), pp.1500-1509.

Q3: “Given the nonlinear architecture of the VAE, claims about the linearity or nonlinearity of cortical readout are confounded and not supported by the results.

The inference of behaviorally relevant signals from raw signals is a nonlinear operation, that is x_r=g(f(x)) is nonlinear function of x. So even when a linear KF is used to decode behavior from the inferred behaviorally relevant signals, the overall decoding from raw signals to predicted behavior (i.e., KF applied to g(f(x))) is nonlinear. Thus, the result that decoding of behavior from inferred behaviorally relevant signals (x_r) using a linear KF and a nonlinear ANN reaches similar accuracy (Fig 2), does not suggest that a "linear readout is performed in the motor cortex", as the authors claim (line 471). The authors acknowledge this confound (line 472) but fail to address it adequately. They perform a simulation analysis where the decoding gap between KF and ANN remains unchanged even when d-VAE is used to infer behaviorally relevant signals in the simulation. However, this analysis is not enough for "eliminating the doubt" regarding the confound. I'm sure the authors can also design simulations where the opposite happens and just like in the data, d-VAE can improve linear decoding to match ANN decoding. An adequate way to address this concern would be to use a fully linear version of the autoencoder where the f(.) and g(.) mappings are fully linear. They can simply replace these two networks in their model with affine mappings, redo the modeling and see if the model still helps the KF decoding accuracy reach that of the ANN decoding. In such a scenario, because the overall KF decoding from original raw signals to predicted behavior (linear d-VAE + KF) is linear, then they could move toward the claim that the readout is linear. Even though such a conclusion would still be impaired by the nonlinear reference (d-VAE + ANN decoding) because the achieved nonlinear decoding performance could always be limited by network design and fitting issues. Overall, the third conclusion highlighted in the abstract is a very difficult claim to prove and is unfortunately not supported by the results.”

We aim to explore the readout mechanism of behaviorally-relevant signals, rather than raw signals. Theoretically, the process of removing irrelevant signals should not be considered part of the inherent decoding mechanisms of the relevant signals. Assuming that the relevant signals we extracted are accurate, the conclusion of linear readout is established. On the synthetic data where the ground truth is known, our distilled signals show a significant improvement in neural similarity to the ground truth when compared to raw signals (refer to Fig. S2l). This observation demonstrates that our distilled signals are accurate approximations of the ground truth. Furthermore, on the three widely-used real datasets, our distilled signals meet the stringent criteria we have proposed (see Fig. 2), also providing strong evidence for their accuracy.

Regarding the assertion that we could create simulations in which d-VAE can make signals that are inherently nonlinearly decodable into linearly decodable ones: In reality, we cannot achieve this, as the second criterion can rule out the selection of such signals. Specifically,z=x+y=n^2+y, where z, x, y, and n denote raw signals, relevant signals, irrelevant signals and latent variables. If the relevant signals obtained by d-VAE are n, then these signals can be linear decoded accurately. However, the corresponding irrelevant signals are n^2-n+z; thus, irrelevant signals will have much information, and these extracted relevant signals will not be selected. Furthermore, our synthetic experiments offer additional evidence supporting the conclusion that d-VAE does not make inherently nonlinearly decodable signals become linearly decodable ones. As depicted in Fig. S11c, there exists a significant performance gap between KF and ANN when decoding the ground truth signals of smaller R2 neurons. KF exhibits notably low performance, leaving substantial room for compensation by d-VAE. However, following processing by d-VAE, KF's performance of distilled signals fails to surpass its already low ground truth performance and remains significantly inferior to ANN's performance. These results collectively confirm that our approach does not convert signals that are inherently nonlinearly decodable into linearly decodable ones, and the conclusion of linear readout is not a by-product by d-VAE.

Regarding the suggestion of using linear d-VAE + KF, as discussed in the Discussion section, removing the irrelevant signals requires a nonlinear operation, and linear d-VAE can not effectively separate relevant and irrelevant signals.

Thank you for your valuable feedback.

Q4: “The authors interpret several results as indications that "behavioral information is distributed in a higher-dimensional subspace than expected from raw signals", which is the second main conclusion highlighted in the abstract. However, several of these arguments do not convincingly support that conclusion.

4.1) The authors observe that behaviorally relevant signals for neurons with small principal components (referred to as secondary) have worse decoding with KF but better decoding with ANN (Fig. 6b,e), which also outperforms ANN decoding from raw signals. This observation is taken to suggest that these secondary behaviorally relevant signals encode behavior information in highly nonlinear ways and in a higher dimensions neural space than expected (lines 424 and 428). These conclusions however are confounded by the fact that A) d-VAE uses nonlinear encoding, so one cannot conclude from ANN outperforming KF that behavior is encoded nonlinearly in the motor cortex (see comment 3 above), and B) d-VAE aggregates information across the population so one cannot conclude that these secondary neurons themselves had as much behavior information (see comment 2 above).

4.2) The authors observe that the addition of the inferred behaviorally relevant signals for neurons with small principal components (referred to as secondary) improves the decoding of KF more than it improves the decoding of ANN (red curves in Fig 6c,f). This again is interpreted similarly as in 4.1, and is confounded for similar reasons (line 439): "These results demonstrate that irrelevant signals conceal the smaller variance PC signals, making their encoded information difficult to be linearly decoded, suggesting that behavioral information exists in a higher-dimensional subspace than anticipated from raw signals". This is confounded by because of the two reasons explained in 4.1. To conclude nonlinear encoding based on the difference in KF and ANN decoding, the authors would need to make the encoding/decoding in their VAE linear to have a fully linear decoder on one hand (with linear d-VAE + KF) and a nonlinear decoder on the other hand (with linear d-VAE + ANN), as explained in comment 3.

4.3) From S Fig 8, where the authors compare cumulative variance of PCs for raw and inferred behaviorally relevant signals, the authors conclude that (line 554): "behaviorally-irrelevant signals can cause an overestimation of the neural dimensionality of behaviorally-relevant responses (Supplementary Fig. S8)." However, this analysis does not really say anything about overestimation of "behaviorally relevant" neural dimensionality since the comparison is done with the dimensionality of "raw" signals. The next sentence is ok though: "These findings highlight the need to filter out relevant signals when estimating the neural dimensionality.", because they use the phrase "neural dimensionality" not "neural dimensionality of behaviorally-relevant responses".”

Questions 4.1 and 4.2 are a combination of Q2 and Q3. Please refer to our responses to Q2 and Q3.

Regarding question 4.3 about “behaviorally-irrelevant signals can cause an overestimation of the neural dimensionality of behaviorally-relevant responses”: Previous studies usually used raw signals to estimate the neural dimensionality of specific behaviors. We mean that using raw signals, which include many irrelevant signals, will cause an overestimation of the neural dimensionality. We have modified this sentence in the revised manuscripts.

Thank you for your valuable feedback.

Q5: “Imprecise use of language in many places leads to inaccurate statements. I will list some of these statements”

5.1) In the abstract: "One solution is to accurately separate behaviorally-relevant and irrelevant signals, but this approach remains elusive due to the unknown ground truth of behaviorally-relevant signals". This statement is not accurate because it implies no prior work does this. The authors should make their statement more specific and also refer to some goal that existing linear (e.g., PSID) and nonlinear (e.g., TNDM) methods for extracting behaviorally relevant signals fail to achieve.

5.2) In the abstract: "we found neural responses previously considered useless encode rich behavioral information" => what does "useless" mean operationally? Low behavior tuning? More precise use of language would be better.

5.3) "... recent studies (Glaser 58 et al., 2020; Willsey et al., 2022) demonstrate nonlinear readout outperforms linear readout." => do these studies show that nonlinear "readout" outperforms linear "readout", or just that nonlinear models outperform linear models?

5.4) Line 144: "The first criterion is that the decoding performance of the behaviorally-relevant signals (red bar, Fig.1) should surpass that of raw signals (the red dotted line, Fig.1).". Do the authors mean linear decoding here or decoding in general? If the latter, how can something extracted from neural surpass decoding of neural data, when the extraction itself can be thought of as part of decoding? The operational definition for this "decoding performance" should be clarified.

5.5) Line 311: "we found that the dimensionality of primary subspace of raw signals (26, 64, and 45 for datasets A, B, and C) is significantly higher than that of behaviorally-relevant signals (7, 13, and 9), indicating that behaviorally-irrelevant signals lead to an overestimation of the neural dimensionality of behaviorally-relevant signals." => here the dimensionality of the total PC space (i.e., primary subspace of raw signals) is being compared with that of inferred behaviorally-relevant signals, so the former being higher does not indicate that neural dimensionality of behaviorally-relevant signals was overestimated. The former is simply not behavioral so this conclusion is not accurate.

5.6) Section "Distilled behaviorally-relevant signals uncover that smaller R2 neurons encode rich behavioral information in complex nonlinear ways". Based on what kind of R2 are the neurons grouped? Behavior decoding R2 from raw signals? Using what mapping? Using KF? If KF is used, the result that small R2 neurons benefit a lot from d-VAE could be somewhat expected, given the nonlinearity of d-VAE: because only ANN would have the capacity to unwrap the nonlinear encoding of d-VAE as needed. If decoding performance that is used to group neurons is based on data, regression to the mean could also partially explain the result: the neurons with worst raw decoding are most likely to benefit from a change in decoder, than neurons that already had good decoding. In any case, the R2 used to partition and sort neurons should be more clearly stated and reminded throughout the text and I Fig 3.

5.7) Line 346 "...it is impossible for our model to add the activity of larger R2 neurons to that of smaller R2 neurons" => Is it really impossible? The optimization can definitely add small-scale copies of behaviorally relevant information to all neurons with minimal increase in the overall optimization loss, so this statement seems inaccurate.

5.8) Line 490: "we found that linear decoders can achieve comparable performance to that of nonlinear decoders, providing compelling evidence for the presence of linear readout in the motor cortex." => inaccurate because no d-VAE decoding is really linear, as explained in comment 3 above.

5.9) Line 578: ". However, our results challenge this idea by showing that signals composed of smaller variance PCs nonlinearly encode a significant amount of behavioral information." => inaccurate as results are confounded by nonlinearity of d-VAE as explained in comment 3 above.

5.10) Line 592: "By filtering out behaviorally-irrelevant signals, our study found that accurate decoding performance can be achieved through linear readout, suggesting that the motor cortex may perform linear readout to generate movement behaviors." => inaccurate because it us confounded by the nonlinearity of d-VAE as explained in comment 3 above.”

Regarding “5.1) In the abstract: "One solution is to accurately separate behaviorally-relevant and irrelevant signals, but this approach remains elusive due to the unknown ground truth of behaviorally-relevant signals". This statement is not accurate because it implies no prior work does this. The authors should make their statement more specific and also refer to some goal that existing linear (e.g., PSID) and nonlinear (e.g., TNDM) methods for extracting behaviorally relevant signals fail to achieve”:

We believe our statement is accurate. Our primary objective is to extract accurate behaviorally-relevant signals that closely approximate the ground truth relevant signals. To achieve this, we strike a balance between the reconstruction and decoding performance of the generated signals, aiming to effectively capture the relevant signals. This crucial aspect of our approach sets it apart from other methods. In contrast, other methods tend to emphasize the extraction of valuable latent neural dynamics. We have provided elaboration on the distinctions between d-VAE and other approaches in the Introduction and Discussion sections.

Thank you for your valuable feedback.

Regarding “5.2) In the abstract: "we found neural responses previously considered useless encode rich behavioral information" => what does "useless" mean operationally? Low behavior tuning? More precise use of language would be better.”:

In the analysis of neural signals, smaller variance PC signals are typically seen as noise and are often discarded. Similarly, smaller R2 neurons are commonly thought to be dominated by noise and are not further analyzed. Given these considerations, we believe that the term "considered useless" is appropriate in this context. Thank you for your valuable feedback.

Regarding “5.3) "... recent studies (Glaser 58 et al., 2020; Willsey et al., 2022) demonstrate nonlinear readout outperforms linear readout." => do these studies show that nonlinear "readout" outperforms linear "readout", or just that nonlinear models outperform linear models?”:

In this paper, we consider the two statements to be equivalent. Thank you for your valuable feedback.

Regarding “5.4) Line 144: "The first criterion is that the decoding performance of the behaviorally-relevant signals (red bar, Fig.1) should surpass that of raw signals (the red dotted line, Fig.1).". Do the authors mean linear decoding here or decoding in general? If the latter, how can something extracted from neural surpass decoding of neural data, when the extraction itself can be thought of as part of decoding? The operational definition for this "decoding performance" should be clarified.”:

We mean the latter, as we said in the section “Framework for defining, extracting, and separating behaviorally-relevant signals”, since raw signals contain too many behaviorally-irrelevant signals, deep neural networks are more prone to overfit raw signals than relevant signals. Therefore the decoding performance of relevant signals should surpass that of raw signals. Thank you for your valuable feedback.

Regarding “5.5) Line 311: "we found that the dimensionality of primary subspace of raw signals (26, 64, and 45 for datasets A, B, and C) is significantly higher than that of behaviorally-relevant signals (7, 13, and 9), indicating that behaviorally-irrelevant signals lead to an overestimation of the neural dimensionality of behaviorally-relevant signals." => here the dimensionality of the total PC space (i.e., primary subspace of raw signals) is being compared with that of inferred behaviorally-relevant signals, so the former being higher does not indicate that neural dimensionality of behaviorally-relevant signals was overestimated. The former is simply not behavioral so this conclusion is not accurate.”: In practice, researchers usually used raw signals to estimate the neural dimensionality. We mean that using raw signals to do this would overestimate the neural dimensionality. Thank you for your valuable feedback.

Regarding “5.6) Section "Distilled behaviorally-relevant signals uncover that smaller R2 neurons encode rich behavioral information in complex nonlinear ways". Based on what kind of R2 are the neurons grouped? Behavior decoding R2 from raw signals? Using what mapping? Using KF? If KF is used, the result that small R2 neurons benefit a lot from d-VAE could be somewhat expected, given the nonlinearity of d-VAE: because only ANN would have the capacity to unwrap the nonlinear encoding of d-VAE as needed. If decoding performance that is used to group neurons is based on data, regression to the mean could also partially explain the result: the neurons with worst raw decoding are most likely to benefit from a change in decoder, than neurons that already had good decoding. In any case, the R2 used to partition and sort neurons should be more clearly stated and reminded throughout the text and I Fig 3.”:

When employing R2 to characterize neurons, it indicates the extent to which neuronal activity is explained by the linear encoding model [1-3]. Smaller R2 neurons have a lower capacity for linearly tuning (encoding) behaviors, while larger R2 neurons have a higher capacity for linearly tuning (encoding) behaviors. Specifically, the approach involves first establishing an encoding relationship from velocity to neural signal using a linear model, i.e., y=f(x), where f represents a linear regression model, x denotes velocity, and y denotes the neural signal. Subsequently, R2 is utilized to quantify the effectiveness of the linear encoding model in explaining neural activity. We have provided a comprehensive explanation in the revised manuscript. Thank you for your valuable feedback.

[1] Collinger, J.L., Wodlinger, B., Downey, J.E., Wang, W., Tyler-Kabara, E.C., Weber, D.J., McMorland, A.J., Velliste, M., Boninger, M.L. and Schwartz, A.B., 2013. High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet, 381(9866), pp.557-564.

[2] Wodlinger, B., et al. "Ten-dimensional anthropomorphic arm control in a human brain− machine interface: difficulties, solutions, and limitations." Journal of neural engineering 12.1 (2014): 016011.

[3] Inoue, Y., Mao, H., Suway, S.B., Orellana, J. and Schwartz, A.B., 2018. Decoding arm speed during reaching. Nature communications, 9(1), p.5243.

Regarding Questions 5.7, 5.8, 5.9, and 5.10:

We believe our conclusions are solid. The reasons can be found in our replies in Q2 and Q3. Thank you for your valuable feedback.

Q6: “Imprecise use of language also sometimes is not inaccurate but just makes the text hard to follow.

6.1) Line 41: "about neural encoding and decoding mechanisms" => what is the definition of encoding/decoding and how do these differ? The definitions given much later in line 77-79 is also not clear.

6.2) Line 323: remind the reader about what R2 is being discussed, e.g., R2 of decoding behavior using KF. It is critical to know if linear or nonlinear decoding is being discussed.

6.3) Line 488: "we found that neural responses previously considered trivial encode rich behavioral information in complex nonlinear ways" => "trivial" in what sense? These phrases would benefit from more precision, for example: "neurons that may seem to have little or no behavior information encoded". The same imprecise word ("trivial") is also used in many other places, for example in the caption of Fig S9.

6.4) Line 611: "The same should be true for the brain." => Too strong of a statement for an unsupported claim suggesting the brain does something along the lines of nonlin VAE + linear readout.

6.5) In Fig 1, legend: what is the operational definition of "generating performance"? Generating what? Neural reconstruction?”

Regarding “6.1) Line 41: "about neural encoding and decoding mechanisms" => what is the definition of encoding/decoding and how do these differ? The definitions given much later in line 77-79 is also not clear.”:

We would like to provide a detailed explanation of neural encoding and decoding. Neural encoding means how neuronal activity encodes the behaviors, that is, y=f(x), where y denotes neural activity and, x denotes behaviors, f is the encoding model. Neural decoding means how the brain decodes behaviors from neural activity, that is, x=g(y), where g is the decoding model. For further elaboration, please refer to [1]. We have included references that discuss the concepts of encoding and decoding in the revised manuscript. Thank you for your valuable feedback.

[1] Kriegeskorte, Nikolaus, and Pamela K. Douglas. "Interpreting encoding and decoding models." Current opinion in neurobiology 55 (2019): 167-179.

Regarding “6.2) Line 323: remind the reader about what R2 is being discussed, e.g., R2 of decoding behavior using KF. It is critical to know if linear or nonlinear decoding is being discussed.”:

This question is the same as Q5.6. Please refer to the response to Q5.6. Thank you for your valuable feedback.

Regarding “6.3) Line 488: "we found that neural responses previously considered trivial encode rich behavioral information in complex nonlinear ways" => "trivial" in what sense? These phrases would benefit from more precision, for example: "neurons that may seem to have little or no behavior information encoded". The same imprecise word ("trivial") is also used in many other places, for example in the caption of Fig S9.”:

We have revised this statement in the revised manuscript. Thanks for your recommendation.

Regarding “6.4) Line 611: "The same should be true for the brain." => Too strong of a statement for an unsupported claim suggesting the brain does something along the lines of nonlin VAE + linear readout.”

We mean that removing the interference of irrelevant signals and decoding the relevant signals should logically be two stages. We have revised this statement in the revised manuscript. Thank you for your valuable feedback.

Regarding “6.5) In Fig 1, legend: what is the operational definition of "generating performance"? Generating what? Neural reconstruction?””:

We have replaced “generating performance” with “reconstruction performance” in the revised manuscript. Thanks for your recommendation.

Q7: “In the analysis presented starting in line 449, the authors compare improvement gained for decoding various speed ranges by adding secondary (small PC) neurons to the KF decoder (Fig S11). Why is this done using the KF decoder, when earlier results suggest an ANN decoder is needed for accurate decoding from these small PC neurons? It makes sense to use the more accurate nonlinear ANN decoder to support the fundamental claim made here, that smaller variance PCs are involved in regulating precise control”

Because when the secondary signal is superimposed on the primary signal, the enhancement in KF performance is substantial. We wanted to explore in which aspect of the behavior the KF performance improvement is mainly reflected. In comparison, the improvement of ANN by the secondary signal is very small, rendering the exploration of the aforementioned questions inconsequential. Thank you for your valuable feedback.

Q8: “A key limitation of the VAE architecture is that it doesn't aggregate information over multiple time samples. This may be why the authors decided to use a very large bin size of 100ms and beyond that smooth the data with a moving average. This limitation should be clearly stated somewhere in contrast with methods that can aggregate information over time (e.g., TNDM, LFADS, PSID) ”

We have added this limitation in the Discussion in the revised manuscript. Thanks for your recommendation.

Q9: “Fig 5c and parts of the text explore the decoding when some neurons are dropped. These results should come with a reminder that dropping neurons from behaviorally relevant signals is not technically possible since the extraction of behaviorally relevant signals with d-VAE is a population level aggregation that requires the raw signal from all neurons as an input. This is also important to remind in some places in the text for example:

• Line 498: "...when one of the neurons is destroyed."

• Line 572: "In contrast, our results show that decoders maintain high performance on distilled signals even when many neurons drop out."”

We want to explore the robustness of real relevant signals in the face of neuron drop-out. The signals our model extracted are an approximation of the ground truth relevant signals and thus serve as a substitute for ground truth to study this problem. Thank you for your valuable feedback.

Q10: “Besides the confounded conclusions regarding the readout being linear (see comment 3 and items related to it in comment 5), the authors also don't adequately discuss prior works that suggest nonlinearity helps decoding of behavior from the motor cortex. Around line 594, a few works are discussed as support for the idea of a linear readout. This should be accompanied by a discussion of works that support a nonlinear encoding of behavior in the motor cortex, for example (Naufel et al. 2019; Glaser et al. 2020), some of which the authors cite elsewhere but don't discuss here.”

We have added this discussion in the revised manuscript. Thanks for your recommendation.

Q11: “Selection of hyperparameters is not clearly explained. Starting line 791, the authors give some explanation for one hyperparameter, but not others. How are the other hyperparameters determined? What is the search space for the grid search of each hyperparameter? Importantly, if hyperparameters are determined only based on the training data of each fold, why is only one value given for the hyperparameter selected in each dataset (line 814)? Did all 5 folds for each dataset happen to select exactly the same hyperparameter based on their 5 different training/validation data splits? That seems unlikely.”

We perform a grid search in {0.001, 0.01,0.1,1} for hyperparameter beta. And we found that 0.001 is the best for all datasets. As for the model parameters, such as hidden neuron numbers, this model capacity has reached saturation decoding performance and does not influence the results.

Regarding “Importantly, if hyperparameters are determined only based on the training data of each fold, why is only one value given for the hyperparameter selected in each dataset (line 814)? Did all 5 folds for each dataset happen to select exactly the same hyperparameter based on their 5 different training/validation data splits”: We selected the hyperparameter based on the average performance of 5 folds data on validation sets. The selected value denotes the one that yields the highest average performance across the 5 folds data.

Thank you for your valuable feedback.

Q12: “d-VAE itself should also be explained more clearly in the main text. Currently, only the high-level idea of the objective is explained. The explanation should be more precise and include the idea of encoding to latent state, explain the relation to pip-VAE, explain inputs and outputs, linearity/nonlinearity of various mappings, etc. Also see comment 1 above, where I suggest adding more details about other methods in the main text.”

Our primary objective is to delve into the encoding and decoding mechanisms using the separated relevant signals. Therefore, providing an excessive amount of model details could potentially distract from the main focus of the paper. In response to your suggestion, we have included a visual representation of d-VAE's structure, input, and output (see Fig. S1) in the revised manuscript, which offers a comprehensive and intuitive overview. Additionally, we have expanded on the details of d-VAE and other methods in the Methods section.

Thank you for your valuable feedback.

Q13: “In Fig 1f and g, shouldn't the performance plots be swapped? The current plots seem counterintuitive. If there is bias toward decoding (panel g), why is the irrelevant residual so good at decoding?”

The placement of the performance plots in Fig. 1f and 1g is accurate. When the model exhibits a bias toward decoding, it prioritizes extracting the most relevant features (latent variables) for decoding purposes. As a consequence, the model predominantly generates signals that are closely associated with these extracted features. This selective signal extraction and generation process may result in the exclusion of other potentially useful information, which will be left in the residuals. To illustrate this concept, consider the example of face recognition: if a model can accurately identify an individual using only the person's eyes (assuming these are the most useful features), other valuable information, such as details of the nose or mouth, will be left in the residuals, which could also be used to identify the individual.

Thank you for your valuable feedback.

Author Response:

The following is the authors’ response to the previous reviews.

We carefully read through the second-round reviews and the additional reviews. To us, the review process is somewhat unusual and very much dominated by referee 2, who aggressively insists that we mixed up the trigeminal nucleus and inferior olive and that as a consequence our results are meaningless. We think the stance of referee 2 and the focus on one single issue (the alleged mix-up of trigeminal nucleus and inferior olive) is somewhat unfortunate, leaves out much of our findings and we debated at length on how to deal with further revisions. In the end, we decided to again give priority to addressing the criticism of referees 2, because it is hard to go on with a heavily attacked paper without resolving the matter at stake. The following is a summary of, what we did:

Additional experimental work:

(1) We checked if the peripherin-antibody indeed reliably identifies climbing fibers.

To this end, we sectioned the elephant cerebellum and stained sections with the peripherin-antibody. We find: (i) the cerebellar white matter is strongly reactive for peripherin-antibodies, (ii) cerebellar peripherin-antibody staining of has an axonal appearance. (iii) Cerebellar Purkinje cell somata appear to be ensheated by peripherin-antibody staining. (iv) We observed that the peripherin-antibody reactivity gradually decreases from Purkinje cell somata to the pia in the cerebellar molecular layer. This work is shown in our revised Figure 2. All these four features align with the distribution of climbing fibers (which arrive through the white matter, are axons, ensheat Purkinje cell somata, and innervate Purkinje cell proximally not reaching the pia). In line with previous work, which showed similar cerebellar staining patterns in several species (Errante et al. 1998), we conclude that elephant climbing fibers are strongly reactive for peripherin-antibodies.

(2) We delineated the elephant olivo-cerebellar tract.

The strong peripherin-antibody reactivity of elephant climbing fibers enabled us to delineate the elephant olivo-cerebellar tract. We find the elephant olivo-cerebellar tract is a strongly peripherin-antibody reactive, well-delineated fiber tract several millimeters wide and about a centimeter in height. The unstained olivo-cerebellar tract has a greyish appearance. In the anterior regions of the olivo-cerebellar tract, we find that peripherin-antibody reactive fibers run in the dorsolateral brainstem and approach the cerebellar peduncle, where the tract gradually diminishes in size, presumably because climbing fibers discharge into the peduncle. Indeed, peripherin-antibody reactive fibers can be seen entering the cerebellar peduncle. Towards the posterior end of the peduncle, the olivo-cerebellar disappears (in the dorsal brainstem directly below the peduncle. We note that the olivo-cerebellar tract was referred to as the spinal trigeminal tract by Maseko et al. 2013. We think the tract in question cannot be the spinal trigeminal tract for two reasons: (i) This tract is the sole brainstem source of peripherin-positive climbing fibers entering the peduncle/ the cerebellum; this is the defining characteristic of the olivo-cerebellar tract. (ii) The tract in question is much smaller than the trigeminal nerve, disappears posterior to where the trigeminal nerve enters the brainstem (see below), and has no continuity with the trigeminal nerve; the continuity with the trigeminal nerve is the defining characteristic of the spinal trigeminal tract, however.

The anterior regions of the elephant olivo-cerebellar tract are similar to the anterior regions of olivo-cerebellar tract of other mammals in its dorsolateral position and the relation to the cerebellar peduncle. In its more posterior parts, the elephant olivo-cerebellar tract continues for a long distance (~1.5 cm) in roughly the same dorsolateral position and enters the serrated nucleus that we previously identified as the elephant inferior olive. The more posterior parts of the elephant olivo-cerebellar tract therefore differ from the more posterior parts of the olivo-cerebellar tract of other mammals, which follows a ventromedial trajectory towards a ventromedially situated inferior olive. The implication of our delineation of the elephant olivo-cerebellar tract is that we correctly identified the elephant inferior olive.

(3) An in-depth analysis of peripherin-antibody reactivity also indicates that the trigeminal nucleus receives no climbing fiber input.

We also studied the peripherin-antibody reactivity in and around the trigeminal nucleus. We had also noted in the previous submission that the trigeminal nucleus is weakly positive for peripherin, but that the staining pattern is uniform and not the type of axon bundle pattern that is seen in the inferior olive of other mammals. To us, this observation already argued against the presence of climbing fibers in the trigeminal nucleus. We also noted that the myelin stripes of the trigeminal nucleus were peripherin-antibody-negative. In the context of our olivo-cerebellar tract tracing we now also scrutinized the surroundings of the trigeminal nucleus for peripherin-antibody reactivity. We find that the ventral brainstem surrounding the trigeminal nucleus is devoid of peripherin-antibody reactivity. Accordingly, no climbing fibers, (which we have shown to be strongly peripherin-antibody-positive, see our point 1) arrive at the trigeminal nucleus. The absence of climbing fiber input indicates that previous work that identified the (trigeminal) nucleus as the inferior olive (Maseko et al 2013) is unlikely to be correct.

(4) We characterized the entry of the trigeminal nerve into the elephant brain.

To better understand how trigeminal information enters the elephant’s brain, we characterized the entry of the trigeminal nerve. This analysis indicated to us that the trigeminal nerve is not continuous with the olivo-cerebellar tract (the spinal trigeminal tract of Maseko et al. 2013) as previously claimed by Maseko et al. 2013. We show some of this evidence in Referee-Figure 1 below. The reason we think the trigeminal nerve is discontinuous with the olivo-cerebellar tract is the size discrepancy between the two structures. We first show this for the tracing data of Maseko et al. 2013. In the Maseko et al. 2013 data the trigeminal nerve (Referee-Figure 1A, their plate Y) has 3-4 times the diameter of the olivocerebellar tract (the alleged spinal trigeminal tract, Referee-Figure 1B, their plate Z). Note that most if not all trigeminal fibers are thought to continue from the nerve into the trigeminal tract (see our rat data below). We plotted the diameter of the trigeminal nerve and diameter of the olivo-cerebellar (the spinal trigeminal tract according to Maseko et al. 2013) from the Maseko et al. 2013 data (Referee-Figure 1C) and we found that the olivocerebellar tract has a fairly consistent diameter (46 ± 9 mm2, mean ± SD). Statistical considerations and anatomical evidence suggest that the tracing of the trigeminal nerve into the olivo-cerebellar (the spinal trigeminal tract according to Maseko et al. 2013) is almost certainly wrong. The most anterior point of the alleged spinal trigeminal tract has a diameter of 51 mm2 which is more than 15 standard deviations different from the most posterior diameter (194 mm2) of the trigeminal tract. For this assignment to be correct three-quarters of trigeminal nerve fibers would have to spontaneously disappear, something that does not happen in the brain. We also made similar observations in the African elephant Bibi, where the trigeminal nerve (Referee-Figure 1D) is much larger in diameter than the olivocerebellar tract (Referee-Figure 1E). We could also show that the olivocerebellar tract disappears into the peduncle posterior to where the trigeminal nerve enters (Referee-Figure 1F). Our data are very similar to Maseko et al. indicating that their outlining of structures was done correctly. What appears to have been oversimplified, is the assignment of structures as continuous. We also quantified the diameter of the trigeminal nerve and the spinal trigeminal tract in rats (from the Paxinos & Watson atlas; Referee-Figure 1D); as expected we found the trigeminal nerve and spinal trigeminal tract diameters are essentially continuous.

In our hands, the trigeminal nerve does not continue into a well-defined tract that could be traced after its entry. In this regard, it differs both from the olivo-cerebellar tract of the elephant or the spinal trigeminal tract of the rodent, both of which are well delineated. We think the absence of a well-delineated spinal trigeminal tract in elephants might have contributed to the putative tracing error highlighted in our Referee-Figure 1A-C.

We conclude that a size mismatch indicates trigeminal fibers do not run in the olivo-cerebellar tract (the spinal trigeminal tract according to Maseko et al. 2013).

Author response image 1.

The trigeminal nerve is discontinuous with the olivo-cerebellar tract (the spinal trigeminal tract according to Maseko et al. 2013). A, Trigeminal nerve (orange) in the brain of African elephant LAX as delineated by Maseko et al. 2013 (coronal section; their plate Y). B, Most anterior appearance of the spinal trigeminal tract of Maseko et al. 2013 (blue; coronal section; their plate Z). Note the much smaller diameter of the spinal trigeminal tract compared to the trigeminal nerve shown in C, which argues against the continuity of the two structures. Indeed, our peripherin-antibody staining showed that the spinal trigeminal tract of Maseko corresponds to the olivo-cerebellar tract and is discontinuous with the trigeminal nerve. C, Plot of the trigeminal nerve and olivo-cerebellar tracts (the spinal trigeminal tract according to Maseko et al. 2013) diameter along the anterior-posterior axis. The trigeminal nerve is much larger in diameter than the olivocerebellar tract (the spinal trigeminal tract according to Maseko et al. 2013). C, D measurements, for which sections are shown in panels C and D respectively. The olivocerebellar tract (the spinal trigeminal tract according to Maseko et al. 2013) has a consistent diameter; data replotted from Maseko et al. 2013. At mm 25 the inferior olive appears. D, Trigeminal nerve entry in the brain of African elephant Bibi; our data, coronal section, the trigeminal nerve is outlined in orange, note the large diameter. E, Most anterior appearance of the olivo-cerebellar tract in the brain of African elephant Bibi; our data, coronal section, approximately 3 mm posterior to the section shown in A, the olivocerebellar tract is outlined in blue. Note the smaller diameter of the olivo-cerebellar tract compared to the trigeminal nerve, which argues against the continuity of the two structures. F, Plot of the trigeminal nerve and olivo-cerebellar tract diameter along the anterior-posterior axis. The nerve and olivo-cerebellar tract are discontinuous and the trigeminal nerve is much larger in diameter than the olivocerebellar tract (the spinal trigeminal tract according to Maseko et al. 2013); our data. D, E measurements, for which sections are shown in panels D and E respectively. At mm 27 the inferior olive appears. G, In the rat the trigeminal nerve is continuous in size with the spinal trigeminal tract. Data replotted from Paxinos and Watson.

Reviewer 2 (Public Review):

As indicated in my previous review of this manuscript (see above), it is my opinion that the authors have misidentified, and indeed switched, the inferior olivary nuclear complex (IO) and the trigeminal nuclear complex (Vsens). It is this specific point only that I will address in this second review, as this is the crucial aspect of this paper - if the identification of these nuclear complexes in the elephant brainstem by the authors is incorrect, the remainder of the paper does not have any scientific validity.

Comment: We agree with the referee that it is most important to sort out, the inferior olivary nuclear complex (IO) and the trigeminal nuclear complex, respectively.Change: We did additional experimental work to resolve this matter as detailed at the beginning of our response. Specifically, we ascertained that elephant climbing fibers are strongly peripherin-positive. Based on elephant climbing fiber peripherin-reactivity we delineated the elephant olivo-cerebellar tract. We find that the olivo-cerebellar connects to the structure we refer to as inferior olive to the cerebellum (the referee refers to this structure as the trigeminal nuclear complex). We also found that the trigeminal nucleus (the structure the referee refers to as inferior olive) appears to receive no climbing fibers. We provide indications that the tracing of the trigeminal nerve into the olivo-cerebellar tract by Maseko et al. 2023 was erroneous (Author response image 1). These novel findings support our ideas but are very difficult to reconcile with the referee’s partitioning scheme.

The authors, in their response to my initial review, claim that I "bend" the comparative evidence against them. They further claim that as all other mammalian species exhibit a "serrated" appearance of the inferior olive, and as the elephant does not exhibit this appearance, that what was previously identified as the inferior olive is actually the trigeminal nucleus and vice versa. 

For convenience, I will refer to IOM and VsensM as the identification of these structures according to Maseko et al (2013) and other authors and will use IOR and VsensR to refer to the identification forwarded in the study under review. <br /> The IOM/VsensR certainly does not have a serrated appearance in elephants. Indeed, from the plates supplied by the authors in response (Referee Fig. 2), the cytochrome oxidase image supplied and the image from Maseko et al (2013) shows a very similar appearance. There is no doubt that the authors are identifying structures that closely correspond to those provided by Maseko et al (2013). It is solely a contrast in what these nuclear complexes are called and the functional sequelae of the identification of these complexes (are they related to the trunk sensation or movement controlled by the cerebellum?) that is under debate.

Elephants are part of the Afrotheria, thus the most relevant comparative data to resolve this issue will be the identification of these nuclei in other Afrotherian species. Below I provide images of these nuclear complexes, labelled in the standard nomenclature, across several Afrotherian species. 

(A) Lesser hedgehog tenrec (Echinops telfairi) 

Tenrecs brains are the most intensively studied of the Afrotherian brains, these extensive neuroanatomical studies undertaken primarily by Heinz Künzle. Below I append images (coronal sections stained with cresol violet) of the IO and Vsens (labelled in the standard mammalian manner) in the lesser hedgehog tenrec. It should be clear that the inferior olive is located in the ventral midline of the rostral medulla oblongata (just like the rat) and that this nucleus is not distinctly serrated. The Vsens is located in the lateral aspect of the medulla skirted laterally by the spinal trigeminal tract (Sp5). These images and the labels indicating structures correlate precisely with that provide by Künzle (1997, 10.1016, see his Figure 1K,L. Thus, in the first case of a related species, there is no serrated appearance of the inferior olive, the location of the inferior olive is confirmed through connectivity with the superior colliculus (a standard connection in mammals) by Künzle (1997), and the location of Vsens is what is considered to be typical for mammals. This is in agreement with the authors, as they propose that ONLY the elephants show the variations they report. 

(B) Giant otter shrew (Potomogale velox) 

The otter shrews are close relatives of the Tenrecs. Below I append images of cresyl violet (left column) and myelin (right column) stained coronal sections through the brainstem with the IO, Vsens and Sp5 labelled as per standard mammalian anatomy. Here we see hints of the serration of the IO as defined by the authors, but we also see many myelin stripes across the IO. Vsens is located laterally and skirted by the Sp5. This is in agreement with the authors, as they propose that ONLY the elephants show the variations they report.

(C) Four-toed sengi (Petrodromus tetradactylus) 

The sengis are close relatives of the Tenrecs and otter shrews, these three groups being part of the Afroinsectiphilia, a distinct branch of the Afrotheria. Below I append images of cresyl violet (left column) and myelin (right column) stained coronal sections through the brainstem with the IO, Vsens and Sp5 labelled as per standard mammalian anatomy. Here we see vague hints of the serration of the IO (as defined by the authors), and we also see many myelin stripes across the IO. Vsens is located laterally and skirted by the Sp5. This is in agreement with the authors, as they propose that ONLY the elephants show the variations they report. 

(D) Rock hyrax (Procavia capensis) 

The hyraxes, along with the sirens and elephants form the Paenungulata branch of the Afrotheria. Below I append images of cresyl violet (left column) and myelin (right column) stained coronal sections through the brainstem with the IO, Vsens and Sp5 labelled as per the standard mammalian anatomy. Here we see hints of the serration of the IO (as defined by the authors), but we also see evidence of a more "bulbous" appearance of subnuclei of the IO (particularly the principal nucleus), and we also see many myelin stripes across the IO. Vsens is located laterally and skirted by the Sp5. This is in agreement with the authors, as they propose that ONLY the elephants show the variations they report. 

(E) West Indian manatee (Trichechus manatus) 

The sirens are the closest extant relatives of the elephants in the Afrotheria. Below I append images of cresyl violet (top) and myelin (bottom) stained coronal sections (taken from the University of Wisconsin-Madison Brain Collection, https://brainmuseum.org, and while quite low in magnification they do reveal the structures under debate) through the brainstem with the IO, Vsens and Sp5 labelled as per standard mammalian anatomy. Here we see the serration of the IO (as defined by the authors). Vsens is located laterally and skirted by the Sp5. This is in agreement with the authors, as they propose that ONLY the elephants show the variations they report.

These comparisons and the structural identification, with which the authors agree as they only distinguish the elephants from the other Afrotheria, demonstrate that the appearance of the IO can be quite variable across mammalian species, including those with a close phylogenetic affinity to the elephants. Not all mammal species possess a "serrated" appearance of the IO. Thus, it is more than just theoretically possible that the IO of the elephant appears as described prior to this study. 

So what about elephants? Below I append a series of images from coronal sections through the African elephant brainstem stained for Nissl, myelin, and immunostained for calretinin. These sections are labelled according to standard mammalian nomenclature. In these complete sections of the elephant brainstem, we do not see a serrated appearance of the IOM (as described previously and in the current study by the authors). Rather the principal nucleus of the IOM appears to be bulbous in nature. In the current study, no image of myelin staining in the IOM/VsensR is provided by the authors. However, in the images I provide, we do see the reported myelin stripes in all stains - agreement between the authors and reviewer on this point. The higher magnification image to the bottom left of the plate shows one of the IOM/VsensR myelin stripes immunostained for calretinin, and within the myelin stripes axons immunopositive for calretinin are seen (labelled with an arrow). The climbing fibres of the elephant cerebellar cortex are similarly calretinin immunopositive (10.1159/000345565). In contrast, although not shown at high magnification, the fibres forming the Sp5 in the elephant (in the Maseko description, unnamed in the description of the authors) show no immunoreactivity to calretinin. 

Comment: We appreciate the referee’s additional comments. We concede the possibility that some relatives of elephants have a less serrated inferior olive than most other mammals. We maintain, however, that the elephant inferior olive (our Figure 1J) has the serrated appearance seen in the vast majority of mammals.

Change: None.

Peripherin Immunostaining 

In their revised manuscript the authors present immunostaining of peripherin in the elephant brainstem. This is an important addition (although it does replace the only staining of myelin provided by the authors which is unusual as the word myelin is in the title of the paper) as peripherin is known to specifically label peripheral nerves. In addition, as pointed out by the authors, peripherin also immunostains climbing fibres (Errante et al., 1998). The understanding of this staining is important in determining the identification of the IO and Vsens in the elephant, although it is not ideal for this task as there is some ambiguity. Errante and colleagues (1998; Fig. 1) show that climbing fibres are peripherin-immunopositive in the rat. But what the authors do not evaluate is the extensive peripherin staining in the rat Sp5 in the same paper (Errante et al, 1998, Fig. 2). The image provided by the authors of their peripherin immunostaining (their new Figure 2) shows what I would call the Sp5 of the elephant to be strongly peripherin immunoreactive, just like the rat shown in Errant et al (1998), and more over in the precise position of the rat Sp5! This makes sense as this is where the axons subserving the "extraordinary" tactile sensitivity of the elephant trunk would be found (in the standard model of mammalian brainstem anatomy). Interestingly, the peripherin immunostaining in the elephant is clearly lamellated...this coincides precisely with the description of the trigeminal sensory nuclei in the elephant by Maskeo et al (2013) as pointed out by the authors in their rebuttal. Errante et al (1998) also point out peripherin immunostaining in the inferior olive, but according to the authors this is only "weakly present" in the elephant IOM/VsensR. This latter point is crucial. Surely if the elephant has an extraordinary sensory innervation from the trunk, with 400 000 axons entering the brain, the VsensR/IOM should be highly peripherin-immunopositive, including the myelinated axon bundles?! In this sense, the authors argue against their own interpretation - either the elephant trunk is not a highly sensitive tactile organ, or the VsensR is not the trigeminal nuclei it is supposed to be. 

Comment: We made sure that elephant climbing fibers are strongly peripherin-positive (our revised Figure 2). As we noted in already our previous ms, we see weak diffuse peripherin-reactivity in the trigeminal nucleus (the inferior olive according to the referee), but no peripherin-reactive axon bundles (i.e. climbing fibers) that are seen in the inferior olive of other species. We also see no peripherin-reactive axon bundles (i.e. the olivo-cerebellar tract) arriving in the trigeminal nucleus as the tissue surrounding the trigeminal nucleus is devoid of peripherin-reactivity. Again, this finding is incompatible with the referee’s ideas. As far as we can tell, the trigeminal fibers are not reactive for peripherin in the elephant, i.e. we did not observe peripherin-reactivity very close to the nerve entry, but unfortunately, we did not stain for peripherin-reactivity into the nerve. As the referee alludes to the absence of peripherin-reactivity in the trigeminal tract is a difference between rodents and elephants.

Change: Our novel Figure 2.

Summary: 

(1) Comparative data of species closely related to elephants (Afrotherians) demonstrates that not all mammals exhibit the "serrated" appearance of the principal nucleus of the inferior olive. 

(2) The location of the IO and Vsens as reported in the current study (IOR and VsensR) would require a significant, and unprecedented, rearrangement of the brainstem in the elephants independently. I argue that the underlying molecular and genetic changes required to achieve this would be so extreme that it would lead to lethal phenotypes. Arguing that the "switcheroo" of the IO and Vsens does occur in the elephant (and no other mammals) and thus doesn't lead to lethal phenotypes is a circular argument that cannot be substantiated. 

(3) Myelin stripes in the subnuclei of the inferior olivary nuclear complex are seen across all related mammals as shown above. Thus, the observation made in the elephant by the authors in what they call the VsensR, is similar to that seen in the IO of related mammals, especially when the IO takes on a more bulbous appearance. These myelin stripes are the origin of the olivocerebellar pathway, and are indeed calretinin immunopositive in the elephant as I show. 

(4) What the authors see aligns perfectly with what has been described previously, the only difference being the names that nuclear complexes are being called. But identifying these nuclei is important, as any functional sequelae, as extensively discussed by the authors, is entirely dependent upon accurately identifying these nuclei. 

(4) The peripherin immunostaining scores an own goal - if peripherin is marking peripheral nerves (as the authors and I believe it is), then why is the VsensR/IOM only "weakly positive" for this stain? This either means that the "extraordinary" tactile sensitivity of the elephant trunk is non-existent, or that the authors have misinterpreted this staining. That there is extensive staining in the fibre pathway dorsal and lateral to the IOR (which I call the spinal trigeminal tract), supports the idea that the authors have misinterpreted their peripherin immunostaining.

(5) Evolutionary expediency. The authors argue that what they report is an expedient way in which to modify the organisation of the brainstem in the elephant to accommodate the "extraordinary" tactile sensitivity. I disagree. As pointed out in my first review, the elephant cerebellum is very large and comprised of huge numbers of morphologically complex neurons. The inferior olivary nuclei in all mammals studied in detail to date, give rise to the climbing fibres that terminate on the Purkinje cells of the cerebellar cortex. It is more parsimonious to argue that, in alignment with the expansion of the elephant cerebellum (for motor control of the trunk), the inferior olivary nuclei (specifically the principal nucleus) have had additional neurons added to accommodate this cerebellar expansion. Such an addition of neurons to the principal nucleus of the inferior olive could readily lead to the loss of the serrated appearance of the principal nucleus of the inferior olive, and would require far less modifications in the developmental genetic program that forms these nuclei. This type of quantitative change appears to be the primary way in which structures are altered in the mammalian brainstem. 

Comment: We still disagree with the referee. We note that our conclusions rest on the analysis of 8 elephant brainstems, which we sectioned in three planes and stained with a variety of metabolic and antibody stains and in which assigned two structures (the inferior olive and the trigeminal nucleus). Most of the evidence cited by the referee stems from a single paper, in which 147 structures were identified based on the analysis of a single brainstem sectioned in one plane and stained with a limited set of antibodies. Our synopsis of the evidence is the following.

(1) We agree with the referee that concerning brainstem position our scheme of a ventromedial trigeminal nucleus and a dorsolateral inferior olive deviates from the usual mammalian position of these nuclei (i.e. a dorsolateral trigeminal nucleus and a ventromedial inferior olive).

(2) Cytoarchitectonics support our partitioning scheme. The compact cellular appearance of our ventromedial trigeminal nucleus is characteristic of trigeminal nuclei. The serrated appearance of our dorsolateral inferior olive is characteristic of the mammalian inferior olive; we acknowledge that the referee claims exceptions here. To our knowledge, nobody has described a mammalian trigeminal nucleus with a serrated appearance (which would apply to the elephant in case the trigeminal nucleus is situated dorsolaterally).

(3) Metabolic staining (Cyto-chrome-oxidase reactivity) supports our partitioning scheme. Specifically, our ventromedial trigeminal nucleus shows intense Cyto-chrome-oxidase reactivity as it is seen in the trigeminal nuclei of trigeminal tactile experts.

(4) Isomorphism. The myelin stripes on our ventromedial trigeminal nucleus are isomorphic to trunk wrinkles. Isomorphism is a characteristic of somatosensory brain structures (barrel, barrelettes, nose-stripes, etc) and we know of no case, where such isomorphism was misleading.

(5) The large-scale organization of our ventromedial trigeminal nuclei in anterior-posterior repeats is characteristic of the mammalian trigeminal nuclei. To our knowledge, no such organization has ever been reported for the inferior olive.

(6) Connectivity analysis supports our partitioning scheme. According to our delineation of the elephant olivo-cerebellar tract, our dorsolateral inferior olive is connected via peripherin-positive climbing fibers to the cerebellum. In contrast, our ventromedial trigeminal nucleus (the referee’s inferior olive) is not connected via climbing fibers to the cerebellum.

Change: As discussed, we advanced further evidence in this revision. Our partitioning scheme (a ventromedial trigeminal nucleus and a dorsolateral inferior olive) is better supported by data and makes more sense than the referee’s suggestion (a dorsolateral trigeminal nucleus and a ventromedial inferior olive). It should be published.

Reviewer #3 (Public Review):

Summary: 

The study claims to investigate trunk representations in elephant trigeminal nuclei located in the brainstem. The researchers identify large protrusions visible from the ventral surface of the brainstem, which they examined using a range of histological methods. However, this ventral location is usually where the inferior olivary complex is found, which challenges the author's assertions about the nucleus under analysis. They find that this brainstem nucleus of elephants contains repeating modules, with a focus on the anterior and largest unit which they define as the putative nucleus principalis trunk module of the trigeminal. The nucleus exhibits low neuron density, with glia outnumbering neurons significantly. The study also utilizes synchrotron X-ray phase contrast tomography to suggest that myelin-stripe-axons traverse this module. The analysis maps myelin-rich stripes in several specimens and concludes that based on their number and patterning that they likely correspond with trunk folds; however this conclusion is not well supported if the nucleus has been misidentified. 

Comment: The referee provides a summary of our work. The referee also notes that the correct identification of the trigeminal nucleus is critical to the message of our paper.

Change: In line with these assessments we focused our revision efforts on the issue of trigeminal nucleus identification, please see our introductory comments and our response to Referee 2.

Strengths: 

The strength of this research lies in its comprehensive use of various anatomical methods, including Nissl staining, myelin staining, Golgi staining, cytochrome oxidase labeling, and synchrotron X-ray phase contrast tomography. The inclusion of quantitative data on cell numbers and sizes, dendritic orientation and morphology, and blood vessel density across the nucleus adds a quantitative dimension. Furthermore, the research is commendable for its high-quality and abundant images and figures, effectively illustrating the anatomy under investigation.

Comment: We appreciate this positive assessment.

Change: None

Weaknesses: 

While the research provides potentially valuable insights if revised to focus on the structure that appears to be inferior olivary nucleus, there are certain additional weaknesses that warrant further consideration. First, the suggestion that myelin stripes solely serve to separate sensory or motor modules rather than functioning as an "axonal supply system" lacks substantial support due to the absence of information about the neuronal origins and the termination targets of the axons. Postmortem fixed brain tissue limits the ability to trace full axon projections. While the study acknowledges these limitations, it is important to exercise caution in drawing conclusions about the precise role of myelin stripes without a more comprehensive understanding of their neural connections. 

Comment: We understand these criticisms and the need for cautious interpretation. As we noted previously, we think that the Elife-publishing scheme, where critical referee commentary is published along with our ms, will make this contribution particularly valuable.

Change: Our additional efforts to secure the correct identification of the trigeminal nucleus.

Second, the quantification presented in the study lacks comparison to other species or other relevant variables within the elephant specimens (i.e., whole brain or brainstem volume). The absence of comparative data to different species limits the ability to fully evaluate the significance of the findings. Comparative analyses could provide a broader context for understanding whether the observed features are unique to elephants or more common across species. This limitation in comparative data hinders a more comprehensive assessment of the implications of the research within the broader field of neuroanatomy. Furthermore, the quantitative comparisons between African and Asian elephant specimens should include some measure of overall brain size as a covariate in the analyses. Addressing these weaknesses would enable a richer interpretation of the study's findings. 

Comment: We understand, why the referee asks for additional comparative data, which would make our study more meaningful. We note that we already published a quantitative comparison of African and Asian elephant facial nuclei (Kaufmann et al. 2022). The quantitative differences between African and Asian elephant facial nuclei are similar in magnitude to what we observed here for the trigeminal nucleus, i.e. African elephants have about 10-15% more facial nucleus neurons than Asian elephants. The referee also notes that data on overall elephant brain size might be important for interpreting our data. We agree with this sentiment and we are preparing a ms on African and Asian elephant brain size. We find – unexpectedly given the larger body size of African elephants – that African elephants have smaller brains than Asian elephants. The finding might imply that African elephants, which have more facial nucleus neurons and more trigeminal nucleus trunk module neurons, are neurally more specialized in trunk control than Asian elephants.

Change: We are preparing a further ms on African and Asian elephant brain size, a first version of this work has been submitted.

Reviewer #4 (Public Review): 

Summary: 

The authors report a novel isomorphism in which the folds of the elephant trunk are recognizably mapped onto the principal sensory trigeminal nucleus in the brainstem. Further, they identifiy the enlarged nucleus as being situated in this species in an unusual ventral midline position. 

Comment: The referee summarizes our work.

Change: None.

Strengths: 

The identity of the purported trigeminal nucleus and the isomorphic mapping with the trunk folds is supported by multiple lines of evidence: enhanced staining for cytochrome oxidase, an enzyme associated with high metabolic activity; dense vascularization, consistent with high metabolic activity; prominent myelinated bundles that partition the nucleus in a 1:1 mapping of the cutaneous folds in the trunk periphery; near absence of labeling for the anti-peripherin antibody, specific for climbing fibers, which can be seen as expected in the inferior olive; and a high density of glia.

Comment: The referee again reviews some of our key findings.

Change: None. 

Weaknesses: 

Despite the supporting evidence listed above, the identification of the gross anatomical bumps, conspicuous in the ventral midline, is problematic. This would be the standard location of the inferior olive, with the principal trigeminal nucleus occupying a more dorsal position. This presents an apparent contradiction which at a minimum needs further discussion. Major species-specific specializations and positional shifts are well-documented for cortical areas, but nuclear layouts in the brainstem have been considered as less malleable. 

Comment: The referee notes that our discrepancy with referee 2, needs to be addressed with further evidence and discussion, given the unusual position of both inferior olive and trigeminal nucleus in the partitioning scheme and that the mammalian brainstem tends to be positionally conservative. We agree with the referee. We note that – based on the immense size of the elephant trigeminal ganglion (50 g), half the size of a monkey brain – it was expected that the elephant trigeminal nucleus ought to be exceptionally large.

Change: We did additional experimental work to resolve this matter: (i) We ascertained that elephant climbing fibers are strongly peripherin-positive. (ii) Based on elephant climbing fiber peripherin-reactivity we delineated the elephant olivo-cerebellar tract. We find that the olivo-cerebellar connects to the structure we refer to as inferior olive to the cerebellum. (iii) We also found that the trigeminal nucleus (the structure the referee refers to as inferior olive) appears to receive no climbing fibers. (iv) We provide indications that the tracing of the trigeminal nerve into the olivo-cerebellar tract by Maseko et al. 2023 was erroneous (Referee-Figure 1). These novel findings support our ideas.

Reviewer #5 (Public Review): 

After reading the manuscript and the concerns raised by reviewer 2 I see both sides of the argument - the relative location of trigeminal nucleus versus the inferior olive is quite different in elephants (and different from previous studies in elephants), but when there is a large disproportionate magnification of a behaviorally relevant body part at most levels of the nervous system (certainly in the cortex and thalamus), you can get major shifting in location of different structures. In the case of the elephant, it looks like there may be a lot of shifting. Something that is compelling is that the number of modules separated but the myelin bands correspond to the number of trunk folds which is different in the different elephants. This sort of modular division based on body parts is a general principle of mammalian brain organization (demonstrated beautifully for the cuneate and gracile nucleus in primates, VP in most of species, S1 in a variety of mammals such as the star nosed mole and duck-billed platypus). I don't think these relative changes in the brainstem would require major genetic programming - although some surely exists. Rodents and elephants have been independently evolving for over 60 million years so there is a substantial amount of time for changes in each l lineage to occur.

I agree that the authors have identified the trigeminal nucleus correctly, although comparisons with more out groups would be needed to confirm this (although I'm not suggesting that the authors do this). I also think the new figure (which shows previous divisions of the brainstem versus their own) allows the reader to consider these issues for themselves. When reviewing this paper, I actually took the time to go through atlases of other species and even look at some of my own data from highly derived species. Establishing homology across groups based only on relative location is tough especially when there appears to be large shifts in relative location of structures. My thoughts are that the authors did an extraordinary amount of work on obtaining, processing and analyzing this extremely valuable tissue. They document their work with images of the tissue and their arguments for their divisions are solid. I feel that they have earned the right to speculate - with qualifications - which they provide. 

Comment: The referee summarizes our work and appears to be convinced by the line of our arguments. We are most grateful for this assessment. We add, again, that the skeptical assessment of referee 2 will be published as well and will give the interested reader the possibility to view another perspective on our work.

Change: None. 

Recommendations for the authors: 

Reviewer #1 (Recommendations For The Authors):

With this manuscript being virtually identical to the previous version, it is possible that some of the definitive conclusions about having identified the elephant trigeminal nucleus and trunk representation should be moderated in a more nuanced manner, especially given the careful and experienced perspective from reviewers with first hand knowledge elephant neuroanatomy.

Comment: We agree that both our first and second revisions were very much centered on the debate of the correct identification of the trigeminal nucleus and that our ms did not evolve as much in other regards. This being said we agree with Referee 2 that we needed to have this debate. We also think we advanced important novel data in this context (the delineation of elephant olivo-cerebellar tract through the peripherin-antibody).

Changes: Our revised Figure 2. 

The peripherin staining adds another level of argument to the authors having identified the trigeminal brainstem instead of the inferior olive, if differential expression of peripherin is strong enough to distinguish one structure from the other.

Comment: We think we showed too little peripherin-antibody staining in our previous revision. We have now addressed this problem.

Changes: Our revised Figure 2, i.e. the delineation of elephant olivo-cerebellar tract through the peripherin-antibody).

There are some minor corrections to be made with the addition of Fig. 2., including renumbering the figures in the manuscript (e.g., 406, 521). 

I continue to appreciate this novel investigation of the elephant brainstem and find it an interesting and thorough study, with the use of classical and modern neuroanatomical methods.

Comment: We are thankful for this positive assessment.

Reviewer #2 (Recommendations For The Authors):

I do realise the authors are very unhappy with me and the reviews I have submitted. I do apologise if feelings have been hurt, and I do understand the authors put in a lot of hard work and thought to develop what they have; however, it is unfortunate that the work and thoughts are not correct. Science is about the search for the truth and sometimes we get it wrong. This is part of the scientific process and why most journals adhere to strict review processes of scientific manuscripts. As I said previously, the authors can use their data to write a paper describing and quantifying Golgi staining of neurons in the principal olivary nucleus of the elephant that should be published in a specialised journal and contextualised in terms of the motor control of the trunk and the large cerebellum of the elephant. 

Comment: We appreciate the referee’s kind words. Also, no hard feelings from our side, this is just a scientific debate. In our experience, neuroanatomical debates are resolved by evidence and we note that we provide evidence strengthening our identification of the trigeminal nucleus and inferior olive. As far as we can tell from this effort and the substantial evidence accumulated, the referee is wrong.

Reviewer #4 (Recommendations For The Authors):

As a new reviewer, I have benefited from reading the previous reviews and Author response, even while having several new comments to add. 

(1) The identification of the inferior olive and trigeminal nuclei is obviously center stage. An enlargement of the trigeminal nuclei is not necessarily problematic, given the published reports on the dramatic enlargement of the trigeminal nerve (Purkart et al., 2022). At issue is the conspicuous relocation of the trigeminal nuclei that is being promoted by Reveyaz et al. Conspicuous rearrangements are not uncommon; for example, primary sensory cortical fields in different species (fig. 1 in H.H.A. Oelschlager for dolphins; S. De Vreese et al. (2023) for cetaceans, L. Krubitzer on various species, in the context of evolution). The difficult point here concerns what looks like a rather conspicuous gross anatomical rearrangement, in BRAINSTEM - the assumption being that the brainstem bauplan is going to be specifically conservative and refractory to gross anatomical rearrangement. 

Comment: We agree with the referee that the brainstem rearrangements are unexpected. We also think that the correct identification of nuclei needs to be at the center of our revision efforts.

Change: Our revision provided further evidence (delineation of the olivo-cerebellar tract, characterization of the trigeminal nerve entry) about the identity of the nuclei we studied.

Why would a major nucleus shift to such a different location? and how? Can ex vivo DTI provide further support of the correct identification? Is there other "disruption" in the brainstem? What occupies the traditional position of the trigeminal nuclei? An atlas-equivalent coronal view of the entire brainstem would be informative. The Authors have assembled multiple criteria to support their argument that the ventral "bumps" are in fact a translocated trigeminal principal nucleus: enhanced CO staining, enhanced vascularization, enhanced myelination (via Golgi stains and tomography), very scant labeling for a climbing fiber specific antibody ( anti-peripherin), vs. dense staining of this in the alternative structure that they identify as IO; and a high density of glia. Admittedly, this should be sufficient, but the proposed translocation (in the BRAINSTEM) is sufficiently startling that this is arguably NOT sufficient. <br /> The terminology of "putative" is helpful, but a more cogent presentation of the results and more careful discussion might succeed in winning over at least some of a skeptical readership. 

Comment: We do not know, what led to the elephant brainstem rearrangements we propose. If the trigeminal nuclei had expanded isometrically in elephants from the ancestral pattern, one would have expected a brain with big lateral bumps, not the elephant brain with its big ventromedial bumps. We note, however, that very likely the expansion of the elephant trigeminal nuclei did not occur isometrically. Instead, the neural representation of the elephant nose expanded dramatically and in rodents the nose is represented ventromedially in the brainstem face representation. Thus, we propose a ‘ventromedial outgrowth model’ according to which the elephant ventromedial trigeminal bumps result from a ventromedially direct outgrowth of the ancestral ventromedial nose representation.

We advanced substantially more evidence to support our partitioning scheme, including the delineation of the olivo-cerebellar tract based on peripherin-reactivity. We also identified problems in previous partitioning schemes, such as the claim that the trigeminal nerve continues into the ~4x smaller olivocerebellar tract (Referee-Figure 1C, D); we think such a flow of fibers, (which is also at odds with peripherin-antibody-reactivity and the appearance of nerve and olivocerebellar tract), is highly unlikely if not physically impossible. With all that we do not think that we overstate our case in our cautiously presented ms.

Change: We added evidence on the identification of elephant trigeminal nuclei and inferior olive.

(2) Role of myelin. While the photos of myelin are convincing, it would be nice to have further documentation. Gallyas? Would antibodies to MBP work? What is the myelin distribution in the "standard" trigeminal nuclei (human? macaque or chimpanzee?). What are alternative sources of the bundles? Regardless, I think it would be beneficial to de-emphasize this point about the role of myelin in demarcating compartments. <br /> I would in fact suggest an alternative (more neutral) title that might highlight instead the isomorphic feature; for example, "An isomorphic representation of Trunk folds in the Elephant Trigeminal Nucleus." The present title stresses myelin, but figure 1 already focuses on CO. Additionally, the folds are actually mentioned almost in passing until later in the manuscript. I recommend a short section on these at the beginning of the Results to serve as a useful framework.

Here I'm inclined to agree with the Reviewer, that the Authors' contention that the myelin stipes serve PRIMARILY to separate trunk-fold domains is not particularly compelling and arguably a distraction. The point can be made, but perhaps with less emphasis. After all, the fact that myelin has multiple roles is well-established, even if frequently overlooked. In addition, the Authors might make better use of an extensive relevant literature related to myelin as a compartmental marker; for example, results and discussion in D. Haenelt....N. Weiskopf (eLife, 2023), among others. Another example is the heavily myelinated stria of Gennari in primate visual cortex, consisting of intrinsic pyramidal cell axons, but where the role of the myelination has still not been elucidated. 

Comment: (1) Documentation of myelin. We note that we show further identification of myelinated fibers by the fluorescent dye fluomyelin in Figure 4B. We also performed additional myelin stains as the gold-myelin stain after the protocol of Schmued (Referee-Figure 2). In the end, nothing worked quite as well to visualize myelin-stripes as the bright-field images shown in Figure 4A and it is only the images that allowed us to match myelin-stripes to trunk folds. Hence, we focus our presentation on these images.

(2) Title: We get why the referee envisions an alternative title. This being said, we would like to stick with our current title, because we feel it highlights the major novelty we discovered.

(3) We agree with many of the other comments of the referee on myelin phenomenology. We missed the Haenelt reference pointed out by the referee and think it is highly relevant to our paper

Change: 1. Review image 2. Inclusion of the Haenelt-reference.

Author response image 2.

Myelin stripes of the elephant trunk module visualized by Gold-chloride staining according to Schmued. A, Low magnification micrograph of the trunk module of African elephant Indra stained with AuCl according to Schmued. The putative finger is to the left, proximal is to the right. Myelin stripes can easily be recognized. The white box indicates the area shown in B. B, high magnification micrograph of two myelin stripes. Individual gold-stained (black) axons organized in myelin stripes can be recognized.

Schmued, L. C. (1990). A rapid, sensitive histochemical stain for myelin in frozen brain sections. Journal of Histochemistry & Cytochemistry,38(5), 717-720.

Are the "bumps" in any way "analogous" to the "brain warts" seen in entorhinal areas of some human brains (G. W. van Hoesen and A. Solodkin (1993)? 

Comment: We think this is a similar phenomenon.

Change: We included the Hoesen and A. Solodkin (1993) reference in our discussion.

At least slightly more background (ie, a separate section or, if necessary, supplement) would be helpful, going into more detail on the several subdivisions of the ION and if these undergo major alterations in the elephant.

Comment: The strength of the paper is the detailed delineation of the trunk module, based on myelin stripes and isomorphism. We don’t think we have strong evidence on ION subdivisions, because it appears the trigeminal tract cannot be easily traced in elephants. Accordingly, we find it difficult to add information here.

Change: None.

Is there evidence from the literature of other conspicuous gross anatomical translocations, in any species, especially in subcortical regions? 

Comment: The best example that comes to mind is the star-nosed mole brainstem. There is a beautiful paper comparing the star-nosed mole brainstem to the normal mole brainstem (Catania et al 2011). The principal trigeminal nucleus in the star-nosed mole is far more rostral and also more medial than in the mole; still, such rearrangements are minor compared to what we propose in elephants.

Catania, Kenneth C., Duncan B. Leitch, and Danielle Gauthier. "A star in the brainstem reveals the first step of cortical magnification." PloS one 6.7 (2011): e22406.

Change: None.

(3) A major point concerns the isomorphism between the putative trigeminal nuclei and the trunk specialization. I think this can be much better presented, at least with more discussion and other examples. The Authors mention about the rodent "barrels," but it seemed strange to me that they do not refer to their own results in pig (C. Ritter et al., 2023) nor the work from Ken Catania, 2002 (star-nosed mole; "fingerprints in the brain") or other that might be appropriate. I concur with the Reviewer that there should be more comparative data. 

Comment: We agree.

Change: We added a discussion of other isomorphisms including the the star-nosed mole to our paper.

(4) Textual organization could be improved. 

The Abstract all-important Introduction is a longish, semi "run-on" paragraph. At a minimum this should be broken up. The last paragraph of the Introduction puts forth five issues, but these are only loosely followed in the Results section. I think clarity and good organization is of the upmost importance in this manuscript. I recommend that the Authors begin the Results with a section on the trunk folds (currently figure 5, and discussion), continue with the several points related to the identification of the trigeminal nuclei, and continue with a parallel description of ION with more parallel data on the putative trigeminal and IO structures (currently referee Table 1, but incorporate into the text and add higher magnification of nucleus-specific cell types in the IO and trigeminal nuclei). Relevant comparative data should be included in the Discussion.

Comment: 1. We agree with the referee that our abstract needed to be revised. 2. We also think that our ms was heavily altered by the insertion of the new Figure 2, which complemented Figure 1 from our first submission and is concerned with the identification of the inferior olive. From a standpoint of textual flow such changes were not ideal, but the revisions massively added to the certainty with which we identify the trigeminal nuclei. Thus, although we are not as content as we were with the flow, we think the ms advanced in the revision process and we would like to keep the Figure sequence as is. 3. We already noted above that we included additional comparative evidence.

Change: 1. We revised our abstract. 2. We added comparative evidence.

Reviewer #5 (Recommendations For The Authors): 

The data is invaluable and provides insights into some of the largest mammals on the planet. 

Comment: We are incredibly thankful for this positive assessment.

Author response:

The following is the authors’ response to the previous reviews.

eLife Assessment

This neuroimaging and electrophysiology study in a small cohort of congenital cataract patients with sight recovery aims to characterize the effects of early visual deprivation on excitatory and inhibitory balance in visual cortex. While contrasting sight-recovery with visually intact controls suggested the existence of persistent alterations in Glx/GABA ratio and aperiodic EEG signals, it provided only incomplete evidence supporting claims about the effects of early deprivation itself. The reported data were considered valuable, given the rare study population. However, the small sample sizes, lack of a specific control cohort and multiple methodological limitations will likely restrict usefulness to scientists working in this particular subfield.

We thank the reviewing editors for their consideration and updated assessment of our manuscript after its first revision.

In order to assess the effects of early deprivation, we included an age-matched, normally sighted control group recruited from the same community, measured in the same scanner and laboratory. This study design is analogous to numerous studies in permanently congenitally blind humans, which typically recruited sighted controls, but hardly ever individuals with a different, e.g. late blindness history. In order to improve the specificity of our conclusions, we used a frontal cortex voxel in addition to a visual cortex voxel (MRS). Analogously, we separately analyzed occipital and frontal electrodes (EEG).

Moreover, we relate our findings in congenital cataract reversal individuals to findings in the literature on permanent congenital blindness. Note, there are, to the best of our knowledge, neither MRS nor resting-state EEG studies in individuals with permanent late blindness.

Our participants necessarily have nystagmus and low visual acuity due to their congenital deprivation phase, and the existence of nystagmus is a recruitment criterion to diagnose congenital cataracts.

It might be interesting for future studies to investigate individuals with transient late blindness. However, such a study would be ill-motivated had we not found differences between the most “extreme” of congenital visual deprivation conditions and normally sighted individuals (analogous to why earlier research on permanent blindness investigated permanent congenitally blind humans first, rather than permanently late blind humans, or both in the same study). Any result of these future work would need the reference to our study, and neither results in these additional groups would invalidate our findings.

Since all our congenital cataract reversal individuals by definition had visual impairments, we included an eyes closed condition, both in the MRS and EEG assessment. Any group effect during the eyes closed condition cannot be due to visual acuity deficits changing the bottom-up driven visual activation.

As we detail in response to review 3, our EEG analyses followed the standards in the field.

Public Reviews:

Reviewer #1 (Public review):

Summary

In this human neuroimaging and electrophysiology study, the authors aimed to characterise effects of a period of visual deprivation in the sensitive period on excitatory and inhibitory balance in the visual cortex. They attempted to do so by comparing neurochemistry conditions ('eyes open', 'eyes closed') and resting state, and visually evoked EEG activity between ten congenital cataract patients with recovered sight (CC), and ten age-matched control participants (SC) with normal sight.

First, they used magnetic resonance spectroscopy to measure in vivo neurochemistry from two locations, the primary location of interest in the visual cortex, and a control location in the frontal cortex. Such voxels are used to provide a control for the spatial specificity of any effects, because the single-voxel MRS method provides a single sampling location. Using MR-visible proxies of excitatory and inhibitory neurotransmission, Glx and GABA+ respectively, the authors report no group effects in GABA+ or Glx, no difference in the functional conditions 'eyes closed' and 'eyes open'. They found an effect of group in the ratio of Glx/GABA+ and no similar effect in the control voxel location. They then perform multiple exploratory correlations between MRS measures and visual acuity, and report a weak positive correlation between the 'eyes open' condition and visual acuity in CC participants.

The same participants then took part in an EEG experiment. The authors selected two electrodes placed in the visual cortex for analysis and report a group difference in an EEG index of neural activity, the aperiodic intercept, as well as the aperiodic slope, considered a proxy for cortical inhibition. Control electrodes in the frontal region did not present with the same pattern. They report an exploratory correlation between the aperiodic intercept and Glx in one out of three EEG conditions.

The authors report the difference in E/I ratio, and interpret the lower E/I ratio as representing an adaptation to visual deprivation, which would have initially caused a higher E/I ratio. Although intriguing, the strength of evidence in support of this view is not strong. Amongst the limitations are the low sample size, a critical control cohort that could provide evidence for higher E/I ratio in CC patients without recovered sight for example, and lower data quality in the control voxel. Nevertheless, the study provides a rare and valuable insight into experience-dependent plasticity in the human brain.

Strengths of study

How sensitive period experience shapes the developing brain is an enduring and important question in neuroscience. This question has been particularly difficult to investigate in humans. The authors recruited a small number of sight-recovered participants with bilateral congenital cataracts to investigate the effect of sensitive period deprivation on the balance of excitation and inhibition in the visual brain using measures of brain chemistry and brain electrophysiology. The research is novel, and the paper was interesting and well written.

Limitations

Low sample size. Ten for CC and ten for SC, and further two SC participants were rejected due to lack of frontal control voxel data. The sample size limits the statistical power of the dataset and increases the likelihood of effect inflation.

In the updated manuscript, the authors have provided justification for their sample size by pointing to prior studies and the inherent difficulties in recruiting individuals with bilateral congenital cataracts. Importantly, this highlights the value the study brings to the field while also acknowledging the need to replicate the effects in a larger cohort.

Lack of specific control cohort. The control cohort has normal vision. The control cohort is not specific enough to distinguish between people with sight loss due to different causes and patients with congenital cataracts with co-morbidities. Further data from a more specific populations, such as patients whose cataracts have not been removed, with developmental cataracts, or congenitally blind participants, would greatly improve the interpretability of the main finding. The lack of a more specific control cohort is a major caveat that limits a conclusive interpretation of the results.

In the updated version, the authors have indicated that future studies can pursue comparisons between congenital cataract participants and cohorts with later sight loss.

MRS data quality differences. Data quality in the control voxel appears worse than in the visual cortex voxel. The frontal cortex MRS spectrum shows far broader linewidth than the visual cortex (Supplementary Figures). Compared to the visual voxel, the frontal cortex voxel has less defined Glx and GABA+ peaks; lower GABA+ and Glx concentrations, lower NAA SNR values; lower NAA concentrations. If the data quality is a lot worse in the FC, then small effects may not be detectable.

In the updated version, the authors have added more information that informs the reader of the MRS quality differences between voxel locations. This increases the transparency of their reporting and enhances the assessment of the results.

Because of the direction of the difference in E/I, the authors interpret their findings as representing signatures of sight improvement after surgery without further evidence, either within the study or from the literature. However, the literature suggests that plasticity and visual deprivation drives the E/I index up rather than down. Decreasing GABA+ is thought to facilitate experience dependent remodelling. What evidence is there that cortical inhibition increases in response to a visual cortex that is over-sensitised to due congenital cataracts? Without further experimental or literature support this interpretation remains very speculative.

The updated manuscript contains key reference from non-human work to justify their interpretation.

Heterogeneity in patient group. Congenital cataract (CC) patients experienced a variety of duration of visual impairment and were of different ages. They presented with co-morbidities (absorbed lens, strabismus, nystagmus). Strabismus has been associated with abnormalities in GABAergic inhibition in the visual cortex. The possible interactions with residual vision and confounds of co-morbidities are not experimentally controlled for in the correlations, and not discussed.

The updated document has addressed this caveat.

Multiple exploratory correlations were performed to relate MRS measures to visual acuity (shown in Supplementary Materials), and only specific ones shown in the main document. The authors describe the analysis as exploratory in the 'Methods' section. Furthermore, the correlation between visual acuity and E/I metric is weak, not corrected for multiple comparisons. The results should be presented as preliminary, as no strong conclusions can be made from them. They can provide a hypothesis to test in a future study.

This has now been done throughout the document and increases the transparency of the reporting.

P.16 Given the correlation of the aperiodic intercept with age ("Age negatively correlated with the aperiodic intercept across CC and SC individuals, that is, a flattening of the intercept was observed with age"), age needs to be controlled for in the correlation between neurochemistry and the aperiodic intercept. Glx has also been shown to negatively correlates with age.

This caveat has been addressed in the revised manuscript.

Multiple exploratory correlations were performed to relate MRS to EEG measures (shown in Supplementary Materials), and only specific ones shown in the main document. Given the multiple measures from the MRS, the correlations with the EEG measures were exploratory, as stated in the text, p.16, and in Fig.4. yet the introduction said that there was a prior hypothesis "We further hypothesized that neurotransmitter changes would relate to changes in the slope and intercept of the EEG aperiodic activity in the same subjects." It would be great if the text could be revised for consistency and the analysis described as exploratory.

This has been done throughout the document and increases the transparency of the reporting.

The analysis for the EEG needs to take more advantage of the available data. As far as I understand, only two electrodes were used, yet far more were available as seen in their previous study (Ossandon et al., 2023). The spatial specificity is not established. The authors could use the frontal cortex electrode (FP1, FP2) signals as a control for spatial specificity in the group effects, or even better, all available electrodes and correct for multiple comparisons. Furthermore, they could use the aperiodic intercept vs Glx in SC to evaluate the specificity of the correlation to CC.

This caveat has been addressed. The authors have added frontal electrodes to their analysis, providing an essential regional control for the visual cortex location.

Comments on the latest version:

The authors have made reasonable adjustments to their manuscript that addressed most of my comments by adding further justification for their methodology, essential literature support, pointing out exploratory analyses, limitations and adding key control analyses. Their revised manuscript has overall improved, providing valuable information, though the evidence that supports their claims is still incomplete.

We thank the reviewer for suggesting ways to improve our manuscript and carefully reassessing our revised manuscript.

Reviewer #2 (Public review):

Summary:

The study examined 10 congenitally blind patients who recovered vision through the surgical removal of bilateral dense cataracts, measuring neural activity and neuro chemical profiles from the visual cortex. The declared aim is to test whether restoring visual function after years of complete blindness impacts excitation/inhibition balance in the visual cortex.

Strengths:

The findings are undoubtedly useful for the community, as they contribute towards characterising the many ways in which this special population differs from normally sighted individuals. The combination of MRS and EEG measures is a promising strategy to estimate a fundamental physiological parameter - the balance between excitation and inhibition in the visual cortex, which animal studies show to be heavily dependent upon early visual experience. Thus, the reported results pave the way for further studies, which may use a similar approach to evaluate more patients and control groups.

Weaknesses:

The main methodological limitation is the lack of an appropriate comparison group or condition to delineate the effect of sight recovery (as opposed to the effect of congenital blindness). Few previous studies suggested that Excitation/Inhibition ratio in the visual cortex is increased in congenitally blind patients; the present study reports that E/I ratio decreases instead. The authors claim that this implies a change of E/I ratio following sight recovery. However, supporting this claim would require showing a shift of E/I after vs. before the sight-recovery surgery, or at least it would require comparing patients who did and did not undergo the sight-recovery surgery (as common in the field).

We thank the reviewer for suggesting ways to improve our manuscript and carefully reassessing our revised manuscript.

Since we have not been able to acquire longitudinal data with the experimental design of the present study in congenital cataract reversal individuals, we compared the MRS and EEG results of congenital cataract reversal individuals  to published work in congenitally permanent blind individuals. We consider this as a resource saving approach. We think that the results of our cross-sectional study now justify the costs and enormous efforts (and time for the patients who often have to travel long distances) associated with longitudinal studies in this rare population.

There are also more technical limitations related to the correlation analyses, which are partly acknowledged in the manuscript. A bland correlation between GLX/GABA and the visual impairment is reported, but this is specific to the patients group (N=10) and would not hold across groups (the correlation is positive, predicting the lowest GLX/GABA ratio values for the sighted controls - opposite of what is found). There is also a strong correlation between GLX concentrations and the EEG power at the lowest temporal frequencies. Although this relation is intriguing, it only holds for a very specific combination of parameters (of the many tested): only with eyes open, only in the patients group.

Given the exploratory nature of the correlations, we do not base the majority of our conclusions on this analysis. There are no doubts that the reported correlations need replication; however, replication is only possible after a first report. Thus, we hope to motivate corresponding analyses in further studies.

It has to be noted that in the present study significance testing for correlations were corrected for multiple comparisons, and that some findings replicate earlier reports (e.g. effects on EEG aperiodic slope, alpha power, and correlations with chronological age).

Conclusions:

The main claim of the study is that sight recovery impacts the excitation/inhibition balance in the visual cortex, estimated with MRS or through indirect EEG indices. However, due to the weaknesses outlined above, the study cannot distinguish the effects of sight recovery from those of visual deprivation. Moreover, many aspects of the results are interesting but their validation and interpretation require additional experimental work.

We interpret the group differences between individuals tested years after congenital visual deprivation and normally sighted individuals as supportive of the E/I ratio being impacted by congenital visual deprivation. In the absence of a sensitive period for the development of an E/I ratio, individuals with a transient phase of congenital blindness might have developed a visual system indistinguishable  from normally sighted individuals. As we demonstrate, this is not so. Comparing the results of congenitally blind humans with those of congenitally permanently blind humans (from previous studies) allowed us to identify changes of E/I ratio, which add to those found for congenital blindness.  

We thank the reviewer for the helpful comments and suggestions related to the first submission and first revision of our manuscript. We are keen to translate some of them into future studies.

Reviewer #3 (Public review):

This manuscript examines the impact of congenital visual deprivation on the excitatory/inhibitory (E/I) ratio in the visual cortex using Magnetic Resonance Spectroscopy (MRS) and electroencephalography (EEG) in individuals whose sight was restored. Ten individuals with reversed congenital cataracts were compared to age-matched, normally sighted controls, assessing the cortical E/I balance and its interrelationship and to visual acuity. The study reveals that the Glx/GABA ratio in the visual cortex and the intercept and aperiodic signal are significantly altered in those with a history of early visual deprivation, suggesting persistent neurophysiological changes despite visual restoration.

First of all, I would like to disclose that I am not an expert in congenital visual deprivation, nor in MRS. My expertise is in EEG (particularly in the decomposition of periodic and aperiodic activity) and statistical methods.

Although the authors addressed some of the concerns of the previous version, major concerns and flaws remain in terms of methodological and statistical approaches along with the (over)interpretation of the results. Specific concerns include:

(1 3.1) Response to Variability in Visual Deprivation<br /> Rather than listing the advantages and disadvantages of visual deprivation, I recommend providing at least a descriptive analysis of how the duration of visual deprivation influenced the measures of interest. This would enhance the depth and relevance of the discussion.

Although Review 2 and Review 3 (see below) pointed out problems in interpreting multiple correlational analyses in small samples, we addressed this request by reporting such correlations between visual deprivation history and measured EEG/MRS outcomes.

Calculating the correlation between duration of visual deprivation and behavioral or brain measures is, in fact, a common suggestion. The existence of sensitive periods, which are typically assumed to not follow a linear gradual decline of neuroplasticity, does not necessary allow predicting a correlation with duration of blindness. Daphne Maurer has additionally worked on the concept of “sleeper effects” (Maurer et al., 2007), that is, effects on the brain and behavior by early deprivation which are observed only later in life when the function/neural circuits matures.

In accordance with this reasoning, we did not observe a significant correlation between duration of visual deprivation and any of our dependent variables.

(2 3.2) Small Sample Size<br /> The issue of small sample size remains problematic. The justification that previous studies employed similar sample sizes does not adequately address the limitation in the current study. I strongly suggest that the correlation analyses should not feature prominently in the main manuscript or the abstract, especially if the discussion does not substantially rely on these correlations. Please also revisit the recommendations made in the section on statistical concerns.

In the revised manuscript, we explicitly mention that our sample size is not atypical for the special group investigated, but that a replication of our results in larger samples would foster their impact. We only explicitly mention correlations that survived stringent testing for multiple comparisons in the main manuscript.

Given the exploratory nature of the correlations, we have not based the majority of our claims on this analysis.

(3 3.3) Statistical Concerns<br /> While I appreciate the effort of conducting an independent statistical check, it merely validates whether the reported statistical parameters, degrees of freedom (df), and p-values are consistent. However, this does not address the appropriateness of the chosen statistical methods.

We did not intend for the statcheck report to justify the methods used for statistics, which we have done in a separate section with normality and homogeneity testing (Supplementary Material S9), and references to it in the descriptions of the statistical analyses (Methods, Page 13, Lines 326-329 and Page 15, Lines 400-402).

Several points require clarification or improvement:<br /> (4) Correlation Methods: The manuscript does not specify whether the reported correlation analyses are based on Pearson or Spearman correlation.

The depicted correlations are Pearson correlations. We will add this information to the Methods.

(5) Confidence Intervals: Include confidence intervals for correlations to represent the uncertainty associated with these estimates.

We have added the confidence intervals for all measured correlations to the second revision of our manuscript.

(6) Permutation Statistics: Given the small sample size, I recommend using permutation statistics, as these are exact tests and more appropriate for small datasets.

Our study focuses on a rare population, with a sample size limited by the availability of participants. Our findings provide exploratory insights rather than make strong inferential claims. To this end, we have ensured that our analysis adheres to key statistical assumptions (Shapiro-Wilk as well as Levene’s tests, Supplementary Material S9), and reported our findings with effect sizes, appropriate caution and context.

(7) Adjusted P-Values: Ensure that reported Bonferroni corrected p-values (e.g., p > 0.999) are clearly labeled as adjusted p-values where applicable.

In the revised manuscript, we have changed Figure 4 to say ‘adjusted p,’  which we indeed reported.

(8) Figure 2C

Figure 2C still lacks crucial information that the correlation between Glx/GABA ratio and visual acuity was computed solely in the control group (as described in the rebuttal letter). Why was this analysis restricted to the control group? Please provide a rationale.

Figure 2C depicts the correlation between Glx/GABA+ ratio and visual acuity in the congenital cataract reversal group, not the control group. This is mentioned in the Figure 2 legend, as well as in the main text where the figure is referred to (Page 18, Line 475).

The correlation analyses between visual acuity and MRS/EEG measures were only performed in the congenital cataract reversal group since the sighed control group comprised of individuals with vision in the normal range; thus this analyses would not make sense. Table 1 with the individual visual acuities for all participants, including the normally sighted controls, shows the low variance in the latter group.  

For variables in which no apiori group differences in variance were predicted, we performed the correlation analyses across groups (see Supplementary Material S12, S15).

We have now highlighted these motivations more clearly in the Methods of the revised manuscript (Page 16, Lines 405-410).

(9 3.4) Interpretation of Aperiodic Signal

Relying on previous studies to interpret the aperiodic slope as a proxy for excitation/inhibition (E/I) does not make the interpretation more robust.

How to interpret aperiodic EEG activity has been subject of extensive investigation. We cite studies which provide evidence from multiple species (monkeys, humans) and measurements (EEG, MEG, ECoG), including studies which pharmacologically manipulated E/I balance.

Whether our findings are robust, in fact, requires a replication study. Importantly, we analyzed the intercept of the aperiodic activity fit as well, and discuss results related to the intercept.

Quote:

“(3.4) Interpretation of aperiodic signal:

- Several recent papers demonstrated that the aperiodic signal measured in EEG or ECoG is related to various important aspects such as age, skull thickness, electrode impedance, as well as cognition. Thus, currently, very little is known about the underlying effects which influence the aperiodic intercept and slope. The entire interpretation of the aperiodic slope as a proxy for E/I is based on a computational model and simulation (as described in the Gao et al. paper).

Apart from the modeling work from Gao et al., multiple papers which have also been cited which used ECoG, EEG and MEG and showed concomitant changes in aperiodic activity with pharmacological manipulation of the E/I ratio (Colombo et al., 2019; Molina et al., 2020; Muthukumaraswamy & Liley, 2018). Further, several prior studies have interpreted changes in the aperiodic slope as reflective of changes in the E/I ratio, including studies of developmental groups (Favaro et al., 2023; Hill et al., 2022; McSweeney et al., 2023; Schaworonkow & Voytek, 2021) as well as patient groups (Molina et al., 2020; Ostlund et al., 2021).

- The authors further wrote: We used the slope of the aperiodic (1/f) component of the EEG spectrum as an estimate of E/I ratio (Gao et al., 2017; Medel et al., 2020; Muthukumaraswamy & Liley, 2018). This is a highly speculative interpretation with very little empirical evidence. These papers were conducted with ECoG data (mostly in animals) and mostly under anesthesia. Thus, these studies only allow an indirect interpretation by what the 1/f slope in EEG measurements is actually influenced.

Note that Muthukumaraswamy et al. (2018) used different types of pharmacological manipulations and analyzed periodic and aperiodic MEG activity in humans, in addition to monkey ECoG (Muthukumaraswamy & Liley, 2018). Further, Medel et al. (now published as Medel et al., 2023) compared EEG activity in addition to ECoG data after propofol administration. The interpretation of our results are in line with a number of recent studies in developing (Hill et al., 2022; Schaworonkow & Voytek, 2021) and special populations using EEG. As mentioned above, several prior studies have used the slope of the 1/f component/aperiodic activity as an indirect measure of the E/I ratio (Favaro et al., 2023; Hill et al., 2022; McSweeney et al., 2023; Molina et al., 2020; Ostlund et al., 2021; Schaworonkow & Voytek, 2021), including studies using scalp-recorded EEG from humans.

In the introduction of the revised manuscript, we have made more explicit that this metric is indirect (Page 3, Line 91), (additionally see Discussion, Page 24, Lines 644-645, Page 25, Lines 650-657).

While a full understanding of aperiodic activity needs to be provided, some convergent ideas have emerged. We think that our results contribute to this enterprise, since our study is, to the best of our knowledge, the first which assessed MRS measured neurotransmitter levels and EEG aperiodic activity. “

(10) Additionally, the authors state:

"We cannot think of how any of the exploratory correlations between neurophysiological measures and MRS measures could be accounted for by a difference e.g. in skull thickness."

(11) This could be addressed directly by including skull thickness as a covariate or visualizing it in scatterplots, for instance, by representing skull thickness as the size of the dots.

We are not aware of any study that would justify such an analysis.

Our analyses were based on previous findings in the literature.

Since to the best of our knowledge, no evidence exists that congenital cataracts go together with changes in skull thickness, and that skull thickness might selectively modulate visual cortex Glx/GABA+ but not NAA measures, we decided against following this suggestion.

Notably, the neurotransmitter concentration reported here is after tissue segmentation of the voxel region. The tissue fraction was shown to not differ between groups in the MRS voxels (Supplementary Material S4). The EEG electrode impedance was lowered to <10 kOhm in every participant (Methods, Page 13, Line 344), and preparation was identical across groups.

(12 3.5) Problems with EEG Preprocessing and Analysis

Downsampling: The decision to downsample the data to 60 Hz "to match the stimulation rate" is problematic. This choice conflates subsequent spectral analyses due to aliasing issues, as explained by the Nyquist theorem. While the authors cite prior studies (Schwenk et al., 2020; VanRullen & MacDonald, 2012) to justify this decision, these studies focused on alpha (8-12 Hz), where aliasing is less of a concern compared of analyzing aperiodic signal. Furthermore, in contrast, the current study analyzes the frequency range from 1-20 Hz, which is too narrow for interpreting the aperiodic signal as E/I. Typically, this analysis should include higher frequencies, spanning at least 1-30 Hz or even 1-45 Hz (not 20-40 Hz).

As previously mentied in the Methods (Page 15 Line 376) and the previous response, the pop_resample function used by EEGLAB applies an anti-aliasing filter, at half the resampling frequency (as per the Nyquist theorem

https://eeglab.org/tutorials/05_Preprocess/resampling.html). The upper cut off of the low pass filter set by EEGlab prior to down sampling (30 Hz) is still far above the frequency of interest in the current study  (1-20 Hz), thus allowing us to derive valid results.

Quote:

“- The authors downsampled the data to 60Hz to "to match the stimulation rate". What is the intention of this? Because the subsequent spectral analyses are conflated by this choice (see Nyquist theorem).

This data were collected as part of a study designed to evoke alpha activity with visual white-noise, which ranged in luminance with equal power at all frequencies from 1-60 Hz, restricted by the refresh rate of the monitor on which stimuli were presented (Pant et al., 2023). This paradigm and method was developed by VanRullen and colleagues (Schwenk et al., 2020; Vanrullen & MacDonald, 2012), wherein the analysis requires the same sampling rate between the presented frequencies and the EEG data. The downsampling function used here automatically applies an anti-aliasing filter (EEGLAB 2019) .”

Moreover, the resting-state data were not resampled to 60 Hz. We have made this clearer in the Methods of the second revision (Page 15, Line 367).

Our consistent results of group differences across all three EEG conditions, thus, exclude any possibility that they were driven by aliasing artifacts.

The expected effects of this anti-aliasing filter can be seen in the attached Author response image 1, showing an example participant’s spectrum in the 1-30 Hz range (as opposed to the 1-20 Hz plotted in the manuscript), clearly showing a 30-40 dB drop at 30 Hz. Any aliasing due to, for example, remaining line noise, would additionally be visible in this figure (as well as Figure 3) as a peak.

Author response image 1.

Power spectral density of one congenital cataract-reversal (CC) participant in the visual stimulation condition across all channels. The reduced power at 30 Hz shows the effects of the anti-aliasing filter applied by EEGLAB’s pop_resample function.

As we stated in the manuscript, and in previous reviews, so far there has been no consensus on the exact range of measuring aperiodic activity. We made a principled decision based on the literature (showing a knee in aperiodic fits of this dataset at 20 Hz) (Medel et al., 2023; Ossandón et al., 2023), data quality (possible contamination by line noise at higher frequencies) and the purpose of the visual stimulation experiment (to look at the lower frequency range by stimulating up to 60 Hz, thereby limiting us to quantifying below 30 Hz), that 1-20 Hz would be the fit range in this dataset.

Quote:

“(3) What's the underlying idea of analyzing two separate aperiodic slopes (20-40Hz and 1-19Hz). This is very unusual to compute the slope between 20-40 Hz, where the SNR is rather low.

"Ossandón et al. (2023), however, observed that in addition to the flatter slope of the aperiodic power spectrum in the high frequency range (20-40 Hz), the slope of the low frequency range (1-19 Hz) was steeper in both, congenital cataract-reversal individuals, as well as in permanently congenitally blind humans."

The present manuscript computed the slope between 1-20 Hz. Ossandón et al. as well as Medel et al. (2023) found a “knee” of the 1/f distribution at 20 Hz and describe further the motivations for computing both slope ranges. For example, Ossandón et al. used a data driven approach and compared single vs. dual fits and found that the latter fitted the data better. Additionally, they found the best fit if a knee at 20 Hz was used. We would like to point out that no standard range exists for the fitting of the 1/f component across the literature and, in fact, very different ranges have been used (Gao et al., 2017; Medel et al., 2023; Muthukumaraswamy & Liley, 2018). “

(13) Baseline Removal: Subtracting the mean activity across an epoch as a baseline removal step is inappropriate for resting-state EEG data. This preprocessing step undermines the validity of the analysis. The EEG dataset has fundamental flaws, many of which were pointed out in the previous review round but remain unaddressed. In its current form, the manuscript falls short of standards for robust EEG analysis. If I were reviewing for another journal, I would recommend rejection based on these flaws.

The baseline removal step from each epoch serves to remove the DC component of the recording and detrend the data. This is a standard preprocessing step (included as an option in preprocessing pipelines recommended by the EEGLAB toolbox, FieldTrip toolbox and MNE toolbox), additionally necessary to improve the efficacy of ICA decomposition (Groppe et al., 2009).

In the previous review round, a clarification of the baseline timing was requested, which we added. Beyond this request, there was no mention of the appropriateness of the baseline removal and/or a request to provide reasons for why it might not undermine the validity of the analysis.

Quote:

“- "Subsequently, baseline removal was conducted by subtracting the mean activity across the length of an epoch from every data point." The actual baseline time segment should be specified.

The time segment was the length of the epoch, that is, 1 second for the resting state conditions and 6.25 seconds for the visual stimulation conditions. This has been explicitly stated in the revised manuscript (Page 13, Line 354).”

Prior work in the time (not frequency) domain on event-related potential (ERP) analysis has suggested that the baselining step might cause spurious effects (Delorme, 2023) (although see (Tanner et al., 2016)). We did not perform ERP analysis at any stage. One recent study suggests spurious group differences in the 1/f signal might be driven by an inappropriate dB division baselining method (Gyurkovics et al., 2021), which we did not perform.

Any effect of our baselining procedure on the FFT spectrum would be below the 1 Hz range, which we did not analyze.  

Each of the preprocessing steps in the manuscript match pipelines described and published in extensive prior work. We document how multiple aspects of our EEG results replicate prior findings (Supplementary Material S15, S18, S19), reports of other experimenters, groups and locations, validating that our results are robust.

We therefore reject the claim of methodological flaws in our EEG analyses in the strongest possible terms.

Quote:

“(3.5) Problems with EEG preprocessing and analysis:

- It seems that the authors did not identify bad channels nor address the line noise issue (even a problem if a low pass filter of below-the-line noise was applied).

As pointed out in the methods and Figure 1, we only analyzed data from two occipital channels, O1 and O2 neither of which were rejected for any participant. Channel rejection was performed for the larger dataset, published elsewhere (Ossandón et al., 2023; Pant et al., 2023). As control sites we added the frontal channels FP1 and Fp2 (see Supplementary Material S14)

Neither Ossandón et al. (2023) nor Pant et al. (2023) considered frequency ranges above 40 Hz to avoid any possible contamination with line noise. Here, we focused on activity between 0 and 20 Hz, definitely excluding line noise contaminations (Methods, Page 14, Lines 365-367). The low pass filter (FIR, 1-45 Hz) guaranteed that any spill-over effects of line noise would be restricted to frequencies just below the upper cutoff frequency.

Additionally, a prior version of the analysis used spectrum interpolation to remove line noise; the group differences remained stable (Ossandón et al., 2023). We have reported this analysis in the revised manuscript (Page 14, Lines 364-357).

Further, both groups were measured in the same lab, making line noise (~ 50 Hz) as an account for the observed group effects in the 1-20 Hz frequency range highly unlikely. Finally, any of the exploratory MRS-EEG correlations would be hard to explain if the EEG parameters would be contaminated with line noise.

- What was the percentage of segments that needed to be rejected due to the 120μV criteria? This should be reported specifically for EO & EC and controls and patients.

The mean percentage of 1 second segments rejected for each resting state condition and the percentage of 6.25 long segments rejected in each group for the visual stimulation condition have been added to the revised manuscript (Supplementary Material S10), and referred to in the Methods on Page 14, Lines 372-373).

- The authors downsampled the data to 60Hz to "to match the stimulation rate". What is the intention of this? Because the subsequent spectral analyses are conflated by this choice (see Nyquist theorem).

This data were collected as part of a study designed to evoke alpha activity with visual white-noise, which changed in luminance with equal power at all frequencies from 1-60 Hz, restricted by the refresh rate of the monitor on which stimuli were presented (Pant et al., 2023). This paradigm and method was developed by VanRullen and colleagues (Schwenk et al., 2020; VanRullen & MacDonald, 2012), wherein the analysis requires the same sampling rate between the presented frequencies and the EEG data. The downsampling function used here automatically applies an anti-aliasing filter (EEGLAB 2019) .

- "Subsequently, baseline removal was conducted by subtracting the mean activity across the length of an epoch from every data point." The actual baseline time segment should be specified.

The time segment was the length of the epoch, that is, 1 second for the resting state conditions and 6.25 seconds for the visual stimulation conditions. This has now been explicitly stated in the revised manuscript (Page 14, Lines 379-380).

- "We excluded the alpha range (8-14 Hz) for this fit to avoid biasing the results due to documented differences in alpha activity between CC and SC individuals (Bottari et al., 2016; Ossandón et al., 2023; Pant et al., 2023)." This does not really make sense, as the FOOOF algorithm first fits the 1/f slope, for which the alpha activity is not relevant.

We did not use the FOOOF algorithm/toolbox in this manuscript. As stated in the Methods, we used a 1/f fit to the 1-20 Hz spectrum in the log-log space, and subtracted this fit from the original spectrum to obtain the corrected spectrum. Given the pronounced difference in alpha power between groups (Bottari et al., 2016; Ossandón et al., 2023; Pant et al., 2023), we were concerned it might drive differences in the exponent values. Our analysis pipeline had been adapted from previous publications of our group and other labs (Ossandón et al., 2023; Voytek et al., 2015; Waschke et al., 2017).

We have conducted the analysis with and without the exclusion of the alpha range, as well as using the FOOOF toolbox both in the 1-20 Hz and 20-40 Hz ranges (Ossandón et al., 2023). The findings of a steeper slope in the 1-20 Hz range as well as lower alpha power in CC vs SC individuals remained stable. In Ossandón et al., the comparison between the piecewise fits and FOOOF fits led the authors to use the former, as it outperformed the FOOOF algorithm for their data.

- The model fits of the 1/f fitting for EO, EC, and both participant groups should be reported.

In Figure 3 of the manuscript, we depicted the mean spectra and 1/f fits for each group.

In the revised manuscript, we added the fit quality metrics (average R² values > 0.91 for each group and condition) (Methods Page 15, Lines 395-396; Supplementary Material S11) and additionally show individual subjects’ fits (Supplementary Material S11). “

(14) The authors mention:

"The EEG data sets reported here were part of data published earlier (Ossandón et al., 2023; Pant et al., 2023)." Thus, the statement "The group differences for the EEG assessments corresponded to those of a larger sample of CC individuals (n=38) " is a circular argument and should be avoided."

The authors addressed this comment and adjusted the statement. However, I do not understand, why not the full sample published earlier (Ossandón et al., 2023) was used in the current study?

The recording of EEG resting state data stated in 2013, while MRS testing could only be set up by the second half of 2019. Moreover, not all subjects who qualify for EEG recording qualify for being scanned (e.g. due to MRI safety, claustrophobia)

References

Bottari, D., Troje, N. F., Ley, P., Hense, M., Kekunnaya, R., & Röder, B. (2016). Sight restoration after congenital blindness does not reinstate alpha oscillatory activity in humans. Scientific Reports. https://doi.org/10.1038/srep24683

Colombo, M. A., Napolitani, M., Boly, M., Gosseries, O., Casarotto, S., Rosanova, M., Brichant, J. F., Boveroux, P., Rex, S., Laureys, S., Massimini, M., Chieregato, A., & Sarasso, S. (2019). The spectral exponent of the resting EEG indexes the presence of consciousness during unresponsiveness induced by propofol, xenon, and ketamine. NeuroImage, 189(September 2018), 631–644. https://doi.org/10.1016/j.neuroimage.2019.01.024

Delorme, A. (2023). EEG is better left alone. Scientific Reports, 13(1), 2372. https://doi.org/10.1038/s41598-023-27528-0

Favaro, J., Colombo, M. A., Mikulan, E., Sartori, S., Nosadini, M., Pelizza, M. F., Rosanova, M., Sarasso, S., Massimini, M., & Toldo, I. (2023). The maturation of aperiodic EEG activity across development reveals a progressive differentiation of wakefulness from sleep. NeuroImage, 277. https://doi.org/10.1016/J.NEUROIMAGE.2023.120264

Gao, R., Peterson, E. J., & Voytek, B. (2017). Inferring synaptic excitation/inhibition balance from field potentials. NeuroImage, 158(March), 70–78. https://doi.org/10.1016/j.neuroimage.2017.06.078

Groppe, D. M., Makeig, S., & Kutas, M. (2009). Identifying reliable independent components via split-half comparisons. NeuroImage, 45(4), 1199–1211. https://doi.org/10.1016/j.neuroimage.2008.12.038

Gyurkovics, M., Clements, G. M., Low, K. A., Fabiani, M., & Gratton, G. (2021). The impact of 1/f activity and baseline correction on the results and interpretation of time-frequency analyses of EEG/MEG data: A cautionary tale. NeuroImage, 237. https://doi.org/10.1016/j.neuroimage.2021.118192

Hill, A. T., Clark, G. M., Bigelow, F. J., Lum, J. A. G., & Enticott, P. G. (2022). Periodic and aperiodic neural activity displays age-dependent changes across early-to-middle childhood. Developmental Cognitive Neuroscience, 54, 101076. https://doi.org/10.1016/J.DCN.2022.101076

Maurer, D., Mondloch, C. J., & Lewis, T. L. (2007). Sleeper effects. In Developmental Science. https://doi.org/10.1111/j.1467-7687.2007.00562.x

McSweeney, M., Morales, S., Valadez, E. A., Buzzell, G. A., Yoder, L., Fifer, W. P., Pini, N., Shuffrey, L. C., Elliott, A. J., Isler, J. R., & Fox, N. A. (2023). Age-related trends in aperiodic EEG activity and alpha oscillations during early- to middle-childhood. NeuroImage, 269, 119925. https://doi.org/10.1016/j.neuroimage.2023.119925

Medel, V., Irani, M., Crossley, N., Ossandón, T., & Boncompte, G. (2023). Complexity and 1/f slope jointly reflect brain states. Scientific Reports, 13(1), 21700. https://doi.org/10.1038/s41598-023-47316-0

Molina, J. L., Voytek, B., Thomas, M. L., Joshi, Y. B., Bhakta, S. G., Talledo, J. A., Swerdlow, N. R., & Light, G. A. (2020). Memantine Effects on Electroencephalographic Measures of Putative Excitatory/Inhibitory Balance in Schizophrenia. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 5(6), 562–568. https://doi.org/10.1016/j.bpsc.2020.02.004

Muthukumaraswamy, S. D., & Liley, D. T. (2018). 1/F electrophysiological spectra in resting and drug-induced states can be explained by the dynamics of multiple oscillatory relaxation processes. NeuroImage, 179(November 2017), 582–595. https://doi.org/10.1016/j.neuroimage.2018.06.068

Ossandón, J. P., Stange, L., Gudi-Mindermann, H., Rimmele, J. M., Sourav, S., Bottari, D., Kekunnaya, R., & Röder, B. (2023). The development of oscillatory and aperiodic resting state activity is linked to a sensitive period in humans. NeuroImage, 275, 120171. https://doi.org/10.1016/J.NEUROIMAGE.2023.120171

Ostlund, B. D., Alperin, B. R., Drew, T., & Karalunas, S. L. (2021). Behavioral and cognitive correlates of the aperiodic (1/f-like) exponent of the EEG power spectrum in adolescents with and without ADHD. Developmental Cognitive Neuroscience, 48, 100931. https://doi.org/10.1016/j.dcn.2021.100931

Pant, R., Ossandón, J., Stange, L., Shareef, I., Kekunnaya, R., & Röder, B. (2023). Stimulus-evoked and resting-state alpha oscillations show a linked dependence on patterned visual experience for development. NeuroImage: Clinical, 103375. https://doi.org/10.1016/J.NICL.2023.103375

Schaworonkow, N., & Voytek, B. (2021). Longitudinal changes in aperiodic and periodic activity in electrophysiological recordings in the first seven months of life. Developmental Cognitive Neuroscience, 47. https://doi.org/10.1016/j.dcn.2020.100895

Schwenk, J. C. B., VanRullen, R., & Bremmer, F. (2020). Dynamics of Visual Perceptual Echoes Following Short-Term Visual Deprivation. Cerebral Cortex Communications, 1(1). https://doi.org/10.1093/TEXCOM/TGAA012

Tanner, D., Norton, J. J. S., Morgan-Short, K., & Luck, S. J. (2016). On high-pass filter artifacts (they’re real) and baseline correction (it’s a good idea) in ERP/ERMF analysis. Journal of Neuroscience Methods, 266, 166–170. https://doi.org/10.1016/j.jneumeth.2016.01.002

Vanrullen, R., & MacDonald, J. S. P. (2012). Perceptual echoes at 10 Hz in the human brain. Current Biology. https://doi.org/10.1016/j.cub.2012.03.050

Voytek, B., Kramer, M. A., Case, J., Lepage, K. Q., Tempesta, Z. R., Knight, R. T., & Gazzaley, A. (2015). Age-related changes in 1/f neural electrophysiological noise. Journal of Neuroscience, 35(38). https://doi.org/10.1523/JNEUROSCI.2332-14.2015

Waschke, L., Wöstmann, M., & Obleser, J. (2017). States and traits of neural irregularity in the age-varying human brain. Scientific Reports 2017 7:1, 7(1), 1–12. https://doi.org/10.1038/s41598-017-17766-4

Author response:

The following is the authors’ response to the previous reviews

Public Reviews:

Reviewer #1 (Public review):

Summary:

For many years, there has been extensive electrophysiological research investigating the relationship between local field potential patterns and individual cell spike patterns in the hippocampus. In this study, using state-ofthe-art imaging techniques, they examined spike synchrony of hippocampal cells during locomotion and immobility states. In contrast to conventional understanding of the hippocampus, the authors demonstrated that hippocampal place cells exhibit prominent synchronous spikes locked to theta oscillations.

Strengths:

The voltage imaging used in this study is a highly novel method that allows recording not only suprathreshold-level spikes but also subthreshold-level activity. With its high frame rate, it offers time resolution comparable to electrophysiological recordings.

We thank the reviewer for a thorough review of our manuscript and for recognizing the strength of our study.

Reviewer #2 (Public review):

Summary:

This study employed voltage imaging in the CA1 region of the mouse hippocampus during the exploration of a novel environment. The authors report synchronous activity, involving almost half of the imaged neurons, occurred during periods of immobility. These events did not correlate with SWRs, but instead, occurred during theta oscillations and were phased locked to the trough of theta. Moreover, pairs of neurons with high synchronization tended to display non-overlapping place fields, leading the authors to suggest these events may play a role in binding a distributed representation of the context.

Strengths:

Technically this is an impressive study, using an emerging approach that allow single-cell resolution voltage imaging in animals, that while head-fixed, can move through a real environment. The paper is written clearly and suggests novel observations about population-level activity in CA1.

We thank the reviewer for a thorough review of our manuscript and for recognizing the strength of our study.

Weaknesses:

The evidence provided is weak, with the authors making surprising population-level claims based on a very sparse data set (5 data sets, each with less than 20 neurons simultaneously recorded) acquired with exciting, but less tested technology. Further, while the authors link these observations to the novelty of the context, both in the title and text, they do not include data from subsequent visits to support this. Detailed comments are below:

(1) My first question for the authors, which is not addressed in the discussion, is why these events have not been observed in the countless extracellular recording experiments conducted in rodent CA1 during exploration of novel environments. Those data sets often have 10x the neurons simultaneously recording compared to these present data, thus the highly synchronous firing should be very hard to miss. Ideally, the authors could confirm their claims via the analysis of publicly available electrophysiology data sets. Further, the claim of high extra-SWR synchrony is complicated by the observation that their recorded neurons fail to spike during the limited number of SWRs recorded during behavior- again, not agreeing with much of the previous electrophysiological recordings.

(2) The authors posit that these events are linked to the novelty of the context, both in the text, as well as in the title and abstract. However they do not include any imaging data from subsequent days to demonstrate the failure to see this synchrony in a familiar environment. If these data are available it would strengthen the proposed link to novelty is they were included.

(3) In the discussion the authors begin by speculating the theta present during these synchronous events may be slower type II or attentional theta. This can be supported by demonstrating a frequency shift in the theta recording during these events/immobility versus the theta recording during movement. (4) The authors mention in the discussion that they image deep layer PCs in CA1, however this is not mentioned in the text or methods. They should include data, such as imaging of a slice of a brain post-recording with immunohistochemistry for a layer specific gene to support this.

Comments on revisions:

I have no further major requests and thank the authors for the additional data and analyses.

We thank the reviewer for recognizing our efforts in revising the manuscript.

Reviewer #3 (Public review):

Summary:

In the present manuscript, the authors use a few minutes of voltage imaging of CA1 pyramidal cells in head-fixed mice running on a track while local field potentials (LFPs) are recorded. The authors suggest that synchronous ensembles of neurons are differentially associated with different types of LFP patterns, theta and ripples. The experiments are flawed in that the LFP is not "local" but rather collected the other side of the brain.

Strengths:

The authors use a cutting-edge technique.

We thank the reviewer for a thoughtful review of our manuscript and for pointing out the technical strength of our study.

Weaknesses:

The two main messages of the manuscript indicated in the title are not supported by the data. The title gives two messages that relate to CA1 pyramidal neurons in behaving head-fixed mice: (1) synchronous ensembles are associated with theta (2) synchronous ensembles are not associated with ripples. The main problem with the work is that the theta and ripple signals were recorded using electrophysiology from the opposite hemisphere to the one in which the spiking was monitored. However, both rhythms exhibit profound differences as a function of location.

Theta phase changes with the precise location along the proximo-distal and dorso-ventral axes, and importantly, even reverses with depth. Because the LFP was recorded using a single-contact tungsten electrode, there is no way to know whether the electrode was exactly in the CA1 pyramidal cell layer, or in the CA1 oriens, CA1 radiatum, or perhaps even CA3 - which exhibits ripples and theta which are weakly correlated and in anti-phase with the CA1 rhythms, respectively. Thus, there is no way to know whether the theta phase used in the analysis is the phase of the local CA1 theta.

Although the occurrence of CA1 ripples is often correlated across parts of the hippocampus, ripples are inherently a locally-generated rhythm. Independent ripples occur within a fraction of a millimeter within the same hemisphere. Ripples are also very sensitive to the precise depth - 100 micrometers up or down, and only a positive deflection/sharp wave is evident. Thus, even if the LFP was recorded from the center of the CA1 pyramidal layer in the contralateral hemisphere, it would not suffice for the claim made in the title.

We thank the reviewer for pointing out the issue regarding the claim made in the title. We have revised the manuscript to clarify that the theta and ripple oscillations referenced in the title refer to specific frequency bands of intracellular and contralaterally recorded field potentials rather than field potentials recorded at the same site as the neuronal activity.

Abstract (line19):

“… Notably, these synchronous ensembles were not associated with contralateral ripple oscillations but were instead phase-locked to theta waves recorded in the contralateral CA1 region. Moreover, the subthreshold membrane potentials of neurons exhibited coherent intracellular theta oscillations with a depolarizing peak at the moment of synchrony.”

Introduction (line68):

“… Surprisingly, these synchronous ensembles occurred outside of contralateral ripples and were phase-locked to intracellular theta oscillations as well as extracellular theta oscillations recorded from the contralateral CA1 region.”

To address concerns about electrode placement, we have now included posthoc histological verification of electrode locations, confirming that they were positioned in the contralateral CA1 pyramidal layer (Author response image 1). 

Author response image 1.

Post-hoc histological section showing the location of a DiI-coated electrode in the contralateral CA1 pyramidal layer. Scale bar: 300 μm.

While we appreciate that theta and ripple oscillations exhibit regional variations in phase and amplitude, previous studies have demonstrated a strong co-occurrence and synchrony of these oscillations between both hippocampi1-3. Given that our primary objective was to examine how neuronal ensembles relate to large-scale hippocampal oscillation states rather than local microcircuit-level fluctuations, we recorded theta and ripple oscillations from the contralateral CA1 region.

However, we acknowledge that contralateral recordings do not capture all ipsilateral-specific dynamics. Theta phases vary with depth and precise location, and local ripple events may be independently generated across small spatial scales. To reflect this, we have now explicitly acknowledged these considerations in the discussion. 

Discussion (line527):

While contralateral LFP recordings reliably capture large-scale hippocampal theta and ripple oscillations, they may not fully account for ipsilateral-specific dynamics, such as variations in theta phase alignment or locally generated ripple events. Although contralateral recordings serve as a well-established proxy for large-scale hippocampal oscillatory states, incorporating simultaneous ipsilateral field potential recordings in future studies could refine our understanding of local-global network interactions. Despite these considerations, our findings provide robust evidence for the existence of synchronous neuronal ensembles and their role in coordinating newly formed place cells. These results advance our understanding of how synchronous neuronal ensembles contribute to spatial memory acquisition and hippocampal network coordination.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The authors have provided sufficient experimental and analytical data addressing my comments, particularly regarding consistency with past electrophysiological data and the exclusion of potential imaging artifacts.

We thank the reviewer for recognizing our efforts in revising the manuscript.

Minor comment: In Figure 2C and Figure 5-figure supplement 1, 'paired Student's t-test' is not entirely appropriate. More precisely, either 'paired t-test' or 'Student's t-test' would better indicate the correct statistical method. Please verify whether these data comparisons are within-group or between-group.

Thank you for the comment. We have revised the manuscript as suggested.

Reviewer #2 (Recommendations for the authors):

I have no further major requests and thank the authors for the additional data and analyses.

We thank the reviewer for recognizing our efforts in revising the manuscript.

Minor points- line 169- typo, correct grant to grand

Thank you for pointing it out. The typo has been corrected.

(1) Buzsaki, G. et al. Hippocampal network patterns of activity in the mouse. Neuroscience 116, 201-211 (2003). https://doi.org:10.1016/s03064522(02)00669-3

(2) Szabo, G. G. et al. Ripple-selective GABAergic projection cells in the hippocampus. Neuron 110, 1959-1977 e1959 (2022). https://doi.org:10.1016/j.neuron.2022.04.002

(3) Huang, Y. C. et al. Dynamic assemblies of parvalbumin interneurons in brain oscillations. Neuron 112, 2600-2613 e2605 (2024). https://doi.org:10.1016/j.neuron.2024.05.015

Author response:

The following is the authors’ response to the previous reviews.

Reviewer #1 (Recommendations For The Authors):

The additional data included in this revision nicely strengthens the major claim.

I apologize that my comment about K+ concentration in the prior review was unclear. The cryoEM structure of KCNQ1 with S4 in the resting state was obtained with lowered K+ relative to the active state. Throughout the results and discussion it seems implied that the change in voltage sensor state is somehow causative of the change in selectivity filter state while the paper that identified the structures attributes the change in selectivity filter state not to voltage sensors, but to the change in [K+] between the 2 structures. Unless there is a flaw in my understanding of the conditions in which the selectivity filter structures used in modeling were generated, it seems misleading to ignore the change in [K+] when referring to the activated vs resting or up vs down structures. My understanding is that the closed conformation adopted in the resting/low [K+] is similar to that observed in low [K+] previously and is more commonly associated with [K+]-dependent inactivation, not resulting from voltage sensor deactivation as implied here. The original article presenting the low [K+] structure also suggests this. When discussing conformational changes in the selectivity filter, I strongly suggest referring to these structures as activated/high [K+] vs resting/low [K+] or something similar, as the [K+] concentration is a salient variable.

There seems to be some major confusion here and we will try to explain how we think. Note that in the Mandela and MacKinnon paper, there is no significant difference in the amino acid positions in the selectivity filter between low and high K+ when S4 is in the activated position (See Mandala and Mackinnon, PNAS Suppl. Fig S5 C and D). There are only fewer K+ in the selectivity filter in low K+. So, the structure with the distorted selectivity filter is not due to low K+ by itself. Note that there is no real difference between macroscopic currents recorded in low and high K+ solutions (except what is expected from changes in driving force) for KCNQ1/KCNE1 channels (Larsen et al., Bioph J 2011), suggesting that low K+ do not promote the non-conductive state (Figure 1). We now include a section in the Discussion about high/low K+ in the structures and the absence of effects of K+ on the function of KCNQ1/KCNE1 channels.

Author response image 1.

Macroscopic KCNQ1/KCNE1 currents recorded in different K+ conditions.  Note that there is no difference between current recorded in low K+ (2 mM) conditions and high (96 mM) K+ conditions (n=3 oocytes). Currents were normalized in respect to high K+.

Note also that, in the previous version of the manuscript, we did not propose that the position of S4 is what determines the state of the selectivity filter. We only reported that the CryoEM structure with S4 resting shows a distorted selectivity filter. It seems like our text confused the reviewer to think that we proposed that S4 determines the state of the selectivity filter, when we did not propose this earlier. We previously did not want to speculate too much about this, but we have now included a section in the Discussion to make our view clear in light of the confusion of the reviewers.

It is clear from our data that the majority of sweeps are empty (which we assume is with S4 up), suggesting that the selectivity filter can be (and is in the majority of sweeps) in the non-conducting state even with S4 up.  We think that the selectivity filter switches between a non-conductive and a conductive conformation both with S4 down and with S4 up. The cryoEM structure in low K+ and S4 down just happened to catch the non-conductive state of the selectivity filter.  We have now added a section in the Discussion to clarify all this and explain how we think it works.

However, S4 in the active conformation seems to stabilize the conductive conformation of the selectivity filter, because during long pulses the channel seems to stay open once opened (See Suppl Fig S2). So, one possibility is that the selectivity filter goes more readily into the non-conductive state when S4 is down (and maybe, or not, low K+ plays a role) and then when S4 moves up the selectivity filter sometimes recovers into the conductive state and stays there. We now have included a section in the Discussion to present our view. Since this whole discussion was initiated and pushed by the reviewer, we hope that the reviewers will not demand more data to support these ideas. We think that this addition makes sense since other readers might have the same questions and ideas as the reviewer, and we would like to prevent any confusion about this topic.

Figure 1

It remains unclear in the manuscript itself what "control" refers to. Are control patched the same patches that later receive LG?

Yes, the control means the same patch before LG. We now indicate that in legends and text throughout.

Supplementary Figure S1

Unclear if any changes occur after addition of LG in left panel and if the LG data on right is paired in any way to data on left.

Yes, in all cases the left and right panel in all figures are from the same patch. We now indicate that in legends and text throughout.

The letter p is used both to represent open probability open probability from the all-point amplitude histogram and as a p-value statistical probability indicator sometime lower case, sometimes upper case. This was confusing.

We have now exclusively use lower case p for statistical probability and Po for open probability.

"This indicates that mutations of residues in the more intracellular region of the selectivity filter do not affect the Gmax increases and that the interactions that stabilize the channel involve only residues located near the external region part of the selectivity filter. "

Seems too strongly worded, it remains possible that mutations of other residues in the more intracellular region of the selectivity filter could affect the Gmax increases.

We have changed the text to: "Mutations of residues in the more intracellular region of the selectivity filter do not affect the Gmax increases, as if the interactions that stabilize the channel involve residues located near the external region part of the selectivity filter. "

Supplementary Figure S7

Please report Boltzmann fit parameters. What are "normalized" uA?

We removed the uA, which was mistakenly inserted. The lines in the graphs are just lines connecting the dots and not Boltzmann fits, since we don’t have saturating curves in all panels to make unique fits.

"We have previously shown that the effects of PUFAs on IKs channels involve the binding of PUFAs to two independent sites." Was binding to the sites actually shown? Suggest changing to: "We have previously proposed models in which the effects of PUFAs..."

We have now changed this as the Reviewer suggested: " We have previously proposed models in which the effects of PUFAs on IKs channels involve the binding of PUFAs to two independent sites."

Statistics used not always clear. Methods refer to multiple statistical tests but it is not clear which is used when.

We use two different tests and it is now explained in figure legends when either was used.

n values confusing. Sometimes # of sweeps used as n. Sometimes # patches used as n. In one instance "The average current during the single channel sweeps was increased by 2.3 {plus minus} 0.33 times (n = 4 patches, p =0.0006)" ...this sems a low p value for this n=4 sample?

We have now more clearly indicated what n stands for in each case. There was an extra 0 in the p value, so now it is p = 0.006. Thanks for catching that error.

Reviewer #2 (Recommendations For The Authors):

I still have some comments for the revised manuscript.

(1) (From the previous minor point #6) Since D317E and T309S did not show statistical significance in Figure 5A, the sentences such as "This data shows that Y315 and D317 are necessary for the ability of Lin-Glycine to increase Gmax" or "the effect of Lin-Glycine on Gmax of the KCNQ1/KCNE1 mutant was noticeably reduced compared to the WT channel showing the this residue contributes to the Gmax effect (Figure 5A)." may need to be toned down. Alternatively, I suggest the authors refer to Supplementary Figure S7 to confirm that Y315 and D317 are critical for increasing Gmax.

We have redone the analysis and statistical evaluation in Fig 5. We no use the more appropriate value of the fitted Gmax (which use the whole dose response curve instead of only the 20 mM value) in the statistical evaluation and now Y315F and D317E are statistically different from wt.

(2) Supplementary Fig. S1. All control diary plots include the green arrows to indicate the timing of lin-glycine (LG) application. It is a bit confusing why they are included. Is it to show that LG application did not have an immediate effect? Are the LG-free plots not available?

Not sure what the Reviewer is asking about? In the previous review round the Reviewers asked specifically for this. The arrow shows when LG was applied and the plot on the right shows the effect of LG from the same patch.

(3) The legend to Supplementary Figure S4, "The side chain of residues ... are highlighted as sticks and colored based on the atomic displacement values, from white to blue to red on a scale of 0 to 9 Å." They look mostly blue (or light blue). Which one is colored white? It might be better to use a different color code. It would also be nice to link the color code to the colors of Supplementary Figure S5, which currently uses a single color.

We have removed “from white to blue to red on a scale of 0 to 9 Å” and instead now include a color scale directly in Fig S4 to show how much each atom moved based on the color.

We feel it is not necessary to include color in Fig S5 since the scale of how much each atom moves is shown on the y axis.

(4) Add unit (pA) to the y-axis of Supplementary Figure S2.

pA has been added.

Reviewer #3 (Recommendations For The Authors):

Some issues on how data support conclusions are identified. Further justifications are suggested.

186: “The decrease in first latency is most likely due to an effect of Lin-Glycine on Site I in the VSD and related to the shift in voltage dependence caused by Lin-Glycine." The results in Fig S1B do not seem to support this statement since the mutation Y315F in the pore helix seemed to have eliminated the effect of Lin-Glycine in reducing first latency. The authors may want to show that a mutation that eliminating Site I would eliminate the effect of Lin-Glycine on first latency. On the other hand, it will be also interesting to examine if another pore mutation, such as P320L (Fig 5) also reduce the effect of Lin-Glycine on first latency.

These experiments are very hard and laborious, and we feel these are outside the scope of this paper which focuses on Site II and the mechanism of increasing Gmax. Further studies of the voltage shift and latency will have to be for a future study.

The mutation D317E did not affect the effect of Lin-Glycine on Gmax significantly (Fig 5A, and Fig S7F comparing with Fig S7A), but the authors conclude that D317 is important for Lin-Glycine association. This conclusion needs a better justification.

We have redone the analysis and statistical evaluation in Fig 5. We no use the more appropriate value of the fitted Gmax (which use the whole dose response curve instead of only the 20 mM value) in the statistical evaluation and now D317E is statistically different from wt

Author response:

The following is the authors’ response to the current reviews.

eLife assessment

This useful manuscript challenges the utility of current paradigms for estimating brain-age with magnetic resonance imaging measures, but presents inadequate evidence to support the suggestion that an alternative approach focused on predicting cognition is more useful. The paper would benefit from a clearer explication of the methods and a more critical evaluation of the conceptual basis of the different models. This work will be of interest to researchers working on brain-age and related models.

Thank you so much for providing high-quality reviews on our manuscript. We revised the manuscript to address all of the reviewers’ comments and provided full responses to each of the comments below. Importantly, in this revision, we clarified that we did not intend to use Brain Cognition as an alternative approach as mentioned by the editor. This is because, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Here we made this point more explicit and further stated that the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. By examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And such quantification is the third aim of this study.

Reviewer #1 (Public Review):

In this paper, the authors evaluate the utility of brain age derived metrics for predicting cognitive decline by performing a 'commonality' analysis in a downstream regression that enables the different contribution of different predictors to be assessed. The main conclusion is that brain age derived metrics do not explain much additional variation in cognition over and above what is already explained by age. The authors propose to use a regression model trained to predict cognition ('brain cognition') as an alternative suited to applications of cognitive decline. While this is less accurate overall than brain age, it explains more unique variance in the downstream regression.

Importantly, in this revision, we clarified that we did not intend to use Brain Cognition as an alternative approach. This is because, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Here we made this point more explicit and further stated that the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. By examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age.

REVISED VERSION: while the authors have partially addressed my concerns, I do not feel they have addressed them all. I do not feel they have addressed the weight instability and concerns about the stacked regression models satisfactorily.

Please see our responses to #3 below

I also must say that I agree with Reviewer 3 about the limitations of the brain age and brain cognition methods conceptually. In particular that the regression model used to predict fluid cognition will by construction explain more variance in cognition than a brain age model that is trained to predict age. This suffers from the same problem the authors raise with brain age and would indeed disappear if the authors had a separate measure of cognition against which to validate and were then to regress this out as they do for age correction. I am aware that these conceptual problems are more widespread than this paper alone (in fact throughout the brain age literature), so I do not believe the authors should be penalised for that. However, I do think they can make these concerns more explicit and further tone down the comments they make about the utility of brain cognition. I have indicated the main considerations about these points in the recommendations section below.

Thank you so much for raising this point. We now have the following statement in the introduction and discussion to address this concern (see below).

Briefly, we made it explicit that, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. That is, the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. More importantly, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And this is the third goal of this present study.

From Introduction:

“Third and finally, certain variation in fluid cognition is related to brain MRI, but to what extent does Brain Age not capture this variation? To estimate the variation in fluid cognition that is related to the brain MRI, we could build prediction models that directly predict fluid cognition (i.e., as opposed to chronological age) from brain MRI data. Previous studies found reasonable predictive performances of these cognition-prediction models, built from certain MRI modalities (Dubois et al., 2018; Pat, Wang, Anney, et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). Analogous to Brain Age, we called the predicted values from these cognition-prediction models, Brain Cognition. The strength of an out-of-sample relationship between Brain Cognition and fluid cognition reflects variation in fluid cognition that is related to the brain MRI and, therefore, indicates the upper limit of Brain Age’s capability in capturing fluid cognition. This is, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Consequently, if we included Brain Cognition, Brain Age and chronological age in the same model to explain fluid cognition, we would be able to examine the unique effects of Brain Cognition that explain fluid cognition beyond Brain Age and chronological age. These unique effects of Brain Cognition, in turn, would indicate the amount of co-variation between brain MRI and fluid cognition that is missed by Brain Age.”

From Discussion:

“Third, by introducing Brain Cognition, we showed the extent to which Brain Age indices were not able to capture the variation in fluid cognition that is related to brain MRI. More specifically, using Brain Cognition allowed us to gauge the variation in fluid cognition that is related to the brain MRI, and thereby, to estimate the upper limit of what Brain Age can do. Moreover, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age.

From our results, Brain Cognition, especially from certain cognition-prediction models such as the stacked models, has relatively good predictive performance, consistent with previous studies (Dubois et al., 2018; Pat, Wang, Anney, et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). We then examined Brain Cognition using commonality analyses (Nimon et al., 2008) in multiple regression models having a Brain Age index, chronological age and Brain Cognition as regressors to explain fluid cognition. Similar to Brain Age indices, Brain Cognition exhibited large common effects with chronological age. But more importantly, unlike Brain Age indices, Brain Cognition showed large unique effects, up to around 11%. As explained above, the unique effects of Brain Cognition indicated the amount of co-variation between brain MRI and fluid cognition that was missed by a Brain Age index and chronological age. This missing amount was relatively high, considering that Brain Age and chronological age together explained around 32% of the total variation in fluid cognition. Accordingly, if a Brain Age index was used as a biomarker along with chronological age, we would have missed an opportunity to improve the performance of the model by around one-third of the variation explained.”

This is a reasonably good paper and the use of a commonality analysis is a nice contribution to understanding variance partitioning across different covariates. I have some comments that I believe the authors ought to address, which mostly relate to clarity and interpretation

Reviewer #1 Public Review #1

First, from a conceptual point of view, the authors focus exclusively on cognition as a downstream outcome. I would suggest the authors nuance their discussion to provide broader considerations of the utility of their method and on the limits of interpretation of brain age models more generally.

Thank you for your comments on this issue.

We now discussed the broader consideration in detail:

(1) the consistency between our findings on fluid cognition and other recent works on brain disorders,

(2) the difference between studies investigating the utility of Brain Age in explaining cognitive functioning, including ours and others (e.g., Butler et al., 2021; Cole, 2020, 2020; Jirsaraie, Kaufmann, et al., 2023) and those explaining neurological/psychological disorders (e.g., Bashyam et al., 2020; Rokicki et al., 2021)

and

(3) suggested solutions we and others made to optimise the utility of Brain Age for both cognitive functioning and brain disorders.

From Discussion:

“This discrepancy between the predictive performance of age-prediction models and the utility of Brain Age indices as a biomarker is consistent with recent findings (for review, see Jirsaraie, Gorelik, et al., 2023), both in the context of cognitive functioning (Jirsaraie, Kaufmann, et al., 2023) and neurological/psychological disorders (Bashyam et al., 2020; Rokicki et al., 2021). For instance, combining different MRI modalities into the prediction models, similar to our stacked models, often leads to the highest performance of age-prediction models, but does not likely explain the highest variance across different phenotypes, including cognitive functioning and beyond (Jirsaraie, Gorelik, et al., 2023).”

“There is a notable difference between studies investigating the utility of Brain Age in explaining cognitive functioning, including ours and others (e.g., Butler et al., 2021; Cole, 2020, 2020; Jirsaraie, Kaufmann, et al., 2023) and those explaining neurological/psychological disorders (e.g., Bashyam et al., 2020; Rokicki et al., 2021). We consider the former as a normative type of study and the latter as a case-control type of study (Insel et al., 2010; Marquand et al., 2016). Those case-control Brain Age studies focusing on neurological/psychological disorders often build age-prediction models from MRI data of largely healthy participants (e.g., controls in a case-control design or large samples in a population-based design), apply the built age-prediction models to participants without vs. with neurological/psychological disorders and compare Brain Age indices between the two groups. On the one hand, this means that case-control studies treat Brain Age as a method to detect anomalies in the neurological/psychological group (Hahn et al., 2021). On the other hand, this also means that case-control studies have to ignore under-fitted models when applied prediction models built from largely healthy participants to participants with neurological/psychological disorders (i.e., Brain Age may predict chronological age well for the controls, but not for those with a disorder). On the contrary, our study and other normative studies focusing on cognitive functioning often build age-prediction models from MRI data of largely healthy participants and apply the built age-prediction models to participants who are also largely healthy. Accordingly, the age-prediction models for explaining cognitive functioning in normative studies, while not allowing us to detect group-level anomalies, do not suffer from being under-fitted. This unfortunately might limit the generalisability of our study into just the normative type of study. Future work is still needed to test the utility of brain age in the case-control case.”

“Next, researchers should not select age-prediction models based solely on age-prediction performance. Instead, researchers could select age-prediction models that explained phenotypes of interest the best. Here we selected age-prediction models based on a set of features (i.e., modalities) of brain MRI. This strategy was found effective not only for fluid cognition as we demonstrated here, but also for neurological and psychological disorders as shown elsewhere (Jirsaraie, Gorelik, et al., 2023; Rokicki et al., 2021). Rokicki and colleagues (2021), for instance, found that, while integrating across MRI modalities led to age-prediction models with the highest age-prediction performance, using only T1 structural MRI gave age-prediction models that were better at classifying Alzheimer’s disease. Similarly, using only cerebral blood flow gave age-prediction models that were better at classifying mild/subjective cognitive impairment, schizophrenia and bipolar disorder.

As opposed to selecting age-prediction models based on a set of features, researchers could also select age-prediction models based on modelling methods. For instance, Jirsaraie and colleagues (2023) compared gradient tree boosting (GTB) and deep-learning brain network (DBN) algorithms in building age-prediction models. They found GTB to have higher age-prediction performance but DBN to have better utility in explaining cognitive functioning. In this case, an algorithm with better utility (e.g., DBN) should be used for explaining a phenotype of interest. Similarly, Bashyam and colleagues (2020) built different DBN-based age-prediction models, varying in age-prediction performance. The DBN models with a higher number of epochs corresponded to higher age-prediction performance. However, DBN-based age-prediction models with a moderate (as opposed to higher or lower) number of epochs were better at classifying Alzheimer’s disease, mild cognitive impairment and schizophrenia. In this case, a model from the same algorithm with better utility (e.g., those DBN with a moderate epoch number) should be used for explaining a phenotype of interest. Accordingly, this calls for a change in research practice, as recently pointed out by Jirasarie and colleagues (2023, p7), “Despite mounting evidence, there is a persisting assumption across several studies that the most accurate brain age models will have the most potential for detecting differences in a given phenotype of interest”. Future neuroimaging research should aim to build age-prediction models that are not necessarily good at predicting age, but at capturing phenotypes of interest.”

Reviewer #1 Public Review #2

Second, from a methods perspective, there is not a sufficient explanation of the methodological procedures in the current manuscript to fully understand how the stacked regression models were constructed. I would request that the authors provide more information to enable the reader to better understand the stacked regression models used to ensure that these models are not overfit.

Thank you for allowing us an opportunity to clarify our stacked model. We made additional clarification to make this clearer (see below). We wanted to confirm that we did not use test sets to build a stacked model in both lower and higher levels of the Elastic Net models. Test sets were there just for testing the performance of the models.

From Methods: “We used nested cross-validation (CV) to build these prediction models (see Figure 7). We first split the data into five outer folds, leaving each outer fold with around 100 participants. This number of participants in each fold is to ensure the stability of the test performance across folds. In each outer-fold CV loop, one of the outer folds was treated as an outer-fold test set, and the rest was treated as an outer-fold training set. Ultimately, looping through the nested CV resulted in a) prediction models from each of the 18 sets of features as well as b) prediction models that drew information across different combinations of the 18 separate sets, known as “stacked models.” We specified eight stacked models: “All” (i.e., including all 18 sets of features), “All excluding Task FC”, “All excluding Task Contrast”, “Non-Task” (i.e., including only Rest FC and sMRI), “Resting and Task FC”, “Task Contrast and FC”, “Task Contrast” and “Task FC”. Accordingly, there were 26 prediction models in total for both Brain Age and Brain Cognition.

To create these 26 prediction models, we applied three steps for each outer-fold loop. The first step aimed at tuning prediction models for each of 18 sets of features. This step only involved the outer-fold training set and did not involve the outer-fold test set. Here, we divided the outer-fold training set into five inner folds and applied inner-fold CV to tune hyperparameters with grid search. Specifically, in each inner-fold CV, one of the inner folds was treated as an inner-fold validation set, and the rest was treated as an inner-fold training set. Within each inner-fold CV loop, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters and applied the estimated model to the inner-fold validation set. After looping through the inner-fold CV, we, then, chose the prediction models that led to the highest performance, reflected by coefficient of determination (R2), on average across the inner-fold validation sets. This led to 18 tuned models, one for each of the 18 sets of features, for each outer fold.

The second step aimed at tuning stacked models. Same as the first step, the second step only involved the outer-fold training set and did not involve the outer-fold test set. Here, using the same outer-fold training set as the first step, we applied tuned models, created from the first step, one from each of the 18 sets of features, resulting in 18 predicted values for each participant. We, then, re-divided this outer-fold training set into new five inner folds. In each inner fold, we treated different combinations of the 18 predicted values from separate sets of features as features to predict the targets in separate “stacked” models. Same as the first step, in each inner-fold CV loop, we treated one out of five inner folds as an inner-fold validation set, and the rest as an inner-fold training set. Also as in the first step, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters from our grid. We tuned the hyperparameters of stacked models using grid search by selecting the models with the highest R2 on average across the inner-fold validation sets. This led to eight tuned stacked models.

The third step aimed at testing the predictive performance of the 18 tuned prediction models from each of the set of features, built from the first step, and eight tuned stacked models, built from the second step. Unlike the first two steps, here we applied the already tuned models to the outer-fold test set. We started by applying the 18 tuned prediction models from each of the sets of features to each observation in the outer-fold test set, resulting in 18 predicted values. We then applied the tuned stacked models to these predicted values from separate sets of features, resulting in eight predicted values.

To demonstrate the predictive performance, we assessed the similarity between the observed values and the predicted values of each model across outer-fold test sets, using Pearson’s r, coefficient of determination (R2) and mean absolute error (MAE). Note that for R2, we used the sum of squares definition (i.e., R2 = 1 – (sum of squares residuals/total sum of squares)) per a previous recommendation (Poldrack et al., 2020). We considered the predicted values from the outer-fold test sets of models predicting age or fluid cognition, as Brain Age and Brain Cognition, respectively.”

Note some previous research, including ours (Tetereva et al., 2022), splits the observations in the outer-fold training set into layer 1 and layer 2 and applies the first and second steps to layers 1 and 2, respectively. Here we decided against this approach and used the same outer-fold training set for both first and second steps in order to avoid potential bias toward the stacked models. This is because, when the data are split into two layers, predictive models built for each separate set of features only use the data from layer 1, while the stacked models use the data from both layers 1 and 2. In practice with large enough data, these two approaches might not differ much, as we demonstrated previously (Tetereva et al., 2022).

Reviewer #1 Public Review #3

Please also provide an indication of the different regression strengths that were estimated across the different models and cross-validation splits. Also, how stable were the weights across splits?

The focus of this article is on the predictions. Still, it is informative for readers to understand how stable the feature importance (i.e., Elastic Net coefficients) is. To demonstrate the stability of feature importance, we now examined the rank stability of feature importance using Spearman’s ρ (see Figure 4). Specifically, we correlated the feature importance between two prediction models of the same features, used in two different outer-fold test sets. Given that there were five outer-fold test sets, we computed 10 Spearman’s ρ for each prediction model of the same features. We found Spearman’s ρ to be varied dramatically in both age-prediction (range=.31-.94) and fluid cognition-prediction (range=.16-.84) models. This means that some prediction models were much more stable in their feature importance than others. This is probably due to various factors such as a) the collinearity of features in the model, b) the number of features (e.g., 71,631 features in functional connectivity, which were further reduced to 75 PCAs, as compared to 19 features in subcortical volume based on the ASEG atlas), c) the penalisation of coefficients either with ‘Ridge’ or ‘Lasso’ methods, which resulted in reduction as a group of features or selection of a feature among correlated features, respectively, and d) the predictive performance of the models. Understanding the stability of feature importance is beyond the scope of the current article. As mentioned by Reviewer 1, “The predictions can be stable when the coefficients are not,” and we chose to focus on the prediction in the current article.

Reviewer #1 Public Review #4

Please provide more details about the task designs, MRI processing procedures that were employed on this sample in addition to the regression methods and bias correction methods used. For example, there are several different parameterisations of the elastic net, please provide equations to describe the method used here so that readers can easily determine how the regularisation parameters should be interpreted.

Thank you for the opportunity for us to provide more methodical details.

First, for the task design, we included the following statements:

From Methods:

“HCP-A collected fMRI data from three tasks: Face Name (Sperling et al., 2001), Conditioned Approach Response Inhibition Task (CARIT) (Somerville et al., 2018) and VISual MOTOR (VISMOTOR) (Ances et al., 2009).

First, the Face Name task (Sperling et al., 2001) taps into episodic memory. The task had three blocks. In the encoding block [Encoding], participants were asked to memorise the names of faces shown. These faces were then shown again in the recall block [Recall] when the participants were asked if they could remember the names of the previously shown faces. There was also the distractor block [Distractor] occurring between the encoding and recall blocks. Here participants were distracted by a Go/NoGo task. We computed six contrasts for this Face Name task: [Encode], [Recall], [Distractor], [Encode vs. Distractor], [Recall vs. Distractor] and [Encode vs. Recall].

Second, the CARIT task (Somerville et al., 2018) was adapted from the classic Go/NoGo task and taps into inhibitory control. Participants were asked to press a button to all [Go] but not to two [NoGo] shapes. We computed three contrasts for the CARIT task: [NoGo], [Go] and [NoGo vs. Go].

Third, the VISMOTOR task (Ances et al., 2009) was designed to test simple activation of the motor and visual cortices. Participants saw a checkerboard with a red square either on the left or right. They needed to press a corresponding key to indicate the location of the red square. We computed just one contrast for the VISMOTOR task: [Vismotor], which indicates the presence of the checkerboard vs. baseline.”

Second, for MRI processing procedures, we included the following statements.

From Methods: “HCP-A provides details of parameters for brain MRI elsewhere (Bookheimer et al., 2019; Harms et al., 2018). Here we used MRI data that were pre-processed by the HCP-A with recommended methods, including the MSMALL alignment (Glasser et al., 2016; Robinson et al., 2018) and ICA-FIX (Glasser et al., 2016) for functional MRI. We used multiple brain MRI modalities, covering task functional MRI (task fMRI), resting-state functional MRI (rsfMRI) and structural MRI (sMRI), and organised them into 19 sets of features.”

“ Sets of Features 1-10: Task fMRI contrast (Task Contrast) Task contrasts reflect fMRI activation relevant to events in each task. Bookheimer and colleagues (2019) provided detailed information about the fMRI in HCP-A. Here we focused on the pre-processed task fMRI Connectivity Informatics Technology Initiative (CIFTI) files with a suffix, “_PA_Atlas_MSMAll_hp0_clean.dtseries.nii.” These CIFTI files encompassed both the cortical mesh surface and subcortical volume (Glasser et al., 2013). Collected using the posterior-to-anterior (PA) phase, these files were aligned using MSMALL (Glasser et al., 2016; Robinson et al., 2018), linear detrended (see https://groups.google.com/a/humanconnectome.org/g/hcp-users/c/ZLJc092h980/m/GiihzQAUAwAJ) and cleaned from potential artifacts using ICA-FIX (Glasser et al., 2016).

To extract Task Contrasts, we regressed the fMRI time series on the convolved task events using a double-gamma canonical hemodynamic response function via FMRIB Software Library (FSL)’s FMRI Expert Analysis Tool (FEAT) (Woolrich et al., 2001). We kept FSL’s default high pass cutoff at 200s (i.e., .005 Hz). We then parcellated the contrast ‘cope’ files, using the Glasser atlas (Gordon et al., 2016) for cortical surface regions and the Freesurfer’s automatic segmentation (aseg) (Fischl et al., 2002) for subcortical regions. This resulted in 379 regions, whose number was, in turn, the number of features for each Task Contrast set of features. “

“ Sets of Features 11-13: Task fMRI functional connectivity (Task FC) Task FC reflects functional connectivity (FC ) among the brain regions during each task, which is considered an important source of individual differences (Elliott et al., 2019; Fair et al., 2007; Gratton et al., 2018). We used the same CIFTI file “_PA_Atlas_MSMAll_hp0_clean.dtseries.nii.” as the task contrasts. Unlike Task Contrasts, here we treated the double-gamma, convolved task events as regressors of no interest and focused on the residuals of the regression from each task (Fair et al., 2007). We computed these regressors on FSL, and regressed them in nilearn (Abraham et al., 2014). Following previous work on task FC (Elliott et al., 2019), we applied a highpass at .008 Hz. For parcellation, we used the same atlases as Task Contrast (Fischl et al., 2002; Glasser et al., 2016). We computed Pearson’s correlations of each pair of 379 regions, resulting in a table of 71,631 non-overlapping FC indices for each task. We then applied r-to-z transformation and principal component analysis (PCA) of 75 components (Rasero et al., 2021; Sripada et al., 2019, 2020). Note to avoid data leakage, we conducted the PCA on each training set and applied its definition to the corresponding test set. Accordingly, there were three sets of 75 features for Task FC, one for each task.

Set of Features 14: Resting-state functional MRI functional connectivity (Rest FC) Similar to Task FC, Rest FC reflects functional connectivity (FC ) among the brain regions, except that Rest FC occurred during the resting (as opposed to task-performing) period. HCP-A collected Rest FC from four 6.42-min (488 frames) runs across two days, leading to 26-min long data (Harms et al., 2018). On each day, the study scanned two runs of Rest FC, starting with anterior-to-posterior (AP) and then with posterior-to-anterior (PA) phase encoding polarity. We used the “rfMRI_REST_Atlas_MSMAll_hp0_clean.dscalar.nii” file that was pre-processed and concatenated across the four runs. We applied the same computations (i.e., highpass filter, parcellation, Pearson’s correlations, r-to-z transformation and PCA) with the Task FC.

Sets of Features 15-18: Structural MRI (sMRI)

sMRI reflects individual differences in brain anatomy. The HCP-A used an established pre-processing pipeline for sMRI (Glasser et al., 2013). We focused on four sets of features: cortical thickness, cortical surface area, subcortical volume and total brain volume. For cortical thickness and cortical surface area, we used Destrieux’s atlas (Destrieux et al., 2010; Fischl, 2012) from FreeSurfer’s “aparc.stats” file, resulting in 148 regions for each set of features. For subcortical volume, we used the aseg atlas (Fischl et al., 2002) from FreeSurfer’s “aseg.stats” file, resulting in 19 regions. For total brain volume, we had five FreeSurfer-based features: “FS_IntraCranial_Vol” or estimated intra-cranial volume, “FS_TotCort_GM_Vol” or total cortical grey matter volume, “FS_Tot_WM_Vol” or total cortical white matter volume, “FS_SubCort_GM_Vol” or total subcortical grey matter volume and “FS_BrainSegVol_eTIV_Ratio” or ratio of brain segmentation volume to estimated total intracranial volume.”

Third, for regression methods and bias correction methods used, we included the following statements:

From Methods:

“For the machine learning algorithm, we used Elastic Net (Zou & Hastie, 2005). Elastic Net is a general form of penalised regressions (including Lasso and Ridge regression), allowing us to simultaneously draw information across different brain indices to predict one target variable. Penalised regressions are commonly used for building age-prediction models (Jirsaraie, Gorelik, et al., 2023). Previously we showed that the performance of Elastic Net in predicting cognitive abilities is on par, if not better than, many non-linear and more-complicated algorithms (Pat, Wang, Bartonicek, et al., 2022; Tetereva et al., 2022). Moreover, Elastic Net coefficients are readily explainable, allowing us the ability to explain how our age-prediction and cognition-prediction models made the prediction from each brain feature (Molnar, 2019; Pat, Wang, Bartonicek, et al., 2022) (see below).

Elastic Net simultaneously minimises the weighted sum of the features’ coefficients. The degree of penalty to the sum of the feature’s coefficients is determined by a shrinkage hyperparameter ‘α’: the greater the α, the more the coefficients shrink, and the more regularised the model becomes. Elastic Net also includes another hyperparameter, ‘l1 ratio’, which determines the degree to which the sum of either the squared (known as ‘Ridge’; l1 ratio=0) or absolute (known as ‘Lasso’; l1 ratio=1) coefficients is penalised (Zou & Hastie, 2005). The objective function of Elastic Net as implemented by sklearn (Pedregosa et al., 2011) is defined as:

where X is the features, y is the target, and β is the coefficient. In our grid search, we tuned two Elastic Net hyperparameters: α using 70 numbers in log space, ranging from .1 and 100, and l_1-ratio using 25 numbers in linear space, ranging from 0 and 1.

To understand how Elastic Net made a prediction based on different brain features, we examined the coefficients of the tuned model. Elastic Net coefficients can be considered as feature importance, such that more positive Elastic Net coefficients lead to more positive predicted values and, similarly, more negative Elastic Net coefficients lead to more negative predicted values (Molnar, 2019; Pat, Wang, Bartonicek, et al., 2022). While the magnitude of Elastic Net coefficients is regularised (thus making it difficult for us to interpret the magnitude itself directly), we could still indicate that a brain feature with a higher magnitude weights relatively stronger in making a prediction. Another benefit of Elastic Net as a penalised regression is that the coefficients are less susceptible to collinearity among features as they have already been regularised (Dormann et al., 2013; Pat, Wang, Bartonicek, et al., 2022).

Given that we used five-fold nested cross validation, different outer folds may have different degrees of ‘α’ and ‘l1 ratio’, making the final coefficients from different folds to be different. For instance, for certain sets of features, penalisation may not play a big part (i.e., higher or lower ‘α’ leads to similar predictive performance), resulting in different ‘α’ for different folds. To remedy this in the visualisation of Elastic Net feature importance, we refitted the Elastic Net model to the full dataset without splitting them into five folds and visualised the coefficients on brain images using Brainspace (Vos De Wael et al., 2020) and Nilern (Abraham et al., 2014) packages. Note, unlike other sets of features, Task FC and Rest FC were modelled after data reduction via PCA. Thus, for Task FC and Rest FC, we, first, multiplied the absolute PCA scores (extracted from the ‘components_’ attribute of ‘sklearn.decomposition.PCA’) with Elastic Net coefficients and, then, summed the multiplied values across the 75 components, leaving 71,631 ROI-pair indices. “

References

Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn. Frontiers in Neuroinformatics, 8, 14. https://doi.org/10.3389/fninf.2014.00014

Ances, B. M., Liang, C. L., Leontiev, O., Perthen, J. E., Fleisher, A. S., Lansing, A. E., & Buxton, R. B. (2009). Effects of aging on cerebral blood flow, oxygen metabolism, and blood oxygenation level dependent responses to visual stimulation. Human Brain Mapping, 30(4), 1120–1132. https://doi.org/10.1002/hbm.20574

Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I. M., Truelove-Hill, M., Srinivasan, D., Mamourian, L., Pomponio, R., Fan, Y., Launer, L. J., Masters, C. L., Maruff, P., Zhuo, C., Völzke, H., Johnson, S. C., Fripp, J., Koutsouleris, N., Satterthwaite, T. D., … on behalf of the ISTAGING Consortium, the P. A. disease C., ADNI, and CARDIA studies. (2020). MRI signatures of brain age and disease over the lifespan based on a deep brain network and 14 468 individuals worldwide. Brain, 143(7), 2312–2324. https://doi.org/10.1093/brain/awaa160

Bookheimer, S. Y., Salat, D. H., Terpstra, M., Ances, B. M., Barch, D. M., Buckner, R. L., Burgess, G. C., Curtiss, S. W., Diaz-Santos, M., Elam, J. S., Fischl, B., Greve, D. N., Hagy, H. A., Harms, M. P., Hatch, O. M., Hedden, T., Hodge, C., Japardi, K. C., Kuhn, T. P., … Yacoub, E. (2019). The Lifespan Human Connectome Project in Aging: An overview. NeuroImage, 185, 335–348. https://doi.org/10.1016/j.neuroimage.2018.10.009

Butler, E. R., Chen, A., Ramadan, R., Le, T. T., Ruparel, K., Moore, T. M., Satterthwaite, T. D., Zhang, F., Shou, H., Gur, R. C., Nichols, T. E., & Shinohara, R. T. (2021). Pitfalls in brain age analyses. Human Brain Mapping, 42(13), 4092–4101. https://doi.org/10.1002/hbm.25533

Cole, J. H. (2020). Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical, lifestyle, and cognitive factors. Neurobiology of Aging, 92, 34–42. https://doi.org/10.1016/j.neurobiolaging.2020.03.014

Destrieux, C., Fischl, B., Dale, A., & Halgren, E. (2010). Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. NeuroImage, 53(1), 1–15. https://doi.org/10.1016/j.neuroimage.2010.06.010

Dormann, C. F., Elith, J., Bacher, S., Buchmann, C., Carl, G., Carré, G., Marquéz, J. R. G., Gruber, B., Lafourcade, B., Leitão, P. J., Münkemüller, T., McClean, C., Osborne, P. E., Reineking, B., Schröder, B., Skidmore, A. K., Zurell, D., & Lautenbach, S. (2013). Collinearity: A review of methods to deal with it and a simulation study evaluating their performance. Ecography, 36(1), 27–46. https://doi.org/10.1111/j.1600-0587.2012.07348.x

Dubois, J., Galdi, P., Paul, L. K., & Adolphs, R. (2018). A distributed brain network predicts general intelligence from resting-state human neuroimaging data. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170284. https://doi.org/10.1098/rstb.2017.0284

Elliott, M. L., Knodt, A. R., Cooke, M., Kim, M. J., Melzer, T. R., Keenan, R., Ireland, D., Ramrakha, S., Poulton, R., Caspi, A., Moffitt, T. E., & Hariri, A. R. (2019). General functional connectivity: Shared features of resting-state and task fMRI drive reliable and heritable individual differences in functional brain networks. NeuroImage, 189, 516–532. https://doi.org/10.1016/j.neuroimage.2019.01.068

Fair, D. A., Schlaggar, B. L., Cohen, A. L., Miezin, F. M., Dosenbach, N. U. F., Wenger, K. K., Fox, M. D., Snyder, A. Z., Raichle, M. E., & Petersen, S. E. (2007). A method for using blocked and event-related fMRI data to study “resting state” functional connectivity. NeuroImage, 35(1), 396–405. https://doi.org/10.1016/j.neuroimage.2006.11.051

Fischl, B. (2012). FreeSurfer. NeuroImage, 62(2), 774–781. https://doi.org/10.1016/j.neuroimage.2012.01.021

Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., van der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, S., Montillo, A., Makris, N., Rosen, B., & Dale, A. M. (2002). Whole Brain Segmentation. Neuron, 33(3), 341–355. https://doi.org/10.1016/S0896-6273(02)00569-X

Glasser, M. F., Smith, S. M., Marcus, D. S., Andersson, J. L. R., Auerbach, E. J., Behrens, T. E. J., Coalson, T. S., Harms, M. P., Jenkinson, M., Moeller, S., Robinson, E. C., Sotiropoulos, S. N., Xu, J., Yacoub, E., Ugurbil, K., & Van Essen, D. C. (2016). The Human Connectome Project’s neuroimaging approach. Nature Neuroscience, 19(9), 1175–1187. https://doi.org/10.1038/nn.4361

Glasser, M. F., Sotiropoulos, S. N., Wilson, J. A., Coalson, T. S., Fischl, B., Andersson, J. L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J. R., Van Essen, D. C., & Jenkinson, M. (2013). The minimal preprocessing pipelines for the Human Connectome Project. NeuroImage, 80, 105–124. https://doi.org/10.1016/j.neuroimage.2013.04.127

Gordon, E. M., Laumann, T. O., Adeyemo, B., Huckins, J. F., Kelley, W. M., & Petersen, S. E. (2016). Generation and Evaluation of a Cortical Area Parcellation from Resting-State Correlations. Cerebral Cortex, 26(1), 288–303. https://doi.org/10.1093/cercor/bhu239

Gratton, C., Laumann, T. O., Nielsen, A. N., Greene, D. J., Gordon, E. M., Gilmore, A. W., Nelson, S. M., Coalson, R. S., Snyder, A. Z., Schlaggar, B. L., Dosenbach, N. U. F., & Petersen, S. E. (2018). Functional Brain Networks Are Dominated by Stable Group and Individual Factors, Not Cognitive or Daily Variation. Neuron, 98(2), 439-452.e5. https://doi.org/10.1016/j.neuron.2018.03.035

Hahn, T., Fisch, L., Ernsting, J., Winter, N. R., Leenings, R., Sarink, K., Emden, D., Kircher, T., Berger, K., & Dannlowski, U. (2021). From ‘loose fitting’ to high-performance, uncertainty-aware brain-age modelling. Brain, 144(3), e31–e31. https://doi.org/10.1093/brain/awaa454

Harms, M. P., Somerville, L. H., Ances, B. M., Andersson, J., Barch, D. M., Bastiani, M., Bookheimer, S. Y., Brown, T. B., Buckner, R. L., Burgess, G. C., Coalson, T. S., Chappell, M. A., Dapretto, M., Douaud, G., Fischl, B., Glasser, M. F., Greve, D. N., Hodge, C., Jamison, K. W., … Yacoub, E. (2018). Extending the Human Connectome Project across ages: Imaging protocols for the Lifespan Development and Aging projects. NeuroImage, 183, 972–984. https://doi.org/10.1016/j.neuroimage.2018.09.060

Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., Sanislow, C., & Wang, P. (2010). Research Domain Criteria (RDoC): Toward a New Classification Framework for Research on Mental Disorders. American Journal of Psychiatry, 167(7), 748–751. https://doi.org/10.1176/appi.ajp.2010.09091379

Jirsaraie, R. J., Gorelik, A. J., Gatavins, M. M., Engemann, D. A., Bogdan, R., Barch, D. M., & Sotiras, A. (2023). A systematic review of multimodal brain age studies: Uncovering a divergence between model accuracy and utility. Patterns, 4(4), 100712. https://doi.org/10.1016/j.patter.2023.100712

Jirsaraie, R. J., Kaufmann, T., Bashyam, V., Erus, G., Luby, J. L., Westlye, L. T., Davatzikos, C., Barch, D. M., & Sotiras, A. (2023). Benchmarking the generalizability of brain age models: Challenges posed by scanner variance and prediction bias. Human Brain Mapping, 44(3), 1118–1128. https://doi.org/10.1002/hbm.26144

Marquand, A. F., Rezek, I., Buitelaar, J., & Beckmann, C. F. (2016). Understanding Heterogeneity in Clinical Cohorts Using Normative Models: Beyond Case-Control Studies. Biological Psychiatry, 80(7), 552–561. https://doi.org/10.1016/j.biopsych.2015.12.023

Molnar, C. (2019). Interpretable Machine Learning. A Guide for Making Black Box Models Explainable. https://christophm.github.io/interpretable-ml-book/

Nimon, K., Lewis, M., Kane, R., & Haynes, R. M. (2008). An R package to compute commonality coefficients in the multiple regression case: An introduction to the package and a practical example. Behavior Research Methods, 40(2), 457–466. https://doi.org/10.3758/BRM.40.2.457

Pat, N., Wang, Y., Anney, R., Riglin, L., Thapar, A., & Stringaris, A. (2022). Longitudinally stable, brain‐based predictive models mediate the relationships between childhood cognition and socio‐demographic, psychological and genetic factors. Human Brain Mapping, hbm.26027. https://doi.org/10.1002/hbm.26027

Pat, N., Wang, Y., Bartonicek, A., Candia, J., & Stringaris, A. (2022). Explainable machine learning approach to predict and explain the relationship between task-based fMRI and individual differences in cognition. Cerebral Cortex, bhac235. https://doi.org/10.1093/cercor/bhac235

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, É. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12(85), 2825–2830.

Poldrack, R. A., Huckins, G., & Varoquaux, G. (2020). Establishment of Best Practices for Evidence for Prediction: A Review. JAMA Psychiatry, 77(5), 534–540. https://doi.org/10.1001/jamapsychiatry.2019.3671

Rasero, J., Sentis, A. I., Yeh, F.-C., & Verstynen, T. (2021). Integrating across neuroimaging modalities boosts prediction accuracy of cognitive ability. PLOS Computational Biology, 17(3), e1008347. https://doi.org/10.1371/journal.pcbi.1008347

Robinson, E. C., Garcia, K., Glasser, M. F., Chen, Z., Coalson, T. S., Makropoulos, A., Bozek, J., Wright, R., Schuh, A., Webster, M., Hutter, J., Price, A., Cordero Grande, L., Hughes, E., Tusor, N., Bayly, P. V., Van Essen, D. C., Smith, S. M., Edwards, A. D., … Rueckert, D. (2018). Multimodal surface matching with higher-order smoothness constraints. NeuroImage, 167, 453–465. https://doi.org/10.1016/j.neuroimage.2017.10.037

Rokicki, J., Wolfers, T., Nordhøy, W., Tesli, N., Quintana, D. S., Alnæs, D., Richard, G., de Lange, A.-M. G., Lund, M. J., Norbom, L., Agartz, I., Melle, I., Nærland, T., Selbæk, G., Persson, K., Nordvik, J. E., Schwarz, E., Andreassen, O. A., Kaufmann, T., & Westlye, L. T. (2021). Multimodal imaging improves brain age prediction and reveals distinct abnormalities in patients with psychiatric and neurological disorders. Human Brain Mapping, 42(6), 1714–1726. https://doi.org/10.1002/hbm.25323

Somerville, L. H., Bookheimer, S. Y., Buckner, R. L., Burgess, G. C., Curtiss, S. W., Dapretto, M., Elam, J. S., Gaffrey, M. S., Harms, M. P., Hodge, C., Kandala, S., Kastman, E. K., Nichols, T. E., Schlaggar, B. L., Smith, S. M., Thomas, K. M., Yacoub, E., Van Essen, D. C., & Barch, D. M. (2018). The Lifespan Human Connectome Project in Development: A large-scale study of brain connectivity development in 5–21 year olds. NeuroImage, 183, 456–468. https://doi.org/10.1016/j.neuroimage.2018.08.050

Sperling, R. A., Bates, J. F., Cocchiarella, A. J., Schacter, D. L., Rosen, B. R., & Albert, M. S. (2001). Encoding novel face-name associations: A functional MRI study. Human Brain Mapping, 14(3), 129–139. https://doi.org/10.1002/hbm.1047

Sripada, C., Angstadt, M., Rutherford, S., Kessler, D., Kim, Y., Yee, M., & Levina, E. (2019). Basic Units of Inter-Individual Variation in Resting State Connectomes. Scientific Reports, 9(1), Article 1. https://doi.org/10.1038/s41598-018-38406-5

Sripada, C., Angstadt, M., Rutherford, S., Taxali, A., & Shedden, K. (2020). Toward a “treadmill test” for cognition: Improved prediction of general cognitive ability from the task activated brain. Human Brain Mapping, 41(12), 3186–3197. https://doi.org/10.1002/hbm.25007

Tetereva, A., Li, J., Deng, J. D., Stringaris, A., & Pat, N. (2022). Capturing brain‐cognition relationship: Integrating task‐based fMRI across tasks markedly boosts prediction and test‐retest reliability. NeuroImage, 263, 119588. https://doi.org/10.1016/j.neuroimage.2022.119588

Vieira, B. H., Pamplona, G. S. P., Fachinello, K., Silva, A. K., Foss, M. P., & Salmon, C. E. G. (2022). On the prediction of human intelligence from neuroimaging: A systematic review of methods and reporting. Intelligence, 93, 101654. https://doi.org/10.1016/j.intell.2022.101654

Vos De Wael, R., Benkarim, O., Paquola, C., Lariviere, S., Royer, J., Tavakol, S., Xu, T., Hong, S.-J., Langs, G., Valk, S., Misic, B., Milham, M., Margulies, D., Smallwood, J., & Bernhardt, B. C. (2020). BrainSpace: A toolbox for the analysis of macroscale gradients in neuroimaging and connectomics datasets. Communications Biology, 3(1), 103. https://doi.org/10.1038/s42003-020-0794-7

Woolrich, M. W., Ripley, B. D., Brady, M., & Smith, S. M. (2001). Temporal Autocorrelation in Univariate Linear Modeling of FMRI Data. NeuroImage, 14(6), 1370–1386. https://doi.org/10.1006/nimg.2001.0931

Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301–320. https://doi.org/10.1111/j.1467-9868.2005.00503.x

The following is the authors’ response to the previous reviews.

eLife assessment

This useful manuscript challenges the utility of current paradigms for estimating brain-age with magnetic resonance imaging measures, but presents inadequate evidence to support the suggestion that an alternative approach focused on predicting cognition is more useful. The paper would benefit from a clearer explication of the methods and a more critical evaluation of the conceptual basis of the different models. This work will be of interest to researchers working on brain-age and related models.

Thank you so much for providing high-quality reviews on our manuscript. We revised the manuscript to address all of the reviewers’ comments and provided full responses to each of the comments below. Importantly, in this revision, we clarified that we did not intend to use Brain Cognition as an alternative approach. This is because, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Here we made this point more explicit and further stated that the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. By examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And such quantification is the third aim of this study.

Public Reviews:

Reviewer 1 (Public Review):

In this paper, the authors evaluate the utility of brain-age-derived metrics for predicting cognitive decline by performing a 'commonality' analysis in a downstream regression that enables the different contribution of different predictors to be assessed. The main conclusion is that brain-age-derived metrics do not explain much additional variation in cognition over and above what is already explained by age. The authors propose to use a regression model trained to predict cognition ("brain-cognition") as an alternative suited to applications of cognitive decline. While this is less accurate overall than brain age, it explains more unique variance in the downstream regression.

(1) I thank the authors for addressing many of my concerns with this revision. However, I do not feel they have addressed them all. In particular I think the authors could do more to address the concern I raised about the instability of the regression coefficients and about providing enough detail to determine that the stacked regression models do not overfit.

Thank you Reviewer 1 for the comment. We addressed them in our response to Reviewer 1 Recommendations For The Authors #1 and #2 (see below).

(2) In considering my responses to the authors revision, I also must say that I agree with Reviewer 3 about the limitations of the brain age and brain cognition methods conceptually. In particular that the regression model used to predict fluid cognition will by construction explain more variance in cognition than a brain age model that is trained to predict age. To be fair, these conceptual problems are more widespread than this paper alone, so I do not believe the authors should be penalised for that. However, I would recommend to make these concerns more explicit in the manuscript

Thank you Reviewer 1 for the comment. We addressed them in our response to Reviewer 1 Recommendations For The Authors #3 (see below).

Reviewer 2 (Public Review):

In this study, the authors aimed to evaluate the contribution of brain-age indices in capturing variance in cognitive decline and proposed an alternative index, brain-cognition, for consideration.

The study employs suitable methods and data to address the research questions, and the methods and results sections are generally clear and easy to follow.

I appreciate the authors' efforts in significantly improving the paper, including some considerable changes, from the original submission. While not all reviewer points were tackled, the majority of them were adequately addressed. These include additional analyses, more clarity in the methods and a much richer and nuanced discussion. While recognising the merits of the revised paper, I have a few additional comments.

(1) Perhaps it would help the reader to note that it might be expected for brain-cognition to account for a significantly larger variance (11%) in fluid cognition, in contrast to brain-age. This stems from the fact that the authors specifically trained brain-cognition to predict fluid cognition, the very variable under consideration. In line with this, the authors later recommend that researchers considering the use of brain-age should evaluate its utility using a regression approach. The latter involves including a brain index (e.g. brain-cognition) previously trained to predict the regression's target variable (e.g. fluid cognition) alongside a brain-age index (e.g., corrected brain-age gap). If the target-trained brain index outperforms the brain-age metric, it suggests that relying solely on brain-age might not be the optimal choice. Although not necessarily the case, is it surprising for the target-trained brain index to demonstrate better performance than brain-age? This harks back to the broader point raised in the initial review: while brain-age may prove useful (though sometimes with modest effect sizes) across diverse outcomes as a generally applicable metric, a brain index tailored for predicting a specific outcome, such as brain-cognition in this case, might capture a considerably larger share of variance in that specific context but could lack broader applicability. The latter aspect needs to be empirically assessed.

Thank you so much for raising this point. Reviewer 1 (Public Review #2/Recommendations For The Authors #3) and Reviewer 3 (Recommendations for the Authors #1) made a similar observation. We now made changes to the introduction and discussion to address this concern (please see our responses to Reviewer 1 Recommendations For The Authors #3 below).

Briefly, as in our 2nd revision, we did not intend to compare Brain Age with Brain Cognition since, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Here we made this point more explicit and further stated that the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. By examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And such quantification is the third aim of this study.

(2) Furthermore, the discussion pertaining to training brain-age models on healthy populations for subsequent testing on individuals with neurological or psychological disorders seems somewhat one-sided within the broader debate. This one-sidedness might potentially confuse readers. It is worth noting that the choice to employ healthy participants in the training model is likely deliberate, serving as a norm against which atypical populations are compared. To provide a more comprehensive understanding, referencing Tim Hans's counterargument to Bashyam's perspective could offer a more complete view (https://academic.oup.com/brain/article/144/3/e31/6214475?login=false).

Thank you Reviewer 2 for bringing up this issue. We have now revised the paragraph in question and added nuances on the usage of Brain Age for normative vs. case-control studies. We also cited Tim Hahn’s article that explained the conceptual foundation of the use of Brain Age in case-control studies. Please see below. Additionally, we also made a statement about our study not being able to address issues about the case-control studies directly in the newly written conclusion (see Reviewer 3 Recommendations for the Authors #3).

Discussion:

“There is a notable difference between studies investigating the utility of Brain Age in explaining cognitive functioning, including ours and others (e.g., Butler et al., 2021; Cole, 2020, 2020; Jirsaraie et al., 2023) and those explaining neurological/psychological disorders (e.g., Bashyam et al., 2020; Rokicki et al., 2021). We consider the former as a normative type of study and the latter as a case-control type of study (Insel et al., 2010; Marquand et al., 2016). Those case-control Brain Age studies focusing on neurological/psychological disorders often build age-prediction models from MRI data of largely healthy participants (e.g., controls in a case-control design or large samples in a population-based design), apply the built age-prediction models to participants without vs. with neurological/psychological disorders and compare Brain Age indices between the two groups. On the one hand, this means that case-control studies treat Brain Age as a method to detect anomalies in the neurological/psychological group (Hahn et al., 2021). On the other hand, this also means that case-control studies have to ignore under-fitted models when applied prediction models built from largely healthy participants to participants with neurological/psychological disorders (i.e., Brain Age may predict chronological age well for the controls, but not for those with a disorder). On the contrary, our study and other normative studies focusing on cognitive functioning often build age-prediction models from MRI data of largely healthy participants and apply the built age-prediction models to participants who are also largely healthy. Accordingly, the age-prediction models for explaining cognitive functioning in normative studies, while not allowing us to detect group-level anomalies, do not suffer from being under-fitted. This unfortunately might limit the generalisability of our study into just the normative type of study. Future work is still needed to test the utility of brain age in the case-control case.”

(3) Overall, this paper makes a significant contribution to the field of brain-age and related brain indices and their utility.

Thank you for the encouragement.

Reviewer 3 (Public Review):

The main question of this article is as follows: "To what extent does having information on brain-age improve our ability to capture declines in fluid cognition beyond knowing a person's chronological age?" This question is worthwhile, considering that there is considerable confusion in the field about the nature of brain-age.

(1) Thank you to the authors for addressing so many of my concerns with this revision. There are a few points that I feel still need addressing/clarifying related to 1) calculating brain cognition, 2) the inevitability of their results, and 3) their continued recommendation to use brain-age metrics.

Thank you Reviewer 3 for the comment. We addressed them in our response to Reviewer 3 Recommendations For The Authors #1-3 (see below).

Recommendations for the authors:

Reviewer 1 (Recommendations For The Authors):

(1) I do not feel the authors have fully addressed the concern I raised about the stacked regression models. Despite the new figure, it is still not entirely clear what the authors are using as the training set in the final step. To be clear, the problem occurs because of the parameters, not the hyperparameters (which the authors now state that they are optimising via nested grid search). in other words, given a regression model y = X*beta, if the X are taken to be predictions from a lower level regression model, then they contain information that is derived from both the training set at the test set for the model that this was trained on. If the split is the same (i.e. the predictions are derived on the same test set as is being used at the second level), then this can lead to overfitting. It is not clear to me whether the authors have done this or not. Please provide additional detail to clarify this point.

Thank you for allowing us an opportunity to clarify our stacked model. We wanted to confirm that we did not use test sets to build a stacked model in both lower and higher levels of the Elastic Net models. Test sets were there just for testing the performance of the models. We made additional clarification to make this clearer (see below). Let us explain what we did and provide the rationales below.

From Methods:

“We used nested cross-validation (CV) to build these prediction models (see Figure 7). We first split the data into five outer folds, leaving each outer fold with around 100 participants. This number of participants in each fold is to ensure the stability of the test performance across folds. In each outer-fold CV loop, one of the outer folds was treated as an outer-fold test set, and the rest was treated as an outer-fold training set. Ultimately, looping through the nested CV resulted in a) prediction models from each of the 18 sets of features as well as b) prediction models that drew information across different combinations of the 18 separate sets, known as “stacked models.” We specified eight stacked models: “All” (i.e., including all 18 sets of features), “All excluding Task FC”, “All excluding Task Contrast”, “Non-Task” (i.e., including only Rest FC and sMRI), “Resting and Task FC”, “Task Contrast and FC”, “Task Contrast” and “Task FC”. Accordingly, there were 26 prediction models in total for both Brain Age and Brain Cognition.

To create these 26 prediction models, we applied three steps for each outer-fold loop. The first step aimed at tuning prediction models for each of 18 sets of features. This step only involved the outer-fold training set and did not involve the outer-fold test set. Here, we divided the outer-fold training set into five inner folds and applied inner-fold CV to tune hyperparameters with grid search. Specifically, in each inner-fold CV, one of the inner folds was treated as an inner-fold validation set, and the rest was treated as an inner-fold training set. Within each inner-fold CV loop, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters and applied the estimated model to the inner-fold validation set. After looping through the inner-fold CV, we, then, chose the prediction models that led to the highest performance, reflected by coefficient of determination (R2), on average across the inner-fold validation sets. This led to 18 tuned models, one for each of the 18 sets of features, for each outer fold.

The second step aimed at tuning stacked models. Same as the first step, the second step only involved the outer-fold training set and did not involve the outer-fold test set. Here, using the same outer-fold training set as the first step, we applied tuned models, created from the first step, one from each of the 18 sets of features, resulting in 18 predicted values for each participant. We, then, re-divided this outer-fold training set into new five inner folds. In each inner fold, we treated different combinations of the 18 predicted values from separate sets of features as features to predict the targets in separate “stacked” models. Same as the first step, in each inner-fold CV loop, we treated one out of five inner folds as an inner-fold validation set, and the rest as an inner-fold training set. Also as in the first step, we used the inner-fold training set to estimate parameters of the prediction model with a particular set of hyperparameters from our grid. We tuned the hyperparameters of stacked models using grid search by selecting the models with the highest R2 on average across the inner-fold validation sets. This led to eight tuned stacked models.

The third step aimed at testing the predictive performance of the 18 tuned prediction models from each of the set of features, built from the first step, and eight tuned stacked models, built from the second step. Unlike the first two steps, here we applied the already tuned models to the outer-fold test set. We started by applying the 18 tuned prediction models from each of the sets of features to each observation in the outer-fold test set, resulting in 18 predicted values. We then applied the tuned stacked models to these predicted values from separate sets of features, resulting in eight predicted values.

To demonstrate the predictive performance, we assessed the similarity between the observed values and the predicted values of each model across outer-fold test sets, using Pearson’s r, coefficient of determination (R2) and mean absolute error (MAE). Note that for R2, we used the sum of squares definition (i.e., R2 = 1 – (sum of squares residuals/total sum of squares)) per a previous recommendation (Poldrack et al., 2020). We considered the predicted values from the outer-fold test sets of models predicting age or fluid cognition, as Brain Age and Brain Cognition, respectively.”

Author response image 1.

Diagram of the nested cross-validation used for creating predictions for models of each set of features as well as predictions for stacked models.

Note some previous research, including ours (Tetereva et al., 2022), splits the observations in the outer-fold training set into layer 1 and layer 2 and applies the first and second steps to layers 1 and 2, respectively. Here we decided against this approach and used the same outer-fold training set for both first and second steps in order to avoid potential bias toward the stacked models. This is because, when the data are split into two layers, predictive models built for each separate set of features only use the data from layer 1, while the stacked models use the data from both layers 1 and 2. In practice with large enough data, these two approaches might not differ much, as we demonstrated previously (Tetereva et al., 2022).

(2) I also do not feel the authors have fully addressed the concern I raised about stability of the regression coefficients over splits of the data. I wanted to see the regression coefficients, not the predictions. The predictions can be stable when the coefficients are not.

The focus of this article is on the predictions. Still, as pointed out by reviewer 1, it is informative for readers to understand how stable the feature importance (i.e., Elastic Net coefficients) is. To demonstrate the stability of feature importance, we now examined the rank stability of feature importance using Spearman’s ρ (see Figure 4). Specifically, we correlated the feature importance between two prediction models of the same features, used in two different outer-fold test sets. Given that there were five outer-fold test sets, we computed 10 Spearman’s ρ for each prediction model of the same features. We found Spearman’s ρ to be varied dramatically in both age-prediction (range=.31-.94) and fluid cognition-prediction (range=.16-.84) models. This means that some prediction models were much more stable in their feature importance than others. This is probably due to various factors such as a) the collinearity of features in the model, b) the number of features (e.g., 71,631 features in functional connectivity, which were further reduced to 75 PCAs, as compared to 19 features in subcortical volume based on the ASEG atlas), c) the penalisation of coefficients either with ‘Ridge’ or ‘Lasso’ methods, which resulted in reduction as a group of features or selection of a feature among correlated features, respectively, and d) the predictive performance of the models. Understanding the stability of feature importance is beyond the scope of the current article. As mentioned by Reviewer 1, “The predictions can be stable when the coefficients are not,” and we chose to focus on the prediction in the current article.

Author response image 2.

Stability of feature importance (i.e., Elastic Net Coefficients) of prediction models. Each dot represents rank stability (reflected by Spearman’s ρ) in the feature importance between two prediction models of the same features, used in two different outer-fold test sets. Given that there were five outer-fold test sets, there were 10 Spearman’s ρs for each prediction model. The numbers to the right of the plots indicate the mean of Spearman’s ρ for each prediction model.

(3) I also must say that I agree with Reviewer 3 about the limitations of the brain-age and brain-cognition methods conceptually. In particular that the regression model used to predict fluid cognition will by construction explain more variance in cognition than a brain-age model that is trained to predict age. This suffers from the same problem the authors raise with brain-age and I agree that this would probably disappear if the authors had a separate measure of cognition against which to validate and were then to regress this out as they do for age correction. I am aware that these conceptual problems are more widespread than this paper alone (in fact throughout the brain-age literature), so I do not believe the authors should be penalised for that. However, I do think they can make these concerns more explicit and further tone down the comments they make about the utility of brain-cognition.

Thank you so much for raising this point. Reviewer 2 (Public Review #1) and Reviewer 3 (Recommendations for the Authors #1) made a similar observation. We now made changes to the introduction and discussion to address this concern (see below).

Briefly, we made it explicit that, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. That is, the relationship between Brain Cognition and fluid cognition indicates the upper limit of Brain Age’s capability in capturing fluid cognition. More importantly, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age. And this is the third goal of this present study.

From Introduction:

“Third and finally, certain variation in fluid cognition is related to brain MRI, but to what extent does Brain Age not capture this variation? To estimate the variation in fluid cognition that is related to the brain MRI, we could build prediction models that directly predict fluid cognition (i.e., as opposed to chronological age) from brain MRI data. Previous studies found reasonable predictive performances of these cognition-prediction models, built from certain MRI modalities (Dubois et al., 2018; Pat et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). Analogous to Brain Age, we called the predicted values from these cognition-prediction models, Brain Cognition. The strength of an out-of-sample relationship between Brain Cognition and fluid cognition reflects variation in fluid cognition that is related to the brain MRI and, therefore, indicates the upper limit of Brain Age’s capability in capturing fluid cognition. This is, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. Consequently, if we included Brain Cognition, Brain Age and chronological age in the same model to explain fluid cognition, we would be able to examine the unique effects of Brain Cognition that explain fluid cognition beyond Brain Age and chronological age. These unique effects of Brain Cognition, in turn, would indicate the amount of co-variation between brain MRI and fluid cognition that is missed by Brain Age.”

From Discussion:

“Third, by introducing Brain Cognition, we showed the extent to which Brain Age indices were not able to capture the variation in fluid cognition that is related to brain MRI. More specifically, using Brain Cognition allowed us to gauge the variation in fluid cognition that is related to the brain MRI, and thereby, to estimate the upper limit of what Brain Age can do. Moreover, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age.

From our results, Brain Cognition, especially from certain cognition-prediction models such as the stacked models, has relatively good predictive performance, consistent with previous studies (Dubois et al., 2018; Pat et al., 2022; Rasero et al., 2021; Sripada et al., 2020; Tetereva et al., 2022; for review, see Vieira et al., 2022). We then examined Brain Cognition using commonality analyses (Nimon et al., 2008) in multiple regression models having a Brain Age index, chronological age and Brain Cognition as regressors to explain fluid cognition. Similar to Brain Age indices, Brain Cognition exhibited large common effects with chronological age. But more importantly, unlike Brain Age indices, Brain Cognition showed large unique effects, up to around 11%. As explained above, the unique effects of Brain Cognition indicated the amount of co-variation between brain MRI and fluid cognition that was missed by a Brain Age index and chronological age. This missing amount was relatively high, considering that Brain Age and chronological age together explained around 32% of the total variation in fluid cognition. Accordingly, if a Brain Age index was used as a biomarker along with chronological age, we would have missed an opportunity to improve the performance of the model by around one-third of the variation explained.”

Reviewer #3 (Recommendations For The Authors):

Thank you to the authors for addressing so many of my concerns with this revision. There are a few points that I feel still need addressing/clarifying related to: 1) calculating brain cognition, 2) the inevitability of their results, and 3) their continued recommendation to use brain age metrics.

(1) I understand your point here. I think the distinction is that it is fine to build predictive models, but then there is no need to go through this intermediate step of "brain-cognition". Just say that brain features can predict cognition XX well, and brain-age (or some related metric) can predict cognition YY well. It creates a confusing framework for the reader that can lead them to believe that "brain-cognition" is not just a predicted value of fluid cognition from a model using brain features to predict cognition. While you clearly state that that is in fact what it is in the text, which is a huge improvement, I do not see what is added by going through brain-cognition instead of simply just obtaining a change in R2 where the first model uses brain features alone to predict cognition, and the second adds on brain-age (or related metrics), or visa versa, depending on the question. Please do this analysis, and either compare and contrast it with going through "brain-cognition" in your paper, or switch to this analysis, as it more directly addresses the question of the incremental predictive utility of brain-age above and beyond brain features.

Thank you so much for raising this point. Reviewer 1 (Public Review #2/Recommendations For The Authors #3) and Reviewer 2 (Public Review #1) made a similar observation. We now made changes to the introduction and discussion to address this concern (see our responses to Reviewer 1 Recommendations For The Authors #3 above).

Briefly, as in our 2nd revision, we made it explicitly clear that we did not intend to compare Brain Age with Brain Cognition since, by design, the variation in fluid cognition explained by Brain Cognition should be higher or equal to that explained by Brain Age. And, by examining what was captured by Brain Cognition, over and above Brain Age and chronological age via the unique effects of Brain Cognition, we were able to quantify the amount of co-variation between brain MRI and fluid cognition that was missed by Brain Age.

We have thought about changing the name Brain Cognition into something along the lines of “predicted values of prediction models predicting fluid cognition based on brain MRI.” However, this made the manuscript hard to follow, especially with the commonality analyses. For instance, the sentence, “Here, we tested Brain Cognition’s unique effects in multiple regression models with a Brain Age index, chronological age and Brain Cognition as regressors to explain fluid cognition” would become “Here, we tested predicted values of prediction models predicting fluid cognition based on brain MRI unique effects in multiple regression models with a Brain Age index, chronological age and predicted values of prediction models predicting fluid cognition based on brain MRI as regressors to explain fluid cognition.” We believe, given our additional explanation (see our responses to Reviewer 1 Recommendations For The Authors #3 above), readers should understand what Brain Cognition is, and that we did not intend to compare Brain Age and Brain Cognition directly.

As for the suggested analysis, “obtaining a change in R2 where the first model uses brain features alone to predict cognition, and the second adds on brain-age (or related metrics), or visa versa,” we have already done this in the form of commonality analysis (Nimon et al., 2008) (see Figure 7 below). That is, to obtain unique and common effects of the regressors, we need to look at all of the possible changes in R2 when all possible subsets of regressors were excluded or included, see equations 12 and 13 below.

From Methods:

“Similar to the above multiple regression model, we had chronological age, each Brain Age index and Brain Cognition as the regressors for fluid cognition:

Fluid Cognitioni = β0 + β1 Chronological Agei + β2 Brain Age Indexi,j + β3 Brain Cognitioni + εi, (12)

Applying the commonality analysis here allowed us, first, to investigate the addictive, unique effects of Brain Cognition, over and above chronological age and Brain Age indices. More importantly, the commonality analysis also enabled us to test the common, shared effects that Brain Cognition had with chronological age and Brain Age indices in explaining fluid cognition. We calculated the commonality analysis as follows (Nimon et al., 2017):

Unique Effectchronological age = ΔR2chronological age = R2chronological age, Brain Age index, Brain Cognition – R2 Brain Age index, Brain Cognition

Unique EffectBrain Age index = ΔR2Brain Age index = R2chronological age, Brain Age index, Brain Cognition – R2 chronological age, Brain Cognition

Unique EffectBrain Cognition = ΔR2Brain Cognition = R2chronological age, Brain Age index, Brain Cognition – R2 chronological age, Brain Age Index

Common Effectchronological age, Brain Age index = R2chronological age, Brain Cognition + R2 Brain Age index, Brain Cognition – R2 Brain Cognition – R2chronological age, Brain Age index, Brain Cognition

Common Effectchronological age, Brain Cognition = R2chronological age, Brain Age Index + R2 Brain Age index, Brain Cognition – R2 Brain Age Index – R2chronological age, Brain Age index, Brain Cognition

Common Effect Brain Age index, Brain Cognition = R2chronological age, Brain Age Index + R2 chronological age, Brain Cognition – R2 chronological age – R2chronological age, Brain Age index, Brain Cognition

Common Effect chronological age, Brain Age index, Brain Cognition = R2 chronological age + R2 Brain Age Index + R2 Brain Cognition – R2chronological age, Brain Age Index – R2 chronological age, Brain Cognition – R2 Brain Age Index, Brain Cognition – R2chronological age, Brain Age index, Brain Cognition , (13)”

(2) I agree that the solution is not to exclude age as a covariate, and that there is a big difference between inevitable and obvious. I simply think a further discussion of the inevitability of the results would be clarifying for the readers. There is a big opportunity in the brain-age literature to be as direct as possible about why you are finding what you are finding. People need to know not only what you found, but why you found what you found.

Thank you. We agreed that we need to make this point more explicit and direct. In the revised manuscript, we had the statements in both Introduction and Discussion (see below) about the tight relationship between Brain Age and chronological age by design, making the small unique effects of Brain Age inevitable.

Introduction:

“Accordingly, by design, Brain Age is tightly close to chronological age. Because chronological age usually has a strong relationship with fluid cognition, to begin with, it is unclear how much Brain Age adds to what is already captured by chronological age.“

Discussion:

“First, Brain Age itself did not add much more information to help us capture fluid cognition than what we had already known from a person’s chronological age. This can clearly be seen from the small unique effects of Brain Age indices in the multiple regression models having Brain Age and chronological age as the regressors. While the unique effects of some Brain Age indices from certain age-prediction models were statistically significant, there were all relatively small. Without Brain Age indices, chronological age by itself already explained around 32% of the variation in fluid cognition. Including Brain Age indices only added around 1.6% at best. We believe the small unique effects of Brain Age were inevitable because, by design, Brain Age is tightly close to chronological age. Therefore, chronological age and Brain Age captured mostly a similar variation in fluid cognition.

Investigating the simple regression models and the commonality analysis between each Brain Age index and chronological age provided additional insights….”

(3) I believe it is very important to critically examine the use of brain-age and related metrics. As part of this process, I think we should be asking ourselves the following questions (among others): Why go through age prediction? Wouldn't the predictions of cognition (or another variable) using the same set of brain features always be as good or better? You still have not justified the use of brain-age. As I said before, if you are going to continue to recommend the use of brain-age, you need a very strong argument for why you are recommending this. What does it truly add? Otherwise, temper your statements to indicate possible better paths forward.

Thank you Reviewer 3 for making an argument against the use of Brain Age. We largely agree with you. However, our work only focuses on one phenotype, fluid cognition, and on the normative situation (i.e., not having a case vs control group). As Reviewer 2 pointed out, Brain Age might still have utility in other cases, not studied here. Still, future studies that focus on other phenotypes may consider using our approach as a template to test the utility of Brain Age in other situations. We added the conclusion statement to reflect this.

From Discussion:

“Altogether, we examined the utility of Brain Age as a biomarker for fluid cognition. Here are the three conclusions. First, Brain Age failed to add substantially more information over and above chronological age. Second, a higher ability to predict chronological age did not correspond to a higher utility to capture fluid cognition. Third, Brain Age missed up to around one-third of the variation in fluid cognition that could have been explained by brain MRI. Yet, given our focus on fluid cognition, future empirical research is needed to test the utility of Brain Age on other phenotypes, especially when Brain Age is used for anomaly detection in case-control studies (e.g., Bashyam et al., 2020; Rokicki et al., 2021). We hope that future studies may consider applying our approach (i.e., using the commonality analysis that includes predicted values from a model that directly predicts the phenotype of interest) to test the utility of Brain Age as a biomarker for other phenotypes.”

References

Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I. M., Truelove-Hill, M., Srinivasan, D., Mamourian, L., Pomponio, R., Fan, Y., Launer, L. J., Masters, C. L., Maruff, P., Zhuo, C., Völzke, H., Johnson, S. C., Fripp, J., Koutsouleris, N., Satterthwaite, T. D., … on behalf of the ISTAGING Consortium, the P. A. disease C., ADNI, and CARDIA studies. (2020). MRI signatures of brain age and disease over the lifespan based on a deep brain network and 14 468 individuals worldwide. Brain, 143(7), 2312–2324. https://doi.org/10.1093/brain/awaa160

Butler, E. R., Chen, A., Ramadan, R., Le, T. T., Ruparel, K., Moore, T. M., Satterthwaite, T. D., Zhang, F., Shou, H., Gur, R. C., Nichols, T. E., & Shinohara, R. T. (2021). Pitfalls in brain age analyses. Human Brain Mapping, 42(13), 4092–4101. https://doi.org/10.1002/hbm.25533

Cole, J. H. (2020). Multimodality neuroimaging brain-age in UK biobank: Relationship to biomedical, lifestyle, and cognitive factors. Neurobiology of Aging, 92, 34–42. https://doi.org/10.1016/j.neurobiolaging.2020.03.014

Dubois, J., Galdi, P., Paul, L. K., & Adolphs, R. (2018). A distributed brain network predicts general intelligence from resting-state human neuroimaging data. Philosophical Transactions of the Royal Society B: Biological Sciences, 373(1756), 20170284. https://doi.org/10.1098/rstb.2017.0284

Hahn, T., Fisch, L., Ernsting, J., Winter, N. R., Leenings, R., Sarink, K., Emden, D., Kircher, T., Berger, K., & Dannlowski, U. (2021). From ‘loose fitting’ to high-performance, uncertainty-aware brain-age modelling. Brain, 144(3), e31–e31. https://doi.org/10.1093/brain/awaa454

Insel, T., Cuthbert, B., Garvey, M., Heinssen, R., Pine, D. S., Quinn, K., Sanislow, C., & Wang, P. (2010). Research Domain Criteria (RDoC): Toward a New Classification Framework for Research on Mental Disorders. American Journal of Psychiatry, 167(7), 748–751. https://doi.org/10.1176/appi.ajp.2010.09091379

Jirsaraie, R. J., Kaufmann, T., Bashyam, V., Erus, G., Luby, J. L., Westlye, L. T., Davatzikos, C., Barch, D. M., & Sotiras, A. (2023). Benchmarking the generalizability of brain age models: Challenges posed by scanner variance and prediction bias. Human Brain Mapping, 44(3), 1118–1128. https://doi.org/10.1002/hbm.26144

Marquand, A. F., Rezek, I., Buitelaar, J., & Beckmann, C. F. (2016). Understanding Heterogeneity in Clinical Cohorts Using Normative Models: Beyond Case-Control Studies. Biological Psychiatry, 80(7), 552–561. https://doi.org/10.1016/j.biopsych.2015.12.023

Nimon, K., Lewis, M., Kane, R., & Haynes, R. M. (2008). An R package to compute commonality coefficients in the multiple regression case: An introduction to the package and a practical example. Behavior Research Methods, 40(2), 457–466. https://doi.org/10.3758/BRM.40.2.457

Pat, N., Wang, Y., Anney, R., Riglin, L., Thapar, A., & Stringaris, A. (2022). Longitudinally stable, brain‐based predictive models mediate the relationships between childhood cognition and socio‐demographic, psychological and genetic factors. Human Brain Mapping, hbm.26027. https://doi.org/10.1002/hbm.26027

Poldrack, R. A., Huckins, G., & Varoquaux, G. (2020). Establishment of Best Practices for Evidence for Prediction: A Review. JAMA Psychiatry, 77(5), 534–540. https://doi.org/10.1001/jamapsychiatry.2019.3671

Rasero, J., Sentis, A. I., Yeh, F.-C., & Verstynen, T. (2021). Integrating across neuroimaging modalities boosts prediction accuracy of cognitive ability. PLOS Computational Biology, 17(3), e1008347. https://doi.org/10.1371/journal.pcbi.1008347

Rokicki, J., Wolfers, T., Nordhøy, W., Tesli, N., Quintana, D. S., Alnæs, D., Richard, G., de Lange, A.-M. G., Lund, M. J., Norbom, L., Agartz, I., Melle, I., Nærland, T., Selbæk, G., Persson, K., Nordvik, J. E., Schwarz, E., Andreassen, O. A., Kaufmann, T., & Westlye, L. T. (2021). Multimodal imaging improves brain age prediction and reveals distinct abnormalities in patients with psychiatric and neurological disorders. Human Brain Mapping, 42(6), 1714–1726. https://doi.org/10.1002/hbm.25323

Sripada, C., Angstadt, M., Rutherford, S., Taxali, A., & Shedden, K. (2020). Toward a “treadmill test” for cognition: Improved prediction of general cognitive ability from the task activated brain. Human Brain Mapping, 41(12), 3186–3197. https://doi.org/10.1002/hbm.25007

Tetereva, A., Li, J., Deng, J. D., Stringaris, A., & Pat, N. (2022). Capturing brain‐cognition relationship: Integrating task‐based fMRI across tasks markedly boosts prediction and test‐retest reliability. NeuroImage, 263, 119588. https://doi.org/10.1016/j.neuroimage.2022.119588

Vieira, B. H., Pamplona, G. S. P., Fachinello, K., Silva, A. K., Foss, M. P., & Salmon, C. E. G. (2022). On the prediction of human intelligence from neuroimaging: A systematic review of methods and reporting. Intelligence, 93, 101654. https://doi.org/10.1016/j.intell.2022.101654

Author response:

The following is the authors’ response to the previous reviews.

eLife assessment

This study presents valuable data on the antigenic properties of neuraminidase proteins of human A/H3N2 influenza viruses sampled between 2009 and 2017. The antigenic properties are found to be generally concordant with genetic groups. Additional analysis have strengthened the revised manuscript, and the evidence supporting the claims is solid.

Public Reviews:

Reviewer #1 (Public Review):

Summary

The authors investigated the antigenic diversity of recent (2009-2017) A/H3N2 influenza neuraminidases (NAs), the second major antigenic protein after haemagglutinin. They used 27 viruses and 43 ferret sera and performed NA inhibition. This work was supported by a subset of mouse sera. Clustering analysis determined 4 antigenic clusters, mostly in concordance with the genetic groupings. Association analysis was used to estimate important amino acid positions, which were shown to be more likely close to the catalytic site. Antigenic distances were calculated and a random forest model used to determine potential important sites.

This revision has addressed many of my concerns of inconsistencies in the methods, results and presentation. There are still some remaining weaknesses in the computational work.

Strengths

(1) The data cover recent NA evolution and a substantial number (43) of ferret (and mouse) sera were generated and titrated against 27 viruses. This is laborious experimental work and is the largest publicly available neuraminidase inhibition dataset that I am aware of. As such, it will prove a useful resource for the influenza community.

(2) A variety of computational methods were used to analyse the data, which give a rounded picture of the antigenic and genetic relationships and link between sequence, structure and phenotype.

(3) Issues raised in the previous review have been thoroughly addressed.

Weaknesses

(1). Some inconsistencies and missing data in experimental methods Two ferret sera were boosted with H1N2, while recombinant NA protein for the others. This, and the underlying reason, are clearly explained in the manuscript. The authors note that boosting with live virus did not increase titres. Additionally, one homologous serum (A/Kansas/14/2017) was not generated, although this would not necessarily have impacted the results.

We agree with the reviewer and this point was addressed in the previous rebuttal.

(2) Inconsistency in experimental results

Clustering of the NA inhibition results identifies three viruses which do not cluster with their phylogenetic group. Again this is clearly pointed out in the paper and is consistent with the two replicate ferret sera. Additionally, A/Kansas/14/2017 is in a different cluster based on the antigenic cartography vs the clustering of the titres

We agree with the reviewer and this point was addressed in the previous rebuttal.

(3) Antigenic cartography plot would benefit from documentation of the parameters and supporting analyses

a. The number of optimisations used

We used 500 optimizations. This information is now included in the Methods section.

b. The final stress and the difference between the stress of the lowest few (e.g. 5) optimisations, or alternatively a graph of the stress of all the optimisations. Information on the stress per titre and per point, and whether any of these were outliers

The stress was obtained from 1, 5, 500, or even 5000 optimizations (resulting in stress values of respectively, 1366.47, 1366.47, 2908.60, and 3031.41). Besides limited variation or non-conversion of the stress values after optimization, the obtained maps were consistent in multiple runs. The map was obtained keeping the best optimization (stress value 1366.47, selected using the keepBestOptimization() function).

Author response image 1.

The stress per point is presented in the heat map below.

The heat map indicates stress per serum (x-axis) and strain (y-axis) in blue to red scale.

c. A measure of uncertainty in position (e.g. from bootstrapping)

Bootstrap was performed using 1000 repeats and 100 optimizations per repeat. The uncertainty is represented in the blob plot below.

Author response image 2.

(4) Random forest

The full dataset was used for the random forest model, including tuning the hyperparameters. It is more robust to have a training and test set to be able to evaluate overfitting (there are 25 features to classify 43 sera).

Explicit cross validation is not necessary for random forests as the out of bag process with multiple trees implicitly covers cross validation. In the random forest function in R this is done by setting the mtry argument (number of variables randomly sampled as candidates at each split). R samples variables with replacement (the same variable can be sampled multiple times) of the candidates from the training set. RF will then automatically take the data that is not selected as candidates as test set. Overfit may happen when all data is used for training but the RF method implicitly does use a test set and does not use all data for training.

Code:

rf <- randomForest(X,y=Y,ntree=1500,mtry=25,keep.forest=TRUE,importance=TRUE)

Reviewer #2 (Public Review):

Summary:

The authors characterized the antigenicity of N2 protein of 43 selected A(H3N2) influenza A viruses isolated from 2009-2017 using ferret and mice immune sera. Four antigenic groups were identified, which the authors claimed to be correlated with their respective phylogenic/ genetic groups. Among 102 amino acids differed by the 44 selected N2 proteins, the authors identified residues that differentiate the antigenicity of the four groups and constructed a machine-learning model that provides antigenic distance estimation. Three recent A(H3N2) vaccine strains were tested in the model but there was no experimental data to confirm the model prediction results.

Strengths:

This study used N2 protein of 44 selected A(H3N2) influenza A viruses isolated from 2009-2017 and generated corresponding panels of ferret and mouse sera to react with the selected strains. The amount of experimental data for N2 antigenicity characterization is large enough for model building.

Weaknesses:

The main weakness is that the strategy of selecting 43 A(H3N2) viruses from 2009-2017 was not explained. It is not clear if they represent the overall genetic diversity of human A(H3N2) viruses circulating during this time. In response to the reviewer's comment, the authors have provided a N2 phylogenetic tree using180 randomly selected N2 sequences from human A(H3N2) viruses from 2009-2017. While the 43 strains seems to scatter across the N2 tree, the four antigenic groups described by the author did not correlated with their respective phylogenic/ genetic groups as shown in Fig. 2. The authors should show the N2 phylogenic tree together with Fig. 2 and discuss the discrepancy observed.

The discrepancies between the provided N2 phylogenetic tree using 180 selected N2 sequences was primarily due to visualization. In the tree presented in Figure 2 the phylogeny was ordered according to branch length in a decreasing way. Further, the tree represented in the rebuttal was built with PhyML 3.0 using JTT substitution model, while the tree in figure 2 was build in CLC Workbench 21.0.5 using Bishop-Friday substitution model. The tree below was built using the same methodology as Figure 2, including branch size ordering. No discrepancies are observed.

Phylogenetic tree representing relatedness of N2 head domain. N2 NA sequences were ordered according to the branch length and phylogenetic clusters are colored as follows: G1: orange, G2: green, G3: blue, and G4: purple. NA sequences that were retained in the breadth panel are named according to the corresponding H3N2 influenza viruses. The other NA sequences are coded.

Author response image 3.

The second weakness is the use of double-immune ferret sera (post-infection plus immunization with recombinant NA protein) or mouse sera (immunized twice with recombinant NA protein) to characterize the antigenicity of the selected A(H3N2) viruses. Conventionally, NA antigenicity is characterized using ferret sera after a single infection. Repeated influenza exposure in ferrets has been shown to enhance antibody binding affinity and may affect the cross-reactivity to heterologous strains (PMID: 29672713). The increased cross-reactivity is supported by the NAI titers shown in Table S3, as many of the double immune ferret sera showed the highest reactivity not against its own homologous virus but to heterologous strains. In response to the reviewer's comment, the authors agreed the use of double-immune ferret sera may be a limitation of the study. It would be helpful if the authors can discuss the potential effect on the use of double-immune ferret sera in antigenicity characterization in the manuscript.

Our study was designed to understand the breadth of the anti-NA response after the incorporation of NA as a vaccine antigens. Our data does not allow to conclude whether increased breadth of protection is merely due to increased antibody titers or whether an NA boost immunization was able to induce antibody responses against epitopes that were not previously recognized by primary response to infection. However, we now mention this possibility in the discussion and cite Kosikova et al. CID 2018, in this context.

Another weakness is that the authors used the newly constructed a model to predict antigenic distance of three recent A(H3N2) viruses but there is no experimental data to validate their prediction (eg. if these viruses are indeed antigenically deviating from group 2 strains as concluded by the authors). In response to the comment, the authors have taken two strains out of the dataset and use them for validation. The results is shown as Fig. R7. However, it may be useful to include this in the main manuscript to support the validity of the model.

The removal of 2 strains was performed to illustrate the predictive performance of the RF modeling. However, Random Forest does not require cross-validation. The reason is that RF modeling already uses an out-of-bag evaluation which, in short, consists of using only a fraction of the data for the creation of the decision trees (2/3 of the data), obviating the need for a set aside the test set:

“…In each bootstrap training set, about one-third of the instances are left out. Therefore, the out-of-bag estimates are based on combining only about one- third as many classifiers as in the ongoing main combination. Since the error rate decreases as the number of combinations increases, the out-of-bag estimates will tend to overestimate the current error rate. To get unbiased out-of-bag estimates, it is necessary to run past the point where the test set error converges. But unlike cross-validation, where bias is present but its extent unknown, the out-of-bag estimates are unbiased…” from https://www.stat.berkeley.edu/%7Ebreiman/randomforest2001.pdf

Reviewer #3 (Public Review):

Summary:

This paper by Portela Catani et al examines the antigenic relationships (measured using monotypic ferret and mouse sera) across a panel of N2 genes from the past 14 years, along with the underlying sequence differences and phylogenetic relationships. This is a highly significant topic given the recent increased appreciation of the importance of NA as a vaccine target, and the relative lack of information about NA antigenic evolution compared with what is known about HA. Thus, these data will be of interest to those studying the antigenic evolution of influenza viruses. The methods used are generally quite sound, though there are a few addressable concerns that limit the confidence with which conclusions can be drawn from the data/analyses.

Strengths:

• The significance of the work, and the (general) soundness of the methods. -Explicit comparison of results obtained with mouse and ferret sera

Weaknesses:

• Approach for assessing influence of individual polymorphisms on antigenicity does not account for potential effects of epistasis (this point is acknowledged by the authors).

We agree with the reviewer and this point was addressed in the previous rebuttal.

• Machine learning analyses neither experimentally validated nor shown to be better than simple, phylogenetic-based inference.

We respectfully disagree with the reviewer. This point was addressed in the previous rebuttal as follows.

This is a valid remark and indeed we have found a clear correlation between NAI cross reactivity and phylogenetic relatedness. However, besides achieving good prediction of the experimental data (as shown in Figure 5 and in FigureR7), machine Learning analysis has the potential to rank or indicate major antigenic divergences based on available sequences before it has consolidated as new clade. ML can also support the selection and design of broader reactive antigens. “

Recommendations for the authors:

Reviewer #2 (Recommendations For The Authors):

(1) Discuss the discrepancy between Fig. 2 and the newly constructed N2 phylogenetic tree with 180 randomly selected N2 sequences of A(H3N2) viruses from 2009-2017. Specifically please explain the antigenic vs. phylogenetic relationship observed in Fig. 2 was not observed in the large N2 phylogenetic tree.

Discrepancies were due to different method and visualization. A new tree was provided.

(2) Include a sentence to discuss the potential effect on the use of double-immune ferret sera in antigenic characterization.

We prefer not to speculate on this.

(3) Include the results of the exercise run (with the use of Swe17 and HK17) in the manuscript as a way to validate the model.

The exercise was performed to illustrate predictive potential of the RF modeling to the reviewer. However, cross-validation is not a usual requirement for random forest, since it uses out-of-bag calculations. We prefer to not include the exercise runs within the main manuscript.

Author response:

The following is the authors’ response to the previous reviews

Reviewer #1:

Comment:

The authors quantified information in gesture and speech, and investigated the neural processing of speech and gestures in pMTG and LIFG, depending on their informational content, in 8 different time-windows, and using three different methods (EEG, HD-tDCS and TMS). They found that there is a time-sensitive and staged progression of neural engagement that is correlated with the informational content of the signal (speech/gesture).

Strengths:

A strength of the paper is that the authors attempted to combine three different methods to investigate speech-gesture processing.

We sincerely appreciate the reviewer’s recognition of our efforts in employing a multi-method approach, which integrates three complementary experimental paradigms, each leveraging distinct neurophysiological techniques to provide converging evidence.

In Experiment 1, we found that the degree of inhibition in the pMTG and LIFG was strongly associated with the overlap in gesture-speech representations, as quantified by mutual information. Experiment 2 revealed the time-sensitive dynamics of the pMTG-LIFG circuit in processing both unisensory (gesture or speech) and multisensory information. Experiment 3, utilizing high-temporal-resolution EEG, independently replicated the temporal dynamics of gesture-speech integration observed in Experiment 2, further validating our findings.

The striking convergence across these methodologically independent approaches significantly bolsters the robustness and generalizability of our conclusions regarding the neural mechanisms underlying multisensory integration.

Comment 1: I thank the authors for their careful responses to my comments. However, I remain not convinced by their argumentation regarding the specificity of their spatial targeting and the time-windows that they used.

The authors write that since they included a sham TMS condition, that the TMS selectively disrupted the IFG-pMTG interaction during specific time windows of the task related to gesture-speech semantic congruency. This to me does not show anything about the specificity of the time-windows itself, nor the selectivity of targeting in the TMS condition.

(1) Selection of brain regions (IFG/pMTG)

We thank the reviewer for their thoughtful consideration. The choice of the left IFG and pMTG as regions of interest (ROIs) was informed by a meta-analysis of fMRI studies on gesture-speech integration, which consistently identified these regions as critical hubs (see Author response table 1 for detailed studies and coordinates).

Author response table 1.

Meta-analysis of previous studies on gesture-speech integration.

Based on the meta-analysis of previous studies, we selected the IFG and pMTG as ROIs for gesture-speech integration. The rationale for selecting these brain regions is outlined in the introduction in Lines 63-66: “Empirical studies have investigated the semantic integration between gesture and speech by manipulating their semantic relationship[15-18] and revealed a mutual interaction between them19-21 as reflected by the N400 latency and amplitude14 as well as common neural underpinnings in the left inferior frontal gyrus (IFG) and posterior middle temporal gyrus (pMTG)[15,22,23].”

And further described in Lines 77-78: “Experiment 1 employed high-definition transcranial direct current stimulation (HD-tDCS) to administer Anodal, Cathodal and Sham stimulation to either the IFG or the pMTG”. And Lines 85-88: ‘Given the differential involvement of the IFG and pMTG in gesture-speech integration, shaped by top-down gesture predictions and bottom-up speech processing [23], Experiment 2 was designed to assess whether the activity of these regions was associated with relevant informational matrices”.

In the Methods section, we clarified the selection of coordinates in Lines 194-200: “Building on a meta-analysis of prior fMRI studies examining gesture-speech integration[22], we targeted Montreal Neurological Institute (MNI) coordinates for the left IFG at (-62, 16, 22) and the pMTG at (-50, -56, 10). In the stimulation protocol for HD-tDCS, the IFG was targeted using electrode F7 as the optimal cortical projection site[36], with four return electrodes placed at AF7, FC5, F9, and FT9. For the pMTG, TP7 was selected as the cortical projection site[36], with return electrodes positioned at C5, P5, T9, and P9.”

The selection of IFG or pMTG as integration hubs for gesture and speech has also been validated in our previous studies. Specifically, Zhao et al. (2018, J. Neurosci) applied TMS to both areas. Results demonstrated that disrupting neural activity in the IFG or pMTG via TMS selectively impaired the semantic congruency effect (reaction time costs due to semantic incongruence), while leaving the gender congruency effect unaffected.

These findings identified the IFG and pMTG as crucial hubs for gesture-speech integration, guiding the selection of brain regions for our subsequent studies.

(2) Selection of time windows

The five key time windows (TWs) analyzed in this study were derived from our previous TMS work (Zhao et al., 2021, J. Neurosci), where we segmented the gesture-speech integration period (0–320 ms post-speech onset) into eight 40-ms windows. This interval aligns with established literature on gesture-speech integration, particularly the 200–300 ms window noted by the reviewer. As detailed in Lines (776-779): “Procedure of Experiment 2. Eight time windows (TWs, duration = 40 ms) were segmented in relative to the speech IP. Among the eight TWs, five (TW1, TW2, TW3, TW6, and TW7) were chosen based on the significant results in our prior study[23]. Double-pulse TMS was delivered over each of the TW of either the pMTG or the IFG”.

In our prior work (Zhao et al., 2021, J. Neurosci), we employed a carefully controlled experimental design incorporating two key factors: (1) gesture-speech semantic congruency (serving as our primary measure of integration) and (2) gesture-speech gender congruency (implemented as a matched control factor). Using a time-locked, double-pulse TMS protocol, we systematically targeted each of the eight predefined time windows (TWs) within the left IFG, left pMTG, or vertex (serving as a sham control condition). Our results demonstrated that a TW-selective disruption of gesture-speech integration, indexed by the semantic congruency effect (i.e., a cost of reaction time because of semantic conflict), when stimulating the left pMTG in TW1, TW2, and TW7 but when stimulating the left IFG in TW3 and TW6. Crucially, no significant effects were observed during either sham stimulation or the controlled gender congruency factor (Figure 3 from Zhao et al., 2021, J. Neurosci).

This triple dissociation - showing effects only for semantic integration, only in active stimulation, and only at specific time points - provides compelling causal evidence that IFG-pMTG connectivity plays a temporally precise role in gesture-speech integration.

Noted that this work has undergone rigorous peer review by two independent experts who both endorsed our methodological approach. Their original evaluations, provided below:

Reviewer 1: “significance: Using chronometric TMS-stimulation the data of this experiment suggests a feedforward information flow from left pMTG to left IFG followed by an information flow from left IFG back to the left pMTG.  The study is the first to provide causal evidence for the temporal dynamics of the left pMTG and left IFG found during gesture-speech integration.”

Reviewer 2: “Beyond the new results the manuscript provides regarding the chronometrical interaction of the left inferior frontal gyrus and middle temporal gyrus in gesture-speech interaction, the study more basically shows the possibility of unfolding temporal stages of cognitive processing within domain-specific cortical networks using short-time interval double-pulse TMS. Although this method also has its limitations, a careful study planning as shown here and an appropiate discussion of the results can provide unique insights into cognitive processing.”

References:

Willems, R.M., Ozyurek, A., and Hagoort, P. (2009). Differential roles for left inferior frontal and superior temporal cortex in multimodal integration of action and language. Neuroimage 47, 1992-2004. 10.1016/j.neuroimage.2009.05.066.

Drijvers, L., Jensen, O., and Spaak, E. (2021). Rapid invisible frequency tagging reveals nonlinear integration of auditory and visual information. Human Brain Mapping 42, 1138-1152. 10.1002/hbm.25282.

Drijvers, L., and Ozyurek, A. (2018). Native language status of the listener modulates the neural integration of speech and iconic gestures in clear and adverse listening conditions. Brain and Language 177, 7-17. 10.1016/j.bandl.2018.01.003.

Drijvers, L., van der Plas, M., Ozyurek, A., and Jensen, O. (2019). Native and non-native listeners show similar yet distinct oscillatory dynamics when using gestures to access speech in noise. Neuroimage 194, 55-67. 10.1016/j.neuroimage.2019.03.032.

Holle, H., and Gunter, T.C. (2007). The role of iconic gestures in speech disambiguation: ERP evidence. J Cognitive Neurosci 19, 1175-1192. 10.1162/jocn.2007.19.7.1175.

Kita, S., and Ozyurek, A. (2003). What does cross-linguistic variation in semantic coordination of speech and gesture reveal?: Evidence for an interface representation of spatial thinking and speaking. J Mem Lang 48, 16-32. 10.1016/S0749-596x(02)00505-3.

Bernardis, P., and Gentilucci, M. (2006). Speech and gesture share the same communication system. Neuropsychologia 44, 178-190. 10.1016/j.neuropsychologia.2005.05.007.

Zhao, W.Y., Riggs, K., Schindler, I., and Holle, H. (2018). Transcranial magnetic stimulation over left inferior frontal and posterior temporal cortex disrupts gesture-speech integration. Journal of Neuroscience 38, 1891-1900. 10.1523/Jneurosci.1748-17.2017.

Zhao, W., Li, Y., and Du, Y. (2021). TMS reveals dynamic interaction between inferior frontal gyrus and posterior middle temporal gyrus in gesture-speech semantic integration. The Journal of Neuroscience, 10356-10364. 10.1523/jneurosci.1355-21.2021.

Hartwigsen, G., Bzdok, D., Klein, M., Wawrzyniak, M., Stockert, A., Wrede, K., Classen, J., and Saur, D. (2017). Rapid short-term reorganization in the language network. Elife 6. 10.7554/eLife.25964.

Jackson, R.L., Hoffman, P., Pobric, G., and Ralph, M.A.L. (2016). The semantic network at work and rest: Differential connectivity of anterior temporal lobe subregions. Journal of Neuroscience 36, 1490-1501. 10.1523/JNEUROSCI.2999-15.2016.

Humphreys, G. F., Lambon Ralph, M. A., & Simons, J. S. (2021). A Unifying Account of Angular Gyrus Contributions to Episodic and Semantic Cognition. Trends in neurosciences, 44(6), 452–463. https://doi.org/10.1016/j.tins.2021.01.006

Bonner, M. F., & Price, A. R. (2013). Where is the anterior temporal lobe and what does it do?. The Journal of neuroscience : the official journal of the Society for Neuroscience, 33(10), 4213–4215. https://doi.org/10.1523/JNEUROSCI.0041-13.2013

Comment 2: It could still equally well be the case that other regions or networks relevant for gesture-speech integration are targeted, and it can still be the case that these timewindows are not specific, and effects bleed into other time periods. There seems to be no experimental evidence here that this is not the case.

The selection of IFG and pMTG as regions of interest was rigorously justified through multiple lines of evidence. First, a comprehensive meta-analysis of fMRI studies on gesture-speech integration consistently identified these regions as central nodes (see response to comment 1). Second, our own previous work (Zhao et al., 2018, JN; 2021, JN) provided direct empirical validation of their involvement. Third, by employing the same experimental paradigm, we minimized the likelihood of engaging alternative networks. Fourth, even if other regions connected to IFG or pMTG might be affected by TMS, the distinct engagement of specific time windows of IFG and pMTG minimizes the likelihood of consistent influence from other regions.

Regarding temporal specificity, our 2021 study (Zhao et al., 2021, JN, see details in response to comment 1) systematically examined the entire 0-320ms integration window and found that only select time windows showed significant effects for gesture-speech semantic congruency, while remaining unaffected during gender congruency processing. This double dissociation (significant effects for semantic integration but not gender processing in specific windows) rules out broad temporal spillover.

Comment 3: To be more specific, the authors write that double-pulse TMS has been widely used in previous studies (as found in their table). However, the studies cited in the table do not necessarily demonstrate the level of spatial and temporal specificity required to disentangle the contributions of tightly-coupled brain regions like the IFG and pMTG during the speech-gesture integration process. pMTG and IFG are located in very close proximity, and are known to be functionally and structurally interconnected, something that is not necessarily the case for the relatively large and/or anatomically distinct areas that the authors mention in their table.

Our methodological approach is strongly supported by an established body of research employing double-pulse TMS (dpTMS) to investigate neural dynamics across both primary motor and higher-order cognitive regions. As documented in Author response table 1, multiple studies have successfully applied this technique to: (1) primary motor areas (tongue and lip representations in M1), and (2) semantic processing regions (including pMTG, PFC, and ATL). Particularly relevant precedents include:

(1) Teige et al. (2018, Cortex): Demonstrated precise spatial and temporal specificity by applying 40ms-interval dpTMS to ATL, pMTG, and mid-MTG across multiple time windows (0-40ms, 125-165ms, 250-290ms, 450-490ms), revealing distinct functional contributions from ATL versus pMTG.

(2) Vernet et al. (2015, Cortex): Successfully dissociated functional contributions of right IPS and DLPFC using 40ms-interval dpTMS, despite their anatomical proximity and functional connectivity.

These studies confirm double-pulse TMS can discriminate interconnected nodes at short timescales. Our 2021 study further validated this for IFG-pMTG.

Author response table 2.

Double-pulse TMS studies on brain regions over 3-60 ms time interval

References:

Teige, C., Mollo, G., Millman, R., Savill, N., Smallwood, J., Cornelissen, P. L., & Jefferies, E. (2018). Dynamic semantic cognition: Characterising coherent and controlled conceptual retrieval through time using magnetoencephalography and chronometric transcranial magnetic stimulation. Cortex, 103, 329-349.

Vernet, M., Brem, A. K., Farzan, F., & Pascual-Leone, A. (2015). Synchronous and opposite roles of the parietal and prefrontal cortices in bistable perception: a double-coil TMS–EEG study. Cortex, 64, 78-88.

Comment 4: But also more in general: The mere fact that these methods have been used in other contexts does not necessarily mean they are appropriate or sufficient for investigating the current research question. Likewise, the cognitive processes involved in these studies are quite different from the complex, multimodal integration of gesture and speech. The authors have not provided a strong theoretical justification for why the temporal dynamics observed in these previous studies should generalize to the specific mechanisms of gesture-speech integration..

The neurophysiological mechanisms underlying double-pulse TMS (dpTMS) are well-characterized. While it is established that single-pulse TMS can produce brief artifacts (typically within 0–10 ms) due to transient cortical depolarization (Romero et al., 2019, NC), the dynamics of double-pulse TMS (dpTMS) involve more intricate inhibitory interactions. Specifically, the first pulse increases membrane conductance via GABAergic shunting inhibition, effectively lowering membrane resistance and attenuating the excitatory impact of the second pulse. This results in a measurable reduction in cortical excitability at the paired-pulse interval, as evidenced by suppressed motor evoked potentials (MEPs) (Paulus & Rothwell, 2016, J Physiol). Importantly, this neurophysiological mechanism is independent of cognitive domain and has been robustly demonstrated across multiple functional paradigms.

In our study, we did not rely on previously reported timing parameters but instead employed a dpTMS protocol using a 40-ms inter-pulse interval. Based on the inhibitory dynamics of this protocol, we designed a sliding temporal window sufficiently broad to encompass the integration period of interest. This approach enabled us to capture and localize the critical temporal window associated with ongoing integrative processing in the targeted brain region.

We acknowledge that the previous phrasing may have been ambiguous, a clearer and more detailed description of the dpTMS protocol has now been provided in Lines 88-92: “To this end, we employed chronometric double-pulse transcranial magnetic stimulation, which is known to transiently reduce cortical excitability at the inter-pulse interval]27]. Within a temporal period broad enough to capture the full duration of gesture–speech integration[28], we targeted specific timepoints previously implicated in integrative processing within IFG and pMTG [23].”

References:

Romero, M.C., Davare, M., Armendariz, M. et al. Neural effects of transcranial magnetic stimulation at the single-cell level. Nat Commun 10, 2642 (2019). https://doi.org/10.1038/s41467-019-10638-7

Paulus W, Rothwell JC. Membrane resistance and shunting inhibition: where biophysics meets state-dependent human neurophysiology. J Physiol. 2016 May 15;594(10):2719-28. doi: 10.1113/JP271452. PMID: 26940751; PMCID: PMC4865581.

Obermeier, C., & Gunter, T. C. (2015). Multisensory Integration: The Case of a Time Window of Gesture-Speech Integration. Journal of Cognitive Neuroscience, 27(2), 292-307. https://doi.org/10.1162/jocn_a_00688

Comment 5: Moreover, the studies cited in the table provided by the authors have used a wide range of interpulse intervals, from 20 ms to 100 ms, suggesting that the temporal precision required to capture the dynamics of gesture-speech integration (which is believed to occur within 200-300 ms; Obermeier & Gunter, 2015) may not even be achievable with their 40 ms time windows.

Double-pulse TMS has been empirically validated across neurocognitive studies as an effective method for establishing causal temporal relationships in cortical networks, with demonstrated sensitivity at timescales spanning 3-60 m. Our selection of a 40-ms interpulse interval represents an optimal compromise between temporal precision and physiological feasibility, as evidenced by its successful application in dissociating functional contributions of interconnected regions including ATL/pMTG (Teige et al., 2018) and IPS/DLPFC (Vernet et al., 2015). This methodological approach combines established experimental rigor with demonstrated empirical validity for investigating the precisely timed IFG-pMTG dynamics underlying gesture-speech integration, as shown in our current findings and prior work (Zhao et al., 2021).

Our experimental design comprehensively sampled the 0-320 ms post-stimulus period, fully encompassing the critical 200-300 ms window associated with gesture-speech integration, as raised by the reviewer. Notably, our results revealed temporally distinct causal dynamics within this period: the significantly reduced semantic congruency effect emerged at IFG at 200-240ms, followed by feedback projections from IFG to pMTG at 240-280ms. This precisely timed interaction provides direct neurophysiological evidence for the proposed architecture of gesture-speech integration, demonstrating how these interconnected regions sequentially contribute to multisensory semantic integration.

Comment 6: I do appreciate the extra analyses that the authors mention. However, my 5th comment is still unanswered: why not use entropy scores as a continous measure?

Analysis with MI and entropy as continuous variables were conducted employing Representational Similarity Analysis (RSA) (Popal et.al, 2019). This analysis aimed to build a model to predict neural responses based on these feature metrics.

To capture dynamic temporal features indicative of different stages of multisensory integration, we segmented the EEG data into overlapping time windows (40 ms in duration with a 10 ms step size). The 40 ms window was chosen based on the TMS protocol used in Experiment 2, which also employed a 40 ms time window. The 10 ms step size (equivalent to 5 time points) was used to detect subtle shifts in neural responses that might not be captured by larger time windows, allowing for a more granular analysis of the temporal dynamics of neural activity.

Following segmentation, the EEG data were reshaped into a four-dimensional matrix (42 channels × 20 time points × 97 time windows × 20 features). To construct a neural similarity matrix, we averaged the EEG data across time points within each channel and each time window. The resulting matrix was then processed using the pdist function to compute pairwise distances between adjacent data points. This allowed us to calculate correlations between the neural matrix and three feature similarity matrices, which were constructed in a similar manner. These three matrices corresponded to (1) gesture entropy, (2) speech entropy, and (3) mutual information (MI). This approach enabled us to quantify how well the neural responses corresponded to the semantic dimensions of gesture and speech stimuli at each time window.

To determine the significance of the correlations between neural activity and feature matrices, we conducted 1000 permutation tests. In this procedure, we randomized the data or feature matrices and recalculated the correlations repeatedly, generating a null distribution against which the observed correlation values were compared. Statistical significance was determined if the observed correlation exceeded the null distribution threshold (p < 0.05). This permutation approach helps mitigate the risk of spurious correlations, ensuring that the relationships between the neural data and feature matrices are both robust and meaningful.

Finally, significant correlations were subjected to clustering analysis, which grouped similar neural response patterns across time windows and channels. This clustering allowed us to identify temporal and spatial patterns in the neural data that consistently aligned with the semantic features of gesture and speech stimuli, thus revealing the dynamic integration of these multisensory modalities across time. Results are as follows:

(1)  Two significant clusters were identified for gesture entropy (Figure 1 left). The first cluster was observed between 60-110 ms (channels F1 and F3), with correlation coefficients (r) ranging from 0.207 to 0.236 (p < 0.001). The second cluster was found between 210-280 ms (channel O1), with r-values ranging from 0.244 to 0.313 (p < 0.001).

(2)  For speech entropy (Figure 1 middle), significant clusters were detected in both early and late time windows. In the early time windows, the largest significant cluster was found between 10-170 ms (channels F2, F4, F6, FC2, FC4, FC6, C4, C6, CP4, and CP6), with r-values ranging from 0.151 to 0.340 (p = 0.013), corresponding to the P1 component (0-100 ms). In the late time windows, the largest significant cluster was observed between 560-920 ms (across the whole brain, all channels), with r-values ranging from 0.152 to 0.619 (p = 0.013).

(3)  For mutual information (MI) (Figure 1 right), a significant cluster was found between 270-380 ms (channels FC1, FC2, FC3, FC5, C1, C2, C3, C5, CP1, CP2, CP3, CP5, FCz, Cz, and CPz), with r-values ranging from 0.198 to 0.372 (p = 0.001).

Author response image 1.

Results of RSA analysis.

These additional findings suggest that even using a different modeling approach, neural responses, as indexed by feature metrics of entropy and mutual information, are temporally aligned with distinct ERP components and ERP clusters, as reported in the current manuscript. This alignment serves to further consolidate the results, reinforcing the conclusion we draw. Considering the length of the manuscript, we did not include these results in the current manuscript.

Reference:

Popal, H., Wang, Y., & Olson, I. R. (2019). A guide to representational similarity analysis for social neuroscience. Social cognitive and affective neuroscience, 14(11), 1243-1253.

Comment 7: In light of these concerns, I do not believe the authors have adequately demonstrated the spatial and temporal specificity required to disentangle the contributions of the IFG and pMTG during the gesture-speech integration process. While the authors have made a sincere effort to address the concerns raised by the reviewers, and have done so with a lot of new analyses, I remain doubtful that the current methodological approach is sufficient to draw conclusions about the causal roles of the IFG and pMTG in gesture-speech integration.

To sum up:

(1) Empirical validation from our prior work (Zhao et al., 2018,2021,JN): The selection of IFG and pMTG as target regions was informed by both: (1) a comprehensive meta-analysis of fMRI studies on gesture-speech integration, and (2) our own prior causal evidence from Zhao et al. (2018, J Neurosci), with detailed stereotactic coordinates provided in the attached Response to Editors and Reviewers letter. The temporal parameters were similarly grounded in empirical data from Zhao et al. (2021, J Neurosci), where we systematically examined eight consecutive 40-ms windows spanning the full integration period (0-320 ms). This study revealed a triple dissociation of effects - occurring exclusively during: (i)semantic integration (but not control tasks), (ii) active stimulation (but not sham), and (iii) specific time windows (but not all time windows)- providing robust causal evidence for the spatiotemporal specificity of IFG-pMTG interactions in gesture-speech processing. Notably, all reviewers recognized the methodological strength of this dpTMS approach in their evaluations (see attached JN assessment for details).

(2) Convergent evidence from Experiment 3: Our study employed a multi-method approach incorporating three complementary experimental paradigms, each utilizing distinct neurophysiological techniques to provide converging evidence. Specifically, Experiment 3 implemented high-temporal-resolution EEG, which independently replicated the time-sensitive dynamics of gesture-speech integration observed in our double-pulse TMS experiments. The remarkable convergence between these methodologically independent approaches -demonstrating consistent temporal staging of IFG-pMTG interactions across both causal (TMS) and correlational (EEG) measures - significantly strengthens the validity and generalizability of our conclusions regarding the neural mechanisms underlying multisensory integration.

(3) Established precedents in double-pulse TMS literature: The double-pulse TMS methodology employed in our study is firmly grounded in established neuroscience research. As documented in our detailed Response to Editors and Reviewers letter (citing 11 representative studies), dpTMS has been extensively validated for investigating causal temporal dynamics in cortical networks, with demonstrated sensitivity at timescales ranging from 3-60 ms. Particularly relevant precedents include: 1. Teige et al. (2018, Cortex) successfully dissociated functional contributions of anatomically proximal regions (ATL vs. pMTG vs.mid-MTG) using 40-ms-interval double-pulse TMS; 2. Vernet et al. (2015, Cortex) effectively distinguished neural processing in interconnected frontoparietal regions (right IPS vs. DLPFC) using 40-ms double-pulse TMS parameters. Both parameters are identical to those employed in our current study.

(4) Neurophysiological Plausibility: The neurophysiological basis for the transient double-pulse TMS effects is well-established through mechanistic studies of TMS-induced cortical inhibition (Romero et al.,2019; Paulus & Rothwell, 2016).

Taking together, we respectfully submit that our methodology provides robust support for our conclusions.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1:

Summary:

The work by Combrisson and colleagues investigates the degree to which reward and punishment learning signals overlap in the human brain using intracranial EEG recordings. The authors used information theory approaches to show that local field potential signals in the anterior insula and the three sub regions of the prefrontal cortex encode both reward and punishment prediction errors, albeit to different degrees. Specifically, the authors found that all four regions have electrodes that can selectively encode either the reward or the punishment prediction errors. Additionally, the authors analyzed the neural dynamics across pairs of brain regions and found that the anterior insula to dorsolateral prefrontal cortex neural interactions were specific for punishment prediction errors whereas the ventromedial prefrontal cortex to lateral orbitofrontal cortex interactions were specific to reward prediction errors. This work contributes to the ongoing efforts in both systems neuroscience and learning theory by demonstrating how two differing behavioral signals can be differentiated to a greater extent by analyzing neural interactions between regions as opposed to studying neural signals within one region.

Strengths:

The experimental paradigm incorporates both a reward and punishment component that enables investigating both types of learning in the same group of subjects allowing direct comparisons.

The use of intracranial EEG signals provides much needed insight into the timing of when reward and punishment prediction errors signals emerge in the studied brain regions.

Information theory methods provide important insight into the interregional dynamics associated with reward and punishment learning and allows the authors to assess that reward versus punishment learning can be better dissociated based on interregional dynamics over local activity alone.

We thank the reviewer for this accurate summary. Please find below our answers to the weaknesses raised by the reviewer.

Weaknesses:

The analysis presented in the manuscript focuses solely on gamma band activity. The presence and potential relevance of other frequency bands is not discussed. It is possible that slow oscillations, which are thought to be important for coordinating neural activity across brain regions could provide additional insight.

We thank the reviewer for pointing us to this missing discussion in the first version of the manuscript. We now made this point clearer in the Methods sections entitled “iEEG data analysis” and “Estimate of single-trial gamma-band activity”:

“Here, we focused solely on broadband gamma for three main reasons. First, it has been shown that the gamma band activity correlates with both spiking activity and the BOLD fMRI signals (Lachaux et al., 2007; Mukamel et al., 2004; Niessing et al., 2005; Nir et al., 2007), and it is commonly used in MEG and iEEG studies to map task-related brain regions (Brovelli et al., 2005; Crone et al., 2006; Vidal et al., 2006; Ball et al., 2008; Jerbi et al., 2009; Darvas et al., 2010; Lachaux et al., 2012; Cheyne and Ferrari, 2013; Ko et al., 2013). Therefore, focusing on the gamma band facilitates linking our results with the fMRI and spiking literatures on probabilistic learning. Second, single-trial and time-resolved high-gamma activity can be exploited for the analysis of cortico-cortical interactions in humans using MEG and iEEG techniques (Brovelli et al., 2015; 2017; Combrisson et al., 2022). Finally, while previous analyses of the current dataset (Gueguen et al., 2021) reported an encoding of PE signals at different frequency bands, the power in lower frequency bands were shown to carry redundant information compared to the gamma band power.”

The data is averaged across all electrodes which could introduce biases if some subjects had many more electrodes than others. Controlling for this variation in electrode number across subjects would ensure that the results are not driven by a small subset of subjects with more electrodes.

We thank the reviewer for raising this important issue. We would like to point out that the gamma activity was not averaged across bipolar recordings within an area, nor measures of connectivity. Instead, we used a statistical approach proposed in a previous paper that combines non-parametric permutations with measures of information (Combrisson et al., 2022). As we explain in the “Statistical analysis” section, mutual information (MI) is estimated between PE signals and single-trial modulations in gamma activity separately for each contact (or for each pair of contacts). Then, a one-sample t-test is computed across all of the recordings of all subjects to form the effect size at the group-level. We will address the point of the electrode number in our answer below.

The potential variation in reward versus punishment learning across subjects is not included in the manuscript. While the time course of reward versus punishment prediction errors is symmetrical at the group level, it is possible that some subjects show faster learning for one versus the other type which can bias the group average. Subject level behavioral data along with subject level electrode numbers would provide more convincing evidence that the observed effects are not arising from these potential confounds.

We thank the reviewer for the two points raised. We performed additional analyses at the single-participant level to address the issues raised by the reviewer. We should note, however, that these results are descriptive and cannot be generalized to account for population-level effects. As suggested by the reviewer, we prepared two new figures. The first supplementary figure summarizes the number of participants that had iEEG contacts per brain region and pair of brain regions (Fig. S1A in the Appendix). It can be seen that the number of participants sampled in different brain regions is relatively constant (left panel) and the number of participants with pairs of contacts across brain regions is relatively homogeneous, ranging from 7 to 11 (right panel). Fig. S1B shows the number of bipolar derivations per subject and per brain region.

Author response image 1.

Single subject anatomical repartition. (A) Number of unique subject per brain region and per pair of brain regions (B) Number of bipolar derivations per subject and per brain region

The second supplementary figure describes the estimated prediction error for rewarding and punishing trials for each subject (Fig. S2). The single-subject error bars represent the 95th percentile confidence interval estimated using a bootstrap approach across the different pairs of stimuli presented during the three to six sessions. As the reviewer anticipated, there are indeed variations across subjects, but we observe that RPE and PPE are relatively symmetrical, even at the subject level, and tend toward zero around trial number 10. These results therefore corroborate the patterns observed at the group-level.

Author response image 2.

Single-subject estimation of predictions errors. Single-subject trial-wise reward PE (RPE - blue) and punishment PE (PPE - red), ± 95% confidence interval.

Finally, to assess the variability of local encoding of prediction errors across participants, we quantified the proportion of subjects having at least one significant bipolar derivation encoding either the RPE or PPE (Fig. S4). As expected, we found various proportions of unique subjects with significant R/PPE encoding per region. The lowest proportion was achieved in the ventromedial prefrontal cortex (vmPFC) and lateral orbitofrontal cortex (lOFC) for encoding PPE and RPE, respectively, with approximately 30% of the subjects having the effect. Conversely, we found highly reproducible encodings in the anterior insula (aINS) and dorsolateral prefrontal cortex (dlPFC) with a maximum of 100% of the 9 subjects having at least one bipolar derivation encoding PPE in the dlPFC.

Author response image 3.

Taken together, we acknowledge a certain variability per region and per condition. Nevertheless, the results presented in the supplementary figures suggest that the main results do not arise from a minority of subjects.

We would like to point out that in order to assess across-subject variability, a much larger number of participants would have been needed, given the low signal-to-noise ratios observed at the single-participant level. We thus prefer to add these results as supplementary material in the Appendix, rather than in the main text.

It is unclear if the findings in Figures 3 and 4 truly reflect the differential interregional dynamics in reward versus punishment learning or if these results arise as a statistical byproduct of the reward vs punishment bias observed within each region. For instance, the authors show that information transfer from anterior insula to dorsolateral prefrontal cortex is specific to punishment prediction error. However, both anterior insula and dorsolateral prefrontal cortex have higher prevalence of punishment prediction error selective electrodes to begin with. Therefore the findings in Fig 3 may simply be reflecting the prevalence of punishment specificity in these two regions above and beyond a punishment specific neural interaction between the two regions. Either mathematical or analytical evidence that assesses if the interaction effect is simply reflecting the local dynamics would be important to make this result convincing.

This is an important point that we partly addressed in the manuscript. More precisely, we investigated whether the synergistic effects observed between the dlPFC and vmPFC encoding global PEs (Fig. 5) could be explained by their respective local specificity. Indeed, since we reported larger proportions of recordings encoding the PPE in the dlPFC and the RPE in the vmPFC (Fig. 2B), we checked whether the synergy between dlPFC and vmPFC could be mainly due to complementary roles where the dlPFC brings information about the PPE only and the vmPFC brings information to the RPE only. To address this point, we selected PPE-specific bipolar derivations from the dlPFC and RPE-specific from the vmPFC and, as the reviewer predicted, we found synergistic II between the two regions probably mainly because of their respective specificity. In addition, we included the II estimated between non-selective bipolar derivations (i.e. recordings with significant encoding for both RPE and PPE) and we observed synergistic interactions (Fig. 5C and Fig. S9). Taken together, the local specificity certainly plays a role, but this is not the only factor in defining the type of interactions.

Concerning the interaction information results (II, Fig. 3), several lines of evidence suggest that local specificity cannot account alone for the II effects. For example, the local specificity for PPE is observed across all four areas (Fig. 2A) and the percentage of bipolar derivations displaying an effect is large (equal or above 10%) for three brain regions (aINS, dlPLF and lOFC). If the local specificity were the main driving cause, we would have observed significant redundancy between all pairs of brain regions. On the other hand, the interaction between the aINS and lOFC displayed no significant redundant effect (Fig. 3B). Another example is the result observed in lOFC: approximately 30% of bipolar derivations display a selectivity for PPE (Fig. 2B, third panel from the left), but do not show clear signs of redundant encoding at the level of within-area interactions (Fig. 3A, bottom-left panel). Similarly, the local encoding for RPE is observed across all four brain regions (Fig. 2A) and the percentage of bipolar derivations displaying an effect is large (equal or above 10%) for three brain regions (aINS, dlPLF and vmPFC). Nevertheless, significant between-regions interactions have been observed only between the lOFC and vmPFC (Fig. 3B bottom right panel).

To further support the reasoning, we performed a simulation to show that it is possible to observe synergistic interactions between two regions with the same specificity. As an example, we may consider one region locally encoding early trials of RPE and a second region encoding the late trials of the RPE. Combining the two with the II would lead to synergistic interactions, because each one of them carries information that is not carried by the other. To illustrate this point, we simulated the data of two regions (x and y). To simulate redundant interactions (first row), each region receives a copy of the prediction (one-to-all) and for the synergy (second row), x and y receive early and late PE trials, respectively (all-to-one). This toy example illustrates that the local specificity is not the only factor determining the type of their interactions. We added the following result to the Appendix.

Author response image 4.

Local specificity does not fully determine the type of interactions. Within-area local encoding of PE using the mutual information (MI, in bits) for regions X and Y and between-area interaction information (II, in bits) leading to (A) redundant interactions and (B) synergistic interactions about the PE

Regarding the information transfer results (Fig. 4), similar arguments hold and suggest that the prevalence is not the main factor explaining the arising transfer entropy between the anterior insula (aINS) and dorsolateral prefrontal cortex (dlPFC). Indeed, the lOFC has a strong local specificity for PPE, but the transfer entropy between the lOFC and aINS (or dlPFC) is shown in Fig. S7 does not show significant differences in encoding between PPE and RPE.

Indeed, such transfer can only be found when there is a delay between the gamma activity of the two regions. In this example, the transfer entropy quantifies the amount of information shared between the past activity of the aINS and the present activity of the dlPFC conditioned on the past activity of the dlPFC. The conditioning ensures that the present activity of the dlPFC is not only explained by its own past. Consequently, if both regions exhibit various prevalences toward reward and punishment but without delay (i.e. at the same timing), the transfer entropy would be null because of the conditioning. As a fact, between 10 to -20% of bipolar recordings show a selectivity to the reward PE (represented by a proportion of 40-60% of subjects, Fig.S4). However, the transfer entropy estimated from the aINS to the dlPFC across rewarding trials is flat and clearly non-significant. If the transfer entropy was a byproduct of the local specificity then we should observe an increase, which is not the case here.

Reviewer #2:

Summary:

Reward and punishment learning have long been seen as emerging from separate networks of frontal and subcortical areas, often studied separately. Nevertheless, both systems are complimentary and distributed representations of rewards and punishments have been repeatedly observed within multiple areas. This raised the unsolved question of the possible mechanisms by which both systems might interact, which this manuscript went after. The authors skillfully leveraged intracranial recordings in epileptic patients performing a probabilistic learning task combined with model-based information theoretical analyses of gamma activities to reveal that information about reward and punishment was not only distributed across multiple prefrontal and insular regions, but that each system showed specific redundant interactions. The reward subsystem was characterized by redundant interactions between orbitofrontal and ventromedial prefrontal cortex, while the punishment subsystem relied on insular and dorsolateral redundant interactions. Finally, the authors revealed a way by which the two systems might interact, through synergistic interaction between ventromedial and dorsolateral prefrontal cortex.

Strengths:

Here, the authors performed an excellent reanalysis of a unique dataset using innovative approaches, pushing our understanding on the interaction at play between prefrontal and insular cortex regions during learning. Importantly, the description of the methods and results is truly made accessible, making it an excellent resource to the community.

This manuscript goes beyond what is classically performed using intracranial EEG dataset, by not only reporting where a given information, like reward and punishment prediction errors, is represented but also by characterizing the functional interactions that might underlie such representations. The authors highlight the distributed nature of frontal cortex representations and propose new ways by which the information specifically flows between nodes. This work is well placed to unify our understanding of the complementarity and specificity of the reward and punishment learning systems.

We thank the reviewer for the positive feedback. Please find below our answers to the weaknesses raised by the reviewer.

Weaknesses:

The conclusions of this paper are mostly supported by the data, but whether the findings are entirely generalizable would require further information/analyses.

First, the authors found that prediction errors very quickly converge toward 0 (less than 10 trials) while subjects performed the task for sets of 96 trials. Considering all trials, and therefore having a non-uniform distribution of prediction errors, could potentially bias the various estimates the authors are extracting. Separating trials between learning (at the start of a set) and exploiting periods could prove that the observed functional interactions are specific to the learning stages, which would strengthen the results.

We thank the reviewer for this question. We would like to note that the probabilistic nature of the learning task does not allow a strict distinction between the exploration and exploitation phases. Indeed, the probability of obtaining the less rewarding outcome was 25% (i.e., for 0€ gain in the reward learning condition and -1€ loss in the punishment learning condition). Thus, participants tended to explore even during the last set of trials in each session. This is evident from the average learning curves shown in Fig. 1B of (Gueguen et al., 2021). Learning curves show rates of correct choice (75% chance of 1€ gain) in the reward condition (blue curves) and incorrect choice (75% chance of 1€ loss) in the punishment condition (red curves).

For what concerns the evolution of PEs, as reviewer #1 suggested, we added a new figure representing the single-subject estimates of the R/PPE (Fig S2). Here, the confidence interval is obtained across all pairs of stimuli presented during the different sessions. We retrieved the general trend of the R/PPE converging toward zero around 10 trials. Both average reward and punishment prediction errors converge toward zero in approximately 10 trials, single-participant curves display large variability, also at the end of each session. As a reminder, the 96 trials represent the total number of trials for one session for the four pairs and the number of trials for each stimulus was only 24.

Author response image 5.

Single-subject estimation of predictions errors. Single-subject trial-wise reward PE (RPE - blue) and punishment PE (PPE - red), ± 95% confidence interval

However, the convergence of the R/PPE is due to the average across the pairs of stimuli. In the figure below, we superimposed the estimated R/PPE, per pair of stimuli, for each subject. It becomes very clear that high values of PE can be reached, even for late trials. Therefore, we believe that the split into early/late trials because of the convergence of PE is far from being trivial.

Author response image 6.

Single-subject estimation of predictions errors per pair of stimuli. Single-subject trial-wise reward PE (RPE - blue) and punishment PE (PPE - red)

Consequently, nonzero PRE and PPE occur during the whole session and separating trials between learning (at the start of a set) and exploiting periods, as suggested by the reviewer, does not allow a strict dissociation between learning vs no-learning. Nevertheless, we tested the analysis proposed by the reviewer, at the local level. We splitted the 24 trials of each pair of stimuli into early, middle and late trials (8 trials each). We then reproduced Fig. 2 by computing the mutual information between the gamma activity and the R/PPE for subsets of trials: early (first row) and late trials (second row). We retrieved significant encoding of both R/PPE in the aINS, dlPFC and lOFC in both early and late trials. The vmPFC also showed significant encoding of both during early trials. The only difference emerges in the late trials of the vmPFC where we found a strong encoding of the RPE only. It should also be noted that here since we are sub-selecting the trials, the statistical analyses are only performed using a third of the trials.

Taken together, the combination of high values of PE achieved even for late trials and the fact that most of the findings are reproduced even with a third of the trials does not justify the split into early and late trials here. Crucially, this latest analysis confirms that the neural correlates of learning that we observed reflect PE signals rather than early versus late trials in the session.

Author response image 7.

MI between gamma activity and R/PPE using early and late trials. Time courses of MI estimated between the gamma power and both RPE (blue) and PPE (red) using either early or late trials (first and second row, respectively). Horizontal thick lines represent significant clusters of information (p<0.05, cluster-based correction, non-parametric randomization across epochs).

Importantly, it is unclear whether the results described are a common feature observed across subjects or the results of a minority of them. The authors should report and assess the reliability of each result across subjects. For example, the authors found RPE-specific interactions between vmPFC and lOFC, even though less than 10% of sites represent RPE or both RPE/PPE in lOFC. It is questionable whether such a low proportion of sites might come from different subjects, and therefore whether the interactions observed are truly observed in multiple subjects. The nature of the dataset obviously precludes from requiring all subjects to show all effects (given the known limits inherent to intracerebral recording in patients), but it should be proven that the effects were reproducibly seen across multiple subjects.

We thank the reviewer for this remark that has also been raised by the first reviewer. This issue was raised by the first reviewer. Indeed, we added a supplementary figure describing the number of unique subjects per brain region and per pair of brain regions (Fig. S1A) such as the number of bipolar derivations per region and per subject (Fig. S1B).

Author response image 8.

Single subject anatomical repartition. (A) Number of unique subject per brain region and per pair of brain regions (B) Number of bipolar derivations per subject and per brain region

Regarding the reproducibility of the results across subjects for the local analysis (Fig. 2), we also added the instantaneous proportion of subjects having at least one bipolar derivation showing a significant encoding of the RPE and PPE (Fig. S4). We found a minimum proportion of approximately 30% of unique subjects having the effect in the lOFC and vmPFC, respectively with the RPE and PPE. On the other hand, both the aINS and dlPFC showed between 50 to 100% of the subjects having the effect. Therefore, local encoding of RPE and PPE was never represented by a single subject.

Author response image 9.

Similarly, we performed statistical analysis on interaction information at the single-subject level and counted the proportion of unique subjects having at least one pair of recordings with significant redundant and synergistic interactions about the RPE and PPE (Fig. S5). Consistently with the results shown in Fig. 3, the proportions of significant redundant and synergistic interactions are negative and positive, respectively. For the within-regions interactions, approximately 60% of the subjects with redundant interactions are about R/PPE in the aINS and about the PPE in the dlPFC and 40% about the RPE in the vmPFC. For the across-regions interactions, 60% of the subjects have redundant interactions between the aINS-dlPFC and dlPFC-lOFC about the PPE, and 30% have redundant interactions between lOFC-vmPFC about the RPE. Globally, we reproduced the main results shown in Fig. 3.

Author response image 10.

Inter-subjects reproducibility of redundant interactions about PE signals. Time-courses of proportion of subjects having at least one pair of bipolar derivation with a significant interaction information (p<0.05, cluster-based correction, non-parametric randomization across epochs) about the RPE (blue) or PPE (red). Data are aligned to the outcome presentation (vertical line at 0 seconds). Proportion of subjects with redundant (solid) and synergistic (dashed) interactions are respectively going downward and upward.

Finally, the timings of the observed interactions between areas preclude one of the authors' main conclusions. Specifically, the authors repeatedly concluded that the encoding of RPE/PPE signals are "emerging" from redundancy-dominated prefrontal-insular interactions. However, the between-region information and transfer entropy between vmPFC and lOFC for example is observed almost 500ms after the encoding of RPE/PPE in these regions, questioning how it could possibly lead to the encoding of RPE/PPE. It is also noteworthy that the two information measures, interaction information and transfer entropy, between these areas happened at non overlapping time windows, questioning the underlying mechanism of the communication at play (see Figures 3/4). As an aside, when assessing the direction of information flow, the authors also found delays between pairs of signals peaking at 176ms, far beyond what would be expected for direct communication between nodes. Discussing this aspect might also be of importance as it raises the possibility of third-party involvement.

The local encoding of RPE in the vmPFC and lOFC is observed in a time interval ranging from approximately 0.2-0.4s to 1.2-1.4s after outcome presentation (blue bars in Fig. 2A). The encoding of RPE by interaction information covers a time interval from approximately 1.1s to 1.5s (blue bars in Fig. 3B, bottom right panel). Similarly, significant TE modulations between the vmPFC and lOFC specific for PPE occur mainly in the 0.7s-1.1s range. Thus, it seems that the local encoding of PPE precedes the effects observed at the level of the neural interactions (II and TE). On the other hand, the modulations in MI, II and TE related to PPE co-occur in a time window from 0.2s to 0.7s after outcome presentation. Thus, we agree with the reviewer that a generic conclusion about the potential mechanisms relating the three levels of analysis cannot be drawn. We thus replaced the term “emerge from” by “occur with” from the manuscript which may be misinterpreted as hinting at a potential mechanism. We nevertheless concluded that the three levels of analysis (and phenomena) co-occur in time, thus hinting at a potential across-scales interaction that needs further study. Indeed, our study suggests that further work, beyond the scope of the current study, is required to better understand the interaction between scales.

Regarding the delay for the conditioning of the transfer entropy, the value of 176 ms reflects the delay at which we observed a maximum of transfer entropy. However, we did not use a single delay for conditioning, we used every possible delay between [116, 236] ms, as explained in the Method section. We would like to stress that transfer entropy is a directed metric of functional connectivity, and it can only be interpreted as quantifying statistical causality defined in terms of predictacìbility according to the Wiener-Granger principle, as detailed in the methods. Thus, it cannot be interpreted in Pearl’s causal terms and as indexing any type of direct communication between nodes. This is a known limitation of the method, which has been stressed in past literature and that we believe does not need to be addressed here.

To account for this, we revised the discussion to make sure this issue is addressed in the following paragraph:

“Here, we quantified directional relationships between regions using the transfer entropy (Schreiber, 2000), which is a functional connectivity measure based on the Granger-Wiener causality principle. Tract tracing studies in the macaque have revealed strong interconnections between the lOFC and vmPFC in the macaque (Carmichael and Price, 1996; Öngür and Price, 2000). In humans, cortico-cortical anatomical connections have mainly been investigated using diffusion magnetic resonance imaging (dMRI). Several studies found strong probabilities of structural connectivity between the anterior insula with the orbitofrontal cortex and dorsolateral part of the prefrontal cortex (Cloutman et al., 2012; Ghaziri et al., 2017), and between the lOFC and vmPFC (Heather Hsu et al., 2020). In addition, the statistical dependency (e.g. coherence) between the LFP of distant areas could be potentially explained by direct anatomical connections (Schneider et al., 2021; Vinck et al., 2023). Taken together, the existence of an information transfer might rely on both direct or indirect structural connectivity. However, here we also reported differences of TE between rewarding and punishing trials given the same backbone anatomical connectivity (Fig. 4). [...] “

Reviewer #3:

Summary:

The authors investigated that learning processes relied on distinct reward or punishment outcomes in probabilistic instrumental learning tasks were involved in functional interactions of two different cortico-cortical gamma-band modulations, suggesting that learning signals like reward or punishment prediction errors can be processed by two dominated interactions, such as areas lOFC-vmPFC and areas aINS-dlPFC, and later on integrated together in support of switching conditions between reward and punishment learning. By performing the well-known analyses of mutual information, interaction information, and transfer entropy, the conclusion was accomplished by identifying directional task information flow between redundancy-dominated and synergy-dominated interactions. Also, this integral concept provided a unifying view to explain how functional distributed reward and/or punishment information were segregated and integrated across cortical areas.

Strengths:

The dataset used in this manuscript may come from previously published works (Gueguen et al., 2021) or from the same grant project due to the methods. Previous works have shown strong evidence about why gamma-band activities and those 4 areas are important. For further analyses, the current manuscript moved the ideas forward to examine how reward/punishment information transfer between recorded areas corresponding to the task conditions. The standard measurements such mutual information, interaction information, and transfer entropy showed time-series activities in the millisecond level and allowed us to learn the directional information flow during a certain window. In addition, the diagram in Figure 6 summarized the results and proposed an integral concept with functional heterogeneities in cortical areas. These findings in this manuscript will support the ideas from human fMRI studies and add a new insight to electrophysiological studies with the non-human primates.

We thank the reviewer for the summary such as for highlighting the strengths. Please find below our answers regarding the weaknesses of the manuscript.

Weaknesses:

After reading through the manuscript, the term "non-selective" in the abstract confused me and I did not actually know what it meant and how it fits the conclusion. If I learned the methods correctly, the 4 areas were studied in this manuscript because of their selective responses to the RPE and PPE signals (Figure 2). The redundancy- and synergy-dominated subsystems indicated that two areas shared similar and complementary information, respectively, due to the negative and positive value of interaction information (Page 6). For me, it doesn't mean they are "non-selective", especially in redundancy-dominated subsystem. I may miss something about how you calculate the mutual information or interaction information. Could you elaborate this and explain what the "non-selective" means?

In the study performed by Gueguen et al. in 2021, the authors used a general linear model (GLM) to link the gamma activity to both the reward and punishment prediction errors and they looked for differences between the two conditions. Here, we reproduced this analysis except that we used measures from the information theory (mutual information) that were able to capture linear and non-linear relationships (although monotonic) between the gamma activity and the prediction errors. The clusters we reported reflect significant encoding of either the RPE and/or the PPE. From Fig. 2, it can be seen that the four regions have a gamma activity that is modulated according to both reward and punishment PE. We used the term “non-selective”, because the regions did not encode either one or the other, but various proportions of bipolar derivations encoding either one or both of them.

The directional information flows identified in this manuscript were evidenced by the recording contacts of iEEG with levels of concurrent neural activities to the task conditions. However, are the conclusions well supported by the anatomical connections? Is it possible that the information was transferred to the target via another area? These questions may remain to be elucidated by using other approaches or animal models. It would be great to point this out here for further investigation.

We thank the reviewer for this interesting question. We added the following paragraph to the discussion to clarify the current limitations of the transfer entropy and the link with anatomical connections :

“Here, we quantified directional relationships between regions using the transfer entropy (Schreiber, 2000), which is a functional connectivity measure based on the Granger-Wiener causality principle. Tract tracing studies in the macaque have revealed strong interconnections between the lOFC and vmPFC in the macaque (Carmichael and Price, 1996; Öngür and Price, 2000). In humans, cortico-cortical anatomical connections have mainly been investigated using diffusion magnetic resonance imaging (dMRI). Several studies found strong probabilities of structural connectivity between the anterior insula with the orbitofrontal cortex and dorsolateral part of the prefrontal cortex (Cloutman et al., 2012; Ghaziri et al., 2017), and between the lOFC and vmPFC (Heather Hsu et al., 2020). In addition, the statistical dependency (e.g. coherence) between the LFP of distant areas could be potentially explained by direct anatomical connections (Schneider et al., 2021). Taken together, the existence of an information transfer might rely on both direct or indirect structural connectivity. However, here we also reported differences of TE between rewarding and punishing trials given the same backbone anatomical connectivity (Fig. 4). Our results are further supported by a recent study involving drug-resistant epileptic patients with resected insula who showed poorer performance than healthy controls in case of risky loss compared to risky gains (Von Siebenthal et al., 2017).”

References

Carmichael ST, Price J. 1996. Connectional networks within the orbital and medial prefrontal cortex of macaque monkeys. J Comp Neurol 371:179–207.

Cloutman LL, Binney RJ, Drakesmith M, Parker GJM, Lambon Ralph MA. 2012. The variation of function across the human insula mirrors its patterns of structural connectivity: Evidence from in vivo probabilistic tractography. NeuroImage 59:3514–3521. oi:10.1016/j.neuroimage.2011.11.016

Combrisson E, Allegra M, Basanisi R, Ince RAA, Giordano BL, Bastin J, Brovelli A. 2022. Group-level inference of information-based measures for the analyses of cognitive brain networks from neurophysiological data. NeuroImage 258:119347. doi:10.1016/j.neuroimage.2022.119347

Ghaziri J, Tucholka A, Girard G, Houde J-C, Boucher O, Gilbert G, Descoteaux M, Lippé S, Rainville P, Nguyen DK. 2017. The Corticocortical Structural Connectivity of the Human Insula. Cereb Cortex 27:1216–1228. doi:10.1093/cercor/bhv308

Gueguen MCM, Lopez-Persem A, Billeke P, Lachaux J-P, Rheims S, Kahane P, Minotti L, David O, Pessiglione M, Bastin J. 2021. Anatomical dissociation of intracerebral signals for reward and punishment prediction errors in humans. Nat Commun 12:3344. doi:10.1038/s41467-021-23704-w

Heather Hsu C-C, Rolls ET, Huang C-C, Chong ST, Zac Lo C-Y, Feng J, Lin C-P. 2020. Connections of the Human Orbitofrontal Cortex and Inferior Frontal Gyrus. Cereb Cortex 30:5830–5843. doi:10.1093/cercor/bhaa160

Lachaux J-P, Fonlupt P, Kahane P, Minotti L, Hoffmann D, Bertrand O, Baciu M. 2007. Relationship between task-related gamma oscillations and BOLD signal: new insights from combined fMRI and intracranial EEG. Hum Brain Mapp 28:1368–1375. doi:10.1002/hbm.20352

Mukamel R, Gelbard H, Arieli A, Hasson U, Fried I, Malach R. 2004. Coupling Between Neuronal Firing, Field Potentials, and fMRI in Human Auditory Cortex. Cereb Cortex 14:881.

Niessing J, Ebisch B, Schmidt KE, Niessing M, Singer W, Galuske RA. 2005. Hemodynamic signals correlate tightly with synchronized gamma oscillations. science 309:948–951.

Nir Y, Fisch L, Mukamel R, Gelbard-Sagiv H, Arieli A, Fried I, Malach R. 2007. Coupling between neuronal firing rate, gamma LFP, and BOLD fMRI is related to interneuronal correlations. Curr Biol 17:1275–1285.

Öngür D, Price JL. 2000. The organization of networks within the orbital and medial prefrontal cortex of rats, monkeys and humans. Cereb Cortex 10:206–219.

Schneider M, Broggini AC, Dann B, Tzanou A, Uran C, Sheshadri S, Scherberger H, Vinck M. 2021. A mechanism for inter-areal coherence through communication based on connectivity and oscillatory power. Neuron 109:4050-4067.e12. doi:10.1016/j.neuron.2021.09.037

Schreiber T. 2000. Measuring information transfer. Phys Rev Lett 85:461.

Von Siebenthal Z, Boucher O, Rouleau I, Lassonde M, Lepore F, Nguyen DK. 2017. Decision-making impairments following insular and medial temporal lobe resection for drug-resistant epilepsy. Soc Cogn Affect Neurosci 12:128–137. doi:10.1093/scan/nsw152

Recommendations for the authors

Reviewer #1

(1) Overall, the writing of the manuscript is dense and makes it hard to follow the scientific logic and appreciate the key findings of the manuscript. I believe the manuscript would be accessible to a broader audience if the authors improved the writing and provided greater detail for their scientific questions, choice of analysis, and an explanation of their results in simpler terms.

We extensively modified the introduction to better describe the rationale and research question.

(2) In the introduction the authors state "we hypothesized that reward and punishment learning arise from complementary neural interactions between frontal cortex regions". This stated hypothesis arrives rather abruptly after a summary of the literature given that the literature summary does not directly inform their stated hypothesis. Put differently, the authors should explicitly state what the contradictions and/or gaps in the literature are, and what specific combinations of findings guide them to their hypothesis. When the authors state their hypothesis the reader is still left asking: why are the authors focusing on the frontal regions? What do the authors mean by complementary interactions? What specific evidence or contradiction in the literature led them to hypothesize that complementary interactions between frontal regions underlie reward and punishment learning?

We extensively modified the introduction and provided a clearer description of the brain circuits involved and the rationale for searching redundant and synergistic interactions between areas.

(3) Related to the above point: when the authors subsequently state "we tested whether redundancy- or synergy dominated interactions allow the emergence of collective brain networks differentially supporting reward and punishment learning", the Introduction (up to the point of this sentence) has not been written to explain the synergy vs. redundancy framework in the literature and how this framework comes into play to inform the authors' hypothesis on reward and punishment learning.

We extensively modified the introduction and provided a clearer description of redundant and synergistic interactions between areas.

(4) The explanation of redundancy vs synergy dominated brain networks itself is written densely and hard to follow. Furthermore, how this framework informs the question on the neural substrates of reward versus punishment learning is unclear. The authors should provide more precise statements on how and why redundancy vs. synergy comes into play in reward and punishment learning. Put differently, this redundancy vs. synergy framework is key for understanding the manuscript and the introduction is not written clearly enough to explain the framework and how it informs the authors' hypothesis and research questions on the neural substrates of reward vs. punishment learning.

Same as above

(5) While the choice of these four brain regions in context of reward and punishment learning does makes sense, the authors do not outline a clear scientific justification as to why these regions were selected in relation to their question.

Same as above

(6) Could the authors explain why they used gamma band power (as opposed to or in addition to the lower frequency bands) to investigate MI. Relatedly, when the authors introduce MI analysis, it would be helpful to briefly explain what this analysis measures and why it is relevant to address the question they are asking.

Please see our answer to the first public comment. We added a paragraph to the discussion section to justify our choice of focusing on the gamma band only. We added the following sentence to the result section to justify our choice for using mutual-information:

The MI allowed us to detect both linear and non-linear relationships between the gamma activity and the PE

An extended explanation justifying our choice for the MI was already present in the method section.

(7) The authors state that "all regions displayed a local "probabilistic" encoding of prediction errors with temporal dynamics peaking around 500 ms after outcome presentation". It would be helpful for the reader if the authors spelled out what they mean by probabilistic in this context as the term can be interpreted in many different ways.

We agree with the reviewer that the term “probabilistic” can be interpreted in different ways. In the revised manuscript we changed “probabilistic” for “mixed”.

(8) The authors should include a brief description of how they compute RPE and PPE in the beginning of the relevant results section.

The explanation of how we estimated the PE is already present in the result section: “We estimated trial-wise prediction errors by fitting a Q-learning model to behavioral data. Fitting the model consisted in adjusting the constant parameters to maximize the likelihood of observed choices etc.”

(9) It is unclear from the Methods whether the authors have taken any measures to address the likely difference in the number of electrodes across subjects. For example, it is likely that some subjects have 10 electrodes in vmPFC while others may have 20. In group analyses, if the data is simply averaged across all electrodes then each subject contributes a different number of data points to the analysis. Hence, a subject with more electrodes can bias the group average. A starting point would be to state the variation in number of electrodes across subjects per brain region. If this variation is rather small, then simple averaging across electrodes might be justified. If the variation is large then one idea would be to average data across electrodes within subjects prior to taking the group average or use a resampling approach where the minimum number of electrodes per brain area is subsampled.

We addressed this point in our public answers. As a reminder, the new version of the manuscript contains a figure showing the number of unique patients per region, the PE at per participant level together with local-encoding at the single participant level.

(10) One thing to consider is whether the reward and punishment in the task is symmetrical in valence. While 1$ increase and 1$ decrease is equivalent in magnitude, the psychological effect of the positive (vs. the negative) outcome may still be asymmetrical and the direction and magnitude of this asymmetry can vary across individuals. For instance, some subjects may be more sensitive to the reward (over punishment) while others are more sensitive to the punishment (over reward). In this scenario, it is possible that the differentiation observed in PPE versus RPE signals may arise from such psychological asymmetry rather than the intrinsic differences in how certain brain regions (and their interactions) may encode for reward vs punishment. Perhaps the authors can comment on this possibility, and/or conduct more in depth behavioral analysis to determine if certain subjects adjust their choice behavior faster in response to reward vs. punishment contexts.

While it could be possible that individuals display different sensitivities vis-à-vis positive and negative prediction errors (and, indeed, a vast body of human reinforcement learning literature seems to point in this direction; Palminteri & Lebreton, 2022), it is unclear to us how such differences would explain into the recruitment of anatomically distinct areas reward and punishment prediction errors. It is important to note here that our design partially orthogonalized positive and reward vs. negative and punishment PEs, because the neutral outcome can generate both positive and negative prediction errors, as a function of the learning context (reward-seeking and punishment avoidance). Back to the main question, for instance, Lefebvre et al (2017) investigated with fMRI the neural correlates of reward prediction errors only and found that inter-individual differences in learning rates for positive and negative prediction errors correlated with differences in the degree of striatal activation and not with the recruitment of different areas. To sum up, while we acknowledge that individuals may display different sensitivity to prediction errors (and reward magnitudes), we believe that such differences should translated in difference in the degree of activation of a given system (the reward systems vs the punishment one) rather than difference in neural system recruitment

(11) As summarized in Fig 6, the authors show that information transfer between aINS to dlPFC was PPE specific whereas the information transfer between vmPFC to lOFC was RPE specific. What is unclear is if these findings arise as an inevitable statistical byproduct of the fact that aINS has high PPE-specificity and that vmPFC has high RPE-specificity. In other words, it is possible that the analysis in Fig 3,4 are sensitive to fact that there is a larger proportion of electrodes with either PPE or RPE sensitivity in aINS and vmPFC respectively - and as such, the II analysis might reflect the dominant local encoding properties above and beyond reflecting the interactions between regions per se. Simply put, could the analysis in Fig 3B turn out in any other way given that there are more PPE specific electrodes in aINS and more RPE specific electrodes in vmPFC? Some options to address this question would be to limit the electrodes included in the analyses (in Fig 3B for example) so that each region has the same number of PPE and RPE specific electrodes included.

Please see the simulation we added to the revised manuscript (Fig. S10) demonstrating that synergistic interactions can emerge between regions with the same specificity.

Regarding the possibility that Fig. 3 and 4 are sensitive to the number of bipolar derivations being R/PPE specific, a counter-example is the vmPFC. The vmPFC has a few recordings specific to punishment (Fig. 2) in almost 30% of the subjects (Fig. S4). However, there is no II about the PPE between recordings of the vmPFC (Fig. 3). The same reasoning also holds for the lOFC. Therefore, the proportion of recordings being RPE or PPE-specific is not sufficient to determine the type of interactions.

(12)  Related to the point above, what would the results presented in Fig 3A (and 3B) look like if the authors ran the analyses on RPE specific and PPE specific electrodes only. Is the vmPFC-vmPFC RPE effect in Fig 3A arising simply due to the high prevalence of RPE specific electrodes in vmPFC (as shown in Fig. 2)?

Please see our answer above.

Reviewer #2:

Regarding Figure 2A, the authors argued that their findings "globally reproduced their previously published findings" (from Gueguen et al, 2021). It is worth noting though that in their original analysis, both aINS and lOFC show differential effects (aINS showing greater punishment compared to reward, and the opposite for lOFC) compared to the current analysis. Although I would be akin to believe that the nonlinear approach used here might explain part of the differences (as the authors discussed), I am very wary of the other argument advanced: "the removal of iEEG sites contaminated with pathological activity". This raised some red flags. Does that mean some of the conclusions observed in Gueguen et al (2021) are only the result of noise contamination, and therefore should be disregarded? The author might want to add a short supplementary figure using the same approach as in Gueguen (2021) but using the subset of contacts used here to comfort potential readers of the validity of their previous manuscript.

We appreciate the reviewer's concerns and understand the request for additional information. However, we would like to point out that the figure suggested by the reviewer is already present in the supplementary files of Gueguen et al. 2021 (see Fig. S2). The results of this study should not be disregarded, as the supplementary figure reproduces the results of the main text after excluding sites with pathological activity. Including or excluding sites contaminated with epileptic activity does not have a significant impact on the results, as analyses are performed at each time-stamp and across trials, and epileptic spikes are never aligned in time across trials.

That being said, there are some methodological differences between the two studies. To extract gamma power, Gueguen et al. filtered and averaged 10 Hz sub-bands, while we used multi-tapers. Additionally, they used a temporal smoothing of 250 ms, while we used less smoothing. However, as explained in the main text, we used information-theoretical approaches to capture the statistical dependencies between gamma power and PE. Despite divergent methodologies, we obtained almost identical results.

The data and code supporting this manuscript should be made available. If raw data cannot be shared for ethical reasons, single-trial gamma activities should at least be provided. Regarding the code used to process the data, sharing it could increase the appeal (and use) of the methods applied.

We thank the reviewer for this suggestion. We added a section entitled “Code and data availability” and gave links to the scripts, notebooks and preprocessed data.

Author response:

The following is the authors’ response to the previous reviews.

Recommendations for the authors:

Reviewer #2 (Recommendations for the authors):

I appreciate the efforts the authors made to clarify and justify their statements and methodology, respectively. I additionally appreciate the efforts they made to provide me with detailed information - including figures - to aid my comprehension. However, there are two things I nevertheless recommend the authors to include in the main manuscript.

(1) Statement about animal wellbeing: The authors state that they were constrained in their imaging session duration not because of a commonly reported technical limitation, such as photobleaching (which I honestly assumed), but rather the general wellbeing of the animals, who exhibited signs of distress after longer imaging periods. I find this to be a critical issue and perhaps the best argument against performing longer imaging experiments (which would have increased the number of trials, thus potentially boosting the performance of their model). To say that they put animal welfare above all other scientific and technical considerations speaks to a strong ethical adherence to animal welfare policy, and I believe this should be somehow incorporated into the methods.

We have now included this at the top of page 26:

“Mice fully recovered from the brief isoflurane anesthesia, showing a clear blinking reflex, whisking and sniffing behaviors and normal body posture and movements, immediately after head fixation. In our experimental conditions, mice were imaged in sessions of up to 25 min since beyond this time we started observing some signs of distress or discomfort. Thus, we avoided longer recording times at the expense of collecting larger trial numbers, in strong adherence of animal welfare and ethics policy. A pilot group of mice were habituated to the head fixed condition in daily 20 min sessions for 3 days, however we did not observe a marked contrast in the behavior of habituated versus unhabituated mice beyond our relatively short 25 min imaging sessions. In consequence imaging sessions never surpassed a maximum of 25 min, after which the mouse was returned to its home cage.”

(2) Author response image 2: I sincerely thank the authors for providing us reviewers with this figure, which compares the performance of the naïve Bayesian classifier their ultimately use in the study with other commonly implemented models. Also here I falsely assumed that other models, which take correlated activity into account, did not generally perform better than their ultimate model of choice. Although dwelling on it would be distractive (and outside the primary scope of the study), I would encourage the authors to include it as a figure supplement (and simply mention these controls en passant when they justify their choice of the naïve Bayesian classifier).

This figure was now included in the revised manuscript as supplemental figure 3.

Page 10 now reads:

“We performed cross-validated, multi-class classification of the single-trial population responses (decoding, Fig. 2A) using a naive Bayes classifier to evaluate the prediction errors as the absolute difference between the stimulus azimuth and the predicted azimuth (Fig. 2A). We chose this classification algorithm over others due to its generally good performance with limited available data. We visualized the cross-validated prediction error distribution in cumulative plots where the observed prediction errors were compared to the distribution of errors for random azimuth sampling (Fig. 2B). When decoding all simultaneously recorded units, the observed classifier output was not significantly better (shifted towards smaller prediction errors) than the chance level distribution (Fig. 2B). The classifier also failed to decode complete DCIC population responses recorded with neuropixels probes (Fig. 3A). Other classifiers performed similarly (Suppl. Fig. 3A).”

The bottom paragraph in page 19 now reads:

“To characterize how the observed positive noise correlations could affect the representation of stimulus azimuth by DCIC top ranked unit population responses, we compared the decoding performance obtained by classifying the single-trial response patterns from top ranked units in the modeled decorrelated datasets versus the acquired data (with noise correlations). With the intention to characterize this with a conservative approach that would be less likely to find a contribution of noise correlations as it assumes response independence, we relied on the naive Bayes classifier for decoding throughout the study. Using this classifier, we observed that the modeled decorrelated datasets produced stimulus azimuth prediction error distributions that were significantly shifted towards higher decoding errors (Fig. 6B, C) and, in our imaging datasets, were not significantly different from chance level (Fig. 6B). Altogether, these results suggest that the detected noise correlations in our simultaneously acquired datasets can help reduce the error of the IC population code for sound azimuth. We observed a similar, but not significant tendency with another classifier that does not assume response independence (KNN classifier), though overall producing larger decoding errors than the Bayes classifier (Suppl. Fig. 3B).”

Reviewer #3 (Recommendations for the authors):

I am generally happy with the response to the reviews.

I find the Author response image 3 quite interesting. The neuropixel data looks somewhat like I expected (especially for mouse #3 and maybe mouse #4). I find the distribution of weights across units in the imaging dataset compared to in the pixel dataset intriguing (though it probably is just the dimensionality of the data being so much higher).

I'm not too familiar with facial movements but is it the case that the DCIC would be more modulated by ipsilateral movement compared to contralateral movements? Are face movements in mice conjugate or do both sides of the face move more or less independently? If not it may be interesting in future work to record bilaterally and see if that provides more information about DCIC responses.

We sincerely thank the editors and reviewers for their careful appraisal, commendation of our effort and helpful constructive feedback which greatly improved the presentation of our study. Below in green font is a point by point reply to the comments provided by the reviewers.

Public Reviews:

Reviewer #1 (Public Review):

Summary: In this study, the authors address whether the dorsal nucleus of the inferior colliculus (DCIC) in mice encodes sound source location within the front horizontal plane (i.e., azimuth). They do this using volumetric two-photon Ca2+ imaging and high-density silicon probes (Neuropixels) to collect single-unit data. Such recordings are beneficial because they allow large populations of simultaneous neural data to be collected. Their main results and the claims about those results are the following:

(1) DCIC single-unit responses have high trial-to-trial variability (i.e., neural noise);

(2) approximately 32% to 40% of DCIC single units have responses that are sensitive to sound source azimuth;

(3) single-trial population responses (i.e., the joint response across all sampled single units in an animal) encode sound source azimuth "effectively" (as stated in title) in that localization decoding error matches average mouse discrimination thresholds;

(4) DCIC can encode sound source azimuth in a similar format to that in the central nucleus of the inferior colliculus (as stated in Abstract);

(5) evidence of noise correlation between pairs of neurons exists;

and (6) noise correlations between responses of neurons help reduce population decoding error.

While simultaneous recordings are not necessary to demonstrate results #1, #2, and #4, they are necessary to demonstrate results #3, #5, and #6.

Strengths:

- Important research question to all researchers interested in sensory coding in the nervous system.

- State-of-the-art data collection: volumetric two-photon Ca2+ imaging and extracellular recording using high-density probes. Large neuronal data sets.

- Confirmation of imaging results (lower temporal resolution) with more traditional microelectrode results (higher temporal resolution).

- Clear and appropriate explanation of surgical and electrophysiological methods. I cannot comment on the appropriateness of the imaging methods.

Strength of evidence for claims of the study:

(1) DCIC single-unit responses have high trial-to-trial variability - The authors' data clearly shows this.

(2) Approximately 32% to 40% of DCIC single units have responses that are sensitive to sound source azimuth - The sensitivity of each neuron's response to sound source azimuth was tested with a Kruskal-Wallis test, which is appropriate since response distributions were not normal. Using this statistical test, only 8% of neurons (median for imaging data) were found to be sensitive to azimuth, and the authors noted this was not significantly different than the false positive rate. The Kruskal-Wallis test was not performed on electrophysiological data. The authors suggested that low numbers of azimuth-sensitive units resulting from the statistical analysis may be due to the combination of high neural noise and relatively low number of trials, which would reduce statistical power of the test. This may be true, but if single-unit responses were moderately or strongly sensitive to azimuth, one would expect them to pass the test even with relatively low statistical power. At best, if their statistical test missed some azimuthsensitive units, they were likely only weakly sensitive to azimuth. The authors went on to perform a second test of azimuth sensitivity-a chi-squared test-and found 32% (imaging) and 40% (e-phys) of single units to have statistically significant sensitivity. This feels a bit like fishing for a lower p-value. The Kruskal-Wallis test should have been left as the only analysis. Moreover, the use of a chi-squared test is questionable because it is meant to be used between two categorical variables, and neural response had to be binned before applying the test.

The determination of what is a physiologically relevant “moderate or strong azimuth sensitivity” is not trivial, particularly when comparing tuning across different relays of the auditory pathway like the CNIC, auditory cortex, or in our case DCIC, where physiologically relevant azimuth sensitivities might be different. This is likely the reason why azimuth sensitivity has been defined in diverse ways across the bibliography (see Groh, Kelly & Underhill, 2003 for an early discussion of this issue). These diverse approaches include reaching a certain percentage of maximal response modulation, like used by Day et al. (2012, 2015, 2016) in CNIC, and ANOVA tests, like used by Panniello et al. (2018) and Groh, Kelly & Underhill (2003) in auditory cortex and IC respectively. Moreover, the influence of response variability and biases in response distribution estimation due to limited sampling has not been usually accounted for in the determination of azimuth sensitivity.

As Reviewer #1 points out, in our study we used an appropriate ANOVA test (KruskalWallis) as a starting point to study response sensitivity to stimulus azimuth at DCIC. Please note that the alpha = 0.05 used for this test is not based on experimental evidence about physiologically relevant azimuth sensitivity but instead is an arbitrary p-value threshold. Using this test on the electrophysiological data, we found that ~ 21% of the simultaneously recorded single units reached significance (n = 4 mice). Nevertheless these percentages, in our small sample size (n = 4) were not significantly different from our false positive detection rate (p = 0.0625, Mann-Whitney, See Author response image 1).  In consequence, for both our imaging (Fig. 3C) and electrophysiological data, we could not ascertain if the percentage of neurons reaching significance in these ANOVA tests were indeed meaningfully sensitive to azimuth or this was due to chance.

Author response image 1.

Percentage of the neuropixels recorded DCIC single units across mice that showed significant median response tuning, compared to false positive detection rate (α = 0.05, chance level).

We reasoned that the observed markedly variable responses from DCIC units, which frequently failed to respond in many trials (Fig. 3D, 4A), in combination with the limited number of trial repetitions we could collect, results in under-sampled response distribution estimations. This under-sampling can bias the determination of stochastic dominance across azimuth response samples in Kruskal-Wallis tests. We would like to highlight that we decided not to implement resampling strategies to artificially increase the azimuth response sample sizes with “virtual trials”, in order to avoid “fishing for a smaller p-value”, when our collected samples might not accurately reflect the actual response population variability.

As an alternative to hypothesis testing based on ranking and determining stochastic dominance of one or more azimuth response samples (Kruskal-Wallis test), we evaluated the overall statistical dependency to stimulus azimuth of the collected responses.  To do this we implement the Chi-square test by binning neuronal responses into categories. Binning responses into categories can reduce the influence of response variability to some extent, which constitutes an advantage of the Chi-square approach, but we note the important consideration that these response categories are arbitrary.

Altogether, we acknowledge that our Chi-square approach to define azimuth sensitivity is not free of limitations and despite enabling the interrogation of azimuth sensitivity at DCIC, its interpretability might not extend to other brain regions like CNIC or auditory cortex. Nevertheless we hope the aforementioned arguments justify why the Kruskal-Wallis test simply could not “have been left as the only analysis”.

(3) Single-trial population responses encode sound source azimuth "effectively" in that localization decoding error matches average mouse discrimination thresholds - If only one neuron in a population had responses that were sensitive to azimuth, we would expect that decoding azimuth from observation of that one neuron's response would perform better than chance. By observing the responses of more than one neuron (if more than one were sensitive to azimuth), we would expect performance to increase. The authors found that decoding from the whole population response was no better than chance. They argue (reasonably) that this is because of overfitting of the decoder modeltoo few trials used to fit too many parameters-and provide evidence from decoding combined with principal components analysis which suggests that overfitting is occurring. What is troubling is the performance of the decoder when using only a handful of "topranked" neurons (in terms of azimuth sensitivity) (Fig. 4F and G). Decoder performance seems to increase when going from one to two neurons, then decreases when going from two to three neurons, and doesn't get much better for more neurons than for one neuron alone. It seems likely there is more information about azimuth in the population response, but decoder performance is not able to capture it because spike count distributions in the decoder model are not being accurately estimated due to too few stimulus trials (14, on average). In other words, it seems likely that decoder performance is underestimating the ability of the DCIC population to encode sound source azimuth.

To get a sense of how effective a neural population is at coding a particular stimulus parameter, it is useful to compare population decoder performance to psychophysical performance. Unfortunately, mouse behavioral localization data do not exist. Therefore, the authors compare decoder error to mouse left-right discrimination thresholds published previously by a different lab. However, this comparison is inappropriate because the decoder and the mice were performing different perceptual tasks. The decoder is classifying sound sources to 1 of 13 locations from left to right, whereas the mice were discriminating between left or right sources centered around zero degrees. The errors in these two tasks represent different things. The two data sets may potentially be more accurately compared by extracting information from the confusion matrices of population decoder performance. For example, when the stimulus was at -30 deg, how often did the decoder classify the stimulus to a lefthand azimuth? Likewise, when the stimulus was +30 deg, how often did the decoder classify the stimulus to a righthand azimuth?

The azimuth discrimination error reported by Lauer et al. (2011) comes from engaged and highly trained mice, which is a very different context to our experimental setting with untrained mice passively listening to stimuli from 13 random azimuths. Therefore we did not perform analyses or interpretations of our results based on the behavioral task from Lauer et al. (2011) and only made the qualitative observation that the errors match for discussion.

We believe it is further important to clarify that Lauer et al. (2011) tested the ability of mice to discriminate between a positively conditioned stimulus (reference speaker at 0º center azimuth associated to a liquid reward) and a negatively conditioned stimulus (coming from one of five comparison speakers positioned at 20º, 30º, 50º, 70 and 90º azimuth, associated to an electrified lickport) in a conditioned avoidance task. In this task, mice are not precisely “discriminating between left or right sources centered around zero degrees”, making further analyses to compare the experimental design of Lauer et al (2011) and ours even more challenging for valid interpretation.

(4) DCIC can encode sound source azimuth in a similar format to that in the central nucleus of the inferior colliculus - It is unclear what exactly the authors mean by this statement in the Abstract. There are major differences in the encoding of azimuth between the two neighboring brain areas: a large majority of neurons in the CNIC are sensitive to azimuth (and strongly so), whereas the present study shows a minority of azimuth-sensitive neurons in the DCIC. Furthermore, CNIC neurons fire reliably to sound stimuli (low neural noise), whereas the present study shows that DCIC neurons fire more erratically (high neural noise).

Since sound source azimuth is reported to be encoded by population activity patterns at CNIC (Day and Delgutte, 2013), we refer to a population activity pattern code as the “similar format” in which this information is encoded at DCIC. Please note that this is a qualitative comparison and we do not claim this is the “same format”, due to the differences the reviewer precisely describes in the encoding of azimuth at CNIC where a much larger majority of neurons show stronger azimuth sensitivity and response reliability with respect to our observations at DCIC. By this qualitative similarity of encoding format we specifically mean the similar occurrence of activity patterns from azimuth sensitive subpopulations of neurons in both CNIC and DCIC, which carry sufficient information about the stimulus azimuth for a sufficiently accurate prediction with regard to the behavioral discrimination ability.

(5) Evidence of noise correlation between pairs of neurons exists - The authors' data and analyses seem appropriate and sufficient to justify this claim.

(6) Noise correlations between responses of neurons help reduce population decoding error - The authors show convincing analysis that performance of their decoder increased when simultaneously measured responses were tested (which include noise correlation) than when scrambled-trial responses were tested (eliminating noise correlation). This makes it seem likely that noise correlation in the responses improved decoder performance. The authors mention that the naïve Bayesian classifier was used as their decoder for computational efficiency, presumably because it assumes no noise correlation and, therefore, assumes responses of individual neurons are independent of each other across trials to the same stimulus. The use of decoder that assumes independence seems key here in testing the hypothesis that noise correlation contains information about sound source azimuth. The logic of using this decoder could be more clearly spelled out to the reader. For example, if the null hypothesis is that noise correlations do not carry azimuth information, then a decoder that assumes independence should perform the same whether population responses are simultaneous or scrambled. The authors' analysis showing a difference in performance between these two cases provides evidence against this null hypothesis.

We sincerely thank the reviewer for this careful and detailed consideration of our analysis approach. Following the reviewer’s constructive suggestion, we justified the decoder choice in the results section at the last paragraph of page 18:

“To characterize how the observed positive noise correlations could affect the representation of stimulus azimuth by DCIC top ranked unit population responses, we compared the decoding performance obtained by classifying the single-trial response patterns from top ranked units in the modeled decorrelated datasets versus the acquired data (with noise correlations). With the intention to characterize this with a conservative approach that would be less likely to find a contribution of noise correlations as it assumes response independence, we relied on the naive Bayes classifier for decoding throughout the study.

Using this classifier, we observed that the modeled decorrelated datasets produced stimulus azimuth prediction error distributions that were significantly shifted towards higher decoding errors (Fig. 5B, C) and, in our imaging datasets, were not significantly different from chance level (Fig. 5B). Altogether, these results suggest that the detected noise correlations in our simultaneously acquired datasets can help reduce the error of the IC population code for sound azimuth.”

Minor weakness:

- Most studies of neural encoding of sound source azimuth are done in a noise-free environment, but the experimental setup in the present study had substantial background noise. This complicates comparison of the azimuth tuning results in this study to those of other studies. One is left wondering if azimuth sensitivity would have been greater in the absence of background noise, particularly for the imaging data where the signal was only about 12 dB above the noise. The description of the noise level and signal + noise level in the Methods should be made clearer. Mice hear from about 2.5 - 80 kHz, so it is important to know the noise level within this band as well as specifically within the band overlapping with the signal.

We agree with the reviewer that this information is useful. In our study, the background R.M.S. SPL during imaging across the mouse hearing range (2.5-80kHz) was 44.53 dB and for neuropixels recordings 34.68 dB. We have added this information to the methods section of the revised manuscript.

Reviewer #2 (Public Review):

In the present study, Boffi et al. investigate the manner in which the dorsal cortex of the of the inferior colliculus (DCIC), an auditory midbrain area, encodes sound location azimuth in awake, passively listening mice. By employing volumetric calcium imaging (scanned temporal focusing or s-TeFo), complemented with high-density electrode electrophysiological recordings (neuropixels probes), they show that sound-evoked responses are exquisitely noisy, with only a small portion of neurons (units) exhibiting spatial sensitivity. Nevertheless, a naïve Bayesian classifier was able to predict the presented azimuth based on the responses from small populations of these spatially sensitive units. A portion of the spatial information was provided by correlated trial-to-trial response variability between individual units (noise correlations). The study presents a novel characterization of spatial auditory coding in a non-canonical structure, representing a noteworthy contribution specifically to the auditory field and generally to systems neuroscience, due to its implementation of state-of-the-art techniques in an experimentally challenging brain region. However, nuances in the calcium imaging dataset and the naïve Bayesian classifier warrant caution when interpreting some of the results.

Strengths:

The primary strength of the study lies in its methodological achievements, which allowed the authors to collect a comprehensive and novel dataset. While the DCIC is a dorsal structure, it extends up to a millimetre in depth, making it optically challenging to access in its entirety. It is also more highly myelinated and vascularised compared to e.g., the cerebral cortex, compounding the problem. The authors successfully overcame these challenges and present an impressive volumetric calcium imaging dataset. Furthermore, they corroborated this dataset with electrophysiological recordings, which produced overlapping results. This methodological combination ameliorates the natural concerns that arise from inferring neuronal activity from calcium signals alone, which are in essence an indirect measurement thereof.

Another strength of the study is its interdisciplinary relevance. For the auditory field, it represents a significant contribution to the question of how auditory space is represented in the mammalian brain. "Space" per se is not mapped onto the basilar membrane of the cochlea and must be computed entirely within the brain. For azimuth, this requires the comparison between miniscule differences between the timing and intensity of sounds arriving at each ear. It is now generally thought that azimuth is initially encoded in two, opposing hemispheric channels, but the extent to which this initial arrangement is maintained throughout the auditory system remains an open question. The authors observe only a slight contralateral bias in their data, suggesting that sound source azimuth in the DCIC is encoded in a more nuanced manner compared to earlier processing stages of the auditory hindbrain. This is interesting, because it is also known to be an auditory structure to receive more descending inputs from the cortex.

Systems neuroscience continues to strive for the perfection of imaging novel, less accessible brain regions. Volumetric calcium imaging is a promising emerging technique, allowing the simultaneous measurement of large populations of neurons in three dimensions. But this necessitates corroboration with other methods, such as electrophysiological recordings, which the authors achieve. The dataset moreover highlights the distinctive characteristics of neuronal auditory representations in the brain. Its signals can be exceptionally sparse and noisy, which provide an additional layer of complexity in the processing and analysis of such datasets. This will be undoubtedly useful for future studies of other less accessible structures with sparse responsiveness.

Weaknesses:                                                                                               

Although the primary finding that small populations of neurons carry enough spatial information for a naïve Bayesian classifier to reasonably decode the presented stimulus is not called into question, certain idiosyncrasies, in particular the calcium imaging dataset and model, complicate specific interpretations of the model output, and the readership is urged to interpret these aspects of the study's conclusions with caution.

I remain in favour of volumetric calcium imaging as a suitable technique for the study, but the presently constrained spatial resolution is insufficient to unequivocally identify regions of interest as cell bodies (and are instead referred to as "units" akin to those of electrophysiological recordings). It remains possible that the imaging set is inadvertently influenced by non-somatic structures (including neuropil), which could report neuronal activity differently than cell bodies. Due to the lack of a comprehensive ground-truth comparison in this regard (which to my knowledge is impossible to achieve with current technology), it is difficult to imagine how many informative such units might have been missed because their signals were influenced by spurious, non-somatic signals, which could have subsequently misled the models. The authors reference the original Nature Methods article (Prevedel et al., 2016) throughout the manuscript, presumably in order to avoid having to repeat previously published experimental metrics. But the DCIC is neither the cortex nor hippocampus (for which the method was originally developed) and may not have the same light scattering properties (not to mention neuronal noise levels). Although the corroborative electrophysiology data largely eleviates these concerns for this particular study, the readership should be cognisant of such caveats, in particular those who are interested in implementing the technique for their own research.

A related technical limitation of the calcium imaging dataset is the relatively low number of trials (14) given the inherently high level of noise (both neuronal and imaging). Volumetric calcium imaging, while offering a uniquely expansive field of view, requires relatively high average excitation laser power (in this case nearly 200 mW), a level of exposure the authors may have wanted to minimise by maintaining a low the number of repetitions, but I yield to them to explain.

We assumed that the levels of heating by excitation light measured at the neocortex in Prevedel et al. (2016), were representative for DCIC also. Nevertheless, we recognize this approximation might not be very accurate, due to the differences in tissue architecture and vascularization from these two brain areas, just to name a few factors. The limiting factor preventing us from collecting more trials in our imaging sessions was that we observed signs of discomfort or slight distress in some mice after ~30 min of imaging in our custom setup, which we established as a humane end point to prevent distress. In consequence imaging sessions were kept to 25 min in duration, limiting the number of trials collected. However we cannot rule out that with more extensive habituation prior to experiments the imaging sessions could be prolonged without these signs of discomfort or if indeed influence from our custom setup like potential heating of the brain by illumination light might be the causing factor of the observed distress. Nevertheless, we note that previous work has shown that ~200mW average power is a safe regime for imaging in the cortex by keeping brain heating minimal (Prevedel et al., 2016), without producing the lasting damages observed by immunohistochemisty against apoptosis markers above 250mW (Podgorski and Ranganathan 2016, https://doi.org/10.1152/jn.00275.2016).

Calcium imaging is also inherently slow, requiring relatively long inter-stimulus intervals (in this case 5 s). This unfortunately renders any model designed to predict a stimulus (in this case sound azimuth) from particularly noisy population neuronal data like these as highly prone to overfitting, to which the authors correctly admit after a model trained on the entire raw dataset failed to perform significantly above chance level. This prompted them to feed the model only with data from neurons with the highest spatial sensitivity. This ultimately produced reasonable performance (and was implemented throughout the rest of the study), but it remains possible that if the model was fed with more repetitions of imaging data, its performance would have been more stable across the number of units used to train it. (All models trained with imaging data eventually failed to converge.) However, I also see these limitations as an opportunity to improve the technology further, which I reiterate will be generally important for volume imaging of other sparse or noisy calcium signals in the brain.

Transitioning to the naïve Bayesian classifier itself, I first openly ask the authors to justify their choice of this specific model. There are countless types of classifiers for these data, each with their own pros and cons. Did they actually try other models (such as support vector machines), which ultimately failed? If so, these negative results (even if mentioned en passant) would be extremely valuable to the community, in my view. I ask this specifically because different methods assume correspondingly different statistical properties of the input data, and to my knowledge naïve Bayesian classifiers assume that predictors (neuronal responses) are assumed to be independent within a class (azimuth). As the authors show that noise correlations are informative in predicting azimuth, I wonder why they chose a model that doesn't take advantage of these statistical regularities. It could be because of technical considerations (they mention computing efficiency), but I am left generally uncertain about the specific logic that was used to guide the authors through their analytical journey.

One of the main reasons we chose the naïve Bayesian classifier is indeed because it assumes that the responses of the simultaneously recorded neurons are independent and therefore it does not assume a contribution of noise correlations to the estimation of the posterior probability of each azimuth. This model would represent the null hypothesis that noise correlations do not contribute to the encoding of stimulus azimuth, which would be verified by an equal decoding outcome from correlated or decorrelated datasets. Since we observed that this is not the case, the model supports the alternative hypothesis that noise correlations do indeed influence stimulus azimuth encoding. We wanted to test these hypotheses with the most conservative approach possible that would be least likely to find a contribution of noise correlations. Other relevant reasons that justify our choice of the naive Bayesian classifier are its robustness against the limited numbers of trials we could collect in comparison to other more “data hungry” classifiers like SVM, KNN, or artificial neuronal nets. We did perform preliminary tests with alternative classifiers but the obtained decoding errors were similar when decoding the whole population activity (Supplemental figure 3A). Dimensionality reduction following the approach described in the manuscript showed a tendency towards smaller decoding errors observed with an alternative classifier like KNN, but these errors were still larger than the ones observed with the naive Bayesian classifier (median error 45º). Nevertheless, we also observe a similar tendency for slightly larger decoding errors in the absence of noise correlations (decorrelated, Supplemental figure 3B). Sentences detailing the logic of classifier choice are now included in the results section at page 10 and at the last paragraph of page 18 (see responses to Reviewer 1).

That aside, there remain other peculiarities in model performance that warrant further investigation. For example, what spurious features (or lack of informative features) in these additional units prevented the models of imaging data from converging?

Considering the amount of variability observed throughout the neuronal responses both in imaging and neuropixels datasets, it is easy to suspect that the information about stimulus azimuth carried in different amounts by individual DCIC neurons can be mixed up with information about other factors (Stringer et al., 2019). In an attempt to study the origin of these features that could confound stimulus azimuth decoding we explored their relation to face movement (Supplemental Figure 2), finding a correlation to snout movements, in line with previous work by Stringer et al. (2019).

In an orthogonal question, did the most spatially sensitive units share any detectable tuning features? A different model trained with electrophysiology data in contrast did not collapse in the range of top-ranked units plotted. Did this model collapse at some point after adding enough units, and how well did that correlate with the model for the imaging data?

Our electrophysiology datasets were much smaller in size (number of simultaneously recorded neurons) compared to our volumetric calcium imaging datasets, resulting in a much smaller total number of top ranked units detected per dataset. This precluded the determination of a collapse of decoder performance due to overfitting beyond the range plotted in Fig 4G.

How well did the form (and diversity) of the spatial tuning functions as recorded with electrophysiology resemble their calcium imaging counterparts? These fundamental questions could be addressed with more basic, but transparent analyses of the data (e.g., the diversity of spatial tuning functions of their recorded units across the population). Even if the model extracts features that are not obvious to the human eye in traditional visualisations, I would still find this interesting.

The diversity of the azimuth tuning curves recorded with calcium imaging (Fig. 3B) was qualitatively larger than the ones recorded with electrophysiology (Fig. 4B), potentially due to the larger sampling obtained with volumetric imaging. We did not perform a detailed comparison of the form and a more quantitative comparison of the diversity of these functions because the signals compared are quite different, as calcium indicator signal is subject to non linearities due to Ca2+ binding cooperativity and low pass filtering due to binding kinetics. We feared this could lead to misleading interpretations about the similarities or differences between the azimuth tuning functions in imaged and electrophysiology datasets. Our model uses statistical response dependency to stimulus azimuth, which does not rely on features from a descriptive statistic like mean response tuning. In this context, visualizing the trial-to-trial responses as a function of azimuth shows “features that are not obvious to the human eye in traditional visualizations” (Fig. 3D, left inset).

Finally, the readership is encouraged to interpret certain statements by the authors in the current version conservatively. How the brain ultimately extracts spatial neuronal data for perception is anyone's guess, but it is important to remember that this study only shows that a naïve Bayesian classifier could decode this information, and it remains entirely unclear whether the brain does this as well. For example, the model is able to achieve a prediction error that corresponds to the psychophysical threshold in mice performing a discrimination task (~30 {degree sign}). Although this is an interesting coincidental observation, it does not mean that the two metrics are necessarily related. The authors correctly do not explicitly claim this, but the manner in which the prose flows may lead a non-expert into drawing that conclusion.

To avoid misleading the non-expert readers, we have clarified in the manuscript that the observed correspondence between decoding error and psychophysical threshold is explicitly coincidental.

Page 13, end of middle paragraph:

“If we consider the median of the prediction error distribution as an overall measure of decoding performance, the single-trial response patterns from subsamples of at least the 7 top ranked units produced median decoding errors that coincidentally matched the reported azimuth discrimination ability of mice (Fig 4G, minimum audible angle = 31º) (Lauer et al., 2011).”

Page 14, bottom paragraph:

“Decoding analysis (Fig. 4F) of the population response patterns from azimuth dependent top ranked units simultaneously recorded with neuropixels probes showed that the 4 top ranked units are the smallest subsample necessary to produce a significant decoding performance that coincidentally matches the discrimination ability of mice (31° (Lauer et al., 2011)) (Fig. 5F, G).”

We also added to the Discussion sentences clarifying that a relationship between these two variables remains to be determined and it also remains to be determined if the DCIC indeed performs a bayesian decoding computation for sound localization.

Page 20, bottom:

“… Concretely, we show that sound location coding does indeed occur at DCIC on the single trial basis, and that this follows a comparable mechanism to the characterized population code at CNIC (Day and Delgutte, 2013). However, it remains to be determined if indeed the DCIC network is physiologically capable of Bayesian decoding computations. Interestingly, the small number of DCIC top ranked units necessary to effectively decode stimulus azimuth suggests that sound azimuth information is redundantly distributed across DCIC top ranked units, which points out that mechanisms beyond coding efficiency could be relevant for this population code.

While the decoding error observed from our DCIC datasets obtained in passively listening, untrained mice coincidentally matches the discrimination ability of highly trained, motivated mice (Lauer et al., 2011), a relationship between decoding error and psychophysical performance remains to be determined. Interestingly, a primary sensory representations should theoretically be even more precise than the behavioral performance as reported in the visual system (Stringer et al., 2021).”

Moreover, the concept of redundancy (of spatial information carried by units throughout the DCIC) is difficult for me to disentangle. One interpretation of this formulation could be that there are non-overlapping populations of neurons distributed across the DCIC that each could predict azimuth independently of each other, which is unlikely what the authors meant. If the authors meant generally that multiple neurons in the DCIC carry sufficient spatial information, then a single neuron would have been able to predict sound source azimuth, which was not the case. I have the feeling that they actually mean "complimentary", but I leave it to the authors to clarify my confusion, should they wish.

We observed that the response patterns from relatively small fractions of the azimuth sensitive DCIC units (4-7 top ranked units) are sufficient to generate an effective code for sound azimuth, while 32-40% of all simultaneously recorded DCIC units are azimuth sensitive. In light of this observation, we interpreted that the azimuth information carried by the population should be redundantly distributed across the complete subpopulation of azimuth sensitive DCIC units.

In summary, the present study represents a significant body of work that contributes substantially to the field of spatial auditory coding and systems neuroscience. However, limitations of the imaging dataset and model as applied in the study muddles concrete conclusions about how the DCIC precisely encodes sound source azimuth and even more so to sound localisation in a behaving animal. Nevertheless, it presents a novel and unique dataset, which, regardless of secondary interpretation, corroborates the general notion that auditory space is encoded in an extraordinarily complex manner in the mammalian brain.

Reviewer #3 (Public Review):

Summary: Boffi and colleagues sought to quantify the single-trial, azimuthal information in the dorsal cortex of the inferior colliculus (DCIC), a relatively understudied subnucleus of the auditory midbrain. They used two complementary recording methods while mice passively listened to sounds at different locations: a large volume but slow sampling calcium-imaging method, and a smaller volume but temporally precise electrophysiology method. They found that neurons in the DCIC were variable in their activity, unreliably responding to sound presentation and responding during inter-sound intervals. Boffi and colleagues used a naïve Bayesian decoder to determine if the DCIC population encoded sound location on a single trial. The decoder failed to classify sound location better than chance when using the raw single-trial population response but performed significantly better than chance when using intermediate principal components of the population response. In line with this, when the most azimuth dependent neurons were used to decode azimuthal position, the decoder performed equivalently to the azimuthal localization abilities of mice. The top azimuthal units were not clustered in the DCIC, possessed a contralateral bias in response, and were correlated in their variability (e.g., positive noise correlations). Interestingly, when these noise correlations were perturbed by inter-trial shuffling decoding performance decreased. Although Boffi and colleagues display that azimuthal information can be extracted from DCIC responses, it remains unclear to what degree this information is used and what role noise correlations play in azimuthal encoding.

Strengths: The authors should be commended for collection of this dataset. When done in isolation (which is typical), calcium imaging and linear array recordings have intrinsic weaknesses. However, those weaknesses are alleviated when done in conjunction with one another - especially when the data largely recapitulates the findings of the other recording methodology. In addition to the video of the head during the calcium imaging, this data set is extremely rich and will be of use to those interested in the information available in the DCIC, an understudied but likely important subnucleus in the auditory midbrain.

The DCIC neural responses are complex; the units unreliably respond to sound onset, and at the very least respond to some unknown input or internal state (e.g., large inter-sound interval responses). The authors do a decent job in wrangling these complex responses: using interpretable decoders to extract information available from population responses.

Weaknesses:

The authors observe that neurons with the most azimuthal sensitivity within the DCIC are positively correlated, but they use a Naïve Bayesian decoder which assume independence between units. Although this is a bit strange given their observation that some of the recorded units are correlated, it is unlikely to be a critical flaw. At one point the authors reduce the dimensionality of their data through PCA and use the loadings onto these components in their decoder. PCA incorporates the correlational structure when finding the principal components and constrains these components to be orthogonal and uncorrelated. This should alleviate some of the concern regarding the use of the naïve Bayesian decoder because the projections onto the different components are independent. Nevertheless, the decoding results are a bit strange, likely because there is not much linearly decodable azimuth information in the DCIC responses. Raw population responses failed to provide sufficient information concerning azimuth for the decoder to perform better than chance. Additionally, it only performed better than chance when certain principal components or top ranked units contributed to the decoder but not as more components or units were added. So, although there does appear to be some azimuthal information in the recoded DCIC populations - it is somewhat difficult to extract and likely not an 'effective' encoding of sound localization as their title suggests.

As described in the responses to reviewers 1 and 2, we chose the naïve Bayes classifier as a decoder to determine the influence of noise correlations through the most conservative approach possible, as this classifier would be least likely to find a contribution of correlated noise. Also, we chose this decoder due to its robustness against limited numbers of trials collected, in comparison to “data hungry” non linear classifiers like KNN or artificial neuronal nets. Lastly, we observed that small populations of noisy, unreliable (do not respond in every trial) DCIC neurons can encode stimulus azimuth in passively listening mice matching the discrimination error of trained mice. Therefore, while this encoding is definitely not efficient, it can still be considered effective.

Although this is quite a worthwhile dataset, the authors present relatively little about the characteristics of the units they've recorded. This may be due to the high variance in responses seen in their population. Nevertheless, the authors note that units do not respond on every trial but do not report what percent of trials that fail to evoke a response. Is it that neurons are noisy because they do not respond on every trial or is it also that when they do respond they have variable response distributions? It would be nice to gain some insight into the heterogeneity of the responses.

The limited number of azimuth trial repetitions that we could collect precluded us from making any quantification of the unreliability (failures to respond) and variability in the response distributions from the units we recorded, as we feared they could be misleading. In qualitative terms, “due to the high variance in responses seen” in the recordings and the limited trial sampling, it is hard to make any generalization. In consequence we referred to the observed response variance altogether as neuronal noise. Considering these points, our datasets are publicly available for exploration of the response characteristics.

Additionally, is there any clustering at all in response profiles or is each neuron they recorded in the DCIC unique?

We attempted to qualitatively visualize response clustering using dimensionality reduction, observing different degrees of clustering or lack thereof across the azimuth classes in the datasets collected from different mice. It is likely that the limited number of azimuth trials we could collect and the high response variance contribute to an inconsistent response clustering across datasets.

They also only report the noise correlations for their top ranked units, but it is possible that the noise correlations in the rest of the population are different.

For this study, since our aim was to interrogate the influence of noise correlations on stimulus azimuth encoding by DCIC populations, we focused on the noise correlations from the top ranked unit subpopulation, which likely carry the bulk of the sound location information.  Noise correlations can be defined as correlation in the trial to trial response variation of neurons. In this respect, it is hard to ascertain if the rest of the population, that is not in the top rank unit percentage, are really responding and showing response variation to evaluate this correlation, or are simply not responding at all and show unrelated activity altogether. This makes observations about noise correlations from “the rest of the population” potentially hard to interpret.

It would also be worth digging into the noise correlations more - are units positively correlated because they respond together (e.g., if unit x responds on trial 1 so does unit y) or are they also modulated around their mean rates on similar trials (e.g., unit x and y respond and both are responding more than their mean response rate). A large portion of trial with no response can occlude noise correlations. More transparency around the response properties of these populations would be welcome.

Due to the limited number of azimuth trial repetitions collected, to evaluate noise correlations we used the non parametric Kendall tau correlation coefficient which is a measure of pairwise rank correlation or ordinal association in the responses to each azimuth. Positive rank correlation would represent neurons more likely responding together. Evaluating response modulation “around their mean rates on similar trials” would require assumptions about the response distributions, which we avoided due to the potential biases associated with limited sample sizes.

It is largely unclear what the DCIC is encoding. Although the authors are interested in azimuth, sound location seems to be only a small part of DCIC responses. The authors report responses during inter-sound interval and unreliable sound-evoked responses. Although they have video of the head during recording, we only see a correlation to snout and ear movements (which are peculiar since in the example shown it seems the head movements predict the sound presentation). Additional correlates could be eye movements or pupil size. Eye movement are of particular interest due to their known interaction with IC responses - especially if the DCIC encodes sound location in relation to eye position instead of head position (though much of eye-position-IC work was done in primates and not rodent). Alternatively, much of the population may only encode sound location if an animal is engaged in a localization task. Ideally, the authors could perform more substantive analyses to determine if this population is truly noisy or if the DCIC is integrating un-analyzed signals.

We unsuccessfully attempted eye tracking and pupillometry in our videos. We suspect that the reason behind this is a generally overly dilated pupil due to the low visible light illumination conditions we used which were necessary to protect the PMT of our custom scope.

It is likely that DCIC population activity is integrating un-analyzed signals, like the signal associated with spontaneous behaviors including face movements (Stringer et al., 2019), which we observed at the level of spontaneous snout movements. However investigating if and how these signals are integrated to stimulus azimuth coding requires extensive behavioral testing and experimentation which is out of the scope of this study. For the purpose of our study, we referred to trial-to-trial response variation as neuronal noise. We note that this definition of neuronal noise can, and likely does, include an influence from un-analyzed signals like the ones from spontaneous behaviors.

Although this critique is ubiquitous among decoding papers in the absence of behavioral or causal perturbations, it is unclear what - if any - role the decoded information may play in neuronal computations. The interpretation of the decoder means that there is some extractable information concerning sound azimuth - but not if it is functional. This information may just be epiphenomenal, leaking in from inputs, and not used in computation or relayed to downstream structures. This should be kept in mind when the authors suggest their findings implicate the DCIC functionally in sound localization.

Our study builds upon previous reports by other independent groups relying on “causal and behavioral perturbations” and implicating DCIC in sound location learning induced experience dependent plasticity (Bajo et al., 2019, 2010; Bajo and King, 2012), which altogether argues in favor of DCIC functionality in sound localization.

Nevertheless, we clarified in the discussion of the revised manuscript that a relationship between the observed decoding error and the psychophysical performance, or the ability of the DCIC network to perform Bayesian decoding computations, both remain to be determined (please see responses to Reviewer #2).

It is unclear why positive noise correlations amongst similarly tuned neurons would improve decoding. A toy model exploring how positive noise correlations in conjunction with unreliable units that inconsistently respond may anchor these findings in an interpretable way. It seems plausible that inconsistent responses would benefit from strong noise correlations, simply by units responding together. This would predict that shuffling would impair performance because you would then be sampling from trials in which some units respond, and trials in which some units do not respond - and may predict a bimodal performance distribution in which some trials decode well (when the units respond) and poor performance (when the units do not respond).

In samples with more that 2 dimensions, the relationship between signal and noise correlations is more complex than in two dimensional samples (Montijn et al., 2016) which makes constructing interpretable and simple toy models of this challenging. Montijn et al. (2016) provide a detailed characterization and model describing how the accuracy of a multidimensional population code can improve when including “positive noise correlations amongst similarly tuned neurons”. Unfortunately we could not successfully test their model based on Mahalanobis distances as we could not verify that the recorded DCIC population responses followed a multivariate gaussian distribution, due to the limited azimuth trial repetitions we could sample.

Significance: Boffi and colleagues set out to parse the azimuthal information available in the DCIC on a single trial. They largely accomplish this goal and are able to extract this information when allowing the units that contain more information about sound location to contribute to their decoding (e.g., through PCA or decoding on top unit activity specifically). The dataset will be of value to those interested in the DCIC and also to anyone interested in the role of noise correlations in population coding. Although this work is first step into parsing the information available in the DCIC, it remains difficult to interpret if/how this azimuthal information is used in localization behaviors of engaged mice.

Author Response:

The following is the authors’ response to the original reviews.

eLife assessment

This study presents potentially useful findings describing how activity in the corticotropin-releasing hormone neurons in the paraventricular nucleus of the hypothalamus modulates sevoflurane anesthesia, as well as a phenomenon the authors term a "general anesthetic stress response". The technical approaches are solid and the data presented are largely clear. However, the primary conclusion, that the PVHCRH neurons are a mechanism of sevoflurane anesthesia, is inadequately supported.

We appreciate the editors and reviewers for their thorough assessment and constructive feedback. We have provided clarifications and updated the manuscripts to better interpret our results, please see below. As for the primary conclusion, we revised it as PVH CRH neurons potently modulate states of anaesthesia in sevoflurane general anesthesia, being a part of anaesthesia regulatory network of sevoflurane.

Combined Public Review:

This study describes a group of CRH-releasing neurons, located in the paraventricular nucleus of the hypothalamus, which, in mice, affects both the state of sevoflurane anesthesia and a grooming behavior observed after it. PVH-CRH neurons showed elevated calcium activity during the post-anesthesia period. Optogenetic activation of these PVH-CRH neurons during sevoflurane anesthesia shifts the EEG from burst-suppression to a seemingly activated state (an apparent arousal effect), although without a behavioral correlate. Chemogenetic activation of the PVH-CRH neurons delays sevoflurane-induced loss of righting reflex (another apparent arousal effect). On the other hand, chemogenetic inhibition of PVH-CRH neurons delays recovery of the righting reflex and decreases sevoflurane-induced stress (an apparent decrease in the arousal effect). The authors conclude that PVH-CRH neurons are a common substrate for sevoflurane-induced anesthesia and stress. The PVH-CRH neurons are related to behavioral stress responses, and the authors claim that these findings provide direct evidence for a relationship between sevoflurane anesthesia and sevoflurane-mediated stress that might exist even when there is no surgical trauma, such as an incision. In its current form, the article does not achieve its intended goal.

Thank you for the detailed review. We have carefully considered your comments and have revised the manuscript to provide a clearer interpretation of our findings. Our findings indicate that PVH CRH neurons integrate the anesthetic effect and post-anesthesia stress response of sevoflurane (GA), providing new evidence for understanding the neuronal regulation of sevoflurane GA and identifying a potential brain target for further investigation into modulating the post-anesthesia stress response. However, we did not propose that there was a direct relationship between sevoflurane anesthesia and sevoflurane-mediated stress in the absence of incision. Our results mainly concluded that PVH CRH neurons integrate the anaesthetic effect and post-anaesthesia stress response of sevoflurane GA, which offers new evidence for the neuronal regulation of sevoflurane GA and provides an important but ignored potential cause of the post-anesthesia stress response.

Strengths:

The manuscript uses targeted manipulation of the PVH-CRH neurons, and is technically sound. Also, the number of experiments is substantial.

Thank you.

Weaknesses:

The most significant weaknesses are a) the lack of consideration and measurement of GABAergic mechanisms of sevoflurane anesthesia, b) the failure to use another anesthetic as a control, c) a failure to document a compelling post-anesthesia stress response to sevoflurane in humans, d) limitations in the novelty of the findings. These weaknesses are related to the primary concerns described below:

Concerns about the primary conclusion, that PVH-CRH neurons mediate "the anesthetic effects and post-anesthesia stress response of sevoflurane GA".

Thanks for the advice. Our responses are as below:

1) Just because the activity of a given neural cell type or neural circuit alters an anesthetic's response, this does not mean that those neurons play a role in how the anesthetic creates its anesthetic state. For example, sevoflurane is commonly used in children. Its primary mechanism of action is through enhancement of GABA-mediated inhibition. Children with ADHD on Ritalin (a dopamine reuptake inhibitor) who take it on the day of surgery can often require increased doses of sevoflurane to achieve the appropriate anesthetic state. The mesocortical pathway through which Ritalin acts is not part of the mechanism of action of sevoflurane. Through this pathway, Ritalin is simply increasing cortical excitability making it more challenging for the inhibitory effects of sevoflurane at GABAergic synapses to be effective. Similarly, here, altering the activity of the PVHCRH neurons and seeing a change in anesthetic response to sevoflurane does not mean that these neurons play a role in the fundamental mechanism of this anesthetic's action. With the current data set, the primary conclusions should be tempered.

Thank you for your comments. Our results adequately uncover PVH CRH neurons that modulate the state of consciousness as well as the stress response in sevoflurane GA, but are insufficient to demonstrate that these neurons play a role in the underlying mechanism of sevoflurane anesthesia. We will revise our conclusions and make them concrete. The primary conclusion has been revised as PVH CRH neurons potently modulate states of anaesthesia in sevoflurane GA, being a part of the anaesthesia regulatory network of sevoflurane.

2) It is important to compare the effects of sevoflurane with at least one other inhaled ether anesthetic. Isoflurane, desflurane, and enflurane are ether anesthetics that are very similar to each other, as well as being similar to sevoflurane. It is important to distinguish whether the effects of sevoflurane pertain to other anesthetics, or, alternatively, relate to unique idiosyncratic properties of this gas that may not be a part of its anesthetic properties.

For example, one study cited by the authors (Marana et al.. 2013) concludes that there is weak evidence for differences in stress-related hormones between sevoflurane and desflurane, with lower levels of cortisol and ACTH observed during the desflurane intraoperative period. It is not clear that this difference in some stress-related hormones is modeled by post-sevoflurane excess grooming in the mice, but using desflurane as a control could help determine this.

Thank you for your suggestions. We completely agree on the importance of determining whether the effects of sevoflurane apply to other anesthetics or arise from unique idiosyncratic attributes separate from its anesthetic properties. However, it is challenging to definitively conclude whether the effects of sevoflurane observed in our study extend to other inhaled anesthetics, even with desflurane as a control. While sevoflurane shares many common anesthetic properties with other inhalation agents, it also exhibits distinct characteristics and potential idiosyncrasies that set it apart from its counterparts. Regarding studies related to desflurane's impact on hormone levels or stress-like behaviors, one study involving 20 women scheduled for elective total abdominal hysterectomy demonstrated that there was no significant correlation between the intra-operative depth of anesthesia achieved with desflurane and the extent of the endocrine-metabolic stress response (as indicated by the concentrations of plasma cortisol, glucose, and lactate)1. Besides, a study conducted with mice suggested the abilities related to sensorimotor functions, anxiety and depression did not undergo significant changes after 7 days of anesthesia administered with 8.0% desflurane for 6 h2. Furthermore, a study involving 50 Caucasian women undergoing laparoscopic surgery for benign ovarian cysts demonstrated that in low stress surgery, desflurane, when compared to sevoflurane, exhibited superior control over the intraoperative cortisol and ACTH response 3. Based on these findings, we propose that the effect we observed in this study is likely attributed to the unique idiosyncratic properties of sevoflurane. We will conduct additional experiments to investigate this proposal with other commonly used anaesthetics in our future studies.

Concerns about the clinical relevance of the experiments

In anesthesiology practice, perioperative stress observed in patients is more commonly related to the trauma of the surgical intervention, with inadequate levels of antinociception or unconsciousness intraoperatively and/or poor post-operative pain control. The authors seem to be suggesting that the anesthetic itself is causing stress, but there is no evidence of this from human patients cited. We were not aware that this is a documented clinical phenomenon. It is important to know whether sevoflurane effectively produces behavioral stress in the recovery room in patients that could be related to the putative stress response (excess grooming) observed in mice. For example, in surgeries or procedures that required only a brief period of unconsciousness that could be achieved by administering sevoflurane alone (comparable to the 30 min administered to the mice), is there clinical evidence of post-operative stress?

Thank you for your question. There is currently no direct evidence available. Studies on sevoflurane in humans primarily focus on its use during surgical interventions, making it difficult to find studies that solely administer sevoflurane, as was done in our study with mice. Generally, a short anesthesia time refers to procedures that last less than one hour, while a long anesthesia time could be considered for procedures lasting several hours or more4. A study published in eLife investigated the patterns of reemerging consciousness and cognitive function in 30 healthy adults who underwent GA for three hours 5. This finding suggests that the cognitive dysfunction observed immediately and persistently after GA in healthy animals may not necessarily apply anesthesia and postoperative neurocognitive disorders could be influenced by factors other than GA, such as surgery or patient comorbidity. Therefore, further studies are needed to verify the post-operative stress in sevoflurane-only short time anesthesia.

Indeed, stress after surgeries can result from multiple factors aside from anesthesia, including pain, anxiety, inflammation, but what we want to illustrate in this study is that anesthesia could be one of these factors that we ignored in previous studies. In our current study, we did not propose that there was a direct relationship between sevoflurane anesthesia and sevoflurane-mediated stress without incision. We observed stress-related behavioural changes after exposure of sevoflurane GA in mouse model, indicating sevoflurane-mediated stress might exist without surgical trauma. Importantly, whether anesthetic administration alone will cause post-operative stress is worth studying in different species especially human.

Patients who receive sevoflurane as the primary anesthetic do not wake up more stressed than if they had had one of the other GABAergic anesthetics. If there were signs of stress upon emergence (increased heart rate, blood pressure, thrashing movements) from general anesthesia, the anesthesiologist would treat this right away. The most likely cause of post-operative stress behaviors in humans is probably inadequate anti-nociception during the procedure, which translates into inadequate post-op analgesia and likely delirium. It is the case that children receiving sevoflurane do have a higher likelihood of post-operative delirium. Perhaps the authors' studies address a mechanism for delirium associated with sevoflurane, but this is not considered. Delirium seems likely to be the closest clinical phenomenon to what was studied.

We agree with your idea. We aim to establish a connection between post-operative delirium in humans and stress-like behaviors observed in mice following sevoflurane anesthesia. Specifically, we have observed that the increased grooming behavior exhibited by mice after sevoflurane anesthesia resembles the fuzzy state of consciousness experienced during post-operative delirium6. In our discussion, we also emphasized the occurrence of sevoflurane-induced emergence agitation, a common phenomenon reported in clinical studies with an incidence of up to 80%. This state is characterized by hyperactivity, confusion, delirium, and emotional agitation 7,8. Meanwhile, in our experimental tests, namely the open field test (OFT) and elevated plus maze (EPM) test, we observed that mice exposed to sevoflurane inhalation displayed reduced movement distances during both the OFT and EPM tests (Figure 7G and I). These findings suggest a decline in behavioral activity similar to what is observed in cases of delirium.

Concerns about the novelty of the findings

CRH is associated with arousal in numerous studies. In fact, the authors' own work, published in eLife in 2021, showed that stimulating the hypothalamic CRH cells leads to arousal and their inhibition promotes hypersomnia. In both papers, the authors use fos expression in CRH cells during a specific event to implicate the cells, then manipulate them and measure EEG responses. In the previous work, the cells were active during wakefulness; here- they were active in the awake state that follows anesthesia (Figure 1). Thus, the findings in the current work are incremental.

Thank you for acknowledging our previous work focusing on the changes in the sleep-wake state of mice when PVH CRH neurons are manipulated. In this study, our primary objective was to identify the neuronal mechanisms mediating the anesthetic effects and post-anesthetic stress response of sevoflurane GA. While our study claims that activation of PVH CRH neurons leads to arousal, it provides evidence that PVH CRH neurons may play a role in the regulation of conscious states in GA. Our current findings uncover that PVH CRH neurons modulate the state of consciousness as well as the stress response in sevoflurane GA, and that the modulation of PVH CRH neurons bidirectionally altered the induction and recovery of sevoflurane GA. This identifies a new brain region involved in sevoflurane GA that goes beyond the arousal-related regions.

The activation of CRH cells in PVN has already been shown to result in grooming by Jaideep Bains (cited as reference 58). Thus, the involvement of these cells in this behavior is expected. The authors perform elaborate manipulations of CRH cells and numerous analyses of grooming and related behaviors. For example, they compare grooming and paw licking after anesthesia with those after other stressors such as forced swim, spraying mice with water, physical attack, and restraint. However, the relevance of these behaviors to humans and generalization to other types of anesthetics is not clear.

The hyperactivity of PVH CRH neurons and behavior (e.g., excessive self-grooming) in mice may partially mirror the observed agitation and underlying mechanisms during emergence from sevoflurane GA in patients. As mentioned in the Discussion section (page 16, lines 371-374), sevoflurane-induced emergence agitation represents a prevalent manifestation of the post-anesthesia stress response. It is frequently observed, with an incidence of up to 80% in clinical reports, and is characterized by hyperactivity, confusion, delirium, and emotional agitation7,8. Our aim in this study is to distinguish the excessive stress responses of patients to sevoflurane GA from stress triggered by other factors. Other stimuli, such as forced swimming, can be considered sources of both physical and emotional stress, which are associated with depression and anxiety in humans.

Regarding generalization to other types of anesthetics, we propose that the stress-related behavioral effects observed in this study might occur in cases of the administration of certain types of anesthetics. For example, one study showed that intravenous ketamine infusion (10 mg/kg, 2 hours) elevated plasma corticosterone and progesterone levels in rats, reducing locomotor activity (sedation) 9. The administration of intravenous anesthesia with propofol combined with sevoflurane caused greater postoperative stress than the single use of propofol10. However, desflurane, a common inhaled ether anesthetic, when compared to sevoflurane, was associated with better control of intraoperative cortisol and ACTH response in low-stress surgeries8. Thus, these behaviors observed after exposure to sevoflurane GA may be related to the post-anesthesia stress response in humans, which might also occur in cases of the administration of certain types of anesthetics.

Recommendations for the authors:

Reviewer 1

1) The CRH-Cre mouse line should be validated. There are several lines of these mice, and their fidelity varies.

The CRH-Cre mouse line we used in this study is from The Jackson Laboratory (https://www.jax.org/strain/012704) with the name B6(Cg)-Crhtm1(cre)Zjh/J (Strain #: 012704). These CRH-ires-CRE knock-in mice have Cre recombinase expression directed to CRH positive neurons by the endogenous promoter/enhancer elements of the corticotropin releasing hormone locus (Crh). We have done standard PCR to validate the mouse line following genotyping protocols provided by the Jackson Laboratory. The protocol primers were: 10574 (SEQUENCE 5' → 3': CTT ACA CAT TTC GTC CTA GCC); 10575 (SEQUENCE 5' → 3': CAC GAC CAG GCT GCG GCT AAC); 10576 (SEQUENCE 5' → 3': CAA TGT ATC TTA TCA TGT CTG GAT CC). The 468-bp CRH-specific PCR product was amplified in mutant (CRH-Cre+/+) mice; in heterozygote (CRH-Cre+/-) mice, both the 468-bp and the 676-bp PCR products were detected; in wild type (WT) mice, only the 676-bp WT allele-specific PCR product was amplified. An example of PCR results is presented below. The heterozygote and mutant mice were included in our study.

Author response image 1.

1. It would be very helpful to validate the CRH antibody. Using any antiserum at 1:800 suggests that it may not be potent or highly specific.

As requested, we used the same CRH antibody at a concentration of 1:800, following the methods described in the Method section. The results are displayed below.

Author response image 2.

1. In Figure 1C, the control sections are out of focus, any cells are blurry, reducing confidence in the analyses (locus ceruleus cells appear confluent in the control?)

Sorry for the confusing figure and we have revised the control section part of Figure 1C:

Author response image 3.

Reviewer 2

1) In the Abstract, to say that "General anesthetics benefit patients undergoing surgeries without consciousness. ..." is a gross understatement of the essential role that general anesthesia plays today to make surgery not only tolerable but humane. This opening sentence should be rewritten. General anesthesia is a fundamental process required to undertake safely and humanely a high fraction of surgeries and invasive diagnostic procedures.

As requested, we rewrote this opening sentence, please see the follows:

GA is a fundamental process required to undertake surgeries and invasive diagnostic procedures safely and humanely. However, the undesired stress response associated with GA can lead to delayed recovery and even increased morbidity in clinical settings.

2) In the Abstract, when discussing the response of the PVN-CRH neurons to chemogenetic inhibition, say exactly what the "opposite effect" is.

Thanks for your insights. We have rewritten our abstract as follows:

Chemogenetic activation of these neurons delayed the induction and accelerated emergence from sevoflurane GA, whereas chemogenetic inhibition of PVH CRH neurons promoted induction and prolonged emergence from sevoflurane GA.

3) In all spectrograms the dynamic range is compressed between 0.5 and 1. Please make use of the full range, as some details might be missed because of this compression.

We are sorry for the incorrect unit of the spectrograms. We have provided the correct one with full range, please see below:

Author response image 4.

Author response image 5.

4) The spectrogram in Figure 2D has several frequency chirps that do not seem physiological.

Thank you for your comments. The frequency chips of the spectrogram during the During and Post 1 phase were caused by recording noises. To avoid confusion, we have deleted the spectrogram in Figure 2D.

5) The 3D plots in Figures 3G and H are not helpful. Thanks for the comment. We'd like to keep the 3D plots as they aid visual comparison of three different features of grooming, which complements other panels in Figure 3.

6) The spectrograms in Figures 5A and B are too small, while the spectra in Figures 5C and D are too large. Please invert this relationship, as it is interesting and important to see the details in the spectrograms. The same happens in Figure 6.

We adjusted the layout of the Figure 5 and Figure 6 as requested, please see below:

Author response image 6.

Author response image 7.

7) In Figure 6H, the authors compute the burst-suppression ratio during a period that seemingly has no bursts or suppressions (Figure 6B).

The burst-suppression ratio was computed from data with the minimum duration of burst and suppression periods set at 0.5 s. Sorry for the confusion. We added a new supplementary figure (Figure 6-figure supplement 8) displaying a 40-second EEG with a burst suppression period to better visualize the burst suppression.

Author response image 8.

8) The data analyses are done in terms of p-values. They should be reported as confidence intervals so that any effect the authors wish to establish is measured along with its uncertainty.

Thank you for your valuable suggestions regarding our manuscript. We appreciate your thoughtful consideration of our work. We understand your concern but we would like to provide some justification for our choice of reporting p-values and explain why we believe they are appropriate for our study. First, the use of p-values for hypothesis testing and significance assessment is a common practice in our field. Many previous studies in our area of research also report results in terms of p-values. For example, Wei Xu11 published in 2020 suggested sevoflurane inhibits MPB neurons through postsynaptic GABAA-Rs and background potassium channels, Ao Y12 demonstrated that activation of the TH:LC-PVT projections is helpful in facilitating the transition from isoflurane anesthesia to an arousal state, using P-value as data analyses. By adhering to this convention, we ensure that our findings are consistent with the existing body of literature. This makes it easier for readers to compare and integrate our results with previous work. Secondly, while confidence intervals can provide a measure of effect size and uncertainty, p-values offer a concise way to communicate statistical significance. They help readers quickly assess whether an effect is statistically significant or not, which is often the primary concern when interpreting research findings. We hope that by providing these reasons for our choice of reporting p-values, we can address your concern while maintaining the integrity and consistency of our study. If you believe there are specific instances where reporting confidence intervals would be more informative, please feel free to highlight those, and we will consider your suggestion on a case-by-case basis. 

References

1. Baldini, G., Bagry, H. & Carli, F. Depth of anesthesia with desflurane does not influence the endocrine-metabolic response to pelvic surgery. Acta Anaesthesiol Scand 52, 99-105, doi:10.1111/j.1399-6576.2007.01470.x (2008).
2. Niikura, R. et al. Exploratory analyses of postanesthetic effects of desflurane using behavioral test battery of mice. Behav Pharmacol 31, 597-609, doi:10.1097/fbp.0000000000000567 (2020).
3. Marana, E. et al. Desflurane versus sevoflurane: a comparison on stress response. Minerva Anestesiol 79, 7-14 (2013).
4. Vutskits, L. & Xie, Z. Lasting impact of general anaesthesia on the brain: mechanisms and relevance. Nat Rev Neurosci 17, 705-717, doi:10.1038/nrn.2016.128 (2016).
5. Mashour, G. A. et al. Recovery of consciousness and cognition after general anesthesia in humans. Elife 10, doi:10.7554/eLife.59525 (2021).
6. Mattison, M. L. P. Delirium. Ann Intern Med 173, Itc49-itc64, doi:10.7326/aitc202010060 (2020).
7. Dahmani, S. et al. Pharmacological prevention of sevoflurane- and desflurane-related emergence agitation in children: a meta-analysis of published studies. Br J Anaesth 104, 216-223, doi:10.1093/bja/aep376 (2010).
8. Lim, B. G. et al. Comparison of the incidence of emergence agitation and emergence times between desflurane and sevoflurane anesthesia in children: A systematic review and meta-analysis. Medicine (Baltimore) 95, e4927, doi:10.1097/MD.0000000000004927 (2016).
9. Radford, K. D. et al. Association between intravenous ketamine-induced stress hormone levels and long-term fear memory renewal in Sprague-Dawley rats. Behav Brain Res 378, 112259, doi:10.1016/j.bbr.2019.112259 (2020).
10. Yang, L., Chen, Z. & Xiang, D. Effects of intravenous anesthesia with sevoflurane combined with propofol on intraoperative hemodynamics, postoperative stress disorder and cognitive function in elderly patients undergoing laparoscopic surgery. Pak J Med Sci 38, 1938-1944, doi:10.12669/pjms.38.7.5763 (2022).
11. Xu, W. et al. Sevoflurane depresses neurons in the medial parabrachial nucleus by potentiating postsynaptic GABA(A) receptors and background potassium channels. Neuropharmacology 181, 108249, doi:10.1016/j.neuropharm.2020.108249 (2020).
12. Ao, Y. et al. Locus Coeruleus to Paraventricular Thalamus Projections Facilitate Emergence From Isoflurane Anesthesia in Mice. Front Pharmacol 12, 643172, doi:10.3389/fphar.2021.643172 (2021).

Author response:

The following is the authors’ response to the current reviews.

eLife assessment:

The manuscript establishes a sophisticated mouse model for acute retinal artery occlusion (RAO) by combining unilateral pterygopalatine ophthalmic artery occlusion (UPOAO) with a silicone wire embolus and carotid artery ligation, generating ischemia-reperfusion injury upon removal of the embolus. This clinically relevant model is useful for studying the cellular and molecular mechanisms of RAO. The data overall are solid, presenting a novel tool for screening pathogenic genes and promoting further therapeutic research in RAO.

Thank you for your thorough evaluation. We are pleased that you find our mouse model for acute retinal artery occlusion to be sophisticated and clinically relevant. Your recognition of the model’s utility in studying the cellular and molecular mechanisms of RAO, as well as its potential for advancing therapeutic research, is highly encouraging and underscores the significance of our work. We are grateful for your supportive feedback.

Public Reviews:

Reviewer #1:

Summary:

Wang, Y. et al. used a silicone wire embolus to definitively and acutely clot the pterygopalatine ophthalmic artery in addition to carotid artery ligation to completely block blood supply to the mouse inner retina, which mimic clinical acute retinal artery occlusion. A detailed characterization of this mouse model determined the time course of inner retina degeneration and associated functional deficits, which closely mimic human patients. Whole retina transcriptome profiling and comparison revealed distinct features associated with ischemia, reperfusion, and different model mechanisms. Interestingly and importantly, this team found a sequential event including reperfusion-induced leukocyte infiltration from blood vessels, residual microglial activation, and neuroinflammation that may lead to neuronal cell death.

Strengths:

Clear demonstration of the surgery procedure with informative illustrations, images, and superb surgical videos.

Two time points of ischemia and reperfusion were studied with convincing histological and in vivo data to demonstrate the time course of various changes in retinal neuronal cell survivals, ERG functions, and inner/outer retina thickness.

The transcriptome comparison among different retinal artery occlusion models provides informative evidence to differentiate these models.

The potential applications of the in vivo retinal ischemia-reperfusion model and relevant readouts demonstrated by this study will certainly inspire further investigation of the dynamic morphological and functional changes of retinal neurons and glial cell responses during disease progression and before and after treatments.

We sincerely appreciate your detailed and positive feedback. These evaluations are invaluable in highlighting the significance and impact of our work. Thank you for your thoughtful and supportive review.

Weaknesses:

The revised manuscript has been significantly improved in clarity and readability. It has addressed all my questions convincingly.

Thank you for your positive feedback. We are pleased to hear that the revisions have significantly improved the manuscript's clarity and readability, and that we have convincingly addressed all your questions. Your encouraging words are of great importance to us.

Reviewer #2 (Public Review):

Summary:

The authors of this manuscript aim to develop a novel animal model to accurately simulate the retinal ischemic process in retinal artery occlusion (RAO). A unilateral pterygopalatine ophthalmic artery occlusion (UPOAO) mouse model was established using silicone wire embolization combined with carotid artery ligation. This manuscript provided data to show the changes of major classes of retinal neural cells and visual dysfunction following various durations of ischemia (30 minutes and 60 minutes) and reperfusion (3 days and 7 days) after UPOAO. Additionally, transcriptomics was utilized to investigate the transcriptional changes and elucidate changes in the pathophysiological process in the UPOAO model post-ischemia and reperfusion. Furthermore, the authors compared transcriptomic differences between the UPOAO model and other retinal ischemic-reperfusion models, including HIOP and UCCAO, and revealed unique pathological processes.

Strengths:

The UPOAO model represents a novel approach for studying retinal artery occlusion. The study is very comprehensive.

Thank you for your positive feedback. We are delighted that you find the UPOAO model to be a novel and comprehensive approach to studying retinal artery occlusion. Your recognition of the depth and significance of our study is highly valuable and encourages us in our ongoing research.

Weaknesses:

Originally, some statements were incorrect and confusing. However, the authors have made clarifications in the revised manuscript to avoid confusion.

We sincerely appreciate your meticulous review of the manuscript. We have thoroughly addressed the inaccuracies identified in the revised version. Additionally, we have polished the article to ensure improved readability. We apologize for any confusion caused by these inaccuracies and genuinely. We appreciate your careful attention to detail, and your patience and meticulous suggestions have significantly improved the clarity and readability of our manuscript.

The following is the authors’ response to the original reviews.

Recommendations for the authors:

Reviewer #1:

The revised manuscript has been significantly improved in clarity and readability. It has addressed all my questions convincingly.

Thank you for your positive feedback. We are pleased to hear that the revisions have significantly improved the manuscript's clarity and readability, and that we have convincingly addressed all your questions. Your encouraging words are of great importance to us.

Reviewer #2:

The authors have revised the manuscript and/or provided answers to the majority of prior comments, which have helped to strengthen the work. However, addressing the following concerns is still necessary to further improve the manuscript.

Thank you for acknowledging our revisions and the improvements made to the manuscript. We appreciate your continued feedback and will address the remaining concerns to further enhance the quality of our work.

The quantification method of RGCs is described in detail in the response letter, but this detailed methodology was not included in the revised manuscript to clarify the quantification process.

Thank you for your helpful recommendations. We have added detailed methodology in the revised manuscript to clarify the quantification process (line 180-188).

The graphs in Fig. 3D b-wave and Fig. 3E-b wave are duplicated.

We apologize for the error in our figures. We have corrected the mistake by replacing the duplicated image in Fig. 3E-b wave with the correct one (line 880). Your careful observation has been very helpful in improving our manuscript. Thank you for bringing this to our attention.

The quantifications of the thickness of retinal layers in HE-stained sections in Figure 4 (IPL) and Response Figure 2 are incorrect. For mice retina, the thickness of the IPL is approximately 50 µm.

Thank you for your meticulous review of the manuscript. We have rectified the inaccuracies in the quantification of retinal layer thickness in HE-stained sections in Figure 4, addressing the initial issue with the scale bar.

We consulted with a microscope engineer and used a microscope microscale to calibrate the scale of the fluorescence microscope (BX63; Olympus, Tokyo, Japan) at the suggestion of the engineer.

We recount the thickness of all layers of the HE-stained retinal section (line 902). The inner retina thickness in Figure 4 has been adjusted under a new scale bar, and the thickness of the outer retinal layers is now displayed in

Author response image 1. However, the IPL thickness of the sham eye in the UPOAO model is still not aligned with the common thickness of 50 µm. Therefore we review the literature within our laboratory, focusing on C57BL/6 mice from the same source, revealed that the inner retina thickness (GCC+INL) in the HE-stained sections of the sham eye in the UPOAO model (around 80 µm) is consistent with previous findings (see Author response image 2) conducted by Kaibao Ji and published in Experimental Eye Research in 2021 [1].

We captured and analyzed the average retinal thickness of each layer over a long range of 200-1100 μm from the optic nerve head (see Author response image 3, highlighted by the green line). The field region has been corrected in the revised manuscript (line 232). Considering the significant variation in retinal thickness from the optic nerve to the periphery, we consulted literature on multi-point measurements of HE-stained retinas. The average thickness of the GCC layer in the control group was approximately 57 µm at 600 µm from the optic nerve head and about 48 µm at 1200 µm from the optic nerve head in the literature [2] (see Author response image 4). The GCC layer thickness of the sham eye in the UPOAO model is around 50 µm, in alignment with existing literature. In future studies, we will pay more attention to the issue of thickness averaging.

We appreciate your thorough review and valuable feedback, which has enabled us to correct errors and enhance the accuracy of our research.

Author response image 1.

Thickness of OPL, ONL, IS/OS+RPE in HE staining. n=3; ns: no significance (p>0.05).

Author response image 2.

Cited from Ji, K., et al., Resveratrol attenuates retinal ganglion cell loss in a mouse model of retinal ischemia reperfusion injury via multiple pathways. Experimental Eye Research, 2021. 209: p. 108683.

Author response image 3.

Schematic diagram illustrating the selection of regions. The figure was captured using a fluorescence microscope (BX63; Olympus, Tokyo, Japan) under a 4X objective. Scale bar=500 µm.

Author response image 4.

Cited from Feng, L., et al., Ripa-56 protects retinal ganglion cells in glutamate-induced retinal excitotoxic model of glaucoma. Sci Rep, 2024. 14(1): p. 3834.

There are some typos in the summary table. For example: 'Amplitudes of a-wave (0.3, 2.0, and 10.0 cd.s/m²)' should be 'Amplitudes of a-wave (0.3, 3.0, and 10.0 cd.s/m²)'; and 'IINL thickness' in HE' should be 'INL thickness'.

Thank you for pointing out the typos in the summary table (line 1073). We have corrected 'Amplitudes of a-wave (0.3, 2.0, and 10.0 cd.s/m²)' to 'Amplitudes of a-wave (0.3, 3.0, and 10.0 cd.s/m²)' and 'IINL thickness' to 'INL thickness'. Your attention to detail is greatly appreciated and has been very helpful in improving our manuscript.

References

(1) Ji, K., et al., Resveratrol attenuates retinal ganglion cell loss in a mouse model of retinal ischemia reperfusion injury via multiple pathways. Experimental Eye Research, 2021. 209: p. 108683.

(2) Feng, L., et al., Ripa-56 protects retinal ganglion cells in glutamate-induced retinal excitotoxic model of glaucoma. Sci Rep, 2024. 14(1): p. 3834.

Author response:

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public review):

Summary:

The authors aim to assess the effect of salt stress on root:shoot ratio, identify the underlying genetic mechanisms, and evaluate their contribution to salt tolerance. To this end, the authors systematically quantified natural variations in salt-induced changes in root:shoot ratio. This innovative approach considers the coordination of root and shoot growth rather than exploring biomass and the development of each organ separately. Using this approach, the authors identified a gene cluster encoding eight paralog genes with a domain-of-unknown-function 247 (DUF247), with the majority of SNPs clustering into SR3G (At3g50160). In the manuscript, the authors utilized an integrative approach that includes genomic, genetic, evolutionary, histological, and physiological assays to functionally assess the contribution of their genes of interest to salt tolerance and root development.

Comments on revisions:

As the authors correctly noted, variations across samples, genotypes, or experiments make achieving statistical significance challenging. Should the authors choose to emphasize trends across experiments to draw biological conclusions, careful revisions of the text, including titles and figure legends, will be necessary to address some of the inconsistencies between figures (see examples below). However, I would caution that this approach may dilute the overall impact of the work on SR3G function and regulation. Therefore, I strongly recommend pursuing additional experimental evidence wherever possible to strengthen the conclusions.

(1) Given the phenotypic differences shown in Figures S17A-B, 10A-C, and 6A, the statement that "SR3G does not play a role in plant development under non-stress conditions" (lines 680-681) requires revision to better reflect the observed data.

Thank you to the reviewer for the comment. We appreciate the acknowledgment that variations among experiments are inherent to biological studies. Figures 6A and S17 represent the same experiment, which initially indicated a phenotype for the sr3g mutant under salt stress. To ensure that growth changes were specifically normalized for stress conditions, we calculated the Stress Tolerance Index (Fig. 6B). In Figure 10, we repeated the experiment including all five genotypes, which supported our original observation that the sr3g mutant exhibited a trend toward reduced lateral root number under 75 mM NaCl compared to Col-0, although this difference was not significant (Fig. 10B). Additionally, we confirmed that the wrky75 mutant showed a significant reduction in main root growth under salt stress compared to Col-0, consistent with findings reported in The Plant Cell by Lu et al. 2023. For both main root length and lateral root number, we demonstrated that the double mutants of wrky75/sr3g displayed growth comparable to wild-type Col-0. This result suggests that the sr3g mutation compensates for the salt sensitivity of the wrky75 mutant.

We completely agree with the reviewer that there is a variation in our results regarding the sr3g phenotype under control conditions, as presented in Fig. 6A/Fig. S17 and Fig. 10A-C. In Fig. 6A/Fig. S17, we did not observe any consistent trends in main root or lateral root length for the sr3g mutant compared to Col-0 under control conditions. However, in Fig. 10A-C, we observed a significant reduction in main root length, lateral root number, and lateral root length for the sr3g mutant under control conditions. We believe this may align with SR3G’s role as a negative regulator of salt stress responses. While loss of this gene benefits plants in coping with salt stress, it might negatively impact overall plant growth under non-stress conditions. This interpretation is further supported by our findings on the root suberization pattern in sr3g mutants under control conditions (Fig. 8B), where increased suberization in root sections 1 to 3, compared to Col-0, could inhibit root growth. While SR3G's role in overall plant fitness is intriguing, it is beyond the scope of this study. We cannot rule out the possibility that SR3G contributes positively to plant growth, particularly root growth. That said, we observed no differences in shoot growth between Col-0 and the sr3g mutant under control conditions (Fig. 7). Additionally, we calculated the Stress Tolerance Index for all aspects of root growth shown in Fig. 10 and presented it in Fig. S25.

To address the reviewer request on rephrasing the lines 680-681 from"SR3G does not play a role in plant development under non-stress conditions" (lines 680-681) statement, this statement is found in lines 652-653 and corresponds to Fig. 7, where we evaluated rosette growth in the WT and sr3g mutant under both control and salt stress conditions. We did not observe any significant differences or even trends between the two genotypes under control conditions, confirming the accuracy of the statement. To clarify further, we have added “SR3G does not play a role in rosette growth and development under non-stress conditions”.

(2) I agree with the authors that detecting expression differences in lowly expressed genes can be challenging. However, as demonstrated in the reference provided (Lu et al., 2023), a significant reduction in WRKY75 expression is observed in T-DNA insertion mutant alleles of WRKY75. In contrast, Fig. 9B in the current manuscript shows no reduction in WRKY75 expression in the two mutant alleles selected by the authors, which suggests that these alleles cannot be classified as loss-of-function mutants (line 745). Additionally, the authors note that the wrky75 mutant exhibits reduced main root length under salt stress, consistent with the phenotype reported by Lu et al. (2023). However, other phenotypic discrepancies exist between the two studies. For example, 1) Lu et al. (2023) report that w¬rky75 root length is comparable to WT under control conditions, whereas the current manuscript shows that wrky75 root growth is significantly lower than WT; 2) under salt stress, Lu et al. (2023) show that wrky75 accumulates higher levels of Na+, whereas the current study finds Na+ levels in wrky75 indistinguishable from WT. To confirm the loss of WRKY75 function in these T-DNA insertion alleles the authors should provide additional evidence (e.g., Western blot analysis).

We sincerely appreciate the reviewer acknowledging the challenge of detecting expression differences in lowly expressed genes, such as transcription factors. Transcription factors are typically expressed at lower levels compared to structural or enzymatic proteins, as they function as regulators where small quantities can have substantial effects on downstream gene expression.

That said, we respectfully disagree with the reviewer’s interpretation that there is no reduction in WRKY75 expression in the two mutant lines tested in Fig. 9C. Among the two independent alleles examined, wrky75-3 showed a clear reduction in expression compared to WT Col-0 under both control and salt stress conditions. Using the Tukey test to compare all groups, we observed distinct changes in the assigned significance letters for each case:

Col/root/control (cd) vs wrky75-3/root/control (cd): Although the same significance letter was assigned, we still observed a clear reduction in WRKY75 transcript abundance. More importantly, the variation in expression is notably lower compared to Col-0.

Col/shoot/control (bcd) vs wrky75-3/shoot/control (a): This is significant reduction compared to Col

Col/root/salt (cd) vs wrky75-3/root/salt (bcd): Once again, the reduction in WRKY75 transcript levels corresponds to changes in the assigned significance letters.

Col/shoot/salt (bc) vs wrky75-3/shoot/salt (ab): Once again, the reduction in WRKY75 transcript levels corresponds to changes in the assigned significance letters.

To address the reviewer’s comment regarding the significant reduction in WRKY75 expression observed in T-DNA insertion mutant alleles of WRKY75 in the reference by Lu et al., 2023, we would like to draw the reviewer’s attention to the following points:

a) Different alleles: The authors in The Plant Cell used different alleles than those used in our study, with one of their alleles targeting regions upstream of the WRKY75 gene. While we identified one of their described alleles (WRKY75-1, SALK_101367) on the T-DNA express website, which targets upstream of WRKY75, the other allele (wrky75-25) appears to have been generated through a different mechanism (possibly an RNAi line) that is not defined in the Plant Cell paper and does not appear on the T-DNA express website. The authors mentioned they have received these seeds as gifts from other labs in the acknowledgement ”We thank Prof. Hongwei Guo (Southern University of Science and Technology, China) and Prof. Diqiu Yu (Yunnan University, China) for kindly providing the WRKY75_pro:GUS, 35S_pro:WRKY75-GFP, wrky75-1, and wrky75-25 seeds. We thank Man-cang Zhang (Electrophysiology platform, Henan University) for performing the NMT experiment”.

However, in our study, we selected two different T-DNAs that target the coding regions. While this may explain slight differences in the observed responses, both studies independently link WRKY75 to salt stress, regardless of the alleles used. For your reference, we have included a screenshot of the different alleles used.

Author response image 1.

b) Different developmental stages: They measured WRKY75 expression in 5-day-old seedlings. In our experiment, we used seedlings grown on 1/2x MS for 4 days, followed by transfer to treatment plates with or without 75 mM NaCl for one week. As a result, we analyzed older plants (12 days old) for gene expression analysis. Despite the difference in developmental stage, we were still able to observe a reduction in gene expression.

c) Different tissues: The authors of The Plant Cell used whole seedlings for gene expression analysis, whereas we separated the roots and shoots and measured gene expression in each tissue type individually. This approach is logical, as WRKY75 is a root cell-specific transcription factor with higher expression in the roots compared to the shoots, as demonstrated in our analysis (Fig. 9C).

Based on the reasoning above, we did work with loss-of-function mutants of WRKY75, particularly wrky75-3. To more accurately reflect the nature of the mutation, we have changed the term "loss-of-function" to "knock-down" in line 717.

The reviewer mentioned phenotypic discrepancies between the two studies. We agree that there are some differences, particularly in the magnitude of responses or expression levels. However, despite variations in the alleles used, developmental stages, and tissue types, both studies reached the same conclusion: WRKY75 is involved in the salt stress response and acts as a positive regulator. We have discussed the differences between our study and The Plant Cell in the section above, summarizing them into three main points: different alleles, different developmental stages, and different tissue types.

To address the reviewer’s comment regarding "Lu et al. (2023) report that wrky75 root length is comparable to WT under control conditions, whereas the current manuscript shows that wrky75 root growth is significantly lower than WT": We evaluated root growth differently than The Plant Cell study. In The Plant Cell (Fig. 5, H-J), root elongation was measured in 10-day-old plants with a single time point measurement. They transferred five-day-old wild-type, wrky75-1, wrky75-25, and WRKY75-OE plants to 1/2× MS medium supplemented with 0 mM or 125 mM NaCl for further growth and photographed them 5 days after transfer. In contrast, our study used 4-day-old seedlings, which were transferred to 1/2 MS with or without 0, 75, or 125 mM salt for additional growth (9 days). Rather than measuring root growth only at the end, we scanned the roots every other day, up to five times, to assess root growth rates. Essentially, the precision of our method is higher as we captured growth changes throughout the developmental process, compared to the approach used in The Plant Cell. We do not underestimate the significance of the work conducted by other colleagues in the field, but we also recognize that each laboratory has its own approach and specific practices. This variation in experimental setup is intrinsic to biology, and we believe it is important to study biological phenomena in different ways. Especially as the common or contrasting conclusions reached by different studies, performed by different labs and using different experimental setups are shedding more light on reproducibility and gene contribution across different conditions, which is intrinsic to phenotypic plasticity, and GxE interactions.

The Plant Cell used a very high salt concentration, starting at 125 mM, while we were more cautious in our approach, as such a high concentration can inhibit and obscure more subtle phenotypic changes.

To address the reviewer’s comment on "Lu et al. (2023) show that wrky75 accumulates higher levels of Na+, whereas the current study finds Na+ levels in wrky75 indistinguishable from WT," we would like to highlight the differences in the methodologies used in both studies. The Plant Cell measured Na+ accumulation in the wrky75 mutant using xylem sap (Supplemental Figure S10), which appears to be a convenient and practical approach in their laboratory. In their experiment, wild-type and wrky75 mutant plants were grown in soil for 3 weeks, watered with either a mock solution or 100 mM NaCl solution for 1 day, and then xylem sap was collected for Na+ content analysis. In contrast, our study employed a different method to measure Na+ and K+ ion content, using Inductively Coupled Plasma Atomic Emission Spectroscopy (ICP-AES) for root and shoot Na+ and K+ measurements. Additionally, we collected samples after two weeks on treatment plates and focused on the Na+/K+ ratio, which we consider more relevant than net Na+ or K+ levels, as the ratio of these ions is a critical determinant of plant salt tolerance. With this in mind, we observed a considerable non-significant increase in the Na+/K+ ratio in the shoots of the wrky75-3 mutant (assigned Tukey’s letter c) compared to the Col-0 WT (assigned Tukey’s letters abc) under 125 mM salt, suggesting that this mutant is salt-sensitive. Importantly, the Na+/K+ ratio in the double wrky75/sr3g mutants was reduced to the WT level under the same salt conditions, further indicating that the salt sensitivity of wrky75 is mitigated by the sr3g mutation.

Based on the reasons mentioned above, we believe that conducting additional experiments, such as Western blot analysis, is unnecessary and would not contribute new insights or alter the context of our findings.

Reviewer #2 (Public review):

Summary:

Salt stress is a significant and growing concern for agriculture in some parts of the world. While the effects of sodium excess have been studied in Arabidopsis and (many) crop species, most studies have focused on Na uptake, toxicity and overall effects on yield, rather than on developmental responses to excess Na, per se. The work by Ishka and colleagues aims to fill this gap.

Working from an existing dataset that exposed a diverse panel of A. thaliana accessions to control, moderate, and severe salt stress, the authors identify candidate loci associated with altering the root:shoot ratio under salt stress. Following a series of molecular assays, they characterize a DUF247 protein which they dub SR3G, which appears to be a negative regulator of root growth under salt stress.

Overall, this is a well-executed study which demonstrates the functional role played by a single gene in plant response to salt stress in Arabidopsis.

Review of revised manuscript:

The authors have addressed my point-by-point comments to my satisfaction. In the cases where they have changed their manuscript language, clarified figures, or added analyses I have no further comment. In some cases, there is a fruitful back-and-forth discussion of methodology which I think will be of interest to readers.

I have nothing to add during this round of review. I think that the paper and associated discussion will make a nice contribution to the field.

We sincerely appreciate the reviewer’s recognition of the significance of our work to the field.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

Lines 518-519: The statement that other DUF247s exhibit similar expression patterns to SR3G, suggesting their responsiveness to salt stress, is not fully supported by Fig. S14. Please clarify the specific similarities (and differences) in the expression patterns of the DUF247s shown in Fig. S14, as their expression appears to be spatially and temporally diverse. Additionally, the scale is missing in Fig. S14.

We thank the reviewer. We fixed the text and added expression scales to Figure S14.

Line 684, Fig. 6A should be 7A.

Thanks. It is fixed.

Line 686, Fig. 7A should be 7B.

Thanks. It is fixed.

Lines 721-723: The signal quantification in Fig. 8B does not support the claim that "in section one,..., sr3g-5 showed more suberization compared to Col-0." Given the variability and noise often associated with histological dyes such as Fluorol Yellow staining, conclusions should be cautiously grounded in robust signal quantification. Additionally, please specify the number of biological replicates used in both Fig. 8B and C.

We thank the reviewer for their comments. We believe the statement in the text accurately reflects our results presented in Figure 8B, where we stated “non-significant, but substantially higher levels of root suberization in sr3g-5 compared to Col-0 in sections one to three of the root under control condition (Fig. 8B).” Therefore, we kept the statement and have included the number of biological replicates in the figure legend.

Lines 731-732: Please provide a more detailed explanation of how the significant changes in suberin monomer levels align with the Fluorol Yellow staining results, and clarify how these findings support the proposed negative role of SR3G in root suberization.

Fluorol Yellow is a lipophilic dye widely used to label suberin in plant tissues, specifically in roots in this study. Given the inherent variability in histological assays, we confirmed the increase in suberization using an alternative method, Gas Chromatography–Mass Spectrometry (GC-MS). Both approaches revealed elevated suberin levels in the sr3g mutant compared to Col-0. Since the overall suberin content was higher in the mutant under both control and salt stress conditions, we proposed that SR3G acts as a negative regulator of root suberization.

Lines 686-688 and Figure S24: The authors calculated water mass as FW-DW. A more standard approach for calculating water content is (FW-DW)/FW x 100. Please update the text or adjust the calculation accordingly. Additionally, if the goal is to test differences between WT and the mutant within each condition, a t-test would be a more appropriate statistical method.

We thank the reviewer. We added water content % to the figure S24. We kept the statistical test as it is as we wanted to be able to observe changes across conditions and genotypes.

Lines 633-635 states that "No significant difference was observed between sr3g-4 and Col-0 (Fig. S18), except for the Stress Tolerance Index (STI) calculated using growth rates of lateral root length and number." However, based on the Figure S18 legend and statistical analysis (i.e., ns), it appears that the sr3g-4 mutant shows no alterations in root system architecture compared to Col-0. Please revise the text to accurately reflect the results of the statistical analysis.

We thank the reviewer. We now fixed the text to reflect the result.

Lines 698-707: The statistical analysis does not support the reported differences in the Na+/K+ ratio for the single and double mutants of sr3g-5 and wrky75-3 (Fig. 10D, where levels connected by the same letters indicate they are not significantly different). Furthermore, the conclusion that "the SR3G mutation indeed compensated for the increased Na+ accumulation observed in the wrky75 mutant under salt stress" is also based on non-significant differences (Fig. S25B). Please revise the text to accurately reflect the results of the statistical analysis. Additionally, since each mutant is compared to the WT, I recommend using Dunnett's test for statistical analysis.

We thank the reviewer for their feedback. We have carefully revised the text to better support our findings. As previously mentioned, variations among samples are evident and are well-reflected across all our datasets. We have presented all data and focused on identifying trends within our samples to guide interpretation.

We observed that the SR3G mutation effectively compensated for the increased Na+ accumulation observed in the wrky75 mutant under salt stress. A closer examination of the shoot Na+/K+ ratio under 125 mM salt shows that the wrky75 single mutant has a higher Na+/K+ ratio (indicated by the letter "c") compared to Col-0 (indicated by "abc") and the two double mutants (also indicated by "abc"). Therefore, we have retained the statistical analysis as originally conducted, and maintain our conclusions as is.

Figure 6: data in panel C present the Na/K ratio, not Na+ content. Based on the statistical analysis of root Na+ levels presented in Fig. S17C, there is no significant difference between sr3g-5 and WT. Please update the title of Fig. 6. In addition, in panel A, the title of the Y-axis and figure legend should be "Lateral root growth rate" without the word length, and in panel C, the statistical analysis is missing.

We thank the reviewer. We updated Fig. 6 title and fixed the Y-axis in panel A, and added statistical letters to panel C. Legend was updated to reflect the changes.

Figure 7: Please clearly label the time points where significant differences between genotypes are observed for both early and late salt treatments. Was there a significant difference recorded between WT and sr3g-5 on day 0 under early salt stress? Such differences may arise from initial variations in plant size within this experiment, as indicated by Fig. 7B, where significant differences in rosette area are evident starting from day 0. Additionally, please indicate the statistical analysis in panel E.

We thank the reviewer for this suggestion. We updated the figure with a statistical test added to the panel E. Although the difference between sr3g mutant and Col-0 is indeed significant in its growth rate at day 0, we would like to draw the attention of the reviewer that this growth rate was calculated over the 24 hours after adding salt stress. Therefore, this difference in growth rate is related to exposure to salt stress. Moreover, the growth rate between Col-0 and sr3g mutant does not differ in two other treatments (Control and Late Salt Stress) further supporting the conclusion that sr3g is affecting rosette size and growth rate only under early salt stress conditions.

We have also added the Salt Tolerance Index calculation to Figure S24 as additional evidence, controlling for potential differences in size between Col-0 and sr3g mutant.

Figure S17: statistical analysis is not indicated in panels A, B, and D.

We thank the reviewer for spotting that. We updated the figure with a statistical test.

Figures S21-23: The quality of these figures is insufficient, hindering the ability to effectively interpret the authors' results and main message. Furthermore, a Dunnett's test, rather than a t-test, is the appropriate statistical method for this analysis.

We thank the reviewer for this observation. We have now added a high resolution figures for all supplemental figures, which should increase the resolution of the figures. As we are comparing all of the genotypes to Col-0 one-by-one - the results of individual t-tests are sufficient for this analysis.

Author response:

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public review):

Summary:

In their manuscript entitled 'The domesticated transposon protein L1TD1 associates with its ancestor L1 ORF1p to promote LINE-1 retrotransposition', Kavaklıoğlu and colleagues delve into the role of L1TD1, an RNA binding protein (RBP) derived from a LINE1 transposon. L1TD1 proves crucial for maintaining pluripotency in embryonic stem cells and is linked to cancer progression in germ cell tumors, yet its precise molecular function remains elusive. Here, the authors uncover an intriguing interaction between L1TD1 and its ancestral LINE-1 retrotransposon.

The authors delete the DNA methyltransferase DNMT1 in a haploid human cell line (HAP1), inducing widespread DNA hypo-methylation. This hypomethylation prompts abnormal expression of L1TD1. To scrutinize L1TD1's function in a DNMT1 knock-out setting, the authors create DNMT1/L1TD1 double knock-out cell lines (DKO). Curiously, while the loss of global DNA methylation doesn't impede proliferation, additional depletion of L1TD1 leads to DNA damage and apoptosis.

To unravel the molecular mechanism underpinning L1TD1's protective role in the absence of DNA methylation, the authors dissect L1TD1 complexes in terms of protein and RNA composition. They unveil an association with the LINE-1 transposon protein L1-ORF1 and LINE-1 transcripts, among others.

Surprisingly, the authors note fewer LINE-1 retro-transposition events in DKO cells compared to DNMT1 KO alone.

Strengths:

The authors present compelling data suggesting the interplay of a transposon-derived human RNA binding protein with its ancestral transposable element. Their findings spur interesting questions for cancer types, where LINE1 and L1TD1 are aberrantly expressed.

Weaknesses:

Suggestions for refinement:

The initial experiment, inducing global hypo-methylation by eliminating DNMT1 in HAP1 cells, is intriguing and warrants more detailed description. How many genes experience misregulation or aberrant expression? What phenotypic changes occur in these cells? Why did the authors focus on L1TD1? Providing some of this data would be helpful to understand the rationale behind the thorough analysis of L1TD1.

The finding that L1TD1/DNMT1 DKO cells exhibit increased apoptosis and DNA damage but decreased L1 retro-transposition is unexpected. Considering the DNA damage associated with retro-transposition and the DNA damage and apoptosis observed in L1TD1/DNMT1 DKO cells, one would anticipate the opposite outcome. Could it be that the observation of fewer transposition-positive colonies stems from the demise of the most transpositionpositive colonies? Further exploration of this phenomenon would be intriguing.

Reviewer #2 (Public review):

In this study, Kavaklıoğlu et al. investigated and presented evidence for a role for domesticated transposon protein L1TD1 in enabling its ancestral relative, L1 ORF1p, to retrotranspose in HAP1 human tumor cells. The authors provided insight into the molecular function of L1TD1 and shed some clarifying light on previous studies that showed somewhat contradictory outcomes surrounding L1TD1 expression. Here, L1TD1 expression was correlated with L1 activation in a hypomethylation dependent manner, due to DNMT1 deletion in HAP1 cell line. The authors then identified L1TD1 associated RNAs using RIPSeq, which display a disconnect between transcript and protein abundance (via Tandem Mass Tag multiplex mass spectrometry analysis). The one exception was for L1TD1 itself, is consistent with a model in which the RNA transcripts associated with L1TD1 are not directly regulated at the translation level. Instead, the authors found L1TD1 protein associated with L1-RNPs and this interaction is associated with increased L1 retrotransposition, at least in the contexts of HAP1 cells. Overall, these results support a model in which L1TD1 is restrained by DNA methylation, but in the absence of this repressive mark, L1TD1 is expression, and collaborates with L1 ORF1p (either directly or through interaction with L1 RNA, which remains unclear based on current results), leads to enhances L1 retrotransposition. These results establish feasibility of this relationship existing in vivo in either development or disease, or both.

Comments on revised version:

In general, the authors did an acceptable job addressing the major concerns throughout the manuscript. This revision is much clearer and has improved in terms of logical progression.

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The authors have addressed all my questions in the revised version of the manuscript.

Reviewer #2 (Recommendations for the authors):

Revised comments:

A few points we'd like to see addressed are our comments about the model (Figure S7C), as this is important for the readership to understand this complex finding. Please try to apply some quantification, if possible (question 8). Please do your best to tone down the direct relationship of these findings to embryology (question 11). Based on both reviewer comments, we believe addressing reviewer #1s "Suggestions for refinement" (2 points), would help us change our view of solid to convincing.

Responses to changes:

Major

(1) The study only used one knockout (KO) cell line generated by CRISPR/Cas9.

Considering the possibility of an off-target effect, I suggest the authors attempt one or both of these suggestions.

A)  Generate or acquire a similar DMNT1 deletion that uses distinct sgRNAs, so that the likelihood of off-targets is negligible. A few simple experiments such as qRT-PCR would be sufficient to suggest the same phenotype.

B)  Confirm the DNMT1 depletion also by siRNA/ASO KD to phenocopy the KO effect.

(2) In addition to the strategies to demonstrate reproducibility, a rescue experiment restoring DNMT1 to the KO or KD cells would be more convincing. (Partial rescue would suffice in this case, as exact endogenous expression levels may be hard to replicate).

We have undertook several approaches to study the effect of DNMT1 loss or inactivation: As described above, we have generated a conditional KO mouse with ablation of DNMT1 in the epidermis. DNMT1-deficient keratinocytes isolated from these mice show a significant increase in L1TD1 expression. In addition, treatment of primary human keratinocytes and two squamous cell carcinoma cell lines with the DNMT inhibitor aza-deoxycytidine led to upregulation of L1TD1 expression. Thus, the derepression of L1TD1 upon loss of DNMT1 expression or activity is not a clonal effect.

Also, the spectrum of RNAs identified in RIP experiments as L1TD1-associated transcripts in HAP1 DNMT1 KO cells showed a strong overlap with the RNAs isolated by a related yet different method in human embryonic stem cells. When it comes to the effect of L1TD1 on L1-1 retrotranspostion, a recent study has reported a similar effect of L1TD1 upon overexpression in HeLa cells [4].

All of these points together help to convince us that our findings with HAP1 DNMT KO are in agreement with results obtained in various other cell systems and are therefore not due to off-target effects. With that in mind, we would pursue the suggestion of Reviewer 1 to analyze the effects of DNA hypomethylation upon DNMT1 ablation.

Thank you for addressing this concern. The reference to Beck 2021 and the additional cells lines (R2: keratinocytes and R3: squamous cell carcinoma) provides sufficient evidence that this result is unlikely to be a result of clonal expansion or off targets.

Question: Was the human ES Cell RIP Experiment shown here? What is the overlap?

We refer to the recently published study by Jin et al. (PMID: 38165001). As stated in the Discussion, the majority of L1TD1-associated transcripts in HAP1 cells (69%) identified in our study were also reported as L1TD1 targets in hESCs suggesting a conserved binding affinity of this domesticated transposon protein across different cell types.  

(3) As stated in the introduction, L1TD1 and ORF1p share "sequence resemblance" (Martin 2006). Is the L1TD1 antibody specific or do we see L1 ORF1p if Fig 1C were uncropped?

(6) Is it possible the L1TD1 antibody binds L1 ORF1p? This could make Figure 2D somewhat difficult to interpret. Some validation of the specificity of the L1TD1 antibody would remove this concern (see minor concern below).

This is a relevant question. We are convinced that the L1TD1 antibody does not crossreact with L1 ORF1p for the following reasons: Firstly, the antibody does not recognize L1 ORF1p (40 kDa) in the uncropped Western blot for Figure 1C (Figure R4A). Secondly, the L1TD1 antibody gives only background signals in DKO cells in the indirect immunofluorescence experiment shown in Figure 1E of the manuscript.

Thirdly, the immunogene sequence of L1TD1 that determines the specificity of the antibody was checked in the antibody data sheet from Sigma Aldrich. The corresponding epitope is not present in the L1 ORF1p sequence.

Finally, we have shown that the ORF1p antibody does not cross-react with L1TD1 (Figure R4B).

Response: Thank you for sharing these images. These full images relieve concerns about specificity. The increase of ORF1P in R4B and Main figure 3C is interesting and pointed out in the manuscript. Not for the purposes of this review, but the observation of reduced transposition despite increased ORF1P could be an interesting follow up to this study (combined with the similar UPF1 result could indicate a complex of some kind).

(4) In abstract (P2), the authors mentioned that L1TD1 works as an RNA chaperone, but in the result section (P13), they showed that L1TD1 associates with L1 ORF1p in an RNA independent manner. Those conclusions appear contradictory. Clarification or revision is required.

Our findings that both proteins bind L1 RNA, and that L1TD1 interacts with ORF1p are compatible with a scenario where L1TD1/ORF1p heteromultimers bind to L1 RNA. The additional presence of L1TD1 might thereby enhance the RNA chaperone function of ORF1p. This model is visualized now in Suppl. Figure S7C.

Response: Thank you for the model. To further clarify, do you mean that L1TD1 can bind L1 RNA, but this is not needed for the effect, however this "bonus" binding (that is enabled by heteromultimerization) appears to enhance the retrotransposition frequency? Do you think L1TD1 is binding L1 RNA in this context or simply "stabilizing" ORF1P (Trimer) RNP?

Based on our data, L1TD1 associates with L1 RNA and interacts with L1 ORF1p. Both features might contribute to the enhanced retrotransposition frequency. Interestingly, the L1TD1 protein shares with its ancestor L1 ORF1p the non-canonical RNA recognition motif and the coiled-coil motif required for the trimerization but has two copies instead of one of the C-terminal domain (CTD), a structure with RNA binding and chaperone function. We speculate that the presence of an additional CTD within the L1TD1 protein might thereby enhance the RNA binding and chaperone function of L1TD1/ORF1p heteromultimers.

(5) Figure 2C fold enrichment for L1TD1 and ARMC1 is a bit difficult to fully appreciate. A 100 to 200-fold enrichment does not seem physiological. This appears to be a "divide by zero" type of result, as the CT for these genes was likely near 40 or undetectable. Another qRT-PCR based approach (absolute quantification) would be a more revealing experiment. This is the validation of the RIP experiments and the presentation mode is specifically developed for quantification of RIP assays (Sigma Aldrich RIP-qRT-PCR: Data Analysis Calculation Shell). The unspecific binding of the transcript in the absence of L1TD1 in DNMT1/L1TD1 DKO cells is set to 1 and the value in KO cells represents the specific binding relative the unspecific binding. The calculation also corrects for potential differences in the abundance of the respective transcript in the two cell lines. This is not a physiological value but the quantification of specific binding of transcripts to L1TD1. GAPDH as negative control shows no enrichment, whereas specifically associated transcripts show strong enrichement. We have explained the details of RIPqRT-PCR evaluation in Materials and Methods (page 14) and the legend of Figure 2C in the revised manuscript.

Response: Thank you for the clarification and additional information in the manuscript.

(6) Is it possible the L1TD1 antibody binds L1 ORF1p? This could make Figure 2D somewhat difficult to interpret. Some validation of the specificity of the L1TD1 antibody would remove this concern (see minor concern below).

See response to (3).

Response: Thanks.

(7) Figure S4A and S4B: There appear to be a few unusual aspects of these figures that should be pointed out and addressed. First, there doesn't seem to be any ORF1p in the Input (if there is, the exposure is too low). Second, there might be some L1TD1 in the DKO (lane 2) and lane 3. This could be non-specific, but the size is concerning. Overexposure would help see this.

The ORF1p IP gives rise to strong ORF1p signals in the immunoprecipitated complexes even after short exposure. Under these conditions ORF1p is hardly detectable in the input. Regarding the faint band in DKO HAP1 cells, this might be due to a technical problem during Western blot loading. Therefore, the input samples were loaded again on a Western blot and analyzed for the presence of ORF1p, L1TD1 and beta-actin (as loading control) and shown as separate panel in Suppl. Figure S4A.

The enhanced image is clearer. Thanks.

S4A and S4B now appear to the S6A and S6B, is that correct? (This is due to the addition of new S1 and S2, but please verify image orders were not disturbed).

Yes, the input is shown now as a separate panel in Suppl. Figure S6A.

(8) Figure S4C: This is related to our previous concerns involving antibody cross-reactivity. Figure 3E partially addresses this, where it looks like the L1TD1 "speckles" outnumber the ORF1p puncta, but overlap with all of them. This might be consistent with the antibody crossreacting. The western blot (Figure 3C) suggests an upregulation of ORF1p by at least 23x in the DKO, but the IF image in 3E is hard to tell if this is the case (slightly more signal, but fewer foci). Can you return to the images and confirm the contrast are comparable? Can you massively overexpose the red channel in 3E to see if there is residual overlap? In Figure 3E the L1TD1 antibody gives no signal in DNMT1/L1TD1 DKO cells confirming that it does not recognize ORF1p. In agreement with the Western blot in Figure 3C the L1 ORF1p signal in Figure 3E is stronger in DKO cells. In DNMT1 KO cells the L1 ORF1p antibody does not recognize all L1TD1 speckles. This result is in agreement with the Western blot shown above in Figure R4B and indicates that the L1 ORF1p antibody does not recognize the L1TD1 protein. The contrast is comparable and after overexposure there are still L1TD1 specific speckles. This might be due to differences in abundance of the two proteins.

Response: Suggestion: Would it be possible to use a program like ImageJ to supplement the western blot observation? Qualitatively, In figure 3E, it appears that there is more signal in the DKO, but this could also be due to there being multiple cells clustered together or a particularly nicely stained region. Could you randomly sample 20-30 cells across a few experiments to see if this holds up. I am interested in whether the puncta in the KO image(s) is a very highly concentrated region and in the DKO this is more disperse. Also, the representative DKO seems to be cropped slightly wrong. (Please use puncta as a guide to make the cropping more precise)

As suggested by the reviewer we have quantified the signals of 60 KO cells and 56 DKO cells in three different IF experiments by ImageJ. We measured a 1.4-fold higher expression level of L1 ORF1p in DKO cells. However, the difference is not statistically significant. This is most probably due to the change in cell size and protein content during the cell cycle with increasing protein contents from G1 to G2. Western blot analysis provides signals of comparable protein amounts representing an average expression levels over ten thousands of cells. Nevertheless, the quantification results reflect in principle the IF pictures shown in Figure 3E but IF is probably not the best method to quantify protein amounts. We have also corrected Figure 3E.

Author response image 1.

(9) The choice of ARMC1 and YY2 is unclear. What are the criteria for the selection?

ARMC1 was one of the top hits in a pilot RIP-seq experiment (IP versus input and IP versus IgG IP). In the actual RIP-seq experiment with DKO HAP1 cells instead of IgG IP as a negative control, we found ARMC1 as an enriched hit, although it was not among the top 5 hits. The results from the 2nd RIP-seq further confirmed the validity of ARMC1 as an L1TD1interacting transcript. YY2 was of potential biological relevance as an L1TD1 target due to the fact that it is a processed pseudogene originating from YY1 mRNA as a result of retrotransposition. This is mentioned on page 6 of the revised manuscript.

Response: Appreciated!

(10) (P16) L1 is the only protein-coding transposon that is active in humans. This is perhaps too generalized of a statement as written. Other examples are readily found in the literature.

Please clarify.

We will tone down this statement in the revised manuscript.

Response: Appreciated! To further clarify, the term "active" when it comes to transposable elements, has not been solidified. It can span "retrotransposition competent" to "transcripts can be recovered". There are quite a few reports of GAG transcripts and protein from various ERV/LTR subfamilies in various cells and tissues (in mouse and human at least), however whether they contribute to new insertions is actively researched.

(11) In both the abstract and last sentence in the discussion section (P17), embryogenesis is mentioned, but this is not addressed at all in the manuscript. Please refrain from implying normal biological functions based on the results of this study unless appropriate samples are used to support them.

Much of the published data on L1TD1 function are related to embryonic stem cells [3- 7].

Therefore, it is important to discuss our findings in the context of previous reports.

Response: It is well established that embryonic stem cells are not a perfect or direct proxies for the inner cell mass of embryos, as multiple reports have demonstrated transcriptomic, epigenetic, chromatin accessibility differences. The exact origin of ES cells is also considered controversial. We maintain that the distinction between embryos/embryogenesis and the results presented in the manuscript are not yet interchangeable. An important exception would be complex models of embryogenesis such as embryoids, (or synthetic/artificial embryo models that have been carefully been termed as such so as to not suggest direct implications to embryos). https://www.nature.com/articles/ncb2965  

https://link.springer.com/article/10.1007/s00018-018-2965-y  

https://www.cell.com/developmental-cell/abstract/S1534-5807(24)00363-0?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS1534580724003630%3Fshowall%3Dtrue

We have deleted the corresponding paragraph in the Discussion.

(12) Figure 3E: The format of Figures 1A and 3E are internally inconsistent. Please present similar data/images in a cohesive way throughout the manuscript. We show now consistent IF Figures in the revised manuscript.

Response: Thanks

Minor:

In general:

Still need checking for typos, mostly in Materials and Methods section; Please keep a consistent writing style throughout the whole manuscript. If you use L1 ORF1p, then please use L1 instead of LINE-1, or if you keep LINE-1 in your manuscript, then you should use LINE-1 ORF1p.

A lab member from the US checked again the Materials and Methods section for typos. We keep the short version L1 ORF1p.

(1) Intro:

- Is L1Td1 in mice and Humans? How "conserved" is it and does this suggest function? Murine and human L1TD1 proteins share 44% identity on the amino acid level and it was suggested that the corresponding genes were under positive selection during evolution with functions in transposon control and maintenance of pluripotency [8].

- Why HAP1? (Haploid?) The importance of this cell line is not clear.

HAP1 is a nearly haploid human cancer cell line derived from the KBM-7 chronic myelogenous leukemia (CML) cell line [9, 10]. Due to its haploidy is perfectly suited and widely used for loss-of-function screens and gene editing. After gene editing cells can be used in the nearly haploid or in the diploid state. We usually perform all experiments with diploid HAP1 cell lines. Importantly, in contrast to other human tumor cell lines, this cell line tolerates ablation of DNMT1. We have included a corresponding explanation in the revised manuscript on page 5, first paragraph.

- Global methylation status in DNMT1 KO? (Methylations near L1 insertions, for example?)

The HAP1 DNMT1 KO cell line with a 20 bp deletion in exon 4 used in our study was validated in the study by Smits et al. [11]. The authors report a significant reduction in overall DNA methylation. However, we are not aware of a DNA methylome study on this cell line. We show now data on the methylation of L1 elements in HAP1 cells and upon DNMT1 deletion in the revised manuscript in Suppl. Figure S1B.

Response: Looks great!

(2) Figure 1:

- Figure 1C. Why is LMNB used instead of Actin (Fig1D)?

We show now beta-actin as loading control in the revised manuscript.

- Figure 1G shows increased Caspase 3 in KO, while the matching sentence in the result section skips over this. It might be more accurate to mention this and suggest that the single KO has perhaps an intermediate phenotype (Figure 1F shows a slight but not significant trend).

We fully agree with the reviewer and have changed the sentence on page 6, 2nd paragraph accordingly.

- Would 96 hrs trend closer to significance? An interpretation is that L1TD1 loss could speed up this negative consequence.

We thank the reviewer for the suggestion. We have performed a time course experiment with 6 biological replicas for each time point up to 96 hours and found significant changes in the viability upon loss of DNMT1 and again significant reduction in viability upon additional loss of L1TD1 (shown in Figure 1F). These data suggest that as expected loss of DNMT1 leads to significant reduction viability and that additional ablation of L1TD1 further enhances this effect.

Response: Looks good!

- What are the "stringent conditions" used to remove non-specific binders and artifacts (negative control subtraction?)

Yes, we considered only hits from both analyses, L1TD1 IP in KO versus input and L1TD1 IP in KO versus L1TD1 IP in DKO. This is now explained in more detail in the revised manuscript on page 6, 3rd paragraph.

(3) Figure 2:

- Figure 2A is a bit too small to read when printed.

We have changed this in the revised manuscript.

- Since WT and DKO lack detectable L1TD1, would you expect any difference in RIP-Seq results between these two?

Due to the lack of DNMT1 and the resulting DNA hypomethylation, DKO cells are more similar to KO cells than WT cells with respect to the expressed transcripts.

- Legend says selected dots are in green (it appears blue to me). We have changed this in the revised manuscript.

- Would you recover L1 ORF1p and its binding partners in the KO? (Is the antibody specific in the absence of L1TD1 or can it recognize L1?) I noticed an increase in ORF1p in the KO in Figure 3C.

Thank you for the suggestion. Yes, L1 ORF1p shows slightly increased expression in the proteome analysis and we have marked the corresponding dot in the Volcano plot (Figure 3A).

- Should the figure panel reference near the (Rosspopoff & Trono) reference instead be Sup S1C as well? Otherwise, I don't think S1C is mentioned at all.

- What are the red vs. green dots in 2D? Can you highlight ERV and ALU with different colors?

We added the reference to Suppl. Figure S1C (now S3C) in the revised manuscript. In Figure 2D L1 elements are highlighted in green, ERV elements in yellow, and other associated transposon transcripts in red.

Response: Much better, thanks!

- Which L1 subfamily from Figure 2D is represented in the qRT-PCR in 2E "LINE-1"? Do the primers match a specific L1 subfamily? If so, which? We used primers specific for the human L1.2 subfamily.

- Pulling down SINE element transcripts makes some sense, as many insertions "borrow" L1 sequences for non-autonomous retro transposition, but can you speculate as to why ERVs are recovered? There should be essentially no overlap in sequence.

In the L1TD1 evolution paper [8], a potential link between L1TD1 and ERV elements was discussed:

"Alternatively, L1TD1 in sigmodonts could play a role in genome defense against another element active in these genomes. Indeed, the sigmodontine rodents have a highly active family of ERVs, the mysTR elements [46]. Expansion of this family preceded the death of L1s, but these elements are very active, with 3500 to 7000 speciesspecific insertions in the L1-extinct species examined [47]. This recent ERV amplification in Sigmodontinae contrasts with the megabats (where L1TD1 has been lost in many species); there are apparently no highly active DNA or RNA elements in megabats [48]. If L1TD1 can suppress retroelements other than L1s, this could explain why the gene is retained in sigmodontine rodents but not in megabats."

Furthermore, Jin et al. report the binding of L1TD1 to repetitive sequences in transcripts [12]. It is possible that some of these sequences are also present in ERV RNAs.

Response: Interesting, thanks for sharing

- Is S2B a screenshot? (the red underline).

No, it is a Powerpoint figure, and we have removed the red underline.

(4) Figure 3:

- Text refers to Figure 3B as a western blot. Figure 3B shows a volcano plot. This is likely 3C but would still be out of order (3A>3C>3B referencing). I think this error is repeated in the last result section.

- Figure and legends fail to mention what gene was used for ddCT method (actin, gapdh, etc.).

- In general, the supplemental legends feel underwritten and could benefit from additional explanations. (Main figures are appropriate but please double-check that all statistical tests have been mentioned correctly).

Thank you for pointing this out. We have corrected these errors in the revised manuscript.

(5) Discussion:

- Aluy connection is interesting. Is there an "Alu retrotransposition reporter assay" to test whether L1TD1 enhances this as well?

Thank you for the suggestion. There is indeed an Alu retrotransposition reporter assay reported be Dewannieux et al. [13]. The assay is based on a Neo selection marker. We have previously tested a Neo selection-based L1 retrotransposition reporter assay, but this system failed to properly work in HAP1 cells, therefore we switched to a blasticidin based L1 retrotransposition reporter assay. A corresponding blasticidin-based Alu retrotransposition reporter assay might be interesting for future studies (mentioned in the Discussion, page 11 paragraph 4 of the revised manuscript.

(6) Material and Methods :

- The number of typos in the materials and methods is too numerous to list. Instead, please refer to the next section that broadly describes the issues seen throughout the manuscript.

Writing style

(1) Keep a consistent style throughout the manuscript: for example, L1 or LINE-1 (also L1 ORF1p or LINE-1 ORF1p); per or "/"; knockout or knock-out; min or minute; 3 times or three times; media or medium. Additionally, as TE naming conventions are not uniform, it is important to maintain internal consistency so as to not accidentally establish an imprecise version.

(2) There's a period between "et al" and the comma, and "et al." should be italic.

(3) The authors should explain what the key jargon is when it is first used in the manuscript, such as "retrotransposon" and "retrotransposition".

(4) The authors should show the full spelling of some acronyms when they use it for the first time, such as RNA Immunoprecipitation (RIP).

(5) Use a space between numbers and alphabets, such as 5 μg. (6) 2.0 × 105 cells, that's not an "x".

(7) Numbers in the reference section are lacking (hard to parse).

(8) In general, there are a significant number of typos in this draft which at times becomes distracting. For example, (P3) Introduction: Yet, co-option of TEs thorough (not thorough, it should be through) evolution has created so-called domesticated genes beneficial to the gene network in a wide range of organisms. Please carefully revise the entire manuscript for these minor issues that collectively erode the quality of this submission. Thank you for pointing out these mistakes. We have corrected them in the revised manuscript. A native speaker from our research group has carefully checked the paper. In summary, we have added Supplementary Figure S7C and have changed Figures 1C, 1E, 1F, 2A, 2D, 3A, 4B, S3A-D, S4B and S6A based on these comments.

Response: Thank you for taking these comments on board!

Author response:

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public Review): 

Summary: 

The manuscript focuses on the role of the deubiquitinating enzyme UPS-50/USP8 in endosome maturation. The authors aimed to clarify how this enzyme drives the conversion of early endosomes into late endosomes. Overall, they did achieve their aims in shedding light on the precise mechanisms by which UPS-50/USP8 regulates endosome maturation. The results support their conclusions that UPS-50 acts by disassociating RABX-5 from early endosomes to deactivate RAB-5 and by recruiting SAND-1/Mon1 to activate RAB-7. This work is commendable and will have a significant impact on the field. The methods and data presented here will be useful to the community in advancing our understanding of endosome maturation and identifying potential therapeutic targets for diseases related to endosomal dysfunction. It is worth noting that further investigation is required to fully understand the complexities of endosome maturation. However, the findings presented in this manuscript provide a solid foundation for future studies. 

We thank this reviewer for the instructive suggestions and encouragement.

Strengths: 

The major strengths of this work lie in the well-designed experiments used to examine the effects of UPS-50 loss. The authors employed confocal imaging to obtain a picture of the aftermath of the USP-50 loss. Their findings indicated enlarged early endosomes and MVB-like structures in cells deficient in USP-50/USP8. 

We thank this reviewer for the instructive suggestions and encouragement.

Weaknesses: 

Specifically, there is a need for further investigation to accurately characterize the anomalous structures detected in the usp-50 mutant. Also, the correlation between the presence of these abnormal structures and ESCRT-0 is yet to be addressed, and the current working model needs to be revised to prevent any confusion between enlarged early endosomes and MVBs. 

Excellent suggestions. USP8 has been identified as a protein associated with ESCRT components, which are crucial for endosomal membrane deformation and scission, leading to the formation of intraluminal vesicles (ILVs) within multivesicular bodies (MVBs). In usp-50 mutants, we observed a significant reduction in the punctate signals of HGRS-1::GFP and STAM-1 (Figure 1G and H; and Figure1-figure supplement 1B), indicating a disruption in ESCRT-0 complex localization (Author response image 1). Additionally, lysosomal structures are markedly reduced in these mutants. In contrast, we found that early endosomes, as marked by FYVE, RAB-5, RABEX5, and EEA1, are significantly enlarged in usp-50 mutants. Electron microscopy (EM) imaging further revealed an increase in large cellular vesicles containing various intraluminal structures. Given the reduction in lysosomal structures and the enlargement of early endosomes in usp-50 mutants, these enlarged vesicles are likely aberrant early endosomes rather than late endosomal or lysosomal structures. To address potential confusion, we have revised the manuscript according to the reviewer's comments and updated the model to accurately reflect these observations.

Reviewer #2 (Public Review): 

Summary: 

In this study, the authors study how the deubiquitinase USP8 regulates endosome maturation in C. elegans and mammalian cells. The authors have isolated USP8 mutant alleles in C. elegans and used multiple in vivo reporter lines to demonstrate the impact of USP8 loss-of-function on endosome morphology and maturation. They show that in USP8 mutant cells, the early endosomes and MVB-like structures are enlarged while the late endosomes and lysosomal compartments are reduced. They elucidate that USP8 interacts with Rabx5, a guanine nucleotide exchange factor (GEF) for Rab5, and show that USP8 likely targets specific lysine residue of Rabx5 to dissociate it from early endosomes. They also find that the localization of USP8 to early endosomes is disrupted in Rabx5 mutant cells. They observe that in both Rabx5 and USP8 mutant cells, the Rab7 GEF SAND-1 puncta which likely represents late endosomes are diminished, although Rabex5 is accumulated in USP8 mutant cells. The authors provide evidence that USP8 regulates endosomal maturation in a similar fashion in mammalian cells. Based on their observations they propose that USP8 dissociates Rabex5 from early endosomes and enhances the recruitment of SAND-1 to promote endosome maturation. 

We thank this reviewer for the instructive suggestions and encouragement.

Strengths: 

The major highlights of this study include the direct visualization of endosome dynamics in a living multi-cellular organism, C. elegans. The high-quality images provide clear in vivo evidence to support the main conclusions. The authors have generated valuable resources to study mechanisms involved in endosome dynamics regulation in both the worm and mammalian cells, which would benefit many members of the cell biology community. The work identifies a fascinating link between USP8 and the Rab5 guanine nucleotide exchange factor Rabx5, which expands the targets and modes of action of USP8. The findings make a solid contribution toward the understanding of how endosomal trafficking is controlled. 

We thank this reviewer for the instructive suggestions and encouragement.

Weaknesses: 

- The authors utilized multiple fluorescent protein reporters, including those generated by themselves, to label endosomal vesicles. Although these are routine and powerful tools for studying endosomal trafficking, these results cannot tell whether the endogenous proteins (Rab5, Rabex5, Rab7, etc.) are affected in the same fashion. 

Good suggestion. Indeed, to test whether the endogenous proteins (Rab5, Rabex5, Rab7, etc.) are affected in the same fashion as fluorescent protein reporters, we supplemented our approach with the utilization of endogenous markers. These markers, including Rab5, RAB-5, Rabex5, RABX-5, and EEA1 for early endosomes, as well as RAB-7, Mon1a, and Mon1b for late endosomes, were instrumental in our investigations (refer to Figure 3, Figure 6, Figure 5-figure supplement 1, Figure 5-figure supplement 2, and Figure 6-figure supplement 1). Our comprehensive analysis, employing various methodologies such as tissue-specific fused proteins, CRISPR/Cas9 knock-in, and antibody staining, consistently highlights the critical role of USP8 in early-to-late endosome conversion.

- The authors clearly demonstrated a link between USP8 and Rabx5, and they showed that cells deficient in both factors displayed similar defects in late endosomes/lysosomes. However, the authors didn't confirm whether and/or to which extent USP8 regulates endosome maturation through Rabx5. Additional genetic and molecular evidence might be required to better support their working model. 

Excellent point. To test whether USP-50 regulates endosome maturation through RABX-5, we performed additional genetic analyses. In rabx-5(null) mutant animals, the morphology of 2xFYVE-labeled early endosomes is comparable to that of wild-type controls (Figure 4H and I). Introducing the rabx-5(null) mutation into usp-50(xd413) backgrounds resulted in a significant suppression of the enlarged early endosome phenotype characteristic of usp-50(xd413) mutants (Figure 4H and I). These findings suggest that USP-50 may modulate the size of early endosomes through its interaction with RABX-5.

Reviewer #3 (Public Review): 

Summary: 

The authors were trying to elucidate the role of USP8 in the endocytic pathway. Using C. elegans epithelial cells as a model, they observed that when USP8 function is lost, the cells have a decreased number and size in lysosomes. Since USP8 was already known to be a protein linked to ESCRT components, they looked into what role USP8 might play in connecting lysosomes and multivesicular bodies (MVB). They observed fewer ESCRT-associated vesicles but an increased number of "abnormal" enlarged vesicles when USP8 function was lost. At this specific point, it's not clear what the objective of the authors was. What would have been their hypothesis addressing whether the reduced lysosomal structures in USP8 (-) animals were linked to MVB formation? Then they observed that the abnormally enlarged vesicles, marked by the PI3P biosensor YFP-2xFYVE, are bigger but in the same number in USP8 (-) compared to wild-type animals, suggesting homotypic fusion. They confirmed this result by knocking down USP8 in a human cell line, and they observed enlarged vesicles marked by YFP-2xFYVE as well. At this point, there is quite an important issue. The use of YFP-2xFYVE to detect early endosomes requires the transfection of the cells, which has already been demonstrated to produce differences in the distribution, number, and size of PI3P-positive vesicles (doi.org/10.1080/15548627.2017.1341465). The enlarged vesicles marked by YFP-2xFYVE would not necessarily be due to the loss of UPS8. In any case, it appears relatively clear that USP8 localizes to early endosomes, and the authors claim that this localization is mediated by Rabex-5 (or Rabx-5). They finally propose that USP8 dissociates Rabx-5 from early endosomes facilitating endosome maturation. 

Weaknesses: 

The weaknesses of this study are, on one side, that the results are almost exclusively dependent on the overexpression of fusion proteins. While useful in the field, this strategy does not represent the optimal way to dissect a cell biology issue. On the other side, the way the authors construct the rationale for each approximation is somehow difficult to follow. Finally, the use of two models, C. elegans and a mammalian cell line, which would strengthen the observations, contributes to the difficulty in reading the manuscript. 

The findings are useful but do not clearly support the idea that USP8 mediates Rab5-Rab7 exchange and endosome maturation, In contrast, they appear to be incomplete and open new questions regarding the complexity of this process and the precise role of USP8 within it. 

We thank this reviewer for the insightful comments. Fluorescence-fused proteins serve as potent tools for visualizing subcellular organelles both in vivo and in live settings. Specifically, in epidermal cells of worms, the tissue-specific expression of these fused proteins is indispensable for studying organelle dynamics within living organisms. This approach is necessitated by the inherent limitations of endogenously tagged proteins, whose fluorescence signals are often weak and unsuitable for live imaging or genetic screening purposes. Acknowledging concerns raised by the reviewer regarding potential alterations in organelle morphology due to overexpression of certain fused proteins, we supplemented our approach with the utilization of endogenous markers. These markers, including Rab5, RAB-5, Rabex5, RABX-5, and EEA1 for early endosomes, as well as RAB-7, Mon1a, and Mon1b for late endosomes, were instrumental in our investigations (refer to Figure 3, Figure 6, Figure 5-figure supplement 1, Figure 5-figure supplement 2, and Figure 6-figure supplement 1). Our comprehensive analysis, employing various methodologies such as tissue-specific fused proteins, CRISPR/Cas9 knock-in, and antibody staining, consistently highlights the critical role of USP8 in early-to-late endosome conversion. Specifically, we discovered that the recruitment of USP-50/USP8 to early endosomes is depending on Rabex5. However, instead of stabilizing Rabex5, the recruitment of USP-50/USP8 leads to its dissociation from endosomes, concomitantly facilitating the recruitment of the Rab7 GEF SAND-1/Mon1. In cells with loss-of-function mutations in usp-50/usp8, we observed enhanced RABX-5/Rabex5 signaling and mis-localization of SAND-1/Mon1 proteins from endosomes. Consequently, this disruption impairs endolysosomal trafficking, resulting in the accumulation of enlarged vesicles containing various intraluminal contents and rudimentary lysosomal structures.

Through an unbiased genetic screen, verified by cultured mammalian cell studies, we observed that loss-of-function mutations in usp-50/usp8 result in diminished lysosome/late endosomes. Electron microscopy (EM) analysis indicated that usp-50 mutation leads to abnormally enlarged vesicles containing various intraluminal structures in worm epidermal cells. USP8 is known to regulate the endocytic trafficking and stability of numerous transmembrane proteins. Given that lysosomes receive and degrade materials generated by endocytic pathways, we hypothesized that the abnormally enlarged vesicular structures observed in usp-50 or usp8 mutant cells correspond to the enlarged vesicles coated by early endosome markers. Indeed, in the absence of usp8/usp-50, the endosomal Rab5 signal is enhanced, while early endosomes are significantly enlarged. Given that Rab5 guanine nucleotide exchange factor (GEF), Rabex5, is essential for Rab5 activation, we further investigated its dynamics. Additional analyses conducted in both worm hypodermal cells and cultured mammalian cells revealed an increase of endosomal Rabex5 in response to usp8/usp-50 loss-of-function. Live imaging studies further demonstrated active recruitment of USP8 to newly formed Rab5-positive vesicles, aligning spatiotemporally with Rabex5 regulation. Through systematic exploration of putative USP-50 binding partners on early endosomes, we identified its interaction with Rabex5. Comprehensive genetics and biochemistry experiments demonstrated that USP8 acts through K323 site de-ubiquitination to dissociate Rabex5 from early endosomes and promotes the recruitment of the Rab7 GEF SAND-1/Mon1. In summary, our study began with an unbiased genetic screen and subsequent examination of established theories, leading to the formulation of our own hypothesis. Through multifaceted approaches, we unveiled a novel function of USP8 in early-to-late endosome conversion.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

(1) Within Figures 1K-N, diverse anomalous structures were detected in the usp-50 mutant. Further scrutiny is needed to definitively characterize these structures, particularly as the images in Figures 1M and 1L exhibit notable similarities to lamellar bodies.

We thank the reviewer for the insightful question regarding the resemblance between the vesicles observed in our study and lamellar bodies (LBs). Lamellar bodies are specialized organelles involved in lipid storage and secretion1, prominently studied in keratinocytes of the skin and alveolar type II (ATII) epithelial cells in the lung2. These organelles contain not only lipids but also cell-type specific proteins and lytic enzymes. Due to their acidic pH and functional similarities, LBs are classified as lysosome-related organelles (LROs) or secretory lysosomes3,4. In usp-50 mutants, we observed a considerable number of abnormal vesicles, some of which contain threadlike membrane structures and exhibit morphological similarities to LBs (Figure 2O). However, further analysis with a comprehensive panel of lysosome-related markers demonstrated a significant reduction in lysosomal structures within these mutants. In contrast, vesicles marked by early endosome markers, such as FYVE, RAB-5, RABX-5, and EEA1, were notably enlarged. These results suggest that the enlarged vesicles observed in usp-50 mutants are more likely aberrant early endosomes rather than true lamellar bodies. We have revised the manuscript to reflect these findings and to clearly differentiate between these structures and lysosome-related organelles.

(2) The correlation between the presence of these abnormal structures and ESCRT-0 remains unaddressed, thus the assertion that UPS-50 regulates endolysosome trafficking in conjunction with ESCRT-0 lacks empirical support.

We thank the reviewer for the valuable suggestions. We apologize for any confusion and appreciate the opportunity to clarify our findings. The ESCRT machinery is essential for driving endosomal membrane deformation and scission, which leads to the formation of intraluminal vesicles (ILVs) within multivesicular bodies (MVBs). Recent research has shown that the absence of ESCRT components results in a reduction of ILVs in worm gut cells5. In wild type animals, the ESCRT-0 components HGRS-1 and STAM-1 display a distinct punctate distribution (Figure 1G and H). However, in usp-50 mutants, the punctate signals of HGRS-1::GFP and STAM-1::GFP are significantly reduced (Figure 1G and H; and Figure 1-figure supplement 1B), indicating a role for USP-50 in stabilizing the ESCRT-0 complex. Our TEM analysis revealed an accumulation of abnormally enlarged vesicles containing intraluminal structures in usp-50 mutants. When we examined a panel of early endosome and late endosome/lysosome markers, we found that early endosomes are significantly enlarged, while late endosomal/lysosomal structures are markedly reduced in these mutants. This suggests that the abnormal structures observed in usp-50 mutants are likely enlarged early endosomes rather than classical MVBs. To further investigate whether the reduction in ESCRT components contributes to the late endosome/lysosome defects, we analyzed stam-1 mutants. In these mutants, the size of RAB-7-coated vesicles was reduced (Author response image 1C), and the lysosomal marker LAAT-1 indicated a reduction in lysosomal structures (Author response image 1B). These results highlight the importance of the ESCRT complex in late endosome/lysosome formation. However, the morphology of early endosomes, as marked by 2xFYVE, remained similar to that of wild type in stam-1 mutants (Author response image 1A). Therefore, while reduced ESCRT-0 components may contribute to the late endosome/lysosome defects observed in usp-50 mutants, the enlargement of early endosomes in these mutants may involve additional mechanisms. We have revised the manuscript to incorporate these insights and to address the reviewer's comments more comprehensively.

Author response image 1.

(A) Confocal fluorescence images of hypodermis expressing YFP::2xFYVE to detect EEs in L4 stage animals in wild type and stam-1(ok406) mutants. Scale bar: 5 μm. (B) Confocal fluorescence images of hypodermal cell 7 (hyp7) expressing the LAAT-1::GFP marker to highlight lysosome structures in 3-day-old adult animals. Compared to wild type, LAAT-1::GFP signal is reduced in stam-1(ok406) animals. Scale bar, 5 μm. (C) The reduction of punctate endogenous GFP::RAB-7 signals in stam-1(ok406) animals. Scale bar: 10 μm.

(3) Endosomal dysfunction typically leads to significant alterations in the spatial arrangement of marker proteins across distinct endosomes. In the manuscript, the authors examined the distribution and morphology of early endosomes, multivesicular bodies (MVBs), late endosomes, and lysosomes in a usp-50 deficient background primarily through single-channel confocal imaging. By employing two color images showing RAB-5 and RAB-7, in conjunction with HGRS-1, a more comprehensive picture of the aftermath of USP-50 loss can be obtained.

Good suggestions. We have conducted a double-labeling analysis to examine the distribution of RAB-5 and RAB-7 in conjunction with HGRS-1. In wild type animals, HGRS-1 exhibits a punctate distribution that is partially co-localized with both RAB-5 and RAB-7. In contrast, in usp-50 mutants, the punctate signal of HGRS-1 is significantly reduced, along with its co-localization with RAB-5 and RAB-7 (Author response image 2). These results suggest that, in the absence of USP-50, the stabilization of ESCRT-0 components on endosomes is compromised.

Author response image 2.

ESCRT-0 is adjacent to both early endosomes and late endosomes. (A) Confocal fluorescence images of wild-type and usp-50(xd413) hypodermis at L4 stage co-expressing HGRS-1::GFP (hgrs-1 promoter) and endogenous wrmScarlet::RAB-5. (B) HGRS-1 and RAB-5 puncta were analyzed to produce Manders overlap coefficient M1 (HGRS-1/RAB-5) and M2 (RAB-5/HGRS-1) (N=10). (C) Confocal fluorescence images of wild-type and usp-50(xd413) hypodermis at L4 stage co-expressing HGRS-1::GFP (hgrs-1 promoter) and endogenous wrmScarlet::RAB-7. (D) HGRS-1 and RAB-7 puncta were analyzed to produce Manders overlap coefficient M1 (HGRS-1/RAB-7) and M2 (RAB-7/HGRS-1) (N=10). Scale bar: 10 μm for (A) and (C).

(4) The authors observed enlarged early endosomes in cells depleted of usp-50/usp8, along with enlarged MVB-like structures identified through TEM. The potential identity of these structures as the same organelle could be determined using CLEM.

We thank the reviewer for the valuable suggestion. Our TEM analysis identified a large number of abnormally enlarged vesicles with various intraluminal structures accumulated in usp-50 mutants. As the reviewer correctly noted, CLEM (correlative light and electron microscopy) would be an ideal approach to further characterize these structures. We have been attempting to implement CLEM in C. elegans for a few years. Given that CLEM relies on fluorescence markers, in this study we focused on two tagged proteins, RAB-5 and RABX-5, which show enlargement in their vesicles in usp-50 mutants. Unfortunately, we encountered significant challenges with this approach, as the GFP-tagged RAB-5 and RABX-5 signals did not survive the electron microscopy procedure. Attempts to align EM sections with residual GFP signaling yielded results that were not convincing. Consequently, we concentrated our analysis on a panel of molecular markers, including 2xFYVE, RAB-5, RABX-5, RAB-7, and LAAT-1. These markers consistently indicated that early endosomes are specifically enlarged in usp-50 mutants, while late endosomal/lysosomal structures are notably reduced. Thus, the abnormal structures identified in usp-50 mutants via TEM are likely to be enlarged early endosomes rather than the classical view of MVBs. We have revised the manuscript to reflect these findings and to clarify this point.

(5) The working model depicted in Figure 6 Y (right) requires revision, as it has the potential to mislead authors into mistaking enlarged early endosomes for multivesicular bodies (MVBs).

We thank the reviewer for the excellent suggestion. We have revised the model to clarify that it is the enlarged early endosomes, rather than MVBs, that are observed in usp-50 mutants.

Reviewer #2 (Recommendations For The Authors):

(1) Is there any change of Rabx5 protein level in USP8/USP50 mutant cells?

Good question. In the absence of usp-50/usp8, we indeed observed a noticeable increase in the signal of Rabex5 on endosomes. To determine whether usp-50/usp8 affects the protein level of Rabex5, we investigated the endogenous levels of RABX-5 using the RABX-5::GFP knock-in line. Compared to wild-type controls, we found an elevated protein level of RABX-5::GFP in the knock-in line (Author response image 3). This suggests that USP-50 may play a role in the destabilization of RABX-5/Rabex5 in vivo.

Author response image 3.

The endogenous RABX-5 protein level is increased in usp-50 mutants. (A) The RABX-5::GFP KI protein level is increased in usp-50(xd413). (B) Quantification of endogenous RABX-5::GFP protein level in wild type and usp-50(xd413) mutant animals.

(2) It is interesting that "The rabx-5(null) animals are healthy and fertile and do not display obvious morphological or behavioral defects.", which seems contrary to its role in regulating USP8 localization and endosome maturation.

It has been previously documented that rabx-5 functions redundantly with rme-6, another RAB-5 GEF in C. elegans, to regulate RAB-5 localization in oocytes6. RNA interference (RNAi) targeting rabx-5 in a rme-6 mutant background results in synthetic lethality, whereas neither rabx-5 nor rme-6 single mutants are essential for worm viability. RME-6 co-localizes with clathrin-coated pits, while Rabex-5 is localized to early endosomes. Rabex-5 forms a stable complex with Rabaptin-5 and is part of a large EEA1-positive complex on early endosomes, whereas RME-6 does not interact with Rabaptin-5 (RABN-5) or EEA-1. These findings suggest that while RME-6 and RABX-5 may function redundantly, they likely play distinct roles in regulating intracellular trafficking processes. In the absence of RABX-5, USP-50 appears to lose its endosomal localization, although the size of the early endosome remains comparable to that of wild type. This observation contrasts with the phenotype associated with USP-50 loss-of-function, in which the early endosome is notably enlarged. These results suggest that residual USP-50 present in the endosomes is sufficient to maintain its role in the endocytic pathway. Conversely, the complete absence of USP-50 likely disrupts the transition of early endosomes to late endosomes, indicating a crucial role of USP-50 in this conversion process. It is also noteworthy that, although loss-of-function of rabx-5 does not result in obvious changes to early endosomes, increasing the gene expression level of rabx-5/Rabex-5 alone is sufficient to cause enlargement of early endosomes (Author response image 4) . Indeed, we observed that loss-of-function mutations in u_sp-50/usp_8 lead to abnormally enlarged early endosomes, accompanied by an enhanced signal of endosomal RABX-5. When the rabx-5(null) mutation was introduced into usp-50 mutant animals, the enlarged early endosome phenotype seen in usp-50 mutants was significantly suppressed (Figure 4H and I). This implies that maintaining a lower level of Rab5 GEF may be crucial for endolysosomal trafficking.

(3) Does Rabx5 mutation has any impact on early endosomes?

To address the question, we utilized the CRISPR/Cas9 technique to create a molecular null for rabx-5 (Figure 4E). In the rabx-5(null) mutant animals, we found that the 2xFYVE-labeled early endosomes are indistinguishable from wild type (Figure 4H and 4I). Given that r_abx-5_ functions redundantly with rme-6, another RAB-5 GEF in C. elegans, it is likely that the regulation of early endosome size involves a cooperative interaction between RABX-5 and RME-6.

(4) The authors observed a reduction of ESCRT-0 components in USP8 mutant cells, could this contribute to the late endosome/lysosome defects?

Good suggestion. In wild-type animals, the two ESCRT-0 components, HGRS-1 and STAM-1, exhibit a distinct punctate distribution (Figure 1G and H). However, in usp-50 mutants, the punctate signals of HGRS-1::GFP and STAM-1::GFP are significantly diminished (Figure 1G and H; and Figure 1-figure supplement 1B), which aligns with the role of USP-50 in stabilizing the ESCRT-0 complex. To investigate whether the reduction in ESCRT components might contribute to defects in late endosome/lysosome formation, we examined stam-1 mutants. In stam-1 mutants, we observed a reduction in the size of RAB-7-coated vesicles (Author response image 1). Further, when we introduced the lysosomal marker LAAT-1::GFP into stam-1 mutants, we found a substantial decrease in lysosomal structures compared to wild-type animals (Author response image 1). This suggests that the ESCRT complex is essential for proper late endosome/lysosome formation. In contrast, the morphology of early endosomes, as indicated by the 2xFYVE marker, appeared normal in stam-1 mutants, similar to wild-type animals (Author response image 1). This implies that while a reduction in ESCRT-0 components may contribute to the late endosome/lysosome defects observed in usp-50 mutants, the early endosome enlargement phenotype in _usp-5_0 mutants may involve additional mechanisms.

(5) Rabx5 is accumulated in USP8 mutant cells, I am very curious about the phenotype of USP8-Rabx5 double mutants. Could over-expression of Rabx5 (wild type or mutant forms) cause any defects?

Excellent suggestions. To address the question, we employed the CRISPR/Cas9 technique to create a molecular null for rabx-5 (Figure 4E). In the rabx-5(null) mutant animals, we observed that the punctate USP-50::GFP signal became diffusely distributed (Figure 4F and G). This suggests that rabx-5 is necessary for the endosomal localization of USP-50. Interestingly, in rabx-5(null) mutant animals, the 2xFYVE-labeled early endosomes appeared similar to those in wild-type animals (Figure 4H and I). When rabx-5(null) was introduced into usp-50 mutant animals, the enlarged early endosome phenotype observed in usp-50 was significantly suppressed (Figure 4H and I). This finding indicates that usp-50 indeed functions through rabx-5 to regulate early endosome size. Additionally, we constructed strains overexpressing either wild-type or K323R mutant RABX-5. Our results showed that overexpression of wild-type RABX-5 led to early endosome enlargement (as indicated by YFP::2xFYVE labeling) (Author response image 4A, B and D). In contrast, overexpression of the K323R mutant RABX-5 did not result in noticeable early endosome enlargement (Author response image 4A, C and D). Together, these data are in consistent with our model that USP-50 may regulate RABX-5 by deubiquitinating the K323 site.

Author response image 4.

(A-C) Over-expression wild type RABX-5 causes enlarged EEs (labeled by YFP::2xFYVE) while RABX-5(K323R) mutant form does not. (D) Quantification of the volume of individual YFP::2xFYVE vesicles. Data are presented as mean ± SEM. ****P<0.0001. ns, not significant. One-way ANOVA with Tukey’s test.

(6) Rabx5 could be ubiquitinated at K88 and K323, and Rabx5-K323R showed different activity when compared with the wild-type protein in USP8 mutant cells. Could the authors provide evidence that USP8 could remove the ubiquitin modification from K323 in Rabx5 protein?

We appreciate the reviewer's insightful suggestions. To explore the potential of USP-50 in removing ubiquitin modifications from lysine 323 on the RABX-5 protein, we undertook a series of experiments. Initially, we sought to determine whether USP-50 influences the ubiquitination level of RABX-5 in vivo. However, due to the low expression levels of USP-50, we encountered challenges in obtaining adequate amounts of USP-50 protein from worm lysates. To overcome this, we expressed USP-50::4xFLAG in HEK293 cells for subsequent affinity purification. Concurrently, we utilized anti-GFP agarose beads to purify RABX-5::GFP from worms expressing the rabx-5::gfp construct. We then incubated RABX-5::GFP with USP-50::4xFLAG for varying durations and performed immunoblotting with an anti-ubiquitin antibody. As shown in Author response image 5A, our results revealed a decrease in the ubiquitination level of RABX-5 in the presence of USP-50, suggesting that USP-50 directly deubiquitinates RABX-5. Previous studies have indicated that only a minor fraction of recombinant RABX-5 undergoes ubiquitination in HeLa cells, which is believed to have functional significance7. Our findings are consistent with this observation, as only a small fraction of RABX-5 in worms is ubiquitinated. Rabex-5 is known to interact with both K63- and K48-linked poly-ubiquitin chains. To further elucidate whether USP-50 specifically targets K48 or K63-linked ubiquitination at the K323 site of RABX-5, we incubated various HA-tagged ubiquitin mutants with either wild-type or K323R mutant RABX-5 protein. Our results indicated that the K323R mutation reduces K63-linked ubiquitination of RABX-5 (Author response image 5). This experiment was repeated multiple times with consistent results. Additionally, while overexpression of wild-type RABX-5 led to an enlargement of early endosomes, as evidenced by YFP::2xFYVE labeling, overexpression of the K323R mutant did not produce a noticeable effect on endosome size (Author response image 4). Collectively, this finding indicates that RABX-5 is subject to ubiquitin modification in vivo and that USP-50 plays a significant role in regulating this modification at the K323 site.

Author response image 5.

(A) RABX-5::GFP protein was purified from worm lysates using anti-GFP antibody. FLAG-tagged USP-50 was purified from HEK293T cells using anti-FLAG antibody. Purified RABX-5::GFP was incubated with USP-50::4FLAG for indicated times (0, 15, 30, 60 mins), followed by immunoblotting using antibody against ubiquitin, FLAG or GFP. In the presence of USP-50::4xFLAG, the ubiquitination level of RABX-5::GFP is decreased. (B) Quantification of RABX-5::GFP ubiquitination level from three independent experiments. (C) HEK293T cells were transfected with HA-Ub or indicated mutants and 4xFLAG tagged RABX-5 or RABX-5 K323R mutant for 48h. The cells were subjected to pull down using the FLAG beads, followed by immunoblotting using antibody against HA or Flag.

(7) The authors described "the almost identical phenotype of usp-50/usp8 and sand-1/Mon1 mutants", found protein-protein interaction between USP8 and sand-1, and showed that sand1-GFP signal is diminished in USP8 mutant cells. These observations fit with the possibility that USP8 regulates the stability of sand-1 to promote endosomal maturation. Could this be tested and integrated into the current model?

are grateful for the insightful comments provided by the reviewer. Rab5, known to be activated by Rabex-5, plays a crucial role in the homotypic fusion of early endosomes. Rab5 effectors also include the Rab7 GEF SAND-1/Mon1–Ccz1 complex. Rab7 activation by SAND-1/Mon1-Ccz1 complex is essential for the biogenesis and positioning of late endosomes (LEs) and lysosomes, and for the fusion of endosomes and autophagosomes with lysosomes. The Mon1-Ccz1 complex is able to interact with Rabex5, causing dissociation of Rabex5 from the membrane, which probably terminates the positive feedback loop of Rab5 activation and then promotes the recruitment and activation of Rab7 on endosomes. In our study, we identified an interaction between USP-50 and the Rab5 GEF, RABX-5. In the absence of USP-50, we observed an increased endosomal localization of RABX-5 and the formation of abnormally enlarged early endosomes. This phenotype is reminiscent of that seen in sand-1 loss-of-function mutants, which also exhibit enlarged early endosomes and a concomitant reduction in late endosomes/lysosomes. Notably, USP-50 also interacts with SAND-1, suggesting a potential role in regulating its localization. We could propose several models to elucidate how USP-50 might influence SAND-1 localization, including:

(1) USP-50 may stabilize SAND-1 through direct de-ubiquitination.

(2) In the absence of USP-50, the sustained presence of RABX-5 could lead to continuous Rab5 activation, which might hinder or delay the recruitment of SAND-1.

(3) USP-50 could facilitate SAND-1 recruitment by promoting the dissociation of RABX-5.

We are actively investigating these models in our laboratory. Due to space constraints, a more detailed exploration of how USP-50 regulates SAND-1 stability will be presented in a separate publication.

References:

(1) Schmitz, G., and Müller, G. (1991). Structure and function of lamellar bodies, lipid-protein complexes involved in storage and secretion of cellular lipids. J Lipid Res 32, 1539-1570.

(2) Dietl, P., and Frick, M. (2021). Channels and Transporters of the Pulmonary Lamellar Body in Health and Disease. Cells-Basel 11. https://doi.org/10.3390/cells11010045.

(3) Raposo, G., Marks, M.S., and Cutler, D.F. (2007). Lysosome-related organelles: driving post-Golgi compartments into specialisation. Current opinion in cell biology 19, 394-401. https://doi.org/10.1016/j.ceb.2007.05.001.

(4) Weaver, T.E., Na, C.L., and Stahlman, M. (2002). Biogenesis of lamellar bodies, lysosome-related organelles involved in storage and secretion of pulmonary surfactant. Semin Cell Dev Biol 13, 263-270. https://doi.org/10.1016/s1084952102000551.

(5) Ott, D.P., Desai, S., Solinger, J.A., Kaech, A., and Spang, A. (2024). Coordination between ESCRT function and Rab conversion during endosome maturation. bioRxiv, 2024.2005.2014.594104. https://doi.org/10.1101/2024.05.14.594104.

(6) Sato, M., Sato, K., Fonarev, P., Huang, C.J., Liou, W., and Grant, B.D. (2005). Caenorhabditis elegans RME-6 is a novel regulator of RAB-5 at the clathrin-coated pit. Nature cell biology 7, 559-569. https://doi.org/10.1038/ncb1261.

(7) Mattera, R., Tsai, Y.C., Weissman, A.M., and Bonifacino, J.S. (2006). The Rab5 guanine nucleotide exchange factor Rabex-5 binds ubiquitin (Ub) and functions as a Ub ligase through an atypical Ub-interacting motif and a zinc finger domain. The Journal of biological chemistry 281, 6874-6883. https://doi.org/10.1074/jbc.M509939200.

Author Response

The following is the authors’ response to the previous reviews.

We thank the reviewers for truly valuable advice and comments. We have made multiple corrections and revisions to the original pre-print accordingly per the following comments:

1. Pro1153Leu is extremely common in the general population (allele frequency in gnomAD is 0.5). Further discussion is warranted to justify the possibility that this variant contributes to a phenotype documented in 1.5-3% of the population. Is it possible that this variant is tagging other rare SNPs in the COL11A1 locus, and could any of the existing exome sequencing data be mined for rare nonsynonymous variants?

One possible avenue for future work is to return to any existing exome sequencing data to query for rare variants at the COL11A1 locus. This should be possible for the USA MO case-control cohort. Any rare nonsynonymous variants identified should then be subjected to mutational burden testing, ideally after functional testing to diminish any noise introduced by rare benign variants in both cases and controls. If there is a significant association of rare variation in AIS cases, then they should consider returning to the other cohorts for targeted COL11A1 gene sequencing or whole exome sequencing (whichever approach is easier/less expensive) to demonstrate replication of the association.

Response: Regarding the genetic association of the common COL11A1 variant rs3753841 (p.(Pro1335Leu)), we do not propose that it is the sole risk variant contributing to the association signal we detected and have clarified this in the manuscript. We concluded that it was worthy of functional testing for reasons described here. Although there were several common variants in the discovery GWAS within and around COL11A1, none were significantly associated with AIS and none were in linkage disequilibrium (R2>0.6) with the top SNP rs3753841. We next reviewed rare (MAF<=0.01) coding variants within the COL11A1 LD region of the associated SNP (rs3753841) in 625 available exomes representing 46% of the 1,358 cases from the discovery cohort. The LD block was defined using Haploview based on the 1KG_CEU population. Within the ~41 KB LD region (chr1:103365089- 103406616, GRCh37) we found three rare missense mutations in 6 unrelated individuals, Table below. Two of them (NM_080629.2: c.G4093A:p.A1365T; NM_080629.2:c.G3394A:p.G1132S), from two individuals, are predicted to be deleterious based on CADD and GERP scores and are plausible AIS risk candidates. At this rate we could expect to find only 4-5 individuals with linked rare coding variants in the total cohort of 1,358 which collectively are unlikely to explain the overall association signal we detected. Of course, there also could be deep intronic variants contributing to the association that we would not detect by our methods. However, given this scenario, the relatively high predicted deleteriousness of rs3753841 (CADD= 25.7; GERP=5.75), and its occurrence in a GlyX-Y triplet repeat, we hypothesized that this variant itself could be a risk allele worthy of further investigation.

Author response table 1.

We also appreciate the reviewer’s suggestion to perform a rare variant burden analysis of COL11A1. We did conduct pilot gene-based analysis in 4534 European ancestry exomes including 797 of our own AIS cases and 3737 controls and tested the burden of rare variants in COL11A1. SKATO P value was not significant (COL11A1_P=0.18), but this could due to lack of power and/or background from rare benign variants that could be screened out using the functional testing we have developed.

1. COL11A1 p.Pro1335Leu is pursued as a direct candidate susceptibility locus, but the functional validation involves both: (a) a complementation assay in mouse GPCs, Figure 5; and (b) cultured rib cartilage cells from Col11a1-Ad5 Cre mice (Figure 4). Please address the following:

2A. Is Pro1335Leu a loss of function, gain of function, or dominant negative variant? Further rationale for modeling this change in a Col11a1 loss of function cell line would be helpful.

Response: Regarding functional testing, by knockdown/knockout cell culture experiments, we showed for the first time that Col11a1 negatively regulates Mmp3 expression in cartilage chondrocytes, an AIS-relevant tissue. We then tested the effect of overexpressing the human wt or variant COL11A1 by lentiviral transduction in SV40-transformed chondrocyte cultures. We deleted endogenous mouse Col11a1 by Cre recombination to remove the background of its strong suppressive effects on Mmp3 expression. We acknowledge that Col11a1 missense mutations could confer gain of function or dominant negative effects that would not be revealed in this assay. However as indicated in our original manuscript we have noted that spinal deformity is described in the cho/cho mouse, a Col11a1 loss of function mutant. We also note the recent publication by Rebello et al. showing that missense mutations in Col11a2 associated with congenital scoliosis fail to rescue a vertebral malformation phenotype in a zebrafish col11a2 KO line. Although the connection between AIS and vertebral malformations is not altogether clear, we surmise that loss of the components of collagen type XI disrupt spinal development. in vivo experiments in vertebrate model systems are needed to fully establish the consequences and genetic mechanisms by which COL11A1 variants contribute to an AIS phenotype.

2B. Expression appears to be augmented compared WT in Fig 5B, but there is no direct comparison of WT with variant.

Response: Expression of the mutant (from the lentiviral expression vector) is increased compared to mutant. We observed this effect in repeated experiments. Sequencing confirmed that the mutant and wildtype constructs differed only at the position of the rs3753841 SNP. At this time, we cannot explain the difference in expression levels. Nonetheless, even when the variant COL11A1 is relatively overexpressed it fails to suppress MMP3 expression as observed for the wildtype form.

2C. How do the authors know that their complementation data in Figure 5 are specific? Repetition of this experiment with an alternative common nonsynonymous variant in COL11A1 (such as rs1676486) would be helpful as a comparison with the expectation that it would be similar to WT.

Response: We agree that testing an allelic series throughout COL11A1 could be informative, but we have shifted our resources toward in vivo experiments that we believe will ultimately be more informative for deciphering the mechanistic role of COL11A1 in MMP3 regulation and spine deformity.

2D. The y-axes of histograms in panel A need attention and clarification. What is meant by power? Do you mean fold change?

Response: Power is directly comparable to fold change but allows comparison of absolute expression levels between different genes.

2E. Figure 5: how many technical and biological replicates? Confirm that these are stated throughout the figures.

Response: Thank you for pointing out this oversight. This information has been added throughout.

1. Figure 2: What does the gross anatomy of the IVD look like? Could the authors address this by showing an H&E of an adjacent section of the Fig. 2 A panels?

Response: Panel 2 shows H&E staining. Perhaps the reviewer is referring to the WT and Pax1 KO images in Figure 3? We have now added H&E staining of WT and Pax1 KO IVD as supplemental Figure 3E to clarify the IVD anatomy.

1. Page 9: "Cells within the IVD were negative for Pax1 staining ..." There seems to be specific PAX1 expression in many cells within the IVD, which is concerning if this is indeed a supposed null allele of Pax1. This data seems to support that the allele is not null.

Response: We have now added updated images for the COL11A1 and PAX1 staining to include negative controls in which we omitted primary antibodies. As can be seen, there is faint autofluorescence in the PAX1 negative control that appears to explain the “specific staining” referred to by the reviewer. These images confirm that the allele is truly a null.

1. There is currently a lack of evidence supporting the claim that "Col11a1 is positively regulated by Pax1 in mouse spine and tail". Therefore, it is necessary to conduct further research to determine the direct regulatory role of Pax1 on Col11a1.

Response: We agree with the reviewer and have clarified that Pax1 may have either a direct or indirect role in Col11a1 regulation.

1. There is no data linking loss of COL11A1 function and spine defects in the mouse model. Furthermore, due to the absence of P1335L point mutant mice, it cannot be confirmed whether P1335L can actually cause AIS, and the pathogenicity of this mutation cannot be directly verified. These limitations need to be clearly stated and discussed. A Col11a1 mouse mutant called chondroysplasia (cho), was shown to be perinatal lethal with severe endochondral defects (https://pubmed.ncbi.nlm.nih.gov/4100752/). This information may help contextualize this study.

Response: We partially agree with the reviewer. Spine defects are reported in the cho mouse (for example, please see reference 36 Hafez et al). We appreciate the suggestion to cite the original Seegmiller et al 1971 reference and have added it to the manuscript.

1. A recent article (PMID37462524) reported mutations in COL11A2 associated with AIS and functionally tested in zebrafish. That study should be cited and discussed as it is directly relevant for this manuscript.

Response: We agree with the reviewer that this study provides important information supporting loss of function I type XI collagen in spinal deformity. Language to this effect has been added to the manuscript and this study is now cited in the paper.

1. Please reconcile the following result on page 10 of the results: "Interestingly, the AISassociated gene Adgrg6 was amongst the most significantly dysregulated genes in the RNA-seq analysis (Figure 3c). By qRT-PCR analysis, expression of Col11a1, Adgrg6, and Sox6 were significantly reduced in female and male Pax1-/- mice compared to wild-type mice (Figure 3d-g)." In Figure 3f, the downregulation of Adgrg6 appears to be modest so how can it possibly be highlighted as one of the most significantly downregulated transcripts in the RNAseq data?

Response: By “significant” we were referring to the P-value significance in RNAseq analysis, not in absolute change in expression. This language was clearly confusing, and we have removed it from the manuscript.

1. It is incorrect to refer to the primary cell culture work as growth plate chondrocytes (GPCs), instead, these are primary costal chondrocyte cultures. These primary cultures have a mixture of chondrocytes at differing levels of differentiation, which may change differentiation status during the culturing on plastic. In sum, these cells are at best chondrocytes, and not specifically growth plate chondrocytes. This needs to be corrected in the abstract and throughout the manuscript. Moreover, on page 11 these cells are referred to as costal cartilage, which is confusing to the reader.

Response: Thank you for pointing out these inconsistencies. We have changed the manuscript to say “costal chondrocytes” throughout.

Minor points

• On 10 of the Results: "These data support a mechanistic link between Pax1 and Col11a1, and the AIS-associated genes Gpr126 and Sox6, in affected tissue of the developing tail." qRT-PCR validation of Sox6, although significant, appears to be very modestly downregulated in KO. Please soften this statement in the text.

Response: We have softened this statement.

• Have you got any information about how the immortalized (SV40) costal cartilage affected chondrogenic differentiation? The expression of SV40 seemed to stimulate Mmp13 expression. Do these cells still make cartilage nodules? Some feedback on this process and how it affects the nature of the culture what be appreciated.

Response: The “+ or –“ in Figure 5 refers to Ad5-cre. Each experiment was performed in SV40-immortalized costal chondrocytes. We have removed SV40 from the figure and have clarified the legend to say “qRT-PCR of human COL11A1 and endogenous mouse Mmp3 in SV40 immortalized mouse costal chondrocytes transduced with the lentiviral vector only (lanes 1,2), human WT COL11A1 (lane 3), or COL11A1P1335L. Otherwise we absolutely agree that understanding Mmp13 regulation during chondrocyte differentiation is important. We plan to study this using in vivo systems.

• Figure 1: is the average Odds ratio, can this be stated in the figure legend?

Response: We are not sure what is being asked here. The “combined odds ratio” is calculated as a weighted average of the log of the odds.

• A more consistent use of established nomenclature for mouse versus human genes and proteins is needed.

Human:GENE/PROTEIN Mouse: Gene/PROTEIN

Response: Thank you for pointing this out. The nomenclature has been corrected throughtout the manuscript.

• There is no Figure 5c, but a reference to results in the main text. Please reconcile. -There is no Figure 5-figure supplement 5a, but there is a reference to it in the main text. Please reconcile.

Response: Figure references have been corrected.

• Please indicate dilutions of all antibodies used when listed in the methods.

Response: Antibody dilutions have been added where missing.

• On page 25, there is a partial sentence missing information in the Histologic methods; "#S36964 Invitrogen, CA, USA)). All images were taken..."

Response: We apologize for the error. It has been removed.

• Table 1: please define all acronyms, including cohort names.

Response: We apologize for the oversight. The legend to the Table has been updated with definitions of all acronyms.

• Figure 2: Indicate that blue staining is DAPI in panel B. Clarify that "-ab" as an abbreviation is primary antibody negative.

Response: A color code for DAPI and COL11A! staining has been added and “-ab” is now defined.

• Page 4: ADGRG6 (also known as GPR126)...the authors set this up for ADGRG6 but then use GPR126 in the manuscript, which is confusing. For clarity, please use the gene name Adgrg6 consistently, rather than alternating with Gpr126.

Response: Thank you for pointing this out. GPR126 has now been changed to ADGRG6 thoughout the manuscript.

• REF 4: Richards, B.S., Sucato, D.J., Johnston C.E. Scoliosis, (Elsevier, 2020). Is this a book, can you provide more clarity in the Reference listing?

Response: Thank you for pointing this out. This reference has been corrected.

• While isolation was addressed, the methods for culturing Rat cartilage endplate and costal chondrocytes are poorly described and should be given more text.

Response: Details about the cartilage endplate and costal chondrocyte isolation and culture have been added to the Methods.

• Page 11: 1st paragraph, last sentence "These results suggest that Mmp3 expression"... this sentence needs attention. As written, I am not clear what the authors are trying to say.

Response: This sentence has been clarified and now reads “These results suggest that Mmp3 expression is negatively regulated by Col11a1 in mouse costal chondrocytes.”

• Page 13: line 4 from the bottom, "ECM-clearing"? This is confusing do you mean ECM degrading?

Response: Yes and thank you. We have changed to “ECM-degrading”.

• Please use version numbers for RefSeq IDs: e.g. NM_080629.3 instead of NM_080629

Response: This change has been made in the revised manuscript.

• It would be helpful for readers if the ethnicity of the discovery case cohort was clearly stated as European ancestry in the Results main text.

Response: “European ancestry” has been added at first description of the discovery cohort in the manuscript.

• Avoid using the term "mutation" and use "variant" instead.

Response: Thank you for pointing this out. “Variant” is now used throughout the manuscript.

• Define error bars for all bar charts throughout and include individual data points overlaid onto bars.

Response: Thank you. Error bars are now clarified in the Figure legends.

Author response:

The following is the authors’ response to the previous reviews.

Reviewer #2 (Public review):

Summary:

The authors reported that mutations were identified in the ZC3H11A gene in four adolescents from 1015 high myopia subjects in their myopia cohort. They further generated Zc3h11a knockout mice utilizing the CRISPR/Cas9 technology.

Comments on revisions:

Chong Chen and colleagues revised the manuscript; however, none of my suggestions from the initial review have been sufficiently addressed.

(1) I indicated that the pathogenicity and novelty of the mutation need to be determined according to established guidelines and databases. However, the conclusion was still drawn without sufficient justification.

Thank you for your valuable feedback on the assessment of mutation pathogenicity and novelty. We regret to inform you that complete familial genetic information required for segregation analysis is currently unavailable in this study. Despite our exhaustive efforts to contact the four mutation carriers and their relatives, we encountered the following uncontrollable limitations: Two patients could not be further traced due to invalid contact information, one patient had relocated to another region, making sample collection logistically unfeasible, the remaining patient explicitly declined family participation in genetic testing due to privacy concerns.

We fully acknowledge that the lack of pedigree data may affect the certainty of pathogenicity evaluation. To address this limitation, we systematically analyzed the four ZC3H11A missense mutations (c.412G>A p.V138I, c.128G>A p.G43E, c.461C>T p.P154L, and c.2239T>A p.S747T) based on ACMG guidelines and database evidence. The key findings are summarized below: All of the identified mutations exhibited very low frequencies or does not exist in the Genome Aggregation Database (gnomAD) and Clinvar, and using pathogenicity prediction software SIFT, PolyPhen2, and CADD, most of them display high pathogenicity levels. Among them, c.412G>A, c.128G>A and c.461C>T were located in or around a domain named zf-CCCH_3 (Figure 1A and B). Furthermore, all of the mutation sites were located in highly conserved amino acids across different species (Figure 1C). The four mutations induced higher structural flexibility and altered the negative charge at corresponding sites, potentially disrupting protein-RNA interactions (Figure 1D and E). Concurrently, overexpression of mutant constructs (ZC3H11A-V138I, ZC3H11A-G43E, ZC3H11A-P154L, and ZC3H11A-S747T) revealed significantly reduced nuclear IκBα mRNA levels compared to the wild-type, suggesting impaired NF-κB pathway regulation (Supplementary Figure 4). Zc3h11a knockout mice also exhibited a myopic phenotype, with alterations in the PI3K-AKT and NF-κB signaling pathways. Integrating this evidence, the mutations meet the following ACMG criteria: PM1 (domain-located mutations), PM2 (extremely low population frequency), PP3 (computational predictions supporting pathogenicity), PS3 (functional validation via experimental assays). Under the ACMG framework, these mutations are classified as "Likely Pathogenic".

Regarding the novelty of this mutation, comprehensive searches in ClinVar, dbSNP, and HGMD databases revealed no prior reports associating this variant with myopia. Similarly, a PubMed literature search identified no direct evidence linking this mutation to myopia. Based on this evidence, we classify this variant as a likely pathogenic and novel mutation.

On the other hand, we acknowledge that the absence of family segregation data may reduce the confidence in pathogenicity assessment. Nevertheless, functional experiments and converging multi-level evidence strongly support the reliability of our conclusion. Future studies will prioritize family-based validation to strengthen the evidence chain. We sincerely appreciate your attention to this matter and kindly request your understanding of the practical limitations inherent to this research.

(2) The phenotype of heterozygous mutant mice is too weak to support the gene's contribution to high myopia. The revised manuscript does not adequately address these discrepancies. Furthermore, no explanation was provided for why conditional gene deletion was not used to avoid embryonic lethality, nor was there any discussion on tissue- or cell-specific mechanistic investigations.

We sincerely appreciate your insightful comments regarding the relationship between murine phenotypes and human disease. We fully acknowledge your concerns about the phenotypic strength of Zc3h11a heterozygous mutant mice and their association with high myopia (HM) pathogenesis. Here we provide point-by-point responses to your valuable comments: Our study demonstrates that Zc3h11a heterozygous mutant mice exhibit myopic refractive phenotypes with upregulated myopia-associated factors (TGF-β1, MMP2, and IL6), although axial elongation did not reach statistical significance. Notably, at 4 and 6 weeks of age, Het mice did display longer axial lengths and vitreous chamber depths compared to WT mice. While these differences did not reach statistical significance at other time points, an increasing trend was still observed. Several technical considerations may explain these findings: The small murine eye size (where 1D refractive change corresponds to only 5-6μm axial length change). The theoretical resolution limit of 6μm for the SD-OCT device used in this study. These factors likely contributed to the marginal statistical significance observed in the subtle changes of vitreous chamber depth and axial length measurements. Additionally, existing research indicates that axial length measurements from frozen sections in age-matched mice tend to be longer than those obtained through in vivo measurements. This phenomenon may reflect species differences between humans and mice - while both show significant refractive power changes, the axial length differences are less pronounced in mice. These results align with previous reports of phenotypic differences between mouse models and human myopia.

To address these issues comprehensively, we have added a dedicated discussion section in the revised manuscript specifically examining these axial length measurement considerations, following your valuable suggestion.

Additionally, we regret to inform you that the currently available floxed ZC3H11A mouse strain requires a minimum of 12-18 months for custom construction, which exceeds our research timeline due to current resource limitations in our team. To address this gap, we have supplemented the discussion section with additional content regarding tissue- and cell-specific mechanisms. Based on your constructive suggestions, we will prioritize the following in our subsequent work: Collaborate with transgenic animal centers to generate Zc3h11a conditional knockout mice. Evaluate the impact of specific knockouts on myopia progression using form-deprivation (FDM) models. While we recognize the limitations of our current study, we believe that by integrating clinical cohort data, phenotypic evidence, and functional experiments, this research provides valuable directional evidence for ZC3H11A's potential role in myopia pathogenesis. Your comments will significantly contribute to improving our future research design, and we sincerely hope you can recognize the exploratory significance of our current findings.

(3) The title, abstract, and main text continue to misrepresent the role of the inflammatory intracellular PI3K-AKT and NF-κB signaling cascade in inducing high myopia. No specific cell types have been identified as contributors to the phenotype. The mice did not develop high myopia, and no relationship between intracellular signaling and myopia progression has been demonstrated in this study.

Thank you for your valuable comments regarding the interpretation of signaling pathways in our study. We fully acknowledge your rigorous concerns about the role of PI3K-AKT and NF-κB signaling cascades in high myopia and recognize that we did not identify specific cell types contributing to the observed phenotype. In response to your feedback, we have removed the hypothetical statement linking genetic changes within inflammatory cells to the development of myopia. The current interpretation is strictly based on experimental evidence of pathway relevance and is supported by the theoretical basis presented in the reference, specifically that loss of Zc3h11a leads to activation of the PI3K-AKT and NF-κB pathways in retinal cells, contributing to the myopic phenotype.

Author response image 1.

Model of the association between inflammation and myopia progression. Activated mAChR3 (M3R) activates phosphoinositide 3-kinase (PI3K)–AKT and mitogen-associated protein kinase (MAPK) signaling pathways, in turn activating NF-κB and AP1 (i.e., the Jun.-Fos heterodimer) and stimulating the expression of the target genes NF-κB, MMP2, TGFβ, IL- 1β and -6, and TNF-α. MMP2 and TGF-β promote tissue remodeling and TNF-α may act in a paracrine feedback loop in the retina or sclera to activate NF-κB during myopia progression.

To address the limitations raised, we will prioritize the following in future studies: Cell-type-specific knockout models to identify key cellular contributors. Mechanistic investigations to establish causal relationships between signaling pathways and myopia progression. We sincerely appreciate your rigorous review, which has significantly improved the scientific accuracy and clarity of our manuscript. We believe the revised version better reflects both the novelty and limitations of our findings. We kindly request your recognition of the study’s contributions while acknowledging its current constraints.

Reviewer #3 (Public review):

Chen et al have identified a new candidate gene for high myopia, ZC3H11A, and using a knock-out mouse model, have attempted to validate it as a myopia gene and explain a potential mechanism. They identified 4 heterozygous missense variants in highly myopic teenagers. These variants are in conserved regions of the protein, and predicted to be damaging, but the only evidence the authors provide that these specific variants affect protein function is a supplement figure showing decreased levels of IκBα after transfection with overexpression plasmids (not specified what type of cells were transfected). This does not prove that these mutations cause loss of function, in fact it implies they have a gain-of-function mechanism. They then created a knock-out mouse. Heterozygotes show myopia at all ages examined but increased axial length only at very early ages. Unfortunately, the authors do not address this point or examine corneal structure in these animals. They show that the mice have decreased B-wave amplitude on electroretinogram (a sign of retinal dysfunction associated with bipolar cells), and decreased expression of a bipolar cell marker, PKCα. On electron microscopy, there are morphologic differences in the outer nuclear layer (where bipolar, amacrine, and horizontal cell bodies reside). Transcriptome analysis identified over 700 differentially expressed genes. The authors chose to focus on the PI3K-AKT and NF-κB signaling pathways and show changes in expression of genes and proteins in those pathways, including PI3K, AKT, IκBα, NF-κB, TGF-β1, MMP-2 and IL-6, although there is very high variability between animals. They propose that myopia may develop in these animals either as a result of visual abnormality (decreased bipolar cell function in the retina) or by alteration of NF-κB signaling. These data provide an interesting new candidate variant for development of high myopia, and provide additional data that MMP2 and IL6 have a role in myopia development. For this revision, none of my previous suggestions have been addressed.

Reviewer #3 (Recommendations for the authors):

None of these suggestions were addressed in the revision:

Major issues:

(1) Figure 2: refraction is more myopic but axial length is not longer - why is this not discussed and explored? The text claims the axial length is longer, but that is not supported by the figure. If this is a measurement issue, that needs to be discussed in the text.

We sincerely appreciate your valuable comments regarding the relationship between refractive status and axial length in our study. In response to your concerns, we have conducted an in-depth analysis and would like to address the issues as follows:

Our data demonstrate significant differences in refractive error between heterozygous (Het) and wild-type (WT) mice during the 4-10 weeks. Notably, at 4 and 6 weeks of age, Het mice did exhibit longer axial lengths and greater vitreous chamber depth compared to WT mice, although these differences did not reach statistical significance at other time points while still showing an increasing trend. Additional measurements of corneal curvature revealed no significant differences between groups. Considering the small size of mouse eyes (where a 1D refractive change corresponds to only 5-6μm axial length change) and the theoretical resolution limit of 6μm for the SD-OCT device used in this study, these technical factors may account for the marginal statistical significance of the observed small changes in vitreous chamber depth and axial length measurements. Furthermore, existing studies have shown that axial length measurements from frozen sections tend to be longer than those obtained from in vivo measurements in age-matched mice. These considerations provide plausible explanations for the apparent discrepancy between refractive changes and axial length parameters. Following your suggestion, we have added a dedicated discussion section addressing these axial length measurement issues in the revised manuscript. We fully understand your concerns regarding data consistency, and your comments have prompted us to conduct more comprehensive and thorough analysis of our results. We believe the revised manuscript now more accurately reflects our findings while providing important technical references for future studies.

(2)  Slipped into the methods is a statement that mice with small eyes or ocular lesions were excluded. How many mice were excluded? Are the authors ignoring another phenotype of these mice?

We appreciate your attention to the exclusion criteria and their implications. Below we provide a detailed clarification: A total of 7 mice (4 Het-KO and 3 WT) with small eyes or ocular lesions were excluded from the observation cohort. These anomalies were consistent with the baseline incidence of spontaneous malformations observed in historical colony data of wild-type C57BL/6J mice (approximately 11%), and were not attributed to the Zc3h11a heterozygous knockout. We have added the above content in the methods section. Your insightful comment has significantly strengthened our reporting rigor. We hope this clarification alleviates your concerns regarding potential selection bias or overlooked phenotypes.

Minor/Word choice issues:

All the figure legends need to be improved so that each figure can be interpreted without having to refer to the text.

Thank you for your valuable comments. We have made modifications to the legend of each graphic, as detailed in the main text.

Abstract: line 24: use refraction, not "vision"

Thank you for your valuable comments. The “Vision” has been changed to “refraction”.

Line 28: re-word "density of bipolar cell-labeled proteins" Do the authors mean density of bipolar cells? Or certain proteins were less abundant in bipolar cells?

Thank you for your rigorous review of this terminology. We acknowledge the need to clarify the precise meaning of the phrase "density of bipolar cell-labeled proteins." In the original text, this term specifically refers to the expression abundance of the bipolar cell-specific marker protein PKCα, which was identified using immunofluorescence labeling techniques. Specifically: We utilized PKCα (a bipolar cell marker) to label bipolar cell populations. The "density" was quantified by measuring the fluorescence signal intensity per unit area in confocal microscopy images, rather than direct cell counting. This metric reflects changes in the expression of the specific marker protein (PKCα) within bipolar cells, which indirectly correlates with alterations in bipolar cell populations. To address ambiguity, we have revised the terminology throughout the manuscript to "bipolar cell-labelled protein PKCα immunofluorescence abundance".

Additionally, since fluorescence intensity quantification is inherently semi-quantitative, we have included Western blot results for PKCα in the revised manuscript (Figure 3I, J) to validate the expression changes observed via immunofluorescence. We sincerely appreciate your feedback, which has significantly improved the precision of our manuscript.

Line 45: axial length, not ocular axis

Thank you for your valuable comments. The “ocular axis” has been changed to “axial length”.

Lines73-75: confusing

Thank you for your valuable comments. The relevant content has been modified to “Multiple zinc finger protein genes (e.g., ZNF644, ZC3H11B, ZFP161, ZENK) are associated with myopia or HM. Of these, ZC3H11B (a human homolog of ZC3H11A) and five GWAS loci (Schippert et al., 2007; Shi et al., 2011; Szczerkowska et al., 2019; Tang et al., 2020; Wang et al., 2004) correlate with AL elongation or HM severity. Proteomic studies further suggest ZC3H11A involvement in the TREX complex, implicating RNA export mechanisms in myopia pathogenesis”

Line 138: what is dark 3.0 and dark 10.0

Thank you for your valuable comments. The relevant content has been modified to “Upon dark adaptation, b-wave amplitudes in seven-week-old Het-KO mice were significantly lower at dark 3.0 (0.48 log cd·s/m²) and dark 10.0 (0.98 log cd·s/m²) compared to WT mice.” A detailed description has been added to the main text methods.

Line 171-175: the GO terms of "biological processes" and "molecular functions" are so broad as to be meaningless.

Thank you for your valuable comments. The relevant content has been modified to “GO enrichment analysis revealed significant enrichment of differentially expressed genes in the following functions: Zinc ion transmembrane transport (GO:0071577) within metal ion homeostasis, associated with retinal photoreceptor maintenance (Ugarte and Osborne, 2001), RNA biosynthesis and metabolism (GO:0006366) in transcriptional regulation, potentially influencing ocular development, negative regulation of NF-κB signaling (GO:0043124) in inflammatory modulation, a pathway involved in scleral remodelling (Xiao et al., 2025), calcium ion binding (GO:0005509), critical for phototransduction (Krizaj and Copenhagen, 2002), zinc ion transmembrane transporter activity (GO:0005385), participating in retinal zinc homeostasis (Figure 5C and D).”

Line 257-259: which results indicated loss of Zc3h11a inhibited translocation of IκBα from nucleus to cytoplasm? Results of this study, or the previously referenced study?

We sincerely appreciate your critical inquiry regarding the mechanistic relationship between Zc3h11a deficiency and IκBα translocation. We are grateful for this opportunity to clarify this important point. The findings regarding Zc3h11a-mediated regulation of IκBα mRNA nuclear export and its impact on NF-κB signaling originate from the study by Darweesh et al. The key experimental evidence demonstrates that: The depletion of Zc3h11a leads to nuclear retention of IκBα mRNA, resulting in failure to maintain normal levels of cytoplasmic IκBα mRNA and protein. This defect in IκBα mRNA export disrupts the essential inhibitory feedback loop on NF-κB activity, causing hyperactivation of this pathway. This manifests as upregulation of numerous innate immune-related mRNAs, including IL-6 and a large group of interferon-stimulated genes.While our study references this mechanism to explain the observed NF-κB dysregulation in Zc3h11a Het-KO mice, the specific nuclear export mechanism was indeed elucidated by Darweesh et al. The reference has been inserted into the corresponding position in the main text. Importantly, our research extends these previous molecular insights into the phenotypic context of myopia.

We sincerely regret any ambiguity in the original text and deeply appreciate your rigorous approach in ensuring proper attribution of these fundamental findings. Your comment has significantly improved the clarity and accuracy of our manuscript.

Figure 6 shows decrease of both mRNA and protein expression, but nothing about translocation.

Thank you for your valuable comments. The research results of Darweesh et al. showed that Zc3h11a protein plays a role in regulation of NF-κB signal transduction. Depletion of Zc3h11a resulted in enhanced NF-κB mediated signaling, with upregulation of numerous innate immune related mRNAs, including IL-6 and a large group of interferon-stimulated genes. IL-6 upregulation in the absence of the Zc3h11a protein correlated with an increased NF-κB transcription factor binding to the IL-6 promoter and decreased IL-6 mRNA decay. The enhanced NF-κB signaling pathway in Zc3h11a deficient cells correlated with a defect in IκBα inhibitory mRNA and protein accumulation. Upon Zc3h11a depletion The IκBα mRNA was retained in the cell nucleus resulting in failure to maintain normal levels of the cytoplasmic IκBα mRNA and protein that is essential for its inhibitory feedback loop on NF-κB activity. These findings demonstrate that ZC3H11A can regulate the NF-κB pathway by controlling the translocation of IκBα mRNA, a mechanism that was indeed elucidated by Darweesh et al. We sincerely apologize for any lack of clarity in our original description and have now inserted the appropriate reference in the relevant section of the main text.

We deeply appreciate your valuable comments in identifying this ambiguity in our manuscript, which have significantly improved the accuracy and clarity of our work.

Line 283: what do you mean "may confer embryonic lethality"? Were they embryonic lethal or not?

We sincerely appreciate your critical request for clarification. Our experimental data from 15 pregnancies of Zc3h11a Het-KO mice intercrosses (n = 15 litters) conclusively confirmed the absence of homozygous knockout (Homo-KO) pups at birth. These findings align with the embryonic lethality of Zc3h11a homozygous deletion as reported by Younis et al. We fully acknowledge the ambiguity in our original phrasing and have revised the text to:“Second, Zc3h11a homozygous KO (Homo-KO) mice were not obtained in our study because homozygous deletion of exons confer embryonic lethality.”Your vigilance in ensuring terminological precision has greatly strengthened the rigor of our manuscript. We hope this clarification fully resolves your concerns.

Line 338: What is meant that Het-KO mice were constructed at 4 weeks of age? Do these mice not have a germline mutation?

Thank you for your valuable comments. We have revised the following content: “The germline heterozygous Zc3h11a knockout (Het-KO) mice were generated by CRISPR/Cas9-mediated gene editing at the embryonic stage on a C57BL/6J background, provided by GemPharmatech Co., Ltd (Nanjing, China). Phenotypic analyses were initiated when the mice reached four weeks of age.”

Line 346-347: how many mice were excluded due to having small eyes or ocular lesions? The methods section should state how refraction and ocular biometrics were measured.

Thank you for your valuable comments. We have added or revised the following content: “To exclude potential confounding effects of spontaneous ocular developmental abnormalities, a total of 7 mice (4 Het-KO and 3 WT) with small eyes or ocular lesions were excluded from the observation cohort. These anomalies were consistent with the baseline incidence of spontaneous malformations observed in historical colony data of wild-type C57BL/6J mice (approximately 11%), and were not attributed to the Zc3h11a heterozygous knockout.

The methods for measuring refraction and ocular biometrics are as follows and have been added to the original method. Refractive measurements were performed by a researcher blinded to the genotypes. Briefly, in a darkroom, mice were gently restrained by tail-holding on a platform facing an eccentric infrared retinoscope (EIR) (Schaeffel et al., 2004; Zhou et al., 2008a). The operator swiftly aligned the mouse position to obtain crisp Purkinje images centered on the pupil using detection software (Schaeffel et al., 2004), enabling axial measurements of refractive state and pupil size. Three repeated measurements per eye were averaged for analysis. The anterior chamber (AC) depth, lens thickness, vitreous chamber (VC) depth, and axial length (AL) of the eye were measured by real-time optical coherence tomography (a custom built OCT) (Zhou et al., 2008b). In simple terms, after anesthesia, each mouse was placed in a cylindrical holder on a positioning stage in front of the optical scanning probe. A video monitoring system was used to observe the eyes during the process. Additionally, by detecting the specular reflection on the corneal apex and the posterior lens apex in the two dimensional OCT image, the optical axis of the mouse eye was aligned with the axis of the probe. Eye dimensions were determined by moving the focal plane with a stepper motor and recording the distance between the interfaces of the eyes. Then, using the designed MATLAB software and appropriate refractive indices, the recorded optical path length was converted into geometric path length. Each eye was scanned three times, and the average value was taken.”

Line 428: what age retinas

Thank you for your meticulous attention to the experimental design details. Regarding the age of retinal samples, we have clarified the following in the revised manuscript:" Retinas were harvested from four-week-old mice for RNA sequencing." This revision enhances the transparency and reproducibility of our methodology. We deeply appreciate your rigorous review.

Figure 3 D-F: these images are too small to adequately assess, please show at higher magnification. Are there fewer bipolar cells, or just decreased expression of PKC? From these images, expression of ZC3H11A does not appear decreased, but the retina appears thinner. Is that true, or are these poorly matched sections?

Thank you for your professional insights regarding image quality and data interpretation. Your rigorous review has significantly enhanced the scientific rigor of our study. We hereby address your concerns point by point: The images in Figures 3D-F were acquired using a Zeiss LSM880 confocal microscope with a 10x eyepiece and 20x objective lens, a standard magnification for retinal section imaging that balances cellular resolution with full-thickness structural preservation. We quantified PKCα immunofluorescence intensity (a bipolar cell-specific marker) to assess changes in bipolar cell populations, rather than direct cell counting. This metric reflects PKCα expression abundance as a proxy for bipolar cell alterations (Figure 3H). To clarify terminology, we have revised the text to "bipolar cell-labelled protein PKCα immunofluorescence abundance" and detailed the methodology in the revised Methods section. Recognizing the semi-quantitative nature of fluorescence intensity analysis, we supplemented these data with Western blot results confirming reduced PKCα protein levels (Figure 3I). Zc3h11a expression was validated both by immunofluorescence intensity (Figure 3G) and Western blot (Figures 6F, H) quantification, confirming reduced expression in Zc3h11a Het-KO retinas. The apparent "retinal thinning" observed in histology sections stems from technical artifacts during tissue processing (fixation, dehydration, sectioning), not biological differences. HE staining, which better preserves sample morphology, showed no structural or thickness differences between Zc3h11a Het-KO mice and wild-type mice (Supplementary Figure 2).

Your expert feedback has driven us to establish a more robust validation framework. We believe the revised data now more accurately reflect the biological reality and sincerely hope these improvements meet your approval.

Figure 3G-J: Relative fluorescence intensity of immunohistochemistry is not a valid measure of protein expression.

We sincerely appreciate your thorough review and valuable comments regarding the immunofluorescence quantification method in Figures 3G-J. In response to your concern that "relative fluorescence intensity is not an effective quantitative measure of protein expression," we have implemented the following improvements to our analysis and validation: To ensure result reliability, all immunofluorescence experiments followed strict protocols: experimental and control samples were fixed, stained, and imaged in the same batch to eliminate inter-batch variability. Imaging was performed using a Zeiss LSM 880 confocal microscope with identical parameters, and the relative fluorescence intensity of specific signals per unit area was measured and statistically analyzed using ZEN software. We fully acknowledge the semi-quantitative nature of relative fluorescence intensity measurements. Therefore, we validated key differentially expressed proteins using Western blot analysis: The Western blot results for Zc3h11a (Figures 6F, H) were completely consistent with the relative fluorescence intensity trends (Figure 3G). Additionally, the newly included Western blot data for PKCα (Figure 3 I) further confirmed the reliability of our relative fluorescence intensity quantification. Your expert advice has significantly enhanced the rigor of our study. Should any additional data or clarification be required, we would be pleased to provide further support.

Figure 4: what are the arrows pointing at? This should be in the Figure legend. What is MB? Why are there no scale bars? What is difference between E and F, not clear from legend.

We sincerely appreciate your thorough review of Figure 4 and your valuable suggestions. In response to your concerns, we have carefully examined and improved the relevant content with the following modifications and clarifications: We sincerely apologize for not clearly indicating the arrow annotations in the original figure legend. In the revised version, we have provided detailed explanations for the arrow indicators: black arrows indicate perinuclear space dilation, blue arrows indicate cytoplasmic edema, and red arrows indicate disorganized and loosely arranged membrane discs. The updated legend has been clearly marked below Figure 4 in the main text. MB represents membrane discs, which are critical subcellular structures in the outer segments of retinal photoreceptor cells (rods and cones). They are responsible for light signal capture and transduction (containing visual pigments such as rhodopsin). The structural integrity of MB is essential for normal visual function. The scale bars in the original figures were located in the lower right corner of each subpanel, with specific parameters as follows: Figures 4A and B: magnification ×1000, scale bar 10 μm; Figures 4C and D: magnification ×700, scale bar 20 μm; Figures 4E and G: magnification ×2000, scale bar 5 μm; Figures 4F and H: magnification ×7000, scale bar 2 μm. Both Figures 4E and 4F show electron microscopy images of membrane discs (MB) in wild-type mouse photoreceptor cells. The only difference lies in the magnification: Figure 4E (×2000) demonstrates the overall arrangement pattern of membrane discs, while Figure 4F (×7000) focuses on ultrastructural details of the membrane discs (such as structural integrity). We have thoroughly checked the consistency between the figures and text, and have supplemented detailed legend descriptions in the main text. Once again, we sincerely appreciate your rigorous review, which has significantly enhanced the scientific rigor and readability of our study. Should you have any further suggestions, we would be happy to incorporate them.

Figure 5A: Why such a large y-axis? Figure legend does not match figure

We sincerely appreciate your careful review of Figure 5A and your valuable suggestions regarding the figure details. In response to your concerns, we have thoroughly examined and improved the relevant content as follows: The Y-axis of the volcano plot represents -log₁₀(p-value), where the magnitude of the values reflects statistical significance. Our RNA-seq data underwent rigorous multiple testing correction, and the adjusted p-values for some genes were extremely small, resulting in large values after -log₁₀ transformation. We have re-examined the data distribution and confirmed that the expanded Y-axis range is solely due to a small number of highly significant genes (as shown in the figure, the majority of genes remain clustered in the lower half of the Y-axis). This result accurately reflects the true data characteristics.

We sincerely apologize for the inadvertent error in the original labeling of "Up/Down" in the figure legend. This has now been corrected, and we strictly adhere to the following threshold criteria: Significantly upregulated (Up): adjusted p-value < 0.05 and log₂(FC) ≥ 1. Significantly downregulated (Down): adjusted p-value < 0.05 and log₂(FC) ≤ -1. To ensure the reliability of our conclusions, we have rechecked the raw data, statistical analysis, and visualization process. We confirmed that all significant genes strictly meet the above threshold criteria and that the visualization accurately reflects the true results. The revised figure has been updated in the manuscript as Figure 5A. We deeply appreciate your valuable feedback, which has helped us correct the errors in the figure and improve its accuracy and readability.

Figure 6F: Based on the western blot, only Zc3h11a appears different.

Thank you for your careful evaluation of the Western blot data in Figure 6F. We fully understand your concerns regarding the visual differences in PI3K and p-AKT/AKT bands and appreciate the opportunity to clarify the quantitative methodology and biological significance of these findings. Below we provide a detailed explanation of the experimental design and data analysis.

First, the data for each group were derived from retinal samples of three independent mice, with all experiments performed in parallel to control for technical variability. Image analysis was conducted using ImageJ software with standardized settings for grayscale quantification. Zc3h11a and PI3K levels were normalized to GAPDH as an internal reference, while p-AKT levels were calculated as a ratio to total AKT. The results showed that Zc3h11a protein levels were significantly reduced (p < 0.01, Figures 6F and H), consistent with the expected effects of heterozygous knockout, with good agreement between visual and statistical results. For PI3K and p-AKT/AKT, the bands appeared visually similar due to: The nonlinear nature of Western blot chemiluminescence signals in the saturation range, which compresses subtle quantitative differences in the images; the fact that p-AKT represents only 5-15% of the total AKT pool, making small proportional changes difficult to discern visually. However, it is important to note that both PI3K and p-AKT/AKT showed statistically significant differences between groups (p < 0.001 and p < 0.01, respectively; Figures 6G and I). Furthermore, signal transduction pathways exhibit cascade amplification effects - in the PI3K-AKT pathway, even small changes in upstream proteins can produce significant downstream effects (e.g., NF-κB activation) through kinase cascades (Figure 6J). Additionally, our RNA-Seq results revealed activation of the PI3K-AKT signaling pathway in Zc3h11a Het-KO mice (Figure 5D), and the qRT-PCR results were consistent with the western blot results (Figure 6A-C). Your expert comments have prompted us to present these data differences with greater biological rigor. Although the visual differences are subtle, based on statistical significance, pathway characteristics, and RNA sequencing, and qRT-PCR data, we believe these changes have biological relevance. We sincerely appreciate your commitment to data rigor and respectfully request your recognition of both the experimental results and the scientific logic of this study.

Figure 8: What is the role of ZC3H11A in this figure? Are the authors proposing that ZC3H11A regulates the translation of IκBα? They have not shown any evidence of that.

Thank you for your insightful exploration of the role of ZC3H11A in Figure 8. We appreciate your critical review and hope to elucidate the mechanistic framework behind our findings. In Figure 8, Zc3h11a is depicted as a regulator of IκBα mRNA nucleocytoplasmic transport, a mechanism originally elucidated by Darweesh et al. Their studies demonstrated that Zc3h11a binds to IκBα mRNA and promotes its nuclear export. Loss of Zc3h11a results in nuclear retention of IκBα mRNA, leading to reduced cytoplasmic IκBα protein levels and subsequent hyperactivation of the NF-κB pathway. While the specific nuclear export mechanism has been elucidated by Darweesh et al., our study demonstrates that Zc3h11a haploinsufficiency results in decreased IκBα mRNA and protein levels in the retina (Figure 7), linking Zc3h11a haploinsufficiency to NF-κB pathway dysregulation in myopia and highlighting that these molecular insights can be extended to a new pathological context (myopia). Your critical comments have enhanced the clarity of our mechanistic concepts and we hope that these descriptions will demonstrate the importance of ZC3H11A as a new candidate gene for myopia.

Author response:

The following is the authors’ response to the previous reviews.

Public Reviews:

Reviewer #1 (Public review):

Summary:

Shen et al. conducted three experiments to study the cortical tracking of the natural rhythms involved in biological motion (BM), and whether these involve audiovisual integration (AVI). They presented participants with visual (dot) motion and/or the sound of a walking person. They found that EEG activity tracks the step rhythm, as well as the gait (2-step cycle) rhythm. The gait rhythm specifically is tracked superadditively (power for A+V condition is higher than the sum of the A-only and V-only condition, Experiments 1a/b), which is independent of the specific step frequency (Experiment 1b). Furthermore, audiovisual integration during tracking of gait was specific to BM, as it was absent (that is, the audiovisual congruency effect) when the walking dot motion was vertically inverted (Experiment 2). Finally, the study shows that an individual's autistic traits are negatively correlated with the BM-AVI congruency effect.

Strengths:

The three experiments are well designed and the various conditions are well controlled. The rationale of the study is clear, and the manuscript is pleasant to read. The analysis choices are easy to follow, and mostly appropriate.

Weaknesses:

There is a concern of double-dipping in one of the tests (Experiment 2, Figure 3: interaction of Upright/Inverted X Congruent/Incongruent). I raised this concern on the original submission, and it has not been resolved properly. The follow-up statistical test (after channel selection using the interaction contrast permutation test) still is geared towards that same contrast, even though the latter is now being tested differently. (Perhaps not explicitly testing the interaction, but in essence still testing the same.) A very simple solution would be to remove the post-hoc statistical tests and simply acknowledge that you're comparing simple means, while the statistical assessment was already taken care of using the permutation test. (In other words: the data appear compelling because of the cluster test, but NOT because of the subsequent t-tests.)

We are sorry that we did not explain this issue clearly before, which might have caused some misunderstanding. When performing the cluster-based permutation test, we only tested whether the audiovisual congruency effect (congruent vs. incongruent) between the upright and inverted conditions was significantly different [i.e., (UprCon – UprInc) vs. (InvCon – InvInc)], without conducting extra statistical analyses on whether the congruency effect was significant in each orientation condition. Such an analysis yielded a cluster with a significant interaction between audiovisual integration and BM orientation for the cortical tracking effect at 1Hz (but not at 2Hz). However, this does not provide valid information about whether the audiovisual congruency effect at this cluster is significant in each orientation condition, given that a significant interaction effect may result from various patterns of data across conditions: such as significant congruency effects in both orientation conditions (Author response image 1a), a significant congruency effect in the upright condition and a non-significant effect in the inverted condition (Author response image 1b), or even non-significant yet opposite effects in the two conditions (Author response image 1c). Here, our results conform to the second pattern, indicating that cortical tracking of the high-order gait cycles involves a domain-specific process exclusively engaged in the AVI of BM. In a similar vein, the non-significant interaction found at 2Hz does not necessarily indicate that the congruency effect is non-significant in each orientation condition (Author response image 1f&e). Indeed, the congruency effect was significant in both the upright and inverted conditions at 2Hz in our study despite the non-significant interaction, suggesting that neural tracking of the lower-order step cycles is associated with a domain-general AVI process mostly driven by temporal correspondence in physical stimuli.

Therefore, we need to perform subsequent t-tests to examine the significance of the simple effects in the two orientation conditions, which do not duplicate the clusterbased permutation test (for interaction only) and cause no double-dipping. Results from interaction and simple effects, put together, provide solid evidence that the cortical tracking of higher-order and lower-order rhythms involves BM-specific and domaingeneral audiovisual processing, respectively.

To avoid ambiguity, we have removed the sentence “We calculated the audiovisual congruency effect for the upright and the inverted conditions” (line 194, which referred to the calculation of the indices rather than any statistical tests) from the manuscript. We have also clarified the meanings of the findings based on the interaction and simple effects together at the two temporal scales, respectively (Lines 205-207; Lines 213-215).

Author response image 1.

Examples of different patterns of data yielding a significant or nonsignificant interaction effect.

Reviewer #2 (Public review):

Summary:

The authors evaluate spectral changes in electroencephalography (EEG) data as a function of the congruency of audio and visual information associated with biological motion (BM) or non-biological motion. The results show supra-additive power gains in the neural response to gait dynamics, with trials in which audio and visual information was presented simultaneously producing higher average amplitude than the combined average power for auditory and visual conditions alone. Further analyses suggest that such supra-additivity is specific to BM and emerges from temporoparietal areas. The authors also find that the BM-specific supra-additivity is negatively correlated with autism traits.

Strengths:

The manuscript is well-written, with a concise and clear writing style. The visual presentation is largely clear. The study involves multiple experiments with different participant groups. Each experiment involves specific considered changes to the experimental paradigm that both replicate the previous experiment's finding yet extend it in a relevant manner.

Weaknesses:

In the revised version of the paper, the manuscript better relays the results and anticipates analyses, and this version adequately resolves some concerns I had about analysis details. Still, it is my view that the findings of the study are basic neural correlate results that do not provide insights into neural mechanisms or the causal relevance of neural effects towards behavior and cognition. The presence of an inversion effect suggests that the supra-additivity is related to cognition, but that leaves open whether any detected neural pattern is actually consequential for multi-sensory integration (i.e., correlation is not causation). In other words, the fact that frequency-specific neural responses to the [audio & visual] condition are stronger than those to [audio] and [visual] combined does not mean this has implications for behavioral performance. While the correlation to autism traits could suggest some relation to behavior and is interesting in its own right, this correlation is a highly indirect way of assessing behavioral relevance. It would be helpful to test the relevance of supra-additive cortical tracking on a behavioral task directly related to the processing of biological motion to justify the claim that inputs are being integrated in the service of behavior. Under either framework, cortical tracking or entrainment, the causal relevance of neural findings toward cognition is lacking.

Overall, I believe this study finds neural correlates of biological motion, and it is possible that such neural correlates relate to behaviorally relevant neural mechanisms, but based on the current task and associated analyses this has not been shown.

Thank you for providing these thoughtful comments regarding the theoretical implications of our neural findings. Previous behavioral evidence highlights the specificity of the audiovisual integration (AVI) of biological motion (BM) and reveals the impairment of such ability in individuals with autism spectrum disorder. However, the neural implementation underlying the AVI of BM, its specificity, and its association with autistic traits remain largely unknown. The current study aimed to address these issues.

It is noteworthy that the operation of multisensory integration does not always depend on specific tasks, as our brains tend to integrate signals from different sensory modalities even when there is no explicit task. Hence, many studies have investigated multisensory integration at the neural level without examining its correlation with behavioral performance. For example, the widely known super-additivity mode for multisensory integration proposed by Perrault and colleagues was based on single-cell recording findings without behavioral tasks (Perrault et al., 2003, 2005). As we mentioned in the manuscript, the super-additive and sub-additive modes indicate non-linear interaction processing, either with potentiated neural activation to facilitate the perception or detection of near-threshold signals (super-additive) or a deactivation mechanism to minimize the processing of redundant information cross-modally (subadditive) (Laurienti et al., 2005; Metzger et al., 2020; Stanford et al., 2005; Wright et al., 2003). Meanwhile, the additive integration mode represents a linear combination between two modalities. Distinguishing among these integration modes helps elucidate the neural mechanism underlying AVI in specific contexts, even though sometimes, the neural-level AVI effects do not directly correspond to a significant behavioral-level AVI effect (Ahmed et al., 2023; Metzger et al., 2020). In the current study, we unveiled the dissociation of multisensory integration modes between neural responses at two temporal scales (Exps. 1a & 1b), which may involve the cooperation of a domain-specific and a domain-general AVI processes (Exp. 2). While these findings were not expected to be captured by a single behavioral index, they revealed the multifaceted mechanism whereby hierarchical cortical activity supports audiovisual BM integration. They also advance our understanding of the emerging view that multi-timescale neural dynamics coordinate multisensory integration (Senkowski & Engel, 2024), especially from the perspective of natural stimuli processing.

Meanwhile, our finding that the cortical tracking of higher-order rhythmic structure in audiovisual BM specifically correlated with individual autistic traits extends previous behavioral evidence that ASD children exhibited reduced orienting to audiovisual synchrony in BM (Falck-Ytter et al., 2018), offering new evidence that individual differences in audiovisual BM processing are present at the neural level and associated with autistic traits. This finding opens the possibility of utilizing the cortical tracking of BM as a potential neural maker to assist the diagnosis of autism spectrum disorder (see more details in our Discussion Lines 334-346).

However, despite the main objective of the current study focusing on the neural processing of BM, we agree with the reviewer that it would be helpful to test the relevance of supra-additive cortical tracking on a behavioral task directly related to BM perception, for further justifying that inputs are being integrated in the service of behavior. In the current study, we adopted a color-change detection task entirely unrelated to audiovisual correspondence but only for maintaining participants’ attention. The advantage of this design is that it allows us to investigate whether and how the human brain integrates audiovisual BM information under task-irrelevant settings, as people in daily life can integrate such information even without a relevant task. However, this advantage is accompanied by a limitation: the task does not facilitate the direct examination of the correlation between neural responses and behavioral performance, since the task performance was generally high (mean accuracy >98% in all experiments). Future research could investigate this issue by introducing behavioral tasks more relevant to BM perception (e.g., Shen et al., 2023). They could also apply advanced neuromodulation techniques to elucidate the causal relevance of the cortical tracking effect to behavior (e.g., Ko sem et al., 2018, 2020).

We have discussed the abovementioned points as a separate paragraph in the revised manuscript (Lines 322-333). In addition, since the scope of the current study does not involve a causal correlation with behavioral performance, we have removed or modified the descriptions related to "functional relevance" in the manuscript (Abstract; Introduction, lines 101-103; Results, lines 239; Discussion, line 336; Supplementary Information, line 794、803). Moreover, we have strengthened the descriptions of the theoretical implications of the current findings in the abstract.

We hope these changes adequately address your concern.

References

Ahmed, F., Nidiffer, A. R., O’Sullivan, A. E., Zuk, N. J., & Lalor, E. C. (2023). The integration of continuous audio and visual speech in a cocktail-party environment depends on attention. Neuroimage, 274, 120143. https://doi.org/10.1016/j.neuroimage.2023.120143

Falck-Ytter, T., Nystro m, P., Gredeba ck, G., Gliga, T., Bo lte, S., & the EASE team. (2018). Reduced orienting to audiovisual synchrony in infancy predicts autism diagnosis at 3 years of age. Journal of Child Psychology and Psychiatry, 59(8), 872–880. https://doi.org/10.1111/jcpp.12863

Ko sem, A., Bosker, H., Jensen, O., Hagoort, P., & Riecke, L. (2020). Biasing the Perception of Spoken Words with Transcranial Alternating Current Stimulation. Journal of Cognitive Neuroscience, 32, 1–10. https://doi.org/10.1162/jocn_a_01579

Ko sem, A., Bosker, H. R., Takashima, A., Meyer, A., Jensen, O., & Hagoort, P. (2018). Neural Entrainment Determines the Words We Hear. Current Biology, 28(18), 2867-2875.e3. https://doi.org/10.1016/j.cub.2018.07.023

Laurienti, P. J., Perrault, T. J., Stanford, T. R., Wallace, M. T., & Stein, B. E. (2005). On the use of superadditivity as a metric for characterizing multisensory integration in functional neuroimaging studies. Experimental Brain Research, 166(3), 289–297. https://doi.org/10.1007/s00221-005-2370-2

Metzger, B. A., Magnotti, J. F., Wang, Z., Nesbitt, E., Karas, P. J., Yoshor, D., & Beauchamp, M. S. (2020). Responses to Visual Speech in Human Posterior Superior Temporal Gyrus Examined with iEEG Deconvolution. The Journal of Neuroscience: The Official Journal of the Society for Neuroscience, 40(36), 6938–6948. https://doi.org/10.1523/JNEUROSCI.0279-20.2020

Perrault, T. J., Vaughan, J. W., Stein, B. E., & Wallace, M. T. (2003). Neuron-Specific Response Characteristics Predict the Magnitude of Multisensory Integration. Journal of Neurophysiology, 90(6), 4022–4026. https://doi.org/10.1152/jn.00494.2003

Perrault, T. J., Vaughan, J. W., Stein, B. E., & Wallace, M. T. (2005). Superior Colliculus Neurons Use Distinct Operational Modes in the Integration of Multisensory Stimuli. Journal of Neurophysiology, 93(5), 2575–2586. https://doi.org/10.1152/jn.00926.2004

Senkowski, D., & Engel, A. K. (2024). Multi-timescale neural dynamics for multisensory integration. Nature Reviews Neuroscience, 25(9), 625–642. https://doi.org/10.1038/s41583-024-00845-7

Shen, L., Lu, X., Wang, Y., & Jiang, Y. (2023). Audiovisual correspondence facilitates the visual search for biological motion. Psychonomic Bulletin & Review, 30(6), 2272–2281. https://doi.org/10.3758/s13423-023-02308-z

Stanford, T. R., Quessy, S., & Stein, B. E. (2005). Evaluating the Operations Underlying Multisensory Integration in the Cat Superior Colliculus. Journal of Neuroscience, 25(28), 6499–6508. https://doi.org/10.1523/JNEUROSCI.5095-04.2005

Wright, T. M., Pelphrey, K. A., Allison, T., McKeown, M. J., & McCarthy, G. (2003). Polysensory Interactions along Lateral Temporal Regions Evoked by Audiovisual Speech. Cerebral Cortex, 13(10), 1034–1043. https://doi.org/10.1093/cercor/13.10.1034