first sighting: in-box Kind of like what's provided "out of the box".

I am surprised I have never heard of this riot or this piece of history.

Oh damn this is kinda sad.

This very feature of these digital literature games that not only tells stories but also allows players to actively engage in them is what builds our understanding and empathy for the video game characters. As players, to fully understand the extent of someone else’s struggles, we must put ourselves into their shoes and this form of game literally allows this metaphor to come to life.

I'm glad the people listened to his words and that he was so calm about it even after probably being upset that he was suspended.

I think the author is building up the charming characteristic for Dupin. To make readers like him more.<br /> Dupin was described as a well mannered man with money.

hänvisa till Figur 6 och bilagan.

Tabellen i bilagan behöver förklaras. Annars tror jag att det blir lite kinesiska.

Varför redovisar vi inte denna del av arbetet som är så relevant för denna rapport?

Thermia Mega XL

om vi integrerar Bilaga 1 inkl tabelltolkning i kapitel 3 så blir dilaga 2 den enda bilagan

punk13, det står att det bör vara lika med 5 minus 6. Stämmer det ?

här emotsäger vi oss själva. Ovan i sektion 1 skrev vi att 1 timme är OK. Så vi plockar bort den texten och håller oss fast till 5 minuter som rekommenderad upplösning.

Eftersom uteluften blir kallare vintertid är det rimligt att maskinerna måste höja framledningstemperaturen, vilket gör att COPvärme minskar. Motsatsen gäller sommartid, att maskinerna jobba med sämre COPkyla när framledningstemperaturen på Kylkretsen behöver vara lägre pga hög utelufttemperatur. Jag tror inte vattenflödet genom förångaren och kondensor ändras runt baslastmaskiner men det är klart något att dubbelkolla.

Här har vi kanske en intressant potentiellt mervärde. Beroende på var i systemet maskinerna kopplas kan man ev lyckas få de att jobba med mer jämnt COP.

Figurerna

ta bort, ersätt med en komma

något

Trenderna

lägg till "som referens" efter uteluften

ta bort

är det "interna" resp "externa" metoden är etablerat begrepp? om inte, ta bort dessa! Annars ok!

menar vi att energimätarna ovan mäter total för alla tre maskiner? bör förtydligas

Apotekaren

kan vi kommentera varför vi inte visar data för vinter resp sommardriftfall?

Denna slutsats måste också fram i början av rapporten

Denna variabel/mätning är lite otydligt. Vad är det som mäts via denna signal? Kompressorfrekvens? hur relaterar vi det till Hertz till exempel? Värt att förklara för läsaren.

men också för att studera hur maskinerna presterar inom olika frekvensintervaller

Detta är en bra slutsats. Har den tagits upp i första kapitlet? Det borde vi göra!

kylvärmepumparna/maskinerna

värmekapacitet

kylvärmepumparna

kylvärmepumpar

Carnot COP?eller menar vi att COP redan multiplicerats med Carnoeffektivitet?

Carnot COP? eller har COPc redan multiplicerats med Carnoeffektivitet?

ta bort?

köld

Dela i två. Vattenflöde genom förångaren och köldmedieflöde genom förångaren

Kondenseringstryck

Förångningstryck

Vattnets fram och returledningstemperatur vid kondensorn

ta bort

. Tillgänlig data inkluderar:

ersätt med "av"

Skriv bara "Bilia"

ändra rubriken till "Pilotprojekt"

proaktiv felsökning tillsammans med ovan åtgärder

Ändra till "Vi föreslår följande plan i eventuella kommande sprintar:" Ta bort "för att uppnå dessa mål".

ersätt med "analysverktyg".

ta bort

BILAGA 1.

Eftersom Bilaga1 är så viktigt så tänker jag att vi tar innehållet där direkt i denna sektion. En förklaring av Bilaga 1ger mer konkret idé på det ev fortsatta arbete.

Även om denna text är bra så säger det ingenting till Fabege. Kanske kan skrivas i slutet av dokumentet

En fråga: går det inte att komma med ett någorlunda kvalificerat påstående i stil med: "vi bedömer att en aktiv och datadriven övervakning av energicentraler som motsvarar Apotekaren och Bilia kan resultera i en förbättring av COP med X procentenheter, samt möjligheten att flytta en effektmängd om Y kW inköpt el och Z kW inköpt fjärrvärme på en axel om Å timmar, för att åstadkomma en plattare effektsignatur för den inköpta elen, såväl som för den inköpta fjärrvärmen."

Jag vet att det är en massa antaganden som behöver göra och att det behövs mycket data och analyser, men nu har jag lyxen att låtsas vara en IT-kille utan koll på byggnadsfysik... om det skulle gå att komma med en sådan hypotes som inte är tagen ur luften, utan tillräckligt väl underbyggd för att hålla för Thomas och Carolines granskning, tror jag att vi har goda förutsättningar för att fortsätta, eftersom de prioriterar effektfrågan inom Fabege.

TA BORT

Denna bilaga blir ett bra säljmaterial om ni förklarar tabellen. Ni kan även koppla det till era observationer i BILIA. Här har vi bra potentiella mervärden som borde lyftas upp i rapporten.

toppen om ni åtminstone kan förklara några av punkterna. T.ex. ett köldmedieläckage skulle visa sig genom en minskad underkylning? ...

gör om till en länk

denna punkt är otydlig. Formulera om om det verkligen är en relevant slutsats så som rubriken indikerar. Det vore toppen om vi kan presentera något relevant i Apotekaren, lik det vi poängterar i Bilia.

när jag klickar här hoppar inte dokumentet till Bilaga 2

system och delsystem

en gång per timme går vi miste om många värmepumpscykler. Jag föreslår att vi endast rekommenderar var 5:e minut.

ersätt med "utvärdering av värmepumpsprestanda i två av Fabeges befintliga anläggningar"

kanske bra en en parentes eller pop up som förklarar att sprint är "en bestämd tidsperiod under vilken ett team arbetar för att uppnå specifika leveranser." (googlade)

har vi ett projektnummer? i så fall borde vi ha uppdragsnummer och dokumentnummer i denna sektion

uppdatera med utskicksdatum

Här ska det stå projektansvarig.

fungerade

Bra, beskriv kort hur

Jag saknar potential med framtidens arbete. Vad kan vi uppnå?

Tänk så här: Thomas vill fortsätta arbete med oss men han måste kunna visa till sina chefer att vårt arbete get ett mervärde. Tänk att så om han ska använda denna rapport för att presentera arbetet till ledningen och försöka få dem att godkänna ett fortsättningsuppdrag. Ledningen har inte superbra koll på tekniska detaljer men förstår begrepp och övergripande koncept.

Inget om superheat och exp.ventilen?

Var det slutsatsen? Inget om att värmepumparna var överdimensionerade t.ex.? Minns inte exakt vad vi sa där

Vi måste kommentera dessa figurer mer. Hur ska Fabege tolka dem? Vilken information visar dem? jag saknar delen men max. prestanda vid 80% max.signal (runt 9-10 VDC).

Tror att det kan vara bra att kort nämna att Carnot effektivitet är en normaliserad COP

Ta bort den delen

This is my first time learning about this I never knew people back in the day could be this cruel.

I learned about this in 7th grade and it was so tragic that so many died because the owners swapped the door. The managers saved themselves but left so many trapped in the building.

Sick.

for - scalable fabrication technique - sustainable building - insulation - windows - beer hops aerogel

scalable fabrication technique - sustainable building - insulation - windows - beer hops aerogel - This fabrication process is highly scalable (Fig. 2a–c) and compatible with roll-to-roll processing, - It combines simple steps: - Moulding to define the volume of the desired hydrogel, - Solvent exchanges at modestly elevated temperatures<br /> - Rolling and drying the gels atop plastic support in rolls - step 1 - Process wood-pulp-derived cellulose nanofibres - via 2,2,6,6-tetramethylpiperidine-1-oxyl radical (TEMPO)-mediated oxidation of native cellulose25,35,36,37 - (Methods and Supplementary Fig. 3). - This process affects the surface charges associated with the carboxylate anion preclude aggregation25,30 of the nanofibres, - which form stable aqueous colloidal dispersions at varying concentrations - step 2 - Pour solution into moulds of desired shapes and sizes - (Methods and Supplementary Figs. 3 and 4) - step 3 - Adding acid interlinks these nanofibres by hydrogen bonds between the carboxyl groups, - transforming the colloidal dispersion into a hydrogel - (Supplementary Fig. 4b) - with a network of sparse nanofibres - step 4 - Exchange the fluid medium within the gel by - replacing water with isopropanol or ethanol - (Fig. 2a,b and Supplementary Fig. 4c,d), - Step 5 - Super critically dry it to form an aerogel - (Fig. 2c and Supplementary Fig. 5a–d)

Pourrait-on les qualifier de littérature ergodique d'après la définition qu'en fait Aspen J. Aarseth (soit une littérature dans laquelle « un effort non-trivial est requis pour permettre au lecteur de traverser le texte » (Cybertext. Perspectives on Ergodic Literature, 1997, p. 2)) ?

I've never heard the phrase "overton window" before, despite this, the effects of overton windows greatly impacts every American. Moving in accordance with the window is exactly expected of lawmakers trying to win the vote of as many Americans as they can.

This is too real...

Author response:

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

We are grateful for the many positive comments. Moreover, we appreciate the recommendations to improve the manuscript; particularly, the important discussion points raised by reviewer 1 and the comments made by reviewer 2 concerning an extended quantification of how near-spike input conductances vary across individual spikes. We have performed several new detailed analyses to address reviewer 2’s comments. In particular, we now provide for all relevant postsynaptic cells the complete distributions of the excitatory and inhibitory input conductance changes that occur right before and after postsynaptic spiking, and we provide corresponding distributions of non-spiking regions as a reference. We performed these analyses separately for different baseline activity levels. Our new results largely support our previous conclusions but provide a much more nuanced picture of the synaptic basis of spiking. To the best of our knowledge, this is the first time that parallel information on input excitation, inhibition and postsynaptic spiking is provided for individual neurons in a biological network. We would argue that our new results further support the fundamental notion that even a reductionist neuronal culture model can give rise to sophisticated network dynamics with spiking – at least partially – triggered by rapid input fluctuations, as predicted by theory. Moreover, it appears that changes in input inhibition are a key mechanism to regulate spiking during spontaneous recurrent network activity. It will be exciting to test whether this holds true for neural circuits in vivo.

In the following section, we address the reviewers’ comments individually.

Reviewer 1:

In this study the authors develop methods to interrogate cultured neuronal networks to learn about the contributions of multiple simultaneously active input neurons to postsynaptic activity. They then use these methods to ask how excitatory and inhibitory inputs combine to result in postsynaptic neuronal firing in a network context.

The study uses a compelling combination of high-density multi-electrode array recordings with patch recordings. They make ingenious use of physiology tricks such as shifting the reversal potential of inhibitory inputs, and identifying inhibitory vs. excitatory neurons through their influence on other neurons, to tease apart the key parameters of synaptic connections.

We thank the reviewer for acknowledging our efforts to develop an approach to investigate the synaptic basis of spiking in biological neurons and for appreciating the technical challenges that needed to be overcome.

The method doesn't have complete coverage of all neurons in the culture, and it appears to work on rather low-density cultures so the size of the networks in the current study is in the low tens.

(1) It would be valuable to see the caveats associated with the small size of the networks examined here.

(2) It would be also helpful if there were a section to discuss how this approach might scale up, and how better network coverage might be achieved.

These are indeed very important points that we should have discussed in more detail. Maximizing the coverage of neurons is critical to our approach, as it determines the number of potential synaptic connections that can be tested. The number of cells that we seeded onto our HD-MEA chip was chosen to achieve monolayer neuronal cultures. As detailed in ‘Materials and Methods -> Electrode selection and long-term extracellular recording of network spiking’, the entire HD-MEA chip (all 26'400 electrodes) was scanned for activity at the beginning of each experiment, and electrodes that recorded spiking activity were subsequently selected. While it is possible that some individual neurons escape detection, since they were not directly adjacent to an electrode, we estimate that a large majority of the active neurons in the culture was covered by our electrode selection method. New generations of CMOS HD-MEAs developed in our laboratory and other groups feature higher electrode densities, larger recording areas, and larger sets of electrodes that can be simultaneously recorded from (e.g., DOI:

10.1109/JSSC.2017.2686580 & 10.1038/s41467-020-18620-4). These features will substantially improve the coverage of the network and also allow for using larger neuronal networks. As suggested by reviewer 1, we added these points to the Discussion section of the revised manuscript.

The authors obtain a number of findings on the conditions in which the dynamics of excitatory and inhibitory inputs permit spiking, and the statistics of connectivity that result in this. This is of considerable interest, and clearly one would like to see how these findings map to larger networks, to non-cortical networks, and ideally to networks in-vivo. The suite of approaches discussed here could potentially serve as a basis for such further development.

(3) It would be useful for the authors to suggest such approaches.

We are confident that our suite of approaches will open important avenues to study the E & I input basis of postsynaptic spiking in other circuits beyond the in vitro cortical networks studied here. In fact, CMOS HD-MEA probes have been successfully combined with patch clamping in vivo (DIO: 10.1101/370080) and, in principle, the strategies and software tools introduced in our study would be equally applicable in an in vivo context. However, currently available in vitro CMOS HD-MEAs still surpass their in vivo counterparts (e.g., Neuropixels probes) in terms of electrode count. Moreover, using in vitro neural networks enables easy access and better network coverage compared to in vivo conditions. These are the main reasons why we chose an in vitro network for our investigation. We added these points to the Discussion section of the revised manuscript.

(4) The authors report a range of synaptic conductance waveforms in time. Not surprisingly, E and I look broadly different. Could the authors comment on the implications of differences in time-course of conductance profiles even within E (or I) synapses? Is this functional or is it an outcome of analysis uncertainty?

We are grateful to the reviewer for raising this interesting point. On the one hand, the onsets of the synaptic conductance waveform estimates were strikingly different between E and I synapses (see Fig. 8D). Furthermore, the rise and decay phases of synaptic currents were distinct for E vs. I inputs (Fig. 4C). We think that these differences are not just due to analysis uncertainty because both these observations are consistent with previously described properties of E and I inputs: Synaptic GABAergic I currents are typically slower compared to Glutamatergic E currents with respect to both rising and decay phase (DOI: 10.1126/science.abj586). Moreover, the relatively small onset latencies for I inputs that we observed are consistent with the well-known local action of inhibition. This finding was also consistent with smaller PRE-POST distances and general differences in neurite characteristics of E compared to I cells (Fig. S2).

One of the challenges in doing such studies in a dish is that the network is simply ticking away without any neural or sensory context to work on, nor any clear idea of what its outputs might mean. Nevertheless, at a single-neuron level one expects that this system might provide a reasonable subset of the kinds of activity an individual cell might have to work on.

(5) Could the authors comment on what subsets of network activity is, and is not, likely to be seen in the culture?

(6) Could they indicate what this would mean for the conclusions about E-I summation, if the in-vivo activity follows different dynamics?

We agree that there are natural limitations to a reductionist model, such as a dissociated cell culture. One may argue that neuronal cultures bear some similarities with neural networks formed during early brain development, where network formation is primarily driven by intrinsic, self-organizational capabilities. While such a self-organization is likely constrained in a 2D culture, it has been shown that several important circuit mechanisms that are observed in vivo are preserved in 2D dissociated cultures. For example, dissociated neuronal cultures can maintain E-I balance and achieve active decorrelation (DOI: 10.1038/nn.4415). In addition, in terms of activity levels, the sequences of heightened and more quiescent network spiking bear similarities with cortical Up-Down state oscillations observed during slow-wave sleep. To what extent individual circuit connectivity motifs and more nuanced network dynamics, found in vivo, can be recapitulated in vitro, is still not clear. However, combining our and previous work (especially DOI: 10.1038/nn.4415), we believe that there is sufficient evidence to justify work such as ours. On the one hand, identifying in simple cell culture models features of network dynamics and microcircuits known (or predicted) to exist in vivo is a testimony of neuronal self-organizing capabilities. On the other hand, our in vitro results will allow for more directed testing of equivalent mechanisms in vivo.

Reviewer 2:

The authors had two aims in this study. First, to develop a tool that lets them quantify the synaptic strength and sign of upstream neurons in a large network of cultured neurons. Second, they aimed at disentangling the contributions of excitatory and inhibitory inputs to spike generation.

For the quantification of synaptic currents, their methods allows them to quantify excitatory and inhibitory currents simultaneously, as the sign of the current is determined by the neuron identity in the high-density extracellular recording. They further made sure that their method works for nonstationary firing rates, and they did a simulation to characterize what kind of connections their analysis does not capture. They did not include the possibility of (dendritic) nonlinearities or gap junctions or any kind of homeostatic processes.

Thank you for the concise summary of our aims and of the features of our method. Indeed, we did not model nonlinear synaptic interactions, short-term plasticity etc., as there is likely a spectrum of possible interaction rules. Importantly, non-linear synaptic interactions were reduced by performing synaptic measurements in voltage-clamp mode.

We do not anticipate that this would impact our connectivity inference per se. However, the presence of a significant number of nonlinear events would imply that some deviations between reconstructed and measured patch current traces were to be expected even if all incoming monosynaptic connections were identified. In the future, it will be exciting to add to our current experimental protocol a simultaneous HD-MEA & patch-clamp recording, in which the membrane potential is measured in current-clamp mode. Following application of our synaptic input-mapping procedure, one could, in this way, directly assess input-sequence dependent non-linear synaptic integration during spontaneous network activity.

I see a clear weakness in the way that they quantify their goodness of fit, as they only report the explained variance, while their data are quite nonstationary. It could help to partition the explained variance into frequency bands, to at least separate the effects of a bias in baseline, the (around 100 Hz) band of synaptic frequencies and whatever high-frequency observation noise there may be. Another weak point is their explanation of unexplained variance by potential activation of extrasynaptic receptors without providing evidence. Given that these cultures are not a tissue and diffusion should be really high, this idea could easily be tested by adding a tiny amount of glutamate to the culture media.

As suggested by the reviewer, we have now partitioned the current traces into frequency bands and separately assessed the goodness-of-fit. We have updated Fig. 3C accordingly:

The following sentence was added to the main text:

“We separately compared slow baseline changes (< 3 Hz), fast synaptic activity (3 - 200 Hz) and putative high-frequency noise (> 200 Hz), yielding a median variance explained of approximately 60% in the 3 - 200 Hz range (Fig. 3C).”

Importantly, the variance explained in the frequency range of synaptic activity remains high. We would also like to point out that, even if all synaptic input connections were identified, one would expect some deviations between measured and reconstructed current trace. This is because the reconstructed trace is based on average input current waveforms and in the measured trace there may be synaptic transmission failures.

We agree that the offered explanation for unexplained variance by activation of extrasynaptic receptors is fairly speculative. As it was not a crucial discussion point, we have therefore removed the statement.

For the contributions of excitation and inhibition to neuronal spiking, the authors found a clear reduction of inhibitory inputs and increase of excitation associated with spiking when averaging across many spikes. And interestingly, the inhibition shows a reversal right after a spike and the timescale is faster during higher network activity. While these findings are great and provide further support that their method is working, they stop at this exciting point where I would really have liked to see more detail.

Thank you for acknowledging our main results concerning the synaptic basis of spiking. We attempted to integrate in one manuscript a suite of new approaches, in addition to the respective applications. We, therefore, tried to strike the appropriate level of detail in presenting our findings. With regard to our analyses of which synaptic input events regulate postsynaptic spiking, we agree with reviewer 2’s assessment that more detail concerning the variability across individual spikes would be helpful. In the following parts, we detail multiple new analyses that we have included in the revised manuscript to address reviewer 2’s comments.

A concern, of course, is that the network bursts in cultures are quite stereotypical, and that might cause averages across many bursts to show strange behaviour. So what I am missing here is a reference or baseline or null hypothesis. How does it look when using inputs from neurons that are not connected? And then, it looks like the E/(E+I) curve has lots of peaks of similar amplitude (that could be quantified...), so why does the neuron spike where it does? If I would compare to the peak (of similar amplitude) right before or right after (as a reference) are there some systematic changes? Is maybe the inhibition merely defining some general scaffold where spikes can happen and the excitation causes the spike as spiking is more irregular?

The averaged trace reveals a different timescale for high and low activity states. But does that reflect a superposition of EPSCs in a single trial or rather a different jittering of a single EPSC across trials? For answering this question, it would be good to know the variance (and whether/ how much it changes over time). Maybe not all spikes are preceded by a decrease in inhibition. Could you quantitify (correlate, scatterplot?) how exactly excitation and inhibition contributions relate for single postsynaptic spikes (or single postsynaptic non-spikes)? After all, this would be the kind of detail that requires the large amount of data that this study provides.

First of all, we are very grateful for the reviewer’s thorough assessment of our work and for the many valuable suggestions to improve it. We are convinced that we have addressed with our new analyses and the updated manuscript all issues raised here. One of the main findings from our original manuscript was that a rapid and brief change in input conductance (and particularly a reduction in inhibition) is an important spike trigger/regulator. We followed the reviewer’s suggestion and now provide scatter plots and distributions of the pre- (and post-spike) changes in input excitation and inhibition for individual postsynaptic spikes. A quantification of the peaks in the noisy E/(E+I) traces was not always trivial, which is why we reasoned that an assessment of the respective E and I changes is better suited. Moreover, as an unbiased reference, we generated separately for each postsynaptic cell a corresponding distribution of changes in input conductance in non-spiking periods (using random time points). We included our new results and updated figures in our responses to the specific reviewer comments below.

For the first part, the authors achieved their goal in developing a tool to study synaptic inputs driving subthreshold activity at the soma, and characterizing such connections. For the second part, they found an effect of EPSCs on firing, but they barely did any quantification of its relevance due to the lack of a reference.

With the availability of Neuropixels probes, there is certainly use for their tool in in vivo applications, and their statistical analysis provides a reference for future studies.

The relevance of excitatory and inhibitory currents on spiking remains to be seen in an updated version of the manuscript.

Thank you. Please see our new analyses below. Our new findings are in agreement with the main conclusions of the original manuscript. We provide evidence that rapid pre-spike changes in input conductance are observed across most individual spikes and that these rapid changes occur significantly more often before measured spikes than in non-spiking periods.

I feel that specifically Figures 6 and 7 lack relevant detail and a consistent representation that would allow the reader to establish links between the different panels. The analysis shows very detailed examples, but then jumps into analyses that show population averages over averaged responses, losing or ignoring the variability across trials. In addition, while their results themselves pass a statistical test, it is crucial to establish some measure of how relevant these results are. For that, I would really want to know how much spiking would actually be restricted by the constraints that would be posed by these results, i.e. would this be reflected in tiny changes in spiking probabilities, or are there times when spiking probabilities are necessarily high, or do we see times when we would almost certainly get a spike, but neurons can fire during other times as well.

I would agree that a detailed, quantitative analysis of this question is beyond the scope of this paper, but a qualitative analysis is feasible and should be done.

Please see our revised Figure 6. We have rearranged some of the original panels and removed one example of mean conductance profiles. Moreover, we removed a panel with analysis results based on mean conductances that is now obsolete, as more detailed analyses are provided (which are in agreement with the original findings). Analyses from panels (A-F) are mostly unchanged. Panels (G-J) show the new results.

The following paragraphs, which were added to the main text of the revised manuscript, describe our new findings:

“For a more nuanced picture of which synaptic events are associated with postsynaptic spiking, we next quantified the changes in input excitation and inhibition that preceded individual postsynaptic spikes. In our analysis, we first focused on periods with high synaptic input activity. As previously discussed, cortical neurons in vivo typically receive and integrate barrages of input activation, similar to the high-activity events that we observed here (e.g., the event depicted in Fig. 6A, right). In Fig. 6G/H, individual pre-spike changes in input conductance are shown for two example postsynaptic neurons (plots labeled ‘spiking’, right). To assess how specific these conductance changes were to spiking periods, we also quantified the changes in input conductance that occurred during non-spiking periods as a reference (we used random time points from high-activity events excluding time points adjacent to measured spike times; we upscaled the number of measured spikes by 10x; the respective plots were labeled ‘non-spiking’). Spikes of both example neurons exhibited – compared to non-spiking regions – significantly more often a pre-spike decrease in inhibition, consistent with the mean conductance profiles. Precisely how an increase (top-right quadrants in Fig. 6G/H) or decrease (bottom-left quadrants) in both I and E conductance influenced the neuronal membrane potential is difficult to predict. However, if rapid changes in input conductance had a significant role in triggering spikes, one would expect that fewer spikes would exhibit a hyperpolarizing pre-spike increase in I and decrease in E (top-left quadrant) compared to the non-spiking period. Conversely, a decrease in I and an increase E (bottom-right quadrants) would likely result in a membrane potential depolarization so that more spikes should feature the corresponding pre-spike conductance changes compared to non-spiking periods. These relative shifts are precisely what can be observed in the plots of the two example neurons (Fig. 6G/H) and, in fact, across recordings (Fig. 6I). Finally, we compared the distributions of pre-spike changes in input inhibition and excitation of each postsynaptic neuron (Fig. 6J). Further indicating a pivotal role of inhibition in triggering spikes, 6 out of 7 neurons exhibited a clear decrease in the mean values (and medians) of pre-spike changes in inhibition compared to non-spiking periods. Interestingly, the 3 out of 7 neurons with an increase in excitation showed the smallest decrease in inhibition (or even an increase in inhibition in case of neuron #7). This latter observation suggests a matching of E and I inputs and cell-specific relative contributions of E and I conductance changes in triggering spikes.

Theoretically, neuronal spiking could be driven by a prolonged suprathreshold depolarization (Petersen and Berg 2016; Renart et al. 2007) or, in more favorable subthreshold regimes, by fast synaptic input fluctuations (Ahmadian and Miller 2021; Amit and Brunel 1997; Brunel 2000; Van Vreeswijk and Sompolinsky 1996). In this section, we demonstrated that the majority of investigated neurons featured – during high-activity periods – a significant number of spikes that were associated with rapid pre-spike changes in input conductances. These findings suggest that even simple neuronal cultures can self-organize to form circuits exhibiting sophisticated spiking dynamics.”

Our new analyses detailed in Fig. 6 show that there are also presumably depolarizing events (e.g., decrease in I and increase in E) in non-spiking regions. In future studies, it will be interesting to examine what distinguishes these events from spike-inducing events of similar magnitude – one possibility is a dependency on specific input-activation sequences.

During the first days and weeks of developing neuronal cultures, spiking activity rapidly shifts from synapse-independent activity patterns to spiking dynamics that do depend on synaptic inputs and are progressively organized in network-wide high-activity events (DOI: 10.1016/j.brainres.2008.06.022). In our study, cultures at days-in-vitro 15-18 were used, and approximately 15% of the spikes occurred during high-activity events with relatively strong E and I input activity. In addition, spikes that occurred during low-activity events were at least partially regulated by synaptic input (see answers below related to Fig. 7).

In the following, I am detailing what I would consider necessary to be done about these two Figures:

Figure 6C is indeed great, though I don't see why the authors would characterize synchrony as low. When comparing with Figure 4B, I'd think that some of these values are quite high. And it wouldn't help me to imagine error bars in panel 6D.

We have removed our characterization as ‘low’ from the text. One important difference between our synchrony measure (STTC) and the quantification of spike-transmission probability (STP) is the ‘lag’ of a few milliseconds for the STP quantification window to account for synaptic delay.

Figure 6B is useful, but could be done better: The autocovariance of a shotnoise process is a convolution of the autocovariance of underlying point process and the autocovariance of the EPSC kernel. So one would want to separate those to obtain a better temporal resolution. But a shotnoise process has well defined peaks, and the time of these local maxima can be estimated quite precisely. Now if I would do a peak triggered average instead of the full convolution, I would do half of the deconvolution and obtain a temporally asymmetric curve of what is expected to happen around an EPSC. Importantly, one could directly see expected excitation after inhibition or expected inhibition after excitation, and this visualization could be much better and more intuitively compared to panel 6E.

We appreciate the reviewer’s suggestion to present these results in a more sophisticated way. We would like to propose to stick with the original analysis to have it comparable with related analyses from the literature (e.g., DOI: 10.1038/nn.2105). Therefore, we hope the reviewer finds it acceptable that we leave the presentation of the data in its original form and potentially follow up in future work with the analysis strategy proposed by the reviewer.

Panel D needs some variability estimate (i.e. standard deviation or interquartile range or even a probability density) for those traces.

Figure 6E: Please use more visible colors. A sensitivity analysis to see traces for 2E/(2E+I) and E/(E+2I) would be great.

Figure 6F: with an updated panel B, we should be able to have a slope for average inhibition after excitation for each of these cells. A second panel / third column showing those slopes would be of interest. It would serve as a reference for what could be expected from E-I interactions alone.

With regard to the variability estimate in D, we now provide multiple panels characterizing the variability. For one, Fig. 6H contains a scatter plot of the pre-spike changes in input conductance across all individual postsynaptic spikes from the example cell shown in D. Moreover, in Fig. 7A, we show from the same example cell the standard deviations associated with the mean conductance traces separately for spikes that occurred during low- and high-activity states. For better visibility and because the separation according to activity states is more informative, we kept the original presentation of panel D (however, removing one example cell). In addition, we show the same mean traces from panel D with the respective standard deviations (across all spikes) in Supplementary Figure S3.

Colors in Fig. 6E are adjusted, as requested.

We have removed panel Fig. 6F as we now provide more detailed analyses at single-spike level (see Fig. 6G-J).

Figure 6G: Could the authors provide an interquartile range here?

With regard to the aligned input-output data from original panel Fig. 6G, now in panel Fig. 6F in the updated figure version, we show all individual traces that were averaged: the E/I traces from panel Fig. 6E and the three action potential waveforms from Supplementary Figure S5. Therefore, we chose to present the means only for better visibility.

Figure 7A: it may be hard to squeeze in variability estimates here, but the information on whether and how much variance might be explained is essential. Maybe add another panel to provide a variability estimate? The variability estimate in panel 7B and 7D only reflect variability across connections, and it would be useful to add panels for the time courses of the variability of g (or E/(E+I) respectively).

We now include the standard deviations across the input conductance traces in the updated Fig. 7A, as requested. We have also simplified Fig. 7 and performed the analysis using the 6 out of 7 neurons that, based on our new analysis (Fig. 6J) displayed a clear reduction in pre-spike inhibition, relative to the reference distribution. For a complete overview of the state-dependent changes in input conductance that are associated with individual postsynaptic spikes, we have included a new supplementary figure (Fig. S6). Fig. S6 also includes a characterization of the changes in input inhibition that occur right after postsynaptic spiking. In addition, Fig. S6D shows the standard deviations corresponding to the mean input conductance traces of all cells – separately for high- and low-activity periods.

We added the following paragraph to the main text of the revised manuscript:

“How can these deviations in the mean conductance profiles be explained? To answer this question, we further quantified – separately for low and high g states – the changes in input inhibition that occurred right before and after individual postsynaptic spikes (Fig. S6). This single-spike analysis suggested that, during high g states, most spikes experienced a post-spike increase and pre-spike decrease in inhibition (see also Fig. 6J). On the other hand, low g states were characterized by sparse synaptic input (e.g., see reconstruction in Fig. 6A). Therefore, many of the spikes that occurred during low g states were associated with little change in input conductance (note medians of approximately zero in Fig. S6A/C). Nevertheless, a considerable fraction of spikes (often > 25%) from low g states were also associated with a post-spike increase and pre-spike drop in inhibition. It, therefore, appears that even the sparse inhibitory inputs of low g states could influence spike timing. Moreover, the post-spike increases in input inhibition during low g states suggest that there were strong regulatory inhibitory circuits in place. However, limited activity levels during low g states presumably introduced an increased jitter of these spike-associated changes in input inhibition.

In summary, the input inhibition of high-conductance states provides reliable and narrow windows-of-spiking opportunity. In addition, even during periods of sparse activity, there are rudimentary synaptic mechanisms in place to regulate spike timing.”

As a suggestion for further analysis, though I am well aware that this is likely beyond the scope of this manuscript, I'd suggest the following analysis:

I would split the data into the high and low activity states. Then I would compute the average of E/(E+I) values for spikes. Assuming that spikes tend to happen for local maxima of E/(E+I) I would find local maxima for periods without spike such that their average is equal to the value for actual spikes. Finally, I would test for a systematic difference in either excitation or inhibition.

If there is no difference, you can make the claim that synaptic input does not guarantee a spike, and compare to a global average of E/(E+I).

We are grateful for the fantastic suggestions for future analysis. We look forward to conducting these analyses in a more detailed follow-up characterization.

In addition to the major alterations detailed above, we performed smaller corrections (e.g., spelling mistakes, inaccuracies) in some parts of the manuscript.

the big realization

After reading this the way I took it was that a windows job when its closed it to keep things out of the house and only bring in the natural light from outside. It seems as she was trying to catch the fly and at that point the fly flies out the open window which is why it says the windows fail. Chasing the fly she looks into the light blinging her and losing the fly

These lines are very symbolic and represent last a few feelings of a person dying. Last sensory/motor skills of the brain. Last connection with nature. Last connection with this living world before pass to soul world. Such as in this these lines represent every single last touch, hearing and seeing sign. Although, the line "I could not see to see" might express the last late hope of living.

"The Windows" in this metaphor refers to the speaker's eyes or ability to see. When the speaker gets closer to passing away, the "Windows failed," signifying his losing vision as his life approaches to an end. By associating the closure of windows with the actual act of dying, such as losing consciousness and sight, this metaphor suggests a barrier separating the individual from the outside world as they approach death. The use of the phrase "I could not see to see" highlights the speaker's complete loss of vision and reinforces their incapacity to notice anything beyond this last instant.

Author response:

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

We are deeply appreciative of the reviewers' insightful comments and constructive feedback on our manuscript. In response, we have implemented substantial revisions to enhance the clarity and impact of our work. Key changes include: 

Reframing: We have shifted our focus from cognitive control to attention and memory processes, aligning more closely with our experimental design. This reframing is reflected throughout the manuscript, including additional citations highlighting the triple network model's involvement in memory processing. To reflect this change, we have updated the title to "Causal dynamics of salience, default mode, and frontoparietal networks during episodic memory formation and recall: A multi-experiment iEEG replication".

Control analyses using resting-state epochs: We have conducted new analyses comparing task periods to resting baseline epochs. These results demonstrate enhanced directed information flow from the anterior insula to both the default mode and frontoparietal networks during encoding and recall periods compared to resting state across all four experiments. This finding underscores the anterior insula's critical role in memory and attention processing.

Control analysis using the inferior frontal gyrus: To address specificity concerns, we performed control analyses using the inferior frontal gyrus as a comparison region. This analysis confirms that the observed directed information flow to the default mode and frontoparietal networks is specific to the anterior insula, rather than a general property of task-engaged brain regions.

These revisions, combined with our rigorous methodologies and comprehensive analyses, provide compelling support for the central claims of our manuscript. We believe these changes significantly enhance the scientific contribution of our work.

Our point-by-point responses to the reviewers' comments are provided below.

Reviewer 1:

-  The authors present results from an impressively sized iEEG sample. For reader context, this type of invasive human data is difficult and time-consuming to collect and many similar studies in high-level journals include 5-20 participants, typically not all of whom have electrodes in all regions of interest. It is excellent that they have been able to leverage open-source data in this way. 

-  Preprocessing of iEEG data also seems sensible and appropriate based on field standards. 

-  The authors tackle the replication issues inherent in much of the literature by replicating findings across task contexts, demonstrating that the principles of network communication evidenced by their results generalize in multiple task memory contexts. Again, the number of iEEG patients who have multiple tasks' worth of data is impressive. 

We thank the reviewer for the encouraging comments and appreciate the positive feedback.  

(1.1) The motivation for investigating the tripartite network during memory tasks is not currently well-elaborated. Though the authors mention, for example, that "the formation of episodic memories relies on the intricate interplay between large-scale brain networks (p. 4)", there are no citations provided for this statement, and the reader is unable to evaluate whether the nodes and networks evidenced to support these processes are the same as networks measured here. 

Recommendation: Detail with citations the motivation for assessing the tripartite network in these tasks. Include work referencing network-level and local effects during encoding and recall.

We appreciate the reviewer's feedback and suggestions for improving our framing. We have substantially expanded and revised the Introduction to elaborate on the motivation for investigating the tripartite network during memory tasks, supported by relevant citations.

We now provide a stronger rationale for examining these networks in the context of episodic memory, emphasizing that while the tripartite network has been extensively studied in cognitive control tasks, growing evidence suggests its relevance to episodic memory as a domain-general network. We cite several key studies that demonstrate the involvement of the salience, default mode, and frontoparietal networks in memory processes, including work by Sestieri et al. (2014) and Vatansever et al. (2021), which show the consistent engagement of these networks during memory tasks. We have also included references to studies examining network-level and local effects during encoding and recall, such as the work by Xie et al. (2012) on disrupted intrinsic connectivity in amnestic mild cognitive impairment, and Le Berre et al. (2017) on the role of insula connectivity in memory awareness (pages 4-5).

Furthermore, we have clarified how our study aims to address gaps in the current understanding by investigating the electrophysiological basis of these network interactions during memory formation and retrieval, which has not been explored in previous research. This expanded framing provides a clearer motivation for our investigation and places our study within the broader context of memory and network neuroscience research (pages 3-6).  

(1.2) In addition, though the tripartite network has been proposed to support cognitive control processes, and the neural basis of cognitive control is the framed focus of this work, the authors do not demonstrate that they have measured cognitive control in addition to simple memory encoding and retrieval processes. Tasks that have investigated cognitive control over memory (such as those cited on p. 13 - Badre et al., 2005; Badre & Wagner, 2007; Wagner et al., 2001; Wagner et al., 2005) generally do not simply include encoding, delay, and recall (as the tasks used here), but tend to include some manipulation that requires participants to engage control processes over memory retrieval, such as task rules governing what choice should be made at recall (e.g., from Badre et al., 2005 Fig. 1: congruency of match, associative strength, number of choices, semantic similarity). Moreover, though there are task-responsive signatures in the nodes of the tripartite networks, concluding that cognitive control is present because cognitive control networks are active would be a reverse inference.

Recommendation: If present, highlight components of the tasks that are known to elicit cognitive control processes and cite relevant literature. If the tasks cannot be argued to elicit cognitive control, reframe the motivation to focus on task-related attention or memory processes. If the latter, reframe the motivation for investigating the tripartite network in this context absent control.

We appreciate the reviewer's insightful comment and recommendation. We acknowledge that our tasks do not include specific manipulations designed to elicit cognitive control processes over memory retrieval. In light of this, we have reframed our motivation and discussion to focus on the role of the tripartite network in attention and memory processes more broadly, rather than cognitive control specifically (pages 3-6).

As noted in Response 1.1, we have revised the Introduction to emphasize the domain-general nature of these networks and their involvement in various cognitive processes, including memory. We also highlight how the salience, default mode, and frontoparietal networks contribute to different aspects of memory formation and retrieval, drawing on relevant literature.

Our revised framing examines the salience network's role in detecting behaviorally relevant stimuli and orienting attention during encoding, the default mode network's involvement in internally-driven processes during recall, and the frontoparietal network's contribution to maintaining and manipulating information in working memory. We now present our study as an investigation into how these networks interact during different phases of memory processing, rather than focusing specifically on cognitive control. This approach aligns better with our experimental design and allows us to explore the broader applicability of the tripartite network model to memory processes. 

This revised reframing provides a more accurate representation of our study's scope and contribution to understanding the role of large-scale brain networks in memory formation and retrieval (pages 3-6). 

(1.3) It is currently unclear if the directed information flow from AI to DMN and FPN nodes truly arises from task-related processes such as cognitive control or if it is a function of static brain network characteristics constrained by anatomy (such as white matter connection patterns, etc.). This is a concern because the authors did not find that influences of AI on DMN or FPN are increased relative to a resting baseline (collected during the task) or that directed information flow differs in successful compared to unsuccessful retrieval. I doubt that this AI influence is 1) supporting a switch between the DMN and FPN via the SN or 2) relevant for behavior if it doesn't differ from baseline-active task or across accuracy conditions. An additional comparison that may help investigate whether this is reflective of static connectivity characteristics would be a baseline comparison during non-task rest or sleep periods.  

Recommendation: As described in the task of the concern, analyze the PTE across the same contacts during sleep or task-free rest periods (if present in the dataset). 

We thank the reviewer for this suggestion. We have now carried out additional analyses using resting-state baseline epochs. We found that directed information flow from the AI to both the DMN and FPN were enhanced during the encoding and recall periods compared to resting-state baseline in all four experiments. These new results have now been included in the revised Results (page 12):    

“Enhanced information flow from the AI to the DMN and FPN during episodic memory processing, compared to resting-state baseline  

We next examined whether directed information flow from the AI to the DMN and FPN nodes during the memory tasks differed from the resting-state baseline. Resting-state baselines were extracted immediately before the start of the task sessions and the duration of task and rest epochs were matched to ensure that differences in network dynamics could not be explained by differences in duration of the epochs. Directed information flow from the AI to both the DMN and FPN were higher during both the memory encoding and recall phases and across the four experiments, compared to baseline in all but two cases (Figures S6, S7). These findings provide strong evidence for enhanced role of AI directed information flow to the DMN and FPN during memory processing compared to the resting state.” 

(1.4) Related to the above concern, it is also questionable how directed information flow from AI facilitates switching between FPN and DMN during both encoding and recall if high gamma activity does not significantly differ in AI versus PCC or mPFC during recall as it does during encoding. It seems erroneous to conclude that the network-level communication is happening or happening with the same effect during both task time points when these effects are decoupled in such a way from the power findings.  

We appreciate the reviewer's insightful observation regarding the apparent discrepancy between our directed information flow findings and the high-gamma activity results. This comment highlights an important distinction in interpreting our results, and we thank the reviewer for the opportunity to address this point.

Our findings demonstrate that directed information flow from the AI to the DMN and FPN persists during both encoding and recall, despite differences in local high-gamma activity patterns. This dissociation suggests that the network-level communication facilitated by the AI may operate independently of local activation levels in individual nodes. It is important to note that our directed connectivity analysis (using phase transfer entropy) was conducted on broadband signals (0.5-80 Hz), while the power analysis focused specifically on the high-gamma band (80-160 Hz). These different frequency ranges may capture distinct aspects of neural processing. The broadband connectivity might reflect more general, sustained network interactions, while high-gamma activity may be more sensitive to specific task demands or cognitive processes.

The phase transfer entropy analysis captures directed interactions over extended time periods, while the high-gamma power analysis provides a more temporally precise measure of local neural activity. The persistent directed connectivity from AI during recall, despite changes in local activity, might reflect the AI's ongoing role in coordinating network interactions, even when its local activation is not significantly different from other regions.

Rather than facilitating "switching" between FPN and DMN, as we may have previously overstated, our results suggest that the AI maintains a consistent pattern of influence on both networks across task phases. This influence might serve different functions during encoding (e.g., orienting attention to external stimuli) and recall (e.g., monitoring and evaluating retrieved information), even if local activation patterns differ.

It is crucial to note that in the three verbal tasks, our analysis of memory recall is time-locked to word production onset. However, the precise timing of the internal recall process initiation is unknown. This limitation may affect our ability to capture the full dynamics of network interactions during recall, particularly in the early stages of memory retrieval. Interestingly, in the spatial memory task WMSM, the PCC/precuneus exhibited an earlier onset and enhanced activity compared to the AI. This task may provide a clearer window into recall processes:

findings align with the view that DMN nodes may play a crucial role in triggering internal recall processes. However, the precise timing of internal retrieval initiation remains a challenge in the three verbal tasks, potentially limiting our ability to capture the full dynamics of regional activity, and its replicability, during early stages of recall.

These observations highlight the need for more detailed investigation of the temporal dynamics of network interactions during memory recall. To further elucidate the relationship between directed connectivity and local activity, future studies could employ time-resolved connectivity analyses and investigate coupling between different frequency bands. This could provide a more precise understanding of how network-level communication relates to local neural dynamics across different task phases.

We have revised the manuscript to more accurately reflect these points and avoid overstating the implications of our findings (pages 15-19). We thank the reviewer for prompting this important clarification, which we believe strengthens the interpretation and discussion of our results.

(1.5) Missing information about the methods used for time-frequency conversion for power calculation and the power normalization/baseline-correction procedure bars a thorough evaluation of power calculation methods and results. 

Recommendation: Include more information about how power was calculated. For example, how were time-series data converted to time-frequency (with complex wavelets, filter-hilbert, etc.)? What settings were used (frequency steps, wavelet length)? How were power values checked for outliers and normalized (decibels, Z-transform)? How was baseline correction applied (subtraction, division)?

We have now included detailed information related to our power calculation and normalization steps as we note on page 28: “We first filtered the signals in the high-gamma (80160 Hz) frequency band (Canolty et al., 2006; Helfrich & Knight, 2016; Kai J. Miller, Weaver, & Ojemann, 2009) using sequential band-pass filters in increments of 10 Hz (i.e., 80–90 Hz, 90– 100 Hz, etc.), using a fourth order two-way zero phase lag Butterworth filter. We used these narrowband filtering processing steps to correct for the 1/f decay of power. We then calculated the amplitude (envelope) of each narrow band signal by taking the absolute value of the analytic signal obtained from the Hilbert transform (Foster, Rangarajan, Shirer, & Parvizi, 2015). Each narrow band amplitude time series was then normalized to its own mean amplitude, expressed as a percentage of the mean. Finally, we calculated the mean of the normalized narrow band amplitude time series, producing a single amplitude time series. Signals were then smoothed using 0.2s windows with 90% overlap (Kwon et al., 2021) and normalized with respect to 0.2s pre-stimulus periods by subtracting the pre-stimulus baseline from the post-stimulus signal.” 

(1.6) If revisions to the manuscript can address concerns about directed information flow possibly being due to anatomical constraints - such as by indicating that directed information flow is not present during non-task rest or sleep - this work may convey important information about the structure and order of communication between these networks during attention to tasks in general. However, the ability of the findings to address cognitive control-specific communication and the nature of neurophysiological mechanisms of this communication - as opposed to the temporal order and structure of recruited networks - may be limited.

We appreciate the reviewer's insightful feedback, which has led to significant improvements in our manuscript. In response, we have made the following key changes. We have shifted our focus from cognitive control to the broader roles of the tripartite network in attention and memory processes. This reframing aligns more closely with our experimental design and the nature of our tasks. We have revised the Introduction, Results, and Discussion sections to reflect this perspective, providing a more accurate representation of our study's scope and contribution. Additionally, to strengthen our findings, we have conducted new analyses comparing task periods to resting-state baselines. These analyses revealed that directed information flow from the anterior insula to both the DMN and FPN was significantly enhanced during memory encoding and recall periods compared to resting-state across all four experiments. This finding provides robust evidence for the specific involvement of these network interactions in memory processing. Please also see Response 1.2 above. 

(1.7) Because phase-transfer entropy is presented as a "causal" analysis in this investigation (PTE), I also believe it is important to highlight for readers recent discussions surrounding the description of "causal mechanisms" in neuroscience (see "Confusion about causation" section from Ross and Bassett, 2024, Nature Neuroscience). A large proportion of neuroscientists (admittedly, myself included) use "causal" only to refer to a mechanism whose modulation or removal (with direct manipulation, such as by lesion or stimulation) is known to change or control a given outcome (such as a successful behavior). As Ross and Bassett highlight, it is debatable whether such mechanistic causality is captured by Granger "causality" (a.k.a. Granger prediction) or the parametric PTE, and the imprecise use of "causation" may be confusing. The authors could consider amending language regarding this analysis if they are concerned about bridging these definitions of causality across a wide audience. 

We thank the reviewer for this suggestion. We would like to clarify here that we define causality in our manuscript as follows: a brain region has a causal influence on a target if knowing the past history of temporal signals in both regions improves the ability to predict the target's signal in comparison to knowing only the target's past, as defined in earlier studies (Granger, 1969; Lobier, Siebenhühner, Palva, & Matias, 2014). We have now included this clarification in the Introduction section (page 6).  

We also agree with the reviewer that to more mechanistically establish a causal link between the neural dynamics and behavior, lesion or brain stimulation studies are necessary. We have now acknowledged this in the revised Discussion as we note: “Although our computational methods suggest causal influences, direct causal manipulations, such as targeted brain stimulation during memory tasks, are needed to establish definitive causal relationships between network nodes.” (page 19). 

Minor additional information that would be helpful to the reader to include: 

(1.8) How exactly was line noise (p. 24) removed? (For example, if notch filtered, how were slight offsets of the line noise from exactly 60.0Hz and harmonics identified and handled?). 

We would like to clarify here that to filter line noise and its harmonics, we used bandstop filters at 57-63 Hz, 117-123 Hz, and 177-183 Hz. To create a band-stop filter, we used a fourth order two-way zero phase lag Butterworth filter. This information has now been included in the revised Methods (page 26). 

(1.9) Why were the alpha and beta bands collapsed for narrowband filtering?

Please note that we did not combine the alpha (8-12 Hz) and beta (12-30 Hz) bands for narrowband filtering, rather these two frequency bands were analyzed separately. However, we combined the delta (0.5-4 Hz) and theta (4-8 Hz) frequency bands into a combined delta-theta (0.5-8 Hz) frequency band for our analysis since previous human electrophysiology studies have not settled on a specific band of frequency (delta or theta) for memory processing. Previous human iEEG (Ekstrom et al., 2005; Ekstrom & Watrous, 2014; Engel & Fries, 2010; Gonzalez et al., 2015; Watrous, Tandon, Conner, Pieters, & Ekstrom, 2013) as well as scalp EEG and MEG studies, have shown that both the delta and theta frequency band oscillations play a prominent role for human memory encoding as well as retrieval (Backus, Schoffelen, Szebényi, Hanslmayr, & Doeller, 2016; Clouter, Shapiro, & Hanslmayr, 2017; Griffiths, Martín-Buro, Staresina, & Hanslmayr, 2021; Guderian & Düzel, 2005; Guderian, Schott, Richardson-Klavehn, & Düzel, 2009).  

Reviewer 2:

In this study, the authors leverage a large public dataset of intracranial EEG (the University of Pennsylvania RAM repository) to examine electrophysiologic network dynamics involving the participation of salience, frontoparietal, and default mode networks in the completion of several episodic memory tasks. They do this through a focus on the anterior insula (AI; salience network), which they hypothesize may help switch engagement between the DMN and FPN in concert with task demands. By analyzing high-gamma spectral power and phase transfer entropy (PTE; a putative measure of information "flow"), they show that the AI shows higher directed PTE towards nodes of both the DMN and FPN, during encoding and recall, across multiple tasks. They further demonstrate that high-gamma power in the PCC/precuneus is decreased relative to the AI during memory encoding. They interpret these results as evidence of "triple-network" control processes in memory tasks, governed by a key role of the AI. 

I commend the authors on leveraging this large public dataset to help contextualize network models of brain function with electrophysiological mechanisms - a key problem in much of the fMRI literature. I also appreciate that the authors emphasized replicability across multiple memory tasks, in an effort to demonstrate conserved or fundamental mechanisms that support a diversity of cognitive processes. However, I believe that their strong claims regarding causal influences within circumscribed brain networks cannot be supported by the evidence as presented. In my efforts to clearly communicate these inadequacies, I will suggest several potential analyses for the authors to consider that might better link the data to their central hypotheses.

We thank the reviewer for the encouraging comments and suggestions for improving the manuscript. Please see our detailed responses and clarifications below. 

(2.1) As a general principle, the effects that the authors show - both in regards to their highgamma power analysis and PTE analysis - do not offer sufficient specificity for a reader to understand whether these are general effects that may be repeated throughout the brain, or whether they reflect unique activity to the networks/regions that are laid out in the Introduction's hypothesis. This lack of specificity manifests in several ways, and is best communicated through examples of control analyses. 

We appreciate the reviewer's insightful comment regarding the specificity of our findings. We agree that additional analyses could provide valuable context for interpreting our results. In response, we have conducted the following additional analyses and made corresponding revisions to the manuscript:

Following the reviewer's suggestion, we have selected the inferior frontal gyrus (IFG, BA 44) as a control region. The IFG serves as an ideal control region due to its anatomical adjacency to the AI, its involvement in a wide range of cognitive control functions including response inhibition (Cai, Ryali, Chen, Li, & Menon, 2014), and its frequent co-activation with the AI in fMRI studies. Furthermore, the IFG has been associated with controlled retrieval of memory (Badre et al., 2005; Badre & Wagner, 2007; Wagner et al., 2001), making it a compelling region for comparison. We repeated our PTE analysis using the IFG as the source region, comparing its directed influence on the DMN and FPN nodes to that of the AI.  

Our analysis revealed a striking contrast between the AI and IFG in their patterns of directed information flow. While the AI exhibited strong causal influences on both the DMN and FPN, the IFG showed the opposite pattern. Specifically, both the DMN and FPN demonstrated higher influence on the IFG than the reverse, during both encoding and recall periods, and across all four memory experiments (Figures S4, S5). 

These findings highlight the unique role of the AI in orchestrating large-scale network dynamics during memory processes. The AI's pattern of directed information flow stands in contrast to that of the IFG, despite their anatomical proximity and shared involvement in cognitive control processes. This dissociation underscores the specificity of the AI's function in coordinating network interactions during memory formation and retrieval. These results have now been included in our revised Results on page 11.  

(2.2) First, the PTE analysis is focused solely on the AI's interactions with nodes of the DMN and FPN; while it makes sense to focus on this putative "switch" region, the fact that the authors report significant PTE from the AI to nodes of both networks, in encoding and retrieval, across all tasks and (crucially) also at baseline, raises questions about the meaningfulness of this statistic. One way to address this concern would be to select a control region that would be expected to have little/no directed causal influence on these networks and repeat the analysis. Alternatively (or additionally), the authors could examine the time course of PTE as it evolves throughout an encoding/retrieval interval, and relate that to the timing of behavioral events or changes in high-gamma power. This would directly address an important idea raised in their own Discussion, "the AI is wellpositioned to dynamically engage and disengage with other brain areas."  

Please see Response 2.1 above for additional analyses related to control region.  

We also appreciate the reviewer's suggestion regarding time-resolved PTE analysis. However, it's important to note that our current methodology does not allow for such fine-grained temporal analysis. This is due to the fact that PTE, which is an information theoretic measure and relies on constructing histograms of occurrences of singles, pairs, or triplets of instantaneous phase estimates from the phase time-series (Hillebrand et al., 2016) (Methods), requires sufficient number of cycles in the phase time-series for its reliable estimation (Lobier et al., 2014). PTE is based on estimating the time-delayed directed influences from one time-series to the other and its estimate is the most accurate when a large number of time-points (cycles) are available (Lobier et al., 2014). Since our encoding and recall epochs in the verbal recall tasks were only 1.6 seconds long, which corresponds to only 800 time-points with a 500 Hz sampling rate, we used the entire encoding and recall epochs for the most efficient estimate of PTE, rather than estimating PTE in a time-resolved manner. Please note that this is consistent with previous literature which have used ~ 225000 time-points (3 minutes of resting-state data with 1250 Hz sampling rate) for estimating PTE, for example, see (Hillebrand et al., 2016). 

This limitation prevents us from examining how directed connectivity evolves throughout the encoding and retrieval intervals on a moment-to-moment basis. Future studies employing longer task epochs or alternative methods for time-resolved connectivity analysis could provide valuable insights into the dynamic engagement and disengagement of the AI with other brain areas based on task demands. Such analyses could potentially reveal task-specific temporal patterns in the AI's influence on DMN and FPN nodes during different phases of memory processing.

Finally, it is crucial to note that in the three verbal tasks, our analysis of memory recall is timelocked to word production onset. However, the precise timing of the internal retrieval process initiation is unknown. This limitation may affect our ability to capture the full dynamics of network interactions during recall, particularly in the early stages of memory retrieval. Interestingly, in the spatial memory task, where this timing issue is less problematic due to the nature of the task, we observe that the PCC/precuneus shows an earlier onset of activity compared to the AI. This process is aligned with the view that DMN nodes may trigger internal recall processes, the full extent and replication of which across verbal and spatial tasks could not be examined in this study.  

We have added a discussion of these limitations and future directions to the manuscript to provide a more nuanced interpretation of our findings and to highlight important areas for further investigation (page 19). 

(2.3) Second, the authors state that high-gamma suppression in the PCC/precuneus relative to the AI is an anatomically specific signature that is not present in the FPN. This claim does not seem to be supported by their own evidence as presented in the Supplemental Data (Figures S2 and S3), which to my eye show clear evidence of relative suppression in the MFG and dPPC (e.g. S2a and S3a, most notably) which are notated as "significant" with green bars. I appreciate that the magnitude of this effect may be greater in the PCC/precuneus, but if this is the claim it should be supported by appropriate statistics and interpretation.  

We thank the reviewer for raising this point. We have now directly compared the high-gamma power of the PCC/precuneus with the dPPC and MFG nodes of the FPN and we note that the suppression effects of the PCC/precuneus are stronger compared to those of the dPPC and MFG during memory encoding (Figures S8, S9). 

(2.4) I commend the authors on emphasizing replicability, but I found their Bayes Factor (BF) analysis to be difficult to interpret and qualitatively inconsistent with the results that they show. For example, the authors state that BF analysis demonstrates "high replicability" of the gamma suppression effect in Figure 3a with that of 3c and 3d. While it does appear that significant effects exist across all three tasks, the temporal structure of high gamma signals appears markedly different between the two in ways that may be biologically meaningful. Moreover, it appears that the BF analysis did not support replicability between VFR and CATVFR, which is very surprising; these are essentially the same tasks (merely differing in the presence of word categories) and would be expected to have the highest degree of concordance, not the lowest. I would suggest the authors try to analytically or conceptually reconcile this surprising finding. 

We appreciate the reviewer's commendation on our emphasis on replicability and thank the reviewer for the opportunity to provide clarification.

First, we would like to clarify the nature of our BF analysis. Bayes factors are calculated as the ratio of the marginal likelihood of the replication data, given the posterior distribution estimated from the original data, and the marginal likelihood for the replication data under the null hypothesis of no effect (Ly, Etz, Marsman, & Wagenmakers, 2019). Specifically, BFs use the posterior distribution from the first experiment as a prior distribution for the replication test of the second experiment to constitute a joint multivariate distribution (i.e., the additional evidence for the alternative hypothesis given what was already observed in the original study) and this joint distribution is dependent on the similarity between the two experiments (Ly et al., 2019).  This analysis revealed that PCC/precuneus suppression, in comparison to the AI during memory encoding, observed in the VFR during memory encoding was detected in two other tasks, PALVCR, and WMSM with high BFs. In the CATVFR task, although there were short time periods of PCC/precuneus suppression (Figure 3), the effects were not strong enough like the three other tasks.  

Regarding the high-gamma suppression effect, our BF analysis indeed supports replicability across the VFR, PALVCR, and WMSM tasks. While we agree with the reviewer that the temporal structure of high-gamma signals appears different across tasks, the BF analysis focuses on the overall presence of the suppression effect rather than its precise temporal profile. The high BFs indicate that the core finding - PCC/precuneus suppression relative to the AI during memory encoding - is replicated across these tasks, despite differences in the timing of this suppression. Moreover, at no time point did responses in the PCC/precuneus exceed that of the AI in any of the four memory encoding tasks. 

The reason for differences in temporal profiles is not clear. While VFR and CATVFR are similar, the addition of categorical structure in CATVFR may have introduced cognitive processes that alter the temporal dynamics of regional responses. Moreover, differences in electrode placements across participants in each experiment may also have contributed to variability in the observed effects. Further studies using within-subjects experimental designs are needed to address this. 

We have updated our Results and Discussion sections to reflect these points and to provide a more nuanced interpretation of the replicability across tasks.  

(2.5) To aid in interpretability, it would be extremely helpful for the authors to assess acrosstask similarity in high-gamma power on a within-subject basis, which they are wellpowered to do. For example, could they report the correlation coefficient between HGP timecourses in paired-associates versus free-recall tasks, to better establish whether these effects are consistent on a within-subject basis? This idea could similarly be extended to the PTE analysis. Across-subject correlations would also be a welcome analysis that may provide readers with better-contextualized effect sizes than the output of a Bayes Factor analysis.  

We thank the reviewer for this suggestion. However, a within-subject analysis was not possible because very few participants participated in multiple memory tasks. 

For example, for the AI-PCC/Pr analysis, only 1 individual participated in both the VFR and PALVCR tasks (Tables S2a, S2c). Similarly, for AI-mPFC analysis, only 3 subjects participated in both the VFR and PALVCR tasks (Tables S2a, S2c).  

Due to these small sample sizes, it was not feasible for us to assess replicability across tasks on a within-subject basis in our dataset. Therefore, for all our analysis, we have pooled electrode pairs across subjects and then subjected these to a linear mixed effects modeling framework for assessing significance and then subsequently assessing replicability of these effects using the Bayes factor (BF) framework.    

Recommendations For The Authors: 

(2.6) I would emphasize manuscript organization in a potential rewrite; it was very difficult to follow which analyses were attempting to show a contrast between effects versus a similarity between effects. Results were grouped by the underlying experimental conditions (e.g. encoding/recall, network identity, etc.) but may be better grouped according to the actual effects that were found. 

We thank the reviewer for this suggestion. We considered this possibility, but we feel that the Results section is best organized in the order of the hypotheses we set out to test, starting from analyzing local brain activity using high-gamma power analysis, and then results related to analyzing brain connectivity using PTE. All these results are systematically ordered by presenting results related to encoding first and then the recall periods as they appear sequentially in our task-design, presenting the results related to the VFR task first and then demonstrating replicability of the results in the three other experiments. Results are furthermore arranged by nodes, where we first discuss results related to the DMN nodes, and then the same for the FPN nodes. This is to ensure systematic, unbiased organization of all our results for the readers to clearly follow the hypotheses, statistical analyses, and the brain regions considered. Therefore, for transparency and ethical reasons, we would respectfully like to present our results as they appear in our current manuscript, rather than presenting the results based on effect sizes. 

However, please note that we indeed have ordered our results in the Discussion section based on actual effects, as suggested by the reviewer.  

(2.7) The absence of a PTE effect when analyzing through the lens of successful vs. unsuccessful memory is an important limitation of the current study and a significant departure from the wealth of subsequent memory effects reported in the literature (which the authors have already done a good job citing, e.g. Burke et al. 2014 Neuroimage). I'm glad that the authors raised this in their Discussion, but it is important that the results of such an analysis actually be shown in the manuscript. 

We thank the reviewer for this suggestion. We have now included the results related to PTE dynamics for successful vs. unsuccessful memory trials in the revised Results section as we note on page 12: 

“Differential information flow from the AI to the DMN and FPN for successfully recalled and forgotten memory trials 

We examined memory effects by comparing PTE between successfully recalled and forgotten memory trials. However, this analysis did not reveal differences in directed influence from the AI on the DMN and FPN or the reverse, between successfully recalled and forgotten memory trials during the encoding as well as recall periods in any of the memory experiments (all ps>0.05).”

(2.8) I believe the claims of causality through the use of the PTE are overstated throughout the manuscript and may contribute to further confusion in the literature regarding how causality in the brain can actually be understood. See Mehler and Kording, 2018 arXiv for an excellent discussion on the topic (https://arxiv.org/abs/1812.03363). My recommendation would be to significantly tone down claims that PTE reflects causal interactions in the brain. 

We thank the reviewer for this suggestion. We would like to clarify here that we define causality in our manuscript as follows: a brain region has a causal influence on a target if knowing the past history of temporal signals in both regions improves the ability to predict the target's signal in comparison to knowing only the target's past, as defined in earlier studies (Granger, 1969; Lobier et al., 2014). We have now included this clarification in the Introduction section (page 6).  

We also agree with the reviewer that to more mechanistically establish a causal link between the neural dynamics and behavior, lesion or brain stimulation studies are necessary. We have now acknowledged this in the revised Discussion as we note: “Although our computational methods suggest causal influences, direct causal manipulations, such as targeted brain stimulation during memory tasks, are needed to establish definitive causal relationships between network nodes.” (page 19). 

Finally, we have now significantly toned down our claims that PTE reflects causal interactions in the brain, in the Introduction, Results, and Discussion sections of our revised manuscript.  

(2.9) Relatedly, it may be useful for the authors to consider a supplemental analysis that uses classic measures of inter-regional synchronization, e.g. the PLV, and compare to their PTE findings. They cite literature to suggest a metric like the PTE may be useful, but this hardly rules out the potential utility of investigating narrowband phase synchronization. 

We thank the reviewer for this suggestion. We have now run new analyses based on PLV to examine phase synchronization between the AI and the DMN and FPN. However, we did not find a significant PLV for the interactions between the AI and DMN and FPN nodes for the different task periods compared to the resting baselines, as we note on page 13: 

“Narrowband phase synchronization between the AI and the DMN and FPN during encoding and recall compared to resting baseline  

We next directly compared the phase locking values (PLVs) (see Methods for details) between the AI and the PCC/precuneus and mPFC nodes of the DMN and also the dPPC and MFG nodes of the FPN for the encoding and the recall periods compared to resting baseline. However, narrowband PLV values did not significantly differ between the encoding/recall vs. rest periods, in any of the delta-theta (0.5-8 Hz), alpha (8-12 Hz), beta (12-30 Hz), gamma (30-80 Hz), and high-gamma (80-160 Hz) frequency bands. These results indicate that PTE, rather than phase synchronization, more robustly captures the AI dynamic interactions with the DMN and the FPN.” 

Please note that phase locking measures such as the PLV or coherence do not probe directed causal influences and cannot address how one region drives another. Instead, our study examined the direction of information flow between the AI and the DMN and FPN using robust estimators of the direction of information flow. PTE assesses with the ability of one time-series to predict future values of other time-series, thus estimating the time-delayed causal influences between the two time-series, whereas PLV or coherence can only estimate “instantaneous” phase synchronization, but not predict the future time-series. 

Additionally, please note that the directed information flow from the AI to both the DMN and FPN were enhanced during the encoding and recall periods compared to resting state across all four experiments, in a new set of analyses that we have carried out in our revised manuscript. Specifically, we have now carried out our task versus rest comparison by using resting baseline epochs before the start of the entire session of the task periods, rather than our previously used rest epochs which were in between the task periods. These new results have now been included in the revised Results as we note on page 12:  

“Enhanced information flow from the AI to the DMN and FPN during episodic memory processing, compared to resting-state baseline

We next examined whether directed information flow from the AI to the DMN and FPN nodes during the memory tasks differed from the resting-state baseline. Resting-state baselines were extracted immediately before the start of the task sessions and the duration of task and rest epochs were matched to ensure that differences in network dynamics could not be explained by differences in duration of the epochs. Directed information flow from the AI to both the DMN and FPN were higher during both the memory encoding and recall phases and across the four experiments, compared to baseline in all but two cases (Figures S6, S7). These findings provide strong evidence for enhanced role of AI directed information flow to the DMN and FPN during memory processing compared to the resting state.”  

References 

Backus, A. R., Schoffelen, J. M., Szebényi, S., Hanslmayr, S., & Doeller, C. F. (2016). Hippocampal-Prefrontal Theta Oscillations Support Memory Integration. Curr Biol, 26(4), 450-457. doi:10.1016/j.cub.2015.12.048

Bastos, A. M., Vezoli, J., Bosman, C. A., Schoffelen, J. M., Oostenveld, R., Dowdall, J. R., . . . Fries, P. (2015). Visual areas exert feedforward and feedback influences through distinct frequency channels. Neuron, 85(2), 390-401. doi:10.1016/j.neuron.2014.12.018

Canolty, R. T., Edwards, E., Dalal, S. S., Soltani, M., Nagarajan, S. S., Kirsch, H. E., . . . Knight, R. T. (2006). High gamma power is phase-locked to theta oscillations in human neocortex. Science, 313(5793), 1626-1628. doi:10.1126/science.1128115

Canolty, R. T., & Knight, R. T. (2010). The functional role of cross-frequency coupling. Trends Cogn Sci, 14(11), 506-515. doi:10.1016/j.tics.2010.09.001

Chen, L., Wassermann, D., Abrams, D. A., Kochalka, J., Gallardo-Diez, G., & Menon, V. (2019). The visual word form area (VWFA) is part of both language and attention circuitry. Nat Commun, 10(1), 5601. doi:10.1038/s41467-019-13634-z

Clouter, A., Shapiro, K. L., & Hanslmayr, S. (2017). Theta Phase Synchronization Is the Glue that Binds Human Associative Memory. Curr Biol, 27(20), 3143-3148.e3146. doi:10.1016/j.cub.2017.09.001

Crone, N. E., Boatman, D., Gordon, B., & Hao, L. (2001). Induced electrocorticographic gamma activity during auditory perception. Brazier Award-winning article, 2001. Clin Neurophysiol, 112(4), 565-582. doi:10.1016/s1388-2457(00)00545-9

Daitch, A. L., & Parvizi, J. (2018). Spatial and temporal heterogeneity of neural responses in human posteromedial cortex. Proc Natl Acad Sci U S A, 115(18), 4785-4790. doi:10.1073/pnas.1721714115

Das, A., de Los Angeles, C., & Menon, V. (2022). Electrophysiological foundations of the human default-mode network revealed by intracranial-EEG recordings during restingstate and cognition. Neuroimage, 250, 118927. doi:10.1016/j.neuroimage.2022.118927

Das, A., & Menon, V. (2020). Spatiotemporal Integrity and Spontaneous Nonlinear Dynamic Properties of the Salience Network Revealed by Human Intracranial Electrophysiology: A Multicohort Replication. Cereb Cortex, 30(10), 5309-5321. doi:10.1093/cercor/bhaa111

Das, A., & Menon, V. (2021). Asymmetric Frequency-Specific Feedforward and Feedback Information Flow between Hippocampus and Prefrontal Cortex during Verbal Memory Encoding and Recall. J Neurosci, 41(40), 8427-8440. doi:10.1523/jneurosci.080221.2021

Das, A., & Menon, V. (2022). Replicable patterns of causal information flow between hippocampus and prefrontal cortex during spatial navigation and spatial-verbal memory formation. Cereb Cortex, 32(23), 5343-5361. doi:10.1093/cercor/bhac018

Das, A., & Menon, V. (2023). Concurrent- and after-effects of medial temporal lobe stimulation on directed information flow to and from prefrontal and parietal cortices during memory formation. J Neurosci, 43(17), 3159-3175. doi:10.1523/jneurosci.1728-22.2023

Edwards, E., Soltani, M., Deouell, L. Y., Berger, M. S., & Knight, R. T. (2005). High gamma activity in response to deviant auditory stimuli recorded directly from human cortex. J Neurophysiol, 94(6), 4269-4280. doi:10.1152/jn.00324.2005

Ekstrom, A. D., Caplan, J. B., Ho, E., Shattuck, K., Fried, I., & Kahana, M. J. (2005). Human hippocampal theta activity during virtual navigation. Hippocampus, 15(7), 881-889. doi:10.1002/hipo.20109

Ekstrom, A. D., & Watrous, A. J. (2014). Multifaceted roles for low-frequency oscillations in bottom-up and top-down processing during navigation and memory. Neuroimage, 85 Pt 2, 667-677. doi:10.1016/j.neuroimage.2013.06.049

Engel, A. K., & Fries, P. (2010). Beta-band oscillations--signalling the status quo? Curr Opin Neurobiol, 20(2), 156-165. doi:10.1016/j.conb.2010.02.015

Foster, B. L., Rangarajan, V., Shirer, W. R., & Parvizi, J. (2015). Intrinsic and task-dependent coupling of neuronal population activity in human parietal cortex. Neuron, 86(2), 578590. doi:10.1016/j.neuron.2015.03.018

Gonzalez, A., Hutchinson, J. B., Uncapher, M. R., Chen, J., LaRocque, K. F., Foster, B. L., . . . Wagner, A. D. (2015). Electrocorticography reveals the temporal dynamics of posterior parietal cortical activity during recognition memory decisions. Proc Natl Acad Sci U S A, 112(35), 11066-11071. doi:10.1073/pnas.1510749112

Granger, C. W. J. (1969). Investigating Causal Relations by Econometric Models and Crossspectral Methods. Econometrica, 37(3), 424-438. doi:10.2307/1912791

Griffiths, B. J., Martín-Buro, M. C., Staresina, B. P., & Hanslmayr, S. (2021). Disentangling neocortical alpha/beta and hippocampal theta/gamma oscillations in human episodic memory formation. Neuroimage, 242, 118454. doi:10.1016/j.neuroimage.2021.118454

Guderian, S., & Düzel, E. (2005). Induced theta oscillations mediate large-scale synchrony with mediotemporal areas during recollection in humans. Hippocampus, 15(7), 901-912. doi:10.1002/hipo.20125

Guderian, S., Schott, B. H., Richardson-Klavehn, A., & Düzel, E. (2009). Medial temporal theta state before an event predicts episodic encoding success in humans. Proc Natl Acad Sci U S A, 106(13), 5365-5370. doi:10.1073/pnas.0900289106

Helfrich, R. F., & Knight, R. T. (2016). Oscillatory Dynamics of Prefrontal Cognitive Control. Trends Cogn Sci, 20(12), 916-930. doi:10.1016/j.tics.2016.09.007

Hillebrand, A., Tewarie, P., van Dellen, E., Yu, M., Carbo, E. W., Douw, L., . . . Stam, C. J. (2016). Direction of information flow in large-scale resting-state networks is frequencydependent. Proc Natl Acad Sci U S A, 113(14), 3867-3872. doi:10.1073/pnas.1515657113

Jeffreys, H. (1998). The Theory of Probability (3rd ed.). Oxford, England: Oxford University Press.

Kwon, H., Kronemer, S. I., Christison-Lagay, K. L., Khalaf, A., Li, J., Ding, J. Z., . . . Blumenfeld, H. (2021). Early cortical signals in visual stimulus detection. Neuroimage, 244, 118608. doi:10.1016/j.neuroimage.2021.118608

Lachaux, J. P., George, N., Tallon-Baudry, C., Martinerie, J., Hugueville, L., Minotti, L., . . . Renault, B. (2005). The many faces of the gamma band response to complex visual stimuli. Neuroimage, 25(2), 491-501. doi:10.1016/j.neuroimage.2004.11.052

Lobier, M., Siebenhühner, F., Palva, S., & Matias, P. J. (2014). Phase transfer entropy: A novel phase-based measure for directed connectivity in networks coupled by oscillatory interactions. Neuroimage, 85, 853-872. doi:10.1016/j.neuroimage.2013.08.056

Ly, A., Etz, A., Marsman, M., & Wagenmakers, E. J. (2019). Replication Bayes factors from evidence updating. Behav Res Methods, 51(6), 2498-2508. doi:10.3758/s13428-0181092-x

Mainy, N., Kahane, P., Minotti, L., Hoffmann, D., Bertrand, O., & Lachaux, J. P. (2007). Neural correlates of consolidation in working memory. Hum Brain Mapp, 28(3), 183-193. doi:10.1002/hbm.20264

Miller, K. J., Leuthardt, E. C., Schalk, G., Rao, R. P., Anderson, N. R., Moran, D. W., . . . Ojemann, J. G. (2007). Spectral changes in cortical surface potentials during motor movement. J Neurosci, 27(9), 2424-2432. doi:10.1523/jneurosci.3886-06.2007

Miller, K. J., Weaver, K. E., & Ojemann, J. G. (2009). Direct electrophysiological measurement of human default network areas. Proceedings of the National Academy of Sciences, 106(29), 12174-12177. doi:10.1073/pnas.0902071106

Nuzzo, R. L. (2017). An Introduction to Bayesian Data Analysis for Correlations. Pm r, 9(12), 1278-1282. doi:10.1016/j.pmrj.2017.11.003

Ray, S., Crone, N. E., Niebur, E., Franaszczuk, P. J., & Hsiao, S. S. (2008). Neural correlates of high-gamma oscillations (60-200 Hz) in macaque local field potentials and their potential implications in electrocorticography. J Neurosci, 28(45), 11526-11536. doi:10.1523/jneurosci.2848-08.2008

Sederberg, P. B., Schulze-Bonhage, A., Madsen, J. R., Bromfield, E. B., Litt, B., Brandt, A., & Kahana, M. J. (2007). Gamma oscillations distinguish true from false memories. Psychol Sci, 18(11), 927-932. doi:10.1111/j.1467-9280.2007.02003.x

Tallon-Baudry, C., Bertrand, O., Hénaff, M. A., Isnard, J., & Fischer, C. (2005). Attention modulates gamma-band oscillations differently in the human lateral occipital cortex and fusiform gyrus. Cereb Cortex, 15(5), 654-662. doi:10.1093/cercor/bhh167

Watrous, A. J., Tandon, N., Conner, C. R., Pieters, T., & Ekstrom, A. D. (2013). Frequencyspecific network connectivity increases underlie accurate spatiotemporal memory retrieval. Nature Neuroscience, 16(3), 349-356. doi:10.1038/nn.3315

Wetzels, R., & Wagenmakers, E. J. (2012). A default Bayesian hypothesis test for correlations and partial correlations. Psychon Bull Rev, 19(6), 1057-1064. doi:10.3758/s13423-0120295-x

Decayed hole=broken city. Empty chapel=decrease in people's faith as their hope as also faded away.

Reviewer #2 (Public review):

Summary:

In this paper, the authors investigate the role of NMDA-receptors in recurrent processing. In doing so, the authors present data from two studies, where they attempt to decode different stimulus features, namely contrast, collinearity, and illusory contours. The latter of which the authors claim relies uniquely on recurrent processing. Therefore, to test whether NMDA receptors are particularly involved in recurrent processing they administer a NMDA-antagonist to see whether the decoding of illusory contours is specifically perturbed, and leaves the decoding of other features intact. They further aim to disentangle the role of NMDA-receptors by manipulating visibility and task relevance of the decoded features

In the first experiment, the authors decode two targets, the first was always presented clearly, the second's visibility was manipulated by presenting it after a short lag rather than a long lag (inducing attentional blink), as well as masking the target on half the trials. First, they find for target 1 clear evidence for the NMDA-receptor increasing (rather than decreasing) decoding performance of illusory contours. They move on to analyse target 2 to explore the manipulations of lag and masking. Here they find that masking reduced decoding of all three stimulus features, but only the lag reduced decoding of illusory contours. Importantly, the NMDA-antagonist improved decoding only in the unmasked, long lag condition, in the cluster analyses. However, the interaction with the lag condition was not significant, and the effect on decoding was primarily present in the later decoding time window, and not significant when exploring the peak of the decoding time window.

The second experiment was highly similar, but got rid of the lag manipulation, and replaced it with a manipulation of task relevance. Notably, masking did not abolish the decoding of illusory contours completely, in contrast to the first experiment. More importantly, they find that the NMDA-receptor now clearly increases decoding of illusory contours, particularly when the illusory contours are not masked. No effect of task relevance is found.

Taken together the authors state that evidence is found for NMDA-receptors role in recurrent processing.

Strengths:

This is an interesting study using state-of-the-art methods in combination with drug manipulation to study recurrent processing. Their analysis methods are state-of-the-art, and the question that they are trying to address is topical and interesting to a wide research audience, encompassing both researchers interested in visual perception and consciousness, as well as those interested in perturbed vision as found in psychiatric disorders.

Weaknesses:

The experimental design is somewhat complicated, which can make it difficult to match the authors' claims to the actual evidence that is provided. I have some reservations about the paper which are born out of a few issues.<br /> (1) The title, abstract, and introduction hide their counterintuitive finding of increased decoding, presumably as it was unexpected.<br /> (2) Their analysis choices are sometimes unclear, making it difficult to assess whether the analyses are sensible.<br /> (3) The appropriate tests for the interactions that the authors claim they found are often lacking.

To start off, I think the reader is being a bit tricked when reading the paper. Perhaps my priors are too strong, but I assumed, just like the authors, that NMDA-receptors would disrupt recurrent processing, in line with previous work. However, due to the continuous use of the ambiguous word 'affected' rather than the more clear increased or perturbed recurrent processing, the reader is left guessing what is actually found. That's until they read the results and discussion finding that decoding is actually improved. This seems like a really big deal, and I strongly urge the authors to reword their title, abstract, and introduction to make clear they hypothesized a disruption in decoding in the illusion condition, but found the opposite, namely an increase in decoding. I want to encourage the authors that this is still a fascinating finding.

Apologies if I have missed it, but it is not clear to me whether participants were given the drug or placebo during the localiser task. If they are given the drug this makes me question the logic of their analysis approach. How can one study the presence of a process, if their very means of detecting that process (the localiser) was disrupted in the first place? If participants were not given a drug during the localiser task, please make that clear. I'll proceed with the rest of my comments assuming the latter is the case. But if the former, please note that I am not sure how to interpret their findings in this paper.

The main purpose of the paper is to study recurrent processing. The extent to which this study achieves this aim is completely dependent to what extent we can interpret decoding of illusory contours as uniquely capturing recurrent processing. While I am sure illusory contours rely on recurrent processing, it does not follow that decoding of illusory contours capture recurrent processing alone. Indeed, if the drug selectively manipulates recurrent processing, it's not obvious to me why the authors find the interaction with masking in experiment 2. Recurrent processing seems to still be happening in the masked condition, but is not affected by the NMDA-receptor here, so where does that leave us in interpreting the role of NMDA-receptors in recurrent processing? If the authors can not strengthen the claim that the effects are completely driven by affecting recurrent processing, I suggest that the paper will shift its focus to making claims about the encoding of illusory contours, rather than making primary claims about recurrent processing.

An additional claim is being made with regards to the effects of the drug manipulation. The authors state that this effect is only present when the stimulus is 1) consciously accessed, and 2) attended. The evidence for claim 1 is not supported by experiment 1, as the masking manipulation did not interact in the cluster-analyses, and the analyses focussing on the peak of the timing window do not show a significant effect either. There is evidence for this claim coming from experiment 2 as masking interacts with the drug condition. Evidence for the second claim (about task relevance) is not presented, as there is no interaction with the task condition. A classical error seems to be made here, where interactions are not properly tested. Instead, the presence of a significant effect in one condition but not the other is taken as sufficient evidence for an interaction, which is not appropriate. I therefore urge the authors to dampen the claim about the importance of attending to the decoded features. Alternatively, I suggest the authors run their interactions of interest on the time-courses and conduct the appropriate cluster-based analyses.

How were the length of the peak-timing windows established in Figure 1E? My understanding is that this forms the training-time window for the further decoding analyses, so it is important to justify why they have different lengths, and how they are determined. The same goes for the peak AUC time windows for the interaction analyses. A number of claims in the paper rely on the interactions found in these post-hoc analyses, so the 223- to 323 time window needs justification.

Author response:

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

Reviewer #1 (Public Review):

Summary:

Deletion of the hrp2 and hrp3 loci in P. falciparum poses an immediate public health threat. This manuscript provides a more complete understanding of the dynamic nature with which these deletions are generated. By delving into the likely mechanisms behind their generation, the authors also provide interesting insight into general Plasmodium biology that can inform our broader understanding of the parasite's genomic evolution.

Strengths:

The sub-telomeric regions of P. falciparum (where hrp2 and hrp3 are located) are notoriously difficult to study with short-read sequence data. The authors take an appropriate, targeted approach toward studying the loci of interest, which includes read-depth analysis and local haplotype reconstruction. They additionally use both long-read and short-read data to validate their major findings. There is an extensive set of supplementary plots, which helps clarify several aspects of the data.

Weaknesses:

In this first version, there are a few factors that hinder a full assessment of the robustness and replicability of the results.

Reviewer #1 (Recommendations For The Authors):

Reviewer comment: First, a number of the analyses lack basic details in the methods; for instance, one must visit the authors' personal website to find some of the tools used.

We have extensively updated the methods to clarify which tools were used and how they were run. All code and results for the analyses have been deposited in Zenodo at https://doi.org/10.5281/zenodo.12167687.

Reviewer comment: Second, there are several tricky methodological points that are not fully documented. Read depths are treated (and plotted) discretely as 0/1/2 without any discussion of how thresholds were used and determined.

We have added to the methods section the full details on how read depth was handled, including rounding to the closest 1 normalized coverage for visualizations. To ensure analysis of only highly confident deleted strains, normalized coverage of 0.1 or more was round to 1 instead of 0. Samples were considered for potential genomic deletion if they had zero coverage after rounding from chromosome 8 1,375,557 to 1,387,982 for pfhrp2, chromosome 13 from 2,841,776 to 2,844,785 for pfhrp3, and from chromosome 11 1,991,347 to 2,003,328. These numbers were chosen after visual inspection of samples with any zero coverage within the genomic region of pfhrp2/3.

Reviewer comment: For read mapping to standard vs hybrid chromosomes, there is no documentation on how assignments were made if partially ambiguous or how final sample calls were determined when some reads were discordant. There is no mention of how missing data were handled. Without this, it is difficult to know when conclusions were based on analyses that were more quantitative (for instance, using pre-determined read thresholds) or more subjective (with patterns being extracted visually).

We have updated several parts of the methods section to explicitly state what thresholds and analysis pipelines to use, making our documentation clearer. For mapping to the hybrid vs standard chromosomes for the long reads, spanning reads across the duplicated region were required to extend 50bp upstream and downstream of the region. These regions are significantly different between chromosomes 11 and 13, so requiring spanning reads to map to these regions prevented multi-mapping reads. Reads that started within the duplicated region were allowed to map to both the hybrid and standard chromosomes for visualization in Figure 4. Importantly, for both HB3 and SD01, no reads spanned from the duplicated region into chromosome 13, showing a complete lack of reads that contained the portion of chromosome 13 that came after the duplicated region. None of the other isolates had any spanning reads across the hybrid chromosomes. Details on deletion calls were based on initial visualization of pfhrp2/3 and then on read thresholds (see above response for details).

Reviewer comment: Third, while a new method is employed for local haplotype reconstruction (PathWeaver), the manuscript does not include details on this approach or benchmarking data with which to evaluate its performance and understand any potential artifacts.

We have added an analysis based on biallelic SNPs to compare to the PathWeaver results, which produced similar results to help validate the PathWeaver results. PathWeaver manuscript is in preparation.

Reviewer #2 (Public Review):

This work investigates the mechanisms, patterns, and geographical distribution of pfhrp2 and pfhrp3 deletions in Plasmodium falciparum. Rapid diagnostic tests (RDTs) detect P. falciparum histidine-rich protein 2 (PfHRP2) and its paralog PfHRP3 located in subtelomeric regions. However, laboratory and field isolates with deletions of pfhrp2 and pfhrp3 that can escape diagnosis by RDTs are spreading in some regions of Africa. They find that pfhrp2 deletions are less common and likely occur through chromosomal breakage with subsequent telomeric healing. Pfhrp3 deletions are more common and show three distinct patterns: loss of chromosome 13 from pfhrp3 to the telomere with evidence of telomere healing at breakpoint (Asia; Pattern 13-); duplication of a chromosome 5 segment containing pfhrp1 on chromosome 13 through non-allelic homologous recombination (NAHR) (Asia; Pattern 13-5++); and the most common pattern, duplication of a chromosome 11 segment on chromosome 13 through NAHR (Americas/Africa; Pattern 13-11++). The loss of these genes impacts the sensitivity of RDTs, and knowing these patterns and geographic distribution makes it possible to make better decisions for malaria control.

Reviewer #3 (Public Review):

Summary:

The study provides a detailed analysis of the chromosomal rearrangements related to the deletions of histidine-rich protein 2 (pfhrp2) and pfhrp3 genes in P. falciparum that have clinical significance since malaria rapid diagnostic tests detect these parasite proteins. A large number of publicly available short sequence reads for the whole genome of the parasite were analyzed, and data on coverage and discordant mapping allowed the authors to identify deletions, duplications, and chromosomal rearrangements related to pfhrp3 deletions. Long-read sequences showed support for the presence of a normal chromosome 11 and a hybrid 13-11 chromosome lacking pfhrp3 in some of the pfhrp3-deleted parasites. The findings support that these translocations have repeatedly occurred in natural populations. The authors discuss the implications of these findings and how they do or do not support previous hypotheses on the emergence of these deletions and the possible selective pressures involved.

Strengths:

The genomic regions where these genes are located are challenging to study since they are highly repetitive and paralogous and the use of long-read sequencing allowed to span the duplicated regions, giving support to the identification of the hybrid 13-11 chromosome.

All publicly available whole-genome sequences of the malaria parasite from around the world were analysed which allowed an overview of the worldwide variability, even though this analysis is biased by the availability of sequences, as the authors recognize.

Despite the reduced sample size, the detailed analysis of haplotypes and identification of the location of breakpoints gives support to a single origin event for the 13-5++ parasites.

The analysis of haplotype variation across the duplicated chromosome-11 segment identified breakpoints at varied locations that support multiple translocation events in natural populations. The authors suggest these translocations may be occurring at high frequency in meiosis in natural populations but are strongly selected against in most circumstances, which remains to be tested.

Weaknesses:

Reviewer comment: Relying on sequence data publicly available, that were collected based on diagnostic test positivity and that are limited by sequencing availability, limits the interpretation of the occurrence and relative frequency of the deletions.

However, we have uncovered more mechanisms than previously detected for hrp2 (involving MDR1) in SEA and South American parasites are likely detected by microscopy as RDTs were never introduced due to the presence of the deletions.

Reviewer comment: In the discussion, caution is needed when identifying the least common and most common mechanisms and their geographical associations. The identification of only one type of deletion pattern for Pfhrp2 may be related to these biases.

We added a section in the Discussion on the limitations of our study, which states the following, “Limitations of this study include the use of publicly available sequencing data that were collected often based on positive rapid diagnostic tests, which limits our interpretation of the occurrence and relative frequency of these deletions. This could introduce regional biases due to different diagnostic methods as well as limit the full range of deletion mechanisms, particularly pfhrp2.”

Reviewer comment: The specific objectives of the study are not stated clearly, and it is sometimes difficult to know which findings are new to this study. Is it the first study analyzing all the worldwide available sequences? Is it the first one to do long-read sequencing to span the entire duplicated region?

In the Introduction, we added, “The objectives of this study were to determine the pfhrp3 deletion patterns along with their geographical associations and sequence and assemble the chromosomes containing the deletions using long-read sequencing.”

We also added in the Discussion, “To the best of our knowledge, no prior studies have performed long-read sequencing to definitively span and assemble the entire segmental duplication involved in the deletions.”

Reviewer comment: Another aspect that should be explained in the introduction is that there was previous information about the association of the deletions to patterns found in chromosomes 5 and 11. In the short-read sequences results, it is not clear if these chromosomes were analysed because of the associations found in this study (and no associations were found to putative duplications or deletions in other chromosomes), or if they were specifically included in the analysis because of the previous information (and the other chromosomes were not analysed).

The former is correct. Chromosomes 5 and 11 were analyzed due to the associations found in this study, not from prior information. We have added the following sentence in the Results: “As a result of our short-read analysis demonstrating these three patterns and discordant reads between the chromosomes involved, chromosomes 5, 11, and 13 were further examined. No other chromosomes had associated discordant reads or changes in read coverage. ”

Reviewer comment: An interesting statement in the discussion is that existing pfhrp3 deletions in a low-transmission environment may provide a genetic background on which less frequent pfhrp2 deletion events can occur. Does it mean that the occurrence of pfhrp3 deletions would favor the pfhrp2 deletion events? How, and is there any evidence for that?

We should have stated more explicitly that selection would better be able to act on the now doubly deleted parasite versus a parasite with HRP3 still intact and weakly detectable by RDTs.Since fully RDT-negative parasites require a two-hit mechanism, where both pfhrp2 and pfhrp3 need to be deleted, and since there appear to be more mechanisms and drivers for pfhrp3 deletions, this would create a population of parasites with one hit already and would only require the additional hit of pfhrp2 deletion to occur to become RDT negative. So the point in the discussion being made is not that the pfhrp3 deletion would favor pfhrp2 deletion but rather that there is a population circulating with one hit already, which would make it more likely that the less frequent pfhrp2 deletion would result in a dual deleted parasite and therefore an RDT-negative parasite. The discussion has been modified to the following to try to make this point more clear. “In the setting of RDT use in a low-transmission environment, a pfhrp2 deletion occurring in the context of an existing pfhrp3 deletion may be more strongly selected for compared to pfhrp2 deletion occurring alone still detectable by RDTs.”

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

Reviewer comment: In the text, clonal propagation is the proposed hypothesis for the presence of near-identical copies of the chromosome 11 duplicated region. Even among the parasites showing variation between chromosomes, Figure 5 shows 3 haplotype groups with multiple sample members, which is also suggestive that these are highly related parasites. In addition to confirming COI status, it would be straightforward to calculate the genome-wide relatedness between/among parasites belonging to the same haplotype group. The assumption is that they are clones or highly related. A different finding would require more thought into potential genomic artifacts driving the pattern.

Thank you for this helpful suggestion. We confirmed the COI of each sample using THE REAL McCOIL. Six samples were not monoclonal, and we removed these samples from the downstream analysis to remove any contribution of polyclonal samples to the downstream haplotype analysis. Then, by using hmmIBD on whole-genome biallelic SNPs, we determined the whole-genome relatedness between the parasites. The haplotype groups do appear clonal though there appear to be several clonal groups within the larger groups of clusters 01 (n=28) and 03 (n=12) which combined with the variation seen within the 15.2kb region on chromosome 11/13, there appears to be different events that then lead to the same duplicated chromosome 11.

Reviewer comment: By way of validating the PathWeaver results, it could be useful to use another comparator method on the samples that are COI=1 or 2.

We have added an analysis based on biallelic SNPs to compare to the PathWeaver results, which produced similar results to help validate the PathWeaver results. We continued to use PathWeaver (Hathaway, in preparation), which is better able to detect variation relative to standard GATK4 analyses due to the refined local alignments from assembled haplotypes.

Questions regarding Methods:

Reviewer comment: Were any metrics of genome quality factored into sample selection?

Yes, samples were removed if there was less than <5x median whole genome coverage. Additionally, several subsets of sWGA samples were removed based on visual inspection. These details have been added to the methods section.

Reviewer comment: How were polyclonal samples treated to ensure they did not produce analysis artifacts?

The read-depth analysis required zero coverage across the regions of pfhrp2/pfhrp3, which made it so that most of the samples analyzed were monoclonal (or polyclonal infections of only deleted strains). We have now used THE REAL McCOIL on whole genome SNPs to determine COIs. Six samples were identified as polyclonal, and we removed them for the analysis and updated the manuscript. Their removal did not significantly impact the results or conclusions.

Reviewer comment: How was local realignment of short-read data performed? Was this step informed by the conserved, non-paralogous genomic regions, or were these only used for downstream variant analysis?

No local realignment of short-read data was performed. The analysis was either read depth or de novo assembly from reads from specific regions. Regarding the de novo assembly, variant calls were replaced by complete local haplotypes, and a region was typed based on the haplotype called for the region.

Reviewer comment: For read-depth estimation, what cutoffs were used to classify windows as deletion, WT, or duplication? How much variability was present in the data? The plot legends imply a continuous scale, but in reality, only 3 discrete colors are used (0, 1, 2), so these must represent the data after rounding.

These have been added to the manuscript. See response to Reviewer #1 questions #2 and #3 above

Reviewer comment: Similarly, what thresholds were used for mapping the long-reads? In Fig S21, it appears there is a high proportion of discordant reads.

Long reads were mapped using minimap2 with default settings. For Figure 21, since it is from the mappings to 3D7 chromosome 11 and hybrid 3D7 13-11 chromosome, the genome from the duplicated region from the blue bar underneath is identical, so reads are expected to map to both since the genome regions are identical. The significance of this figure and Figure 4 is the number of long reads that span the whole chr11/13 duplicated region connection the 3D7 chromosome 11 and the hybrid proving that there are reads that start with chromosome 13 sequence and end with chromosome 11 sequence and the lack of reads that span from chromosome 13 into the 3D7 chromosome 13.

Reviewer comment: The section on the mdr1 breakpoints is too vague.

We have updated the methods section to be more explicit about how these breakpoints were determined.

Reviewer comment: I assume that the "Homologous Genomic Structure" section of the Methods is the number analysis that was alluded to in the Results? As with other sections, this needs more information on exact methods and tools

We have now updated the methods section to include exactly how the nucmer commands were run.

Smaller comments:

Reviewer comment: Introduction sub-header: "Precise *pfhrp2* and..."

We have corrected the sub-header.

Reviewer comment: Results (p.5) cite Table S4 instead of S3

We have corrected this to Table S3.

Reviewer comment: Results (p.5) "We identified 27 parasites with pfhrp2 deletion, 172 with pfhrp3 deletion, and 21 with both pfhrp2 and pfhrp3 deletions." This sentence makes it sound like they are 3 mutually exclusive categories. I'd suggest a rewording like "We identified 27 parasites with pfhrp2 deletion and 172 with pfhrp3 deletion. Of these, 21 contained both deletions."

We have re-worded this sentence to the following: “We identified 26 parasites with pfhrp2 deletion and 168 with pfhrp3 deletion. Twenty field samples contained both deletions; 11 were found in Ethiopia, 6 in Peru, and 3 in Brazil, and all had the 13-11++ pfhrp3 deletion pattern.”

Reviewer comment: The annotations used for the deletions differ between the text and the figures. It would be easier for the reader to harmonize the two if these matched.

The figures have been updated to reflect the annotations of the text.

Reviewer comment: Figure numbering does not match the order they are first referenced in the text

The figure numbers have been updated to match the order in which they are first referenced.

Reviewer comment: Results (p. 8) there is no Table S4

This has been changed to Table S3.

Reviewer comment: Results (p.8) mention a genome-wide number analysis, but I couldn't find these results. The referenced figure is for the duplicated region only.

We have updated to point to the correct location of the nucmer results by adding a supplemental table with the results and updated to point to the correct figure.

Reviewer comment: Discussion typo: "Here, we used publicly available short-read and long-read *short-read sequencing data* from..."

This was not a typo, as we used publicly available PacBio long-read data and then generated new Nanopore long-read data. However, we did clarify this in the sentence.

Reviewer #2 (Recommendations For The Authors):

Introduction

Reviewer comment: "(...) suggesting the genes have important infections in normal infections and their loss is selected against". The word "infections" is in place of "role", etc.

We have changed the word accordingly.

Results

Reviewer comment: In the section "Pfhrp2 and pfhrp3 deletions in the global P. falciparum genomic dataset" it is mentioned the number of parasites with each deletion and where it is more common. "We identified 27 parasites with pfhrp2 deletion, 172 with pfhrp3 deletion, and 21 with both pfhrp2 and pfhrp3 deletions." and "Across all regions, pfhrp3 deletions were more common than pfhrp2 deletions; specifically, pfhrp3 deletions and pfhrp2 deletions were present in Africa in 43 and 12, Asia in 53 and 4, and South America in 76 and 11 parasites." It is not clear where the 21 parasites with both pfhrp2 and pfhrp3 deletions are located.

We have specified the following in the Results section: “We identified 26 parasites with pfhrp2 deletion and 168 with pfhrp3 deletion. Twenty field samples contained both deletions; 11 were found in Ethiopia, 6 in Peru, and 3 in Brazil, and all had the 13-11++ pfhrp3 deletion pattern”

Reviewer comment: "It should be noted that these numbers are not accurate measures of prevalence given that most WGS specimens have been collected based on RDT positivity." This, combined with the fact that subtelomeric regions are difficult to sequence and assembly, means these numbers are underestimated. I believe it should be more stressed in the text.

We have added the following sentence, “Furthermore, subtelomeric regions are difficult to sequence and assemble, meaning these numbers may be significantly underestimated.”

Reviewer comment: In the section "Pattern 13-11++ breakpoint occurs in a segmental duplication of ribosomal genes on chromosomes 11 and 13", Figures 2a and 2b should be mentioned in the text instead of just Figure 2.

We have specified Figures 2a and 2b in the text now.

Figures and Tables:

Reviewer comment: Figure 2: I believe the color scale for percentage of identity is unnecessary given that the goal is to show that the paralogs are highly similar, and not that there is a significant difference between 0.99 and 0.998.

Updated the color scale to represent the number of variants between segments rather than percent identity which ranges between 55-133 so that it represents something more discreet than 0.99 and 0.998.

Reviewer comment: Adjust Figure 2b and the size of supplementary figure legends.Supplementary Figure 5-15: the legends are hard to read.

All legends have been adjusted to be much more readable.

Reviewer #3 (Recommendations For The Authors):

Some minor suggestions:

Reviewer comment: The order of the figures should follow the flow of the text, for example, Figure 5 appears in the text between Figure 1 and Figure 2.

We have reordered the figures according to the order in which they appear in the text.

Reviewer comment: Page 3 - "deleted parasites" - better to use: pfhrp2/3-deleted parasites.

We have edited this accordingly.

Reviewer comment: Define the acronyms the first time they are used, e.g. SEA.

We have defined the acronyms accordingly.

Reviewer comment: In the figures where pfmdr1 appears, indicate the correspondence to the full name of the gene that appears in the legend (multidrug resistance protein 1).

Legends updated.

Reviewer comment: Page 5 - Table S4 is missing.

We apologize for our typo. There is no Table S4. We meant to refer to Table S3, which has been updated accordingly.

Reviewer comment: Page 5 - "We identified 27 parasites with pfhrp2 deletion, 172 with pfhrp3 deletion, and 21 with both pfhrp2 and pfhrp3 deletions" - is it "and 21..." OR "from which, 21..."?

We have reworded the sentence to the following: “We identified 26 parasites with pfhrp2 deletion and 168 with pfhrp3 deletion. Twenty field samples contained both deletions; 11 were found in Ethiopia, 6 in Peru, and 3 in Brazil, and all had the 13-11++ pfhrp3 deletion pattern.”

Reviewer comment: Page 5 - "most WGS specimens have been collected based on RDT positivity." - explain better which tests are done - to detect pfhrp2, pfhrp3 or both?

Co-occurrence is not detected?

We used all publicly available WGS data that spanned over 30 studies, and the exact details of what RDTs were used are not readily available to fully answer this question. Though the exact details of RDTs are not known, this does not affect the deletion patterns found in the genomic data but does limit the ability to comment on how this affects prevalence. We have updated the manuscript to the following to be more explicit that we don’t have the full details: “It should be noted that these numbers are not accurate measures of prevalence, given that the publicly available WGS specimens utilized in this analysis come from locations and time periods that commonly used RDT positivity for collection”

Reviewer comment: Supplementary Figure 1 - Legend for "Pattern" - what is the white?

The “Pattern” refers to pfhrp3 deletion pattern with “white” being no pfhrp3 deletion. The annotation title has been changed to “pfhrp3- Pattern” to make this more clear and added to the text of the legend the following:”Of the 6 parasites without HRP3 deletion (marked as white in pfhrp3- Pattern column for having no pfhrp3 deletion),...”

Reviewer comment: Supplementary Figure 8 - explain the haplotype rank. How was it obtained?

The haplotype rank is based on the prevalence of the haplotype. To clarify this better the following has been added to the caption “Each column contains the haplotypes for that genomic region colored by the haplotype prevalence rank (more prevalent have a lower rank number, with most prevalent having rank 1) at that window/column. Colors are by frequency rank of the haplotypes (most prevalent haplotypes have rank 1 and colored red, 2nd most prevalent haplotypes are rank 2 and colored orange, and so forth. Shared colors between columns do not mean they are the same haplotype. If the column is black, there is no variation at that genomic window.”

Reviewer comment: Figure 1 - Pattern in legend appears 11++13- but in text it is always referenced as 13-11++

Figure legend has been updated to reflect the annotation within the text

Reviewer comment: Page 6 - pattern 13- is which one(s) in Figure 1?

This refers to the 13- with TARE1 sequence detected, the text has been updated to “(pattern 13-TARE1)” and the legend of Figure 1 has been updated so these statements match more closely.

Reviewer comment: Page 7 - states "The 21 parasites with pattern 13-" and refers to Supplementary Figure 3 which presents "50 parasites with deletion pattern 13-". I believe this is pattern 13- unassociated with other rearrangements but it should be made clear in the text and legend of the supplementary figure.

Thank you, you are correct. The manuscript has been updated in two locations for better clarity. The text has been updated to be “The 20 parasites with pattern 13-TARE1 without associated other chromosome rearrangements had deletions of the core genome averaging 19kb (range: 11-31kb). Of these 13-TARE1 deletions, 19 out of 20 had detectable TARE1 (pattern 13-TARE) adjacent to the breakpoint, consistent with telomere healing.” The Supplemental Figure 3 legend has been updated to “for the 48 parasites with pfhrp3 deletions not associated with pattern 13-11++”

Reviewer comment: Supplementary figure 25 - "regions containing the pfhrp genes (lighter blue bars below chromosomes 11 and 13)" - the light blue bars are shown below chromosome 8 and 13; what is the difference between yellow and pink bars (telomere associates repetitive elements in the truncated legend)?

The yellow bars are associated with the telomere-associated repetitive element 3 and the pink bars are telomere-associated repetitive element 1. To add clarity the legend has been updated to be “The yellow (TARE3) and pink (TARE1) bars on the bottom of the chromosomes represent the telomere-associated repetitive elements found at the end of chromosomes.”

Reviewer comment: It would be helpful to have a positioning scale in the figures.

Most plots have y-axis and x-axis with the genomic positioning labeled which can serve as a positioning scale so we opted not to add more to the figures to keep them less crowded. Other plots have regions plotted in genomic order but are all relatively positioned which prevents the usage of a positioning scale, we tried to clarify this by adding more details to the captions of these figures.

Reviewer comment: Legend of Figure 6 - The last paragraph seems to be out of place

We have deleted the last sentence in the legend of Figure 6 accordingly.

begin returning to objects from earlier

Author response:

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

Public Reviews: 

Reviewer #1 (Public Review): 

Strengths: 

The paper clearly presents the resource, including the testing of candidate enhancers identified from various insects in Drosophila. This cross-species analysis, and the inherent suggestion that training datasets generated in flies can predict a cis-regulatory activity in distant insects, is interesting. While I can not be sure this approach will prevail in the future, for example with approaches that leverage the prediction of TF binding motifs, the SCRMShaw tool is certainly useful and worth consideration for the large community of genome scientists working on insects. 

We thank the reviewer for the positive comments, and would just like to point out that we agree: while we cannot of course know if other methods will overtake SCRMshaw for enhancer prediction—we assume they will, at some point (although motif-based approaches have not fared as well in the past)—for now, SCRMshaw provides strong performance and is a useful part of the current toolkit.

Weaknesses: 

While the authors made the effort to provide access to the SCRMShaw annotations via the RedFly database, the usefulness of this resource is somewhat limited at the moment. First, it is possible to generate tables of annotated elements with coordinates, but it would be more useful to allow downloads of the 33 genome annotations in GFF (or equivalent) format, with SCRMshaw predictions appearing as a new feature. Also, I should note that unlike most species some annotations seem to have issues in the current RedFly implementation. For example, Vcar and Jcoen turn empty. 

We have addressed these weaknesses in several ways:

(1) We have created GFF versions of the SCRMshaw predictions and provide them standalone and also merged into the available annotation GFFs for each of the 33 species

(2) We have made these GFF files, and also the original SCRMshaw output files, available for download in a Dryad repository linked to the publication (https://doi.org/10.5061/dryad.3j9kd51t0).

(3) We have added the inadvertently omitted species to the REDfly/SCRMshaw database.

We agree that the database functions are still somewhat limited, but note that database development is ongoing and we expect functionality to increase over time. In the meantime, the Dryad repository ensures that all results reported in this paper are directly available.

Reviewer #2 (Public Review): 

Summary: 

… Upon identification of predicted enhancer regions, the authors perform post-processing step filtering and identify the most likely predicted enhancer candidates based on the proximity of an orthologous target gene. …

We respectfully point out a small misunderstanding here on the part of the reviewer. We stress that putative target gene assignments and identities have no impact at all on our prediction of regulatory sequences, i.e., they are not “based on the proximity of an orthologous target gene.” Predictions are solely based on sequence-dependent SCRMshaw scores, with no regard to the nature or identities of nearby annotated features. Putative target genes are mapped to Drosophila orthologs purely as a convenience to aid in interpreting and prioritizing the predicted regulatory elements. We have added language on page 8 (lines 189ff) to make this more clear in the text.

Weaknesses:

This work provides predicted enhancer annotations across many insect species, with reporter gene analysis being conducted on selected regions to test the predictions. However, the code for the SCRMshaw analysis pipeline used in this work is not made available, making reproducibility of this work difficult. Additionally, while the authors claim the predicted enhancers are available within the REDfly database, the predicted enhancer coordinates are currently not downloadable as Supplementary Material or from a linked resource. 

We have placed all the code for this paper into a GitHub repository “Asma_etal_2024_eLife” (https://github.com/HalfonLab/Asma_etal_2024_eLife) to address this concern. As described in our response to Reviewer 1, above, all results are now available in multiple formats in a linked Dryad repository in addition to the REDfly/SCRMshaw database.

The authors do not validate or benchmark the application of SCRMshaw against other published methods, nor do they seek to apply SCRMshaw under a variety of conditions to confirm the robustness of the returned predicted enhancers across species. Since SCRMshaw relies on an established k-mer enrichment of the training loci, its performance is presumably highly sensitive to the selection of training regions as well as the statistical power of the given k-mer counts. The authors do not justify their selection of training regions by which they perform predictions. 

Our objective in this study was not to provide proof-of-principle for the SCRMshaw method, as we have established the efficacy of the approach at this point in several previous publications. Rather, the objective here was to make use of SCRMshaw to provide an annotation resource for insect regulatory genomics. Note that the training regions we used here are the same as those we have used in earlier work. Naturally, we performed various assessments to establish that the method was working here, but we make no claims in this work about SCRMshaw’s relative efficiency compared to other methods. Some of our prior publications include assessments of the sort the reviewer references, which suggest that SCRMshaw is at least comparable to other enhancer discovery approaches. We note that benchmarking of such methods is in fact extremely complicated due to the fact that there are no established true positive/true negative data sets against which to benchmark (we have explored this in Asma et al. 2019 BMC Bioinformatics).

While there is an attempt made to report and validate the annotated predicted enhancers using previously published data and tools, the validation lacks the depth to conclude with confidence that the predicted set of regions across each species is of high quality. In vivo, reporter assays were conducted to anecdotally confirm the validity of a few selected regions experimentally, but even these results are difficult to interpret. There is no large-scale attempt to assess the conservation of enhancer function across all annotated species. 

We respectfully disagree that there is insufficient validation. We bring several different lines of evidence to bear suggesting that our results fall into the accuracy range—roughly 75%—established both here and in previous work. We are also clear about the fact that these are predictions only and need to be viewed as such (e.g. line 638). Although “large-scale” in vivo validation assays would certainly be both interesting and worthwhile, the necessary resources for such an assessment places it beyond our present capability.

Lastly, it is suggested that predicted regions are derived from the shared presence of sequence features such as transcription factor binding motifs, detected through k-mer enrichment via SCRMshaw. This assumption has not been examined, although there are public motif discovery tools that would be appropriate to discover whether SCRMshaw is assigning predicted regions based on previously understood motif grammar, or due to other sequence patterns captured by k-mer count distributions. Understanding the sequence-derived nature of what drives predictions is within the scope of this work and would boost confidence in the predicted enhancers, even if it is limited to a few training examples for the sake of clarity of interpretation. 

Again, we respectfully disagree that “this assumption has not been examined.” Although we did not undertake this analysis here, we have in the past, where we have shown that known TFBS motifs can be recovered from sets of SCRMshaw predictions (e.g., Kazemian et al. 2014 Genome Biology and Evolution). We return to this point when we address the Comments to Authors, below.

Reviewer #3 (Public Review): 

Weaknesses:  

The rates of predicted true positive enhancer identification vary widely across the genomes included here based on the simulations and comparison to datasets of accessible chromatin in a manner that doesn't map neatly onto phylogenetic distance. At this point, it is unclear why these patterns may arise, although this may become more clear as regulatory annotation is undertaken for more genomes. 

We agree that we do not see clear patterns with respect to phylogenetic distance in our results. However, we note that this initial data set is still fairly small, and not carefully phylogenetically distributed. We are hoping that, as the reviewer suggests, some of these questions become more clear as we add more genomes to our analysis. Fortunately, the list of available genomes with chromosome-level assembly is growing rapidly, and as we move ahead we should have much greater ability to choose informative species.

Functional assessment of predicted enhancers was performed through reporter gene assays primarily in Drosophila melanogaster imaginal discs, a system amenable to transgenics. Unfortunately, this mode of canonical imaginal disc development is only representative of a subset of all holometabolous insects; therefore, it is difficult to interpret reporter gene expression in a fly imaginal disc as evidence of a true positive enhancer that would be active in its native species whose adult appendages develop differently through the larval stage (for example, Coleopteran and Lepidopteran legs). However, the reporter gene assays from other tissues do offer strong evidence of true positive enhancer detection, and constraints on transgenic experiments in other systems mean that this approach is the best available. 

Please see an extensive discussion of this point in our response to Reviewer 3, below.

Recommendations for the authors:

Reviewer #2 (Recommendations For The Authors): 

Major Concerns: 

(1) While the GitHub source code for SCRMshaw is provided, the authors do not provide a repository of manuscriptspecific code and scripts for readers. This is a barrier to reproducibility and the code used to perform the analysis should be made available. Additionally, links to available scripts do not work, see Line 690. Post-processing scripts point to a general lab folder, but again, no specific analysis or code is sourced for the work in this specific manuscript (e.g. Line 637). 

As noted above, we have corrected this oversight and established a specific GitHub repository for this manuscript “Asma_etal_2024_eLife” (https://github.com/HalfonLab/Asma_etal_2024_eLife). 

(2) On lines 479-488, there is a discussion about the annotations being provided on REDfly, though no link is provided. 

We have included a link in the text at this point (now line 515).

Additionally, for transparency, it would be valuable to provide in Supplementary Table 1 the genomic coordinates of the original training sets in addition to their identity. 

These coordinates have been added to Supplementary Table 1 as suggested.

Also, it is suggested to provide genomic coordinates of the predicted enhancers for each training set across all species, perhaps with a column denoting a linked ID of one genomic coordinate in a species to another species (i.e. if there is a linked region found from D. melanogaster to J. coenia, labeling this column in both coordinate sets as blastoderm.mapping1_region1). Providing these annotations directly in the work enhances the transparency of the results. 

We are unsure exactly what the reviewer means here by “a linked region.” It is critical to understanding our approach to recognize that the genome sequences have diverged to the point where there is no alignment of non-coding regions possible. Thus there is no way to directly “link” coordinates of a predicted enhancer from one species to those of a predicted enhancer in another species. The coordinates for each prediction are available on a per-species basis either through the database or in the files now available in the linked Dryad repository; these can be filtered for results from a specific training set. The database will allow users to select all results for a given orthologous locus, from any subset of species. More complex searches will continue to become available as we improve functionality of the database, an ongoing project in collaboration with the REDfly team.

(3) Figure 2B: It is unclear what this figure shows. Are the No Fly Orthologs false positives, Orthology pipeline issues, or interesting biology? 

We have clarified this in the Figure 2 legend. “No Mapped Fly Orthologs” indicates that our orthology mapping pipeline did not identify clear D. melanogaster orthologs. For any given gene, this could reflect either a true lack of a respective ortholog, or failure of our procedure to accurately identify an existing ortholog.

(4) SCRMshaw appears to be a versatile tool, previously published in a variety of works. However, in this manuscript, there is little discussion of the sensitivity of SCRMshaw to different initial parameters, how the selection of training loci can impact outcomes, or how SCRMshaw k-mer discovery methods compare to other similar tools.

- This paper would be strengthened by addressing this weakness. Some specific suggestions below: 

In order to strengthen confidence that SCRMshaw is a reliable predictor of enhancer regions in other species, it is suggested that you benchmark against other k-mer-derived methods to assign enhancers, such as GSK-SVM developed by the Beer Lab in 2016  (https://www.beerlab.org/gkmsvm/, https://www.biorxiv.org/content/10.1101/2023.10.06.561128v1). 

We have established the effectiveness of SCRMshaw as an enhancer discovery method in previous work, and the main goal of this study was to make use of the established method to annotate numerous insect genomes as a community resource. Our claim here is that SCRMshaw works well for this purpose; we do not attempt a strong claim about whether other approaches may work equally well or marginally better (although we do not believe this is the case, based on prior work). Benchmarking enhancer discovery is challenging, as we point out in Asma et al. 2019 (BMC Bioinformatics), and, while important, best left for a dedicated comprehensive study. A major problem is that there are no independent objective “truth” sets for enhancers from the various species we interrogate here. Thus, while we could also run, e.g., GSK-SVM, what criteria would we use to establish which method had better accuracy for a given species? Note that the work from Beer’s lab took advantage of the ability to match human-mouse orthologous (or syntenic) regions and available open-chromatin data to assess whether conserved enhancers were discovered, but this is not possible given the degree of divergence, limited synteny, and relative lack of additional data for the insect genomes we are annotating.

- In Table S1, we see that 7-146 regions are used as training sets, which is a huge variety. Does an increase in training set size provide a greater "rate of return" for predicted regions? Is the opposite true? Addressing this question would allow readers to understand if they wish to use SCRMshaw, a reasonable scope for their own training region selections. 

- Within a training set, does subsampling provide the same outcomes in terms of prediction rates? There is no exploration of how "brittle" the training sets are, and whether the generalized k-mer count distributions that are established in a training set are consistent across randomly selected subgroups. Performing this analysis would raise confidence in the method applied and the resulting annotations. 

These are interesting and important questions, but again we feel they are beyond the scope of this particular study, which is focused primarily on using SCRMshaw and not on optimizing various search parameters. That said, this is of course something we have investigated, although as with other aspects of enhancer discovery, the absence of a true gold standard enhancer set makes evaluation difficult. We have not found a clear correlation between training set size and performance beyond the very general finding that performance appears to be best when training set size is moderate, e.g. 20-40 initial enhancers. We suspect that larger training sets often contain too many members that don’t fit the core regulatory model and thus add noise, whereas sets that are too small may not contain enough signal for best performance (although small sets can still be useful, especially if used in an iterative cycle; see Weinstein et al. 2023 PLoS Genetics). However, establishing this rigorously is highly challenging given the limitations with assessing true and false positive rates at scale.

(5) In Figure 2C, when plotting hexMCD, IMM, pacRC, and then the merged set, it is unclear whether the scorespecific bar allows coordinate redundancy, though this is implied. What might be more useful is a revision of this plot where the hexMCD/IMM/pac-RC-specific loci are plotted, with the merged set alongside as is currently reported. This would give the reader a clearer understanding of the variability between these scoring methods and why this variability occurs. 

We have added the breakdowns between IMM, hexMCD, and pacRC in Supplementary Table S2, and made more complete reference to this in the text (lines 682ff). Both the database and the data files in the Dryad repository allow exploration of the overlap between the different methods and contain both separate and merged (for overlap and redundancy) results.

Additionally, there is no information in the Methods section of these three SCRMshaw scores and what they represent, even colloquially. While SCRMshaw has been applied in several papers previously, it would help with scientific clarity to describe in a sentence or two what each score is meant to represent and why one is different from another. 

We had chosen to err on the side of brevity given prior publication of the SCRMshaw methodology, but we recognize now that we went too far in that direction. We have added more complete descriptions of the methods in both the Results (lines 164-167) and the Methods (lines 667-681) sections.

(6) When describing results in Figure 2, an important question arises: "Is there an anti-correlation between the number of predicted regions and evolutionary distance?" This would be an expected result that could complement Figure 4's point that shared orthology across 16 species is rarer than across 10 species. Visualizing and adding this to Figure 2 or Figure 4 would be a powerful statement that would boost confidence in the returned predicted enhancers and/or orthologous regions. 

This is an important question and one in which we are very interested. Unfortunately, we do not have sufficient data at this time to address this proper statistical rigor. As we remarked above in response to Reviewer 3, “We agree that we do not see clear patterns with respect to phylogenetic distance in our results. However, we note that this initial data set is still fairly small, and not carefully phylogenetically distributed. We are hoping that, as the reviewer suggests, some of these questions become more clear as we add more genomes to our analysis. Fortunately, the list of available genomes with chromosome-level assembly is growing rapidly, and as we move ahead we should have much greater ability to choose informative species.”

(7) In Figure 3, the authors seek to convey that SCRMshaw predicts enhancer regions that are mapped nearby one another, across different loci widths, and that this occurrence of nearby predicted regions occurs more than a randomly selected control. This is presumably meant to validate that SCRMshaw is not providing predictions with low specificity, but rather to highlight the possibility that SCRMshaw is identifying groups of shadow enhancers. However, these plots are extremely difficult to decipher and do not strongly support the claims due to the low resolution and difficult interpretability of the boxplot interquartile distributions.

Additionally, as the majority of predicted regions are around ~750bp, how does that address loci groups of <1000bp? This suggests that predicted regions are overlapping, and therefore cannot be meaningfully interpreted as shadow enhancers. This plot should either be moved to the supplements or reworked to more effectively convey the point that "SCRMshaw is detecting predicted regions that are proximal to one another and that this proximity is not due to chance". 

- A suggestion to rework this plot is to change this instead to a bar plot, where the y-axis instead represents "number of predictions with at least 2 predicted regions proximal to one another" divided by "total number of predictions", separating bar color by simulated/observed values. The x-axis grouping can remain the same. Because this plot is a broad generalization of the statement you're trying to make above, knowing whether a few loci have 2 versus 4 proximal predicted enhancers doesn't enhance your point. 

We agree with the reviewer that these are not the clearest plots, and thank them for the suggestions regarding revision. We tried many variations on visualizing these complex data, including those suggested by the reviewer, and have concluded that despite their weaknesses, these plots are still the best visualization. The main problem is that the observed data cluster heavily around zero, so that the box plots are very squat and mainly only the outlier large values are observed. The key point, however, is that the expected values almost never give values much greater than one, so that the observed outlier points are the only points seen in the upper ranges of the y-axis. This is true across the three species, across the bins of locus sizes, and across training sets (averaged into the box plots). The reviewer is correct as well about the bins where locus size is < 1000. However, inspection of the data shows that this is not a large concern, as very few data points lie in this range and we never see multiple predicted enhancers there. Thus we believe while not the prettiest of graphs, Figure 3 does effectively support the claims made in the text. In keeping with our view that it is preferable to have data in the main paper whenever possible, we choose to keep the figure in place rather than move it to the Supplement.

- Label the species for the reader's understanding of each subplot on the plot. 

We apologize for this oversight and have now labeled each plot with its relevant species.

(8) SCRMshaw operates on k-mer count distributions compared to a genomic background across different species, allowing it to assign predicted regions without prior knowledge of an organism's cis-regulatory sequences. This is powerful and boosts the versatility of the method. However, understanding the cis-regulatory origins of the kinds of kmers that are driving the detection of orthologous regions across species is crucial and absolutely within the scope of the paper, particularly for the justification of the provided annotations. Is SCRMshaw making use of enriched motifs within the training region set to assign regions in other species? One would presume so, but it is necessary to show this. There are many motif discovery tools that are readily available and require little up-front knowledge and little to no use of a CLI, such as MEMESuite (https://meme-suite.org/meme/tools/meme). It is highly recommended that, even for a few training pairs that are well understood (e.g. mesoderm.mapping1, dorsal_ectoderm.mapping1), assess the motif enrichment within the original sequence set, then see whether motif enrichments are reflected in the predicted enhancers. As evolutionary distance increases between D. melanogaster and the species of interest, is the assignment of enriched motifs more sparse? Is there a loss of a key motif? These are the kinds of questions that will allow readers to understand how these annotations are assigned as well as boost confidence in their usage. 

This is a very important point and a subject of significant interest to us. We have demonstrated in earlier work (e.g., Kazemian et al. 2014 Genome Biol. Evol.) that SCRMshaw-predicted enhancers do contain expected TFBS motifs, across multiple species—and that even an overall arrangement of sites is sometimes conserved. Thus we have previously answered, in part, the reviewer’s question. 

What we also learned from our previous work is that filtering out relevant motifs from the noise inherent in motif-finding is both arduous and challenging. As the reviewer is no doubt aware, while using motif discovery tools is simple, interpreting the output is much less so. In response to the reviewer’s comments, we revisited this issue with data from a small sample of training sets. We can discover motifs; we can see that the motif profiles are different between different training sets; and we can observe the presence of expected motifs based on the activity profile of the enhancers (e.g., Single-minded binding sites in our mesectoderm/midline training and result data). However, to do this cleanly and with appropriate statistical rigor is beyond what we feel would be practical for this paper. We hope to return to this important question in the future when we have a larger and phylogenetically more evenly-distributed set of species, and the time and resources to address it appropriately.

(9) Figures 5-7 need to have better descriptions. 

We have added to the figure 6 and 7 legends in response to this comment; please note as well that there is substantial detail provided in the text. If there are specific aspects of the figures that are not clear or which lack sufficient description, we are happy to make additional changes.

Minor Concerns 

(1)  In Figure 1A, it is implied that "k-mer count distributions" are actually only "5-mer count distributions". However, in the published documentation of SCRMshaw, it is suggested that k-mers between 1-6 bp are involved in establishing sequence distributions. Please add a justification for the selection of these criteria. It would be helpful to understand the implications of using up to a 3-mer versus a 12-mer when assessing k-mer counts using SCRMshaw.

We have clarified in the Figure 1 legend that this is just an example, and the k-mers of different sizes are used in the IMM method; we have also increased the description of the basic method in the Methods section. To be clear, the hexMCD sub-method is 6-mer based (5th-order Markov chain), as is pacRC, while the IMM method considers Markov chains of orders 0-5.

(2) Control the y-axis to remove white space from Figure 2D. 

We have amended the figure as suggested.

Additionally, expand in the manuscript on expected results from SCRMshaw. Given training regions of 750 bp, is the expectation that you return predicted enhancers of the same length? This is not explicitly stated, only a description of outliers. 

The scoring is not dependent on the length of the training sequences, and there is no direct expectation of predicted enhancer length. Scores are calculated on 10-bp intervals, and a peak-calling algorithm is used to determine the endpoints of each prediction based on where the scores drop below a cutoff value. Thus there is no explicit minimum prediction length beyond the smallest possible length of 10-bp. That said, the initial scoring takes place over a 500-bp sequence window (for reasons of computational efficiency), which does influence scores away from the smaller end of the possible range. We correct for this in part by reducing scores below a certain threshold to zero, to prevent multiple low-scoring regions from combining to give a low but positive score over a long interval. Indeed, we found that in the original version of SCRMshawHD (Asma et al. 2019), multiple low-scoring but above-threshold intervals would get concatenated together in broad peaks, leading to an unrealistically large average prediction length. In the version used here, described in Supplementary Figure S6, low-scoring windows are now first reset to zero and a new threshold is calculated before overlapping scores are summed. This helps to prevent the broad peak problem, and we find that it results in a median prediction length ~750 bp, more in line with expected enhancer sizes.

Reviewer #3 (Recommendations For The Authors): 

Line 161: Given that the SCRMshaw HD method is the basis for the pipeline, the methodology deserves at least an "in brief" recapitulation in this manuscript. 

As we remark in our response to Reviewer 2, above, “We had chosen to err on the side of brevity given prior publication of the SCRMshaw methodology, but we recognize now that we went too far in that direction. We have added more complete descriptions of the methods in both the Results (lines 164-167) and the Methods (lines 667-681) sections.” 

Line 219: Throughout the reporting of the results, there appeared to be a bit of inconsistency/potential typos regarding whether threshold or exact P values were reported. In lines 219, 222, 265, 696, and 811, the reported values seem to clearly be thresholds (< a standard cutoff), while in lines 291,293, 297,300, values appear to be exact but are reported as thresholds (<). 

This is not an error but rather reflects two different types of analysis. The predictions per locus (originally lines 219, 222 etc) are evaluated using an empirical P-value based on 1000 permutations. As such, they are thresholded at 1/1000. The overlap with open chromatin regions, on the other hand, are based on a z-score with the P-values taken from a standard conversion of z-scores to P-values.

Page 13/Table 2: At face value, it seems surprising that the overlap between Dmel SCRMshaw predictions with open chromatin is so much smaller than the overlap between predictions and open chromatin in other species, both in raw % (Tcas, D plexippus, H. himera) and fold enrichment (Tcas), given that the training sets for SCRMshaw are all derived from Dmel data. The discussion here does not touch on this aspect of the results, and the interpretation of this approach, in general, would be strengthened if the authors could comment on potential reasons why this pattern may be arising here, or at least acknowledge that this is an open question.

There are many variables at play here, as the data are from different species, from different tissues, and from different methods. Thus we think it is difficult to read too much into the precise results from these comparisons—the main take-home is really just that there is a significant amount of overlap. In acknowledgment of this, we have slightly modified the text in this section so that it now notes (line 302ff): “These comparisons are imperfect, as the tissues used to obtain the chromatin data do not precisely correspond to the training sequences used for SCRMshaw, and the data were obtained using a variety of methods.”

Line 318-329: The inferences from the reporter gene assay deserve a more nuanced treatment than they are given here. The important nuance that was not addressed by the discussion here is that the imaginal disc mode of development in Drosophila is not broadly representative of the development of larval/adult epithelial tissues across Holometabola; thus, inference of a true positive validation becomes complicated in cases where predicted enhancers from a species were tested and shown to drive expression in a fly imaginal disc that the native species have no direct disc counterpart to. For example, in line 388 a Tcas enhancer is reported to drive expression in the eye-antennal disc, and in lines 404 and 423 additional Tcas enhancers were reported to drive expression in the leg discs; however, Tribolium larvae do not possess antennal discs or leg discs set aside during embryogenesis in the sense that flies do - instead the homologous epithelial tissues form larval antennae and larval legs external to the body wall that are actively used at this life stage and are starkly different in morphology than an internally invaginated epithelial disc, that will directly give rise to adult tissues in subsequent molts. Is the interpretation of an expression pattern driven in a fly disc as a true positive really as straightforward as it was presented here, when in the native species the expression pattern driven by the enhancer in question would be in the context of an extremely different tissue morphology? That said, I understand and am deeply sympathetic to the constraints on the authors in performing transgenic experiments outside of the model fly; but these divergent modes of development across Holometabola deserve a mention and nuance in the interpretation here. 

This is indeed a very important point, and we greatly appreciate Reviewer 3 pointing out this caveat when interpreting the outcomes of our cross-species reporter assay. Reviewer 3 is correct that the imaginal disc mode of adult tissue (i.e. imaginal) development found in Diptera does not represent the imaginal development across Holometabola. 

In fact, imaginal development is quite diverse among Holometabola. For instance, larval leg and antennal cells appear to directly develop into the adult legs and antennae in Coleoptera (i.e. primordial imaginal cells function as larval appendage cells), while some cells within the larval legs and antennae are set aside during larval development specifically for adult appendages in Lepidopteran species (i.e. imaginal cells exist within the larval appendages but do not contribute to the formation of larval appendages). In contrast, an almost entire set of cells that develop into adult epithelia are set aside as imaginal discs during embryogenesis in Diptera. Furthermore, the imaginal disc mode of development appears to have evolved independently in

Hymenoptera. Therefore, determining how imaginal primordial tissues correspond to each other among Holometabola has been a challenging task and a topic of high interest within the evo-devo and entomology communities.

Nevertheless, despite these differences in mode of imaginal development, decades of evo-devo studies suggest that the gene regulatory networks (GRNs) operating in imaginal primordial tissues appear to be fairly well conserved among holometabolan species (for example, see Tomoyasu et al. 2009 regarding wing development and Angelini et al. 2012 regarding leg development between flies and beetles). These outcomes imply that a significant portion of the transcriptional landscape might be conserved across different modes of imaginal development. Therefore, an enhancer functioning in the Tribolium larval leg tissue (which also functions as adult leg primordium) could be active even in the leg imaginal disc of Drosophila, if the trans factors essential for the activation of the enhancer are conserved between the two imaginal tissues. 

That being said, we fully expect there to be both false negative and false positive results in our cross-species reporter assay. We are optimistic about the biological relevance of the positive outcomes of our crossspecies reporter assay, especially when the enhancer activity recapitulates the expression of the corresponding gene in Drosophila (for example, Am_ex Fig6B and Tc_hth Fig7B). Nonetheless, the biological relevance of these enhancer activities needs to be further verified in the native species through reporter assays, enhancer knock-outs, or similar experiments.

In recognition of the Reviewer’s important point, we added the following caveat in our Discussion (lines 549553): “Furthermore, the unique imaginal disc mode of adult epithelial development in D. melanogaster  might have prevented some enhancers of other species from working properly in D. melanogaster imaginal discs, likely producing additional false negative results. Evaluating enhancer activities in the native species will allow us to address the degree of false negatives produced by the cross-species setting.” We moreover mention this caveat in the Results section when we first introduce the reporter assays (line 342).

Line 580: This is the first time that the weakness of the closest-gene pairing approach is mentioned. This deserves mention earlier in the manuscript, as unfortunately, this is one of the major bottlenecks to this and any other approaches to investigating enhancer function. Could the authors address this earlier, perhaps pages 7-8, and provide citations for current understanding in the field of how often closest-gene pairing approaches correctly match enhancers to target genes? 

We have added text as suggested on p.7-8 acknowledging the shortcomings of the closest-gene approach. We also clarify at the end of that section (lines 173-181) that target gene assignments, while useful for interpretation, have no bearing on the enhancer predictions themselves (which are generated prior to the target gene assignment steps).

Retrieved from: https://montrealgazette.com/news/local-news/history-through-our-eyes/history-through-our-eyes-march-17-1955-the-richard-riot

Welcome back. In this part of the series, I'm going to be primarily covering the transport layer, which is layer 4 of the OSI model.

I'm also going to be touching upon layer 5, which is the session layer of the OSI model, because there is some overlap for certain features, and so it's easier to cover them in one lesson.

The transport layer runs over the top of the network layer and provides most of the functionality, which supports most of the networking, which we use day-to-day on the internet.

The session layer runs on top of the transport layer, and many features, which you might use, are often mixed between these two layers.

Now, as I've already mentioned, it's not generally worth the argument of deciding whether things are covered in layer 4 or layer 5, so I'd only explain both of these layers as one grouping of functionality.

The OSI model is conceptual, after all, and many things exist between or across two different layers.

Now, we've got a lot to cover, so let's jump in and get started.

Before we get started with layer 4, I want to summarize the situation and limitations with layer 3.

Now, we have a functional layer 3, which means that we can communicate between two devices, say a source and destination laptop, using a source and destination IP address.

If both of these use public IP addresses, it doesn't matter where on the internet these devices are, layer 3 and IP routing will ensure that any packets generated and sent from the source laptop will move across any layer 2 networks between the source and destination.

Let's say that using layer 3, the source laptop on the top generates 6 IP packets, and these are all destined for the destination laptop at the bottom right.

The important thing to understand about layer 3 in this context is that each packet is a separate and isolated thing, and it's routed independently over the internet.

It might be logical to assume that the packets arrive in the same state, so the same timing, the same order, and the same quality, but sadly, that's not true.

In ideal conditions, yes, but generally, if you're communicating using only IP, then you're going to have intermittent network conditions, and that can result in a few cases where the arrival condition of packets is different than the condition when they were generated and sent.

One of the first things which we might encounter is out-of-order arrival.

In this case, where packet 3 arrives before packet 2, layer 3, specifically IP, provides no method to ensure the ordering of packet arrival.

For applications which only used IP, this would mean complex logic would need to be built into the application to ensure packets could be sequenced in the same way, and this is not a trivial task.

Because each packet is routed as an independent thing, it's possible packet 2 could have taken a slow, less efficient route, which is why it arrives later.

This is a negative of layer 3, which can be fixed at layer 4.

Another issue with layer 3 is that packets can just go missing.

This can be due to network outages or network conditions, which cause temporary routing loops.

Remember, when I talked about packet structure, I talked about the TTL field, which limited the number of hops a packet could go through.

Well, if the number of hops exceeds this, then it will be discarded.

With IP, there's no reliable method of ensuring packet delivery, and so it's a relatively regular occurrence that packets go missing.

Now, network conditions can also cause delay in delivery, and for any latency-sensitive applications, this can cause significant issues.

The key thing to keep in mind about layer 3, every packet is different.

It's single, it's isolated.

It's a different unit of data which is being routed across a layer 3 network using layer 2 networks as transit.

What happens to one packet might not happen or might happen in a different way to another packet.

Another limitation with layer 3, and this one is probably the one which has the most obvious effect, is that if you think back to the structure of IP packets, they have a source and destination field.

They don't have anything beyond that to distinguish channels of communication.

Packets from a source IP to a destination IP, they're all the same.

You couldn't have two applications running on the source IP, communicating with two applications running on the destination IP, because there's no method of distinguishing between the applications.

Any packet sent by one application would look to be the same as one sent by another.

Think about what you're doing on your device right now.

You might be watching a video.

Do you have a web browser open doing something else?

Do you have an SSH connection or email or any other application which uses the internet?

This means multiple applications, and IP on its own offers no way to separate the packets for individual applications.

This is something which is remedied at layer 4.

Lastly, IP has no flow control.

If a source device is transmitting packets faster than a destination device can receive them, then it can saturate the destination connection and cause loss of data, packets which will be dropped.

Now with only layer 3, we wouldn't have anywhere near the flexibility required to have the internet function in the way that it does.

For that, we need layer 4, and that's what I want to cover in this part of the lesson series.

So what is layer 4 and how does it function?

Let's take a look.

So far, this is what we have network model-wise.

We've discussed the physical layer which is layer 1 at the OSI model.

This relates to how raw bit screen data is transmitted to or received from physical shared media.

We've talked about layer 2 which adds identifiable devices, switches and media access control, but layer 2 ends with isolated layer 2 networks.

In the previous part of this lesson series, I introduced layer 3 which adds IP addressing and routing, so packets can be routed from source to destination across multiple interconnected networks.

Layer 4 builds on top of this.

It adds two new protocols, TCP which stands for transmission control protocol and UDP which stands for user datagram protocol.

Now both of these run on top of IP, and both of them add a collection of features depending on which one of them is used.

Now if you've heard the term TCP/IP, that means TCP running on top of IP.

At a high level, you would pick TCP when you want reliability, error correction and ordering of data.

It's used for most of the important application layer protocols such as HTTP, HTTPS, SSH and so on.

Now TCP is a connection-oriented protocol which means that you need to set up a connection between two devices and once set up, it creates a bidirectional channel of communications.

UDP on your hand is faster because it doesn't have the TCP overhead required for the reliable delivery of data.

This means that it's less reliable.

Now there's a great joke about UDP.

I will tell you about it, but you might not get it.

Anyway, it's a good job my lessons are better than my jokes.

In this lesson, I'm going to spend most of my time talking about TCP because it's used by more of the important protocols that you'll use day-to-day on the internet.

But just know that both TCP and UDP, they both run on top of IP and they're used in the same way.

They use IP as transit.

TCP just offers a more reliable connection-oriented architecture whereas UDP is all about performance.

So there's a simple trade-off.

Now for this lesson series, as I talk about, I'm going to be focusing on TCP because that's what's used for most of the important upper layer protocol.

So let's take a look at exactly how TCP works.

TCP introduces something called segments.

Now a segment is just another container for data like packets and frames before them.

Segments are specific to TCP.

Before we get started talking about the segments themselves, it's important to understand that segments are actually contained in which is known as encapsulated within IP packets.

So let's say that we have a stream of packets.

You know by now that these are all isolated packets.

They're just pieces of data which are routed independently from source to destination.

They're all treated separately.

Well, TCP segments are placed inside packets and the packets carry the segments from their source to their destination.

Segments don't have source or destination IP addresses because they use the IP packets for the transit from source to destination.

This is all handled by layer 3.

In this case, the internet protocol.

TCP segments add additional capabilities to IP packets.

Let's step through the structure of segments so that we can fully understand them.

And I'm going to skip past a few attributes of segments just as I did with layer 3 because there are some parts which are less important or less situational.

So I won't be covering either the options or padding fields within a segment.

The first fields which I want to cover are the source and destination ports.

In addition to the source and destination IP addresses that IP packets provide, TCP segments add source and destination ports.

And this gives the combined TCP/IP protocol the ability to have multiple streams of conversations at the same time between two devices.

When you open the AWS web interface, you're communicating from a port on your local machine to a port on the AWS servers, TCP port 443, which is HTTPS.

Now because of port, you can have multiple streams of communication from your machine.

One to AWS, one to Netflix, and one to this website where you're watching this video.

At the other side, AWS can have multiple streams of communication to their servers.

Each conversation is a unique combination of the source and destination IP, the source port, and the destination port.

All of these four values together identify as a single conversation, a single communications channel.

These two fields ought to allow the internet to function in a flexible way that it does.

It's why SSH and HTTPS can exist on the same EC2 instance and why you can have multiple SSH connections open to the same EC2 instance if you wanted to.

And I'll cover more on how this works as we move through this lesson.

Now next, within the segment, is sequence.

And the sequence number is incremented with each segment that's sent.

And it's unique.

It can be used for error correction if things need to be retransmitted.

You can use to ensure that one IP pass is received and the TCP segments are pulled out.

They can be correctly ordered.

So the sequence number is already uniquely identifying the particular segment within a particular connection so that both sides can make observations about it.

And the way that these observations are done is using app knowledgements.

The app knowledgement field on the right, on the one side, can indicate that it's received up to and including a certain sequence number.

Every segment which is transmitted needs to be acknowledged.

Remember that TCP is a reliable protocol and so if the device is transmitting segment one, two, three, and four to another device, then the other device needs to acknowledge that it's received segment one, two, three, and four.

And this is what the app knowledgement field is for.

So sequence number and app knowledgement are used hand in hand.

Next we have a field called flags and things.

Now within a segment, there is an actual flags component which is nine bits.

And this allows various controls over the TCP segments and the wider connection.

Flags are used to close the connection of the synchronized sequence numbers, but there's also additional things like a data offset and some reserved space.

So I thought this flags and things is essentially the flags plus a number of extra fields which I don't need to go into at this point in the lesson.

Now next we've got the TCP window.

And this is interesting.

This defines the number of bytes that you indicate that you're willing to receive between app knowledgements.

Once reached, the sender will pause until you acknowledge that amount of data.

And this is how flow control is implemented.

It lets the receiver control the rate at which the sender sends data.

If you use a smaller window, it provides additional levels of control over how quickly you're sent data.

Larger windows are more efficient because the header of a TCP segment takes up an amount of space and the smaller the window, the more headers are involved.

So this window setting is quite important if you're using a TCP for practical reasons, but we don't need to go into too much detail in this lesson.

Next we have checks on which is used for error checking.

It means that a TCP layer is able to detect errors and can arrange for retransmission of the data as required.

And then lastly, we have the urgent pointer.

And this is a cool feature.

Imagine if you have a data transfer application where 99% of data is the data being transferred and 1% is control traffic.

So communication between the client and the server, coordinating the actual data transfer.

While setting this field in a segment means that both sides can have separate processing.

So the control traffic always takes priority within the communication.

So any protocols which are latency sensitive and transfer data such as FTP and PellNet can use this field.

Now all of these fields together are known as the TCP header.

And the capacity of a TCP segment which remains is logically enough used for data.

So that's a segment that are placed inside packets and transmitted by one network stack, specifically layer 4 of one network stack and received by another network stack using the layer 4 protocol.

In this case TCP.

Okay so this is the end of part 1 of this lesson.

It was getting a little bit on the long side and so I wanted to add a break.

It's an opportunity just to take a rest or grab a coffee.

Part 2 will be continuing immediately from the end of part 1.

So go ahead, complete video and when you're ready join me in part 2.

Welcome back, this is part two of this lesson.

We're going to continue immediately from the end of part one.

So let's get started.

Now the only thing that remains is just to test out this configuration.

And to do that we're going to launch an EC2 instance into the WebA subnet.

So click on services and just type EC2 to move across to the EC2 console.

Now once we're on the EC2 console, just click on launch instance.

Then you'll be taken to the launch instance console.

Into the name box, just go ahead and type a4l-bastian.

Scroll down and we're going to create a bastion instance using Amazon Linux.

So click on Amazon Linux.

In the dropdown below, go ahead and select the latest version of Amazon Linux.

Just make sure that it does say free tier eligible on the right of this dropdown.

Assuming that's all good, just below that make sure that in the architecture dropdown it's set to 64-bit x86.

Moving down further still, under instance type, just make sure that this is set to a free tier eligible instance.

It should default to T2.micro or T3.micro.

Depending on your region, either of these could be free tier eligible.

In my case it's T2.micro, but whatever your shows, just make sure that it's similar sized and says free tier eligible.

Now directly below that, under key pair, just click in this box.

You should at this point in the course have a key pair creator called a4l.

If you do, go ahead and select that key pair in the box.

If you don't, don't worry, you can just go ahead and click on create new key pair.

Enter a4l into the key pair name, select RSA, and then select PEM for the private key format and click on create key pair.

This will download the key pair to your local machine and then you can continue following along with this video.

So select that from the dropdown.

Directly below, under network settings click on edit.

This instance is going to go into the animals for live VPC.

So click on the VPC dropdown and select a4l-vpc1.

Directly below that, click in the subnet dropdown and we want to go ahead and look for sn-web-a.

So select the weba subnet.

This should change both of the dropdowns below.

So auto assign public IP and auto assign IPv6 IP to enable.

So just make sure that both of those are set to enable.

Directly below this, make sure that create security group is checked.

We're going to create a new security group.

Under security group name, just go ahead and enter a4l-bassian-sg and then put that same text in the description box directly below.

Now all of these defaults should be good.

Just make sure it's set to SSH, source anywhere.

Make sure that 0.0.0.0/0 and double colon 4/0 are both present directly below source.

Everything else looks good.

We can accept the rest of the defaults.

Just go ahead and click on launch instance.

Then click on instances at the top left of the screen.

At this point the instance is launching and we'll see that a4l-bassian is currently running.

We'll see the status check is showing initializing.

So we need to give this instance a few minutes to fully provision.

So go ahead and pause this video and we're going to resume it once this instance is ready to go.

And it has two out of two status checks.

So our instance is now showing two out of two status checks.

And that means everything's good and we're ready to connect.

Now if you select the instance, you'll see in the details pane below how it has a public IP version 4 address, a private IP version 4 address, public and private IP version 4 DNS.

And if we scroll down, you'll see lots of other information about this instance.

Now we're only concerned with the public IP version 4 address.

We're going to go ahead and connect to this instance this time using a local SSH client on our machine.

So right click and then select connect.

Now if we want to quickly connect into this instance, we can choose to use EC2 instance connect, which is a way to connect into the instance using a web console.

Now this does need an instance with a public IP version 4 address, but we have allocated a public address.

So if we wanted to, we can just make sure that the username is correct.

It should be EC2-user.

If we hit connect, it will open up a connection to this instance using a web console.

And this is often much easier to connect to EC2 instances if you don't have access to a local SSH client, or if you just want to quickly connect to perform some administration.

We can also connect with an SSH client.

If we select SSH client, it gives us the commands to run in order to connect to this EC2 instance.

So right at the bottom is an example connect command.

So SSH, we pick the key to use and then we pick the user at and then the public IP version 4 DNS.

So if we copy that into our clipboard and then move across to our terminal or command prompt, move into the folder where you downloaded the SSH keybearant to, in my case downloads, and paste in that command and press enter, that should connect us to the EC2 instance.

We'll have to verify the fingerprint, so we need to verify the authenticity of this host.

For this purpose, we can just go ahead and answer yes and press enter.

Now if it's the first time we're connecting using a particular key, and if you're running either macOS or Linux, you might be informed that the permissions on the key are too open.

In this case, the permissions are 0, 6, 4, 4, which are too open and we get this error.

Now it's possible to correct that if we move back to the AWS console.

It also gives us the command to correct these permissions.

So CHmod, space 400, space and then the name of the key.

So I'm going to copy that into my clipboard and move back to my terminal, paste that in and press enter, and that will correct those permissions.

Now if I get the connection command again, so copy that into my clipboard, and this time I'll paste it in and press enter and now I will be connected to this EC2 instance.

Now if you're doing this demonstration on Windows 10, you probably won't have to correct those permissions.

This is something specific to macOS or Linux.

So whenever you're connecting to EC2 instances which have a public IP version 4 address, you've always got the ability to use either EC2 instance connect or a local SSH client.

Now the third option which is session manager, this is a way that you can connect to instances even if they don't have public IP version 4 addressing.

And I'll be detailing this product fully later on in the course because there is some additional configuration that's required.

Now this bastion host, it's an EC2 instance and it does fall under the free tier.

So because it's a T2.micro or whatever type of instance you picked which falls under the free tier, you're not going to be billed for any usage of this instance in a given month.

Now as a general rule, as you're moving through the course, if you're ever intending to take a break, then you always have the option of deleting all of the infrastructure that you've created within a specific demo lesson.

So most of the more complex demo lessons that you'll have moving through the course, at the end of every demo lesson there will be a brief set of steps where I explain how to clean up the account and return it into the same state as it was at the start of the lesson.

But in certain situations I might tell you that one option is not to delete the infrastructure.

Whether you do delete it or not depends on whether you're intending to complete the next demo straight away or whether you're taking a break.

Now in this particular case I'm going to demonstrate exactly how you can clear up this infrastructure. [background noise] [background noise] In the next demo lesson you're going to be continuing using this structure, but I'm going to demonstrate how you can automate the creation using a CloudFormation template.

To clear up this infrastructure though, go ahead, right click on this bastion host and select terminate instance.

You'll need to click terminate to confirm and that will terminate and delete the instance.

You won't be charged for any further usage of that instance.

We need to wait for that instance to fully terminate, so pause the video and wait for it to move into a terminated state and then we can continue.

So that instance is terminated and now that that's done, we can click on services and move across to the VPC console and we're going to delete the entire Animals for Life VPC.

And don't worry, in the next demo lesson I'll explain how we can automate the creation.

So for now and in the course we're going to be using much more automation so that anything that you've done previously, we're going to automate the creation and focus your valuable time only on the things that you've just learned.

So click on your VPCs.

It should list two VPCs, the default one and the Animals for Life VPC.

Select the Animals for Life VPC, click on Actions and then delete the VPC.

Now this is going to delete all of the resources that are associated with this VPC.

So the internet gateway, the route tables and all of the subnets that you've created as part of the demo lessons to this point in the course.

So go ahead and type delete and then click delete to confirm that process and that will fully tidy up the account and return it into the same state as it was at the start of the VPC section.

Now with that being said, this is the end of this lesson.

You've successfully converted three subnets, so Web A, Web B and Web C to be public and you've done that by creating an internet gateway, associating that with the VPC, creating a route table, associating that with those subnets, adding two routes, pointing those routes at the internet gateway and then configuring those subnets to allocate a public IP version for address to any resources launched into those subnets.

So that's the same set of steps that you'll need to do to make any subnets public from an IP version for perspective in future.

So this is going to be the same tasks that you would use in larger production projects.

Although in production, you would probably automate it and don't just show you how to do that as you move through the course.

Now at this point, you've finished everything that you need to do in this demo lesson, so great job.

You've actually created something that is production ready and production useful.

Over the remainder of this section of the course, we're going to refine the design that we've got and add additional capabilities.

So in the upcoming lessons, I'll be talking about network address translation and how that can be used to give private EC2 instances access to the internet for things like software updates.

We'll be talking about the security of subnets using network access control lists known as knuckles and much, much more.

But you're doing a fantastic job so far.

This is not a trivial thing that you've implemented to this point.

So really great job.

But at this point, just go ahead and complete the video.

And then when you're ready, I look forward to you joining me in the next. [no audio]

Mirrors are for locals and windows are for those not as familiar with the area.

I'm reminded of the time when it was revealed about 10–15 years or so ago that when Apple switched to x86 from PowerPC, it wasn't the result of a big porting effort. They'd been maintaining portability all along—doing private builds internally that just never saw the light of day. When this came to light, the reaction was huge. People were awed.

A few years ago, when this piece of trivia was brought back to the forefront of my mind again after having not thought about it for years, I was struck by how silly that reaction was. Of course it makes sense that they'd been maintaining portability. There was nothing stupendous about this.

I think this is the one time when I saw and felt the effects of Apple's legendary reality distortion field firsthand. (In every other instance, I hadn't been close enough and so only perceived it from afar and only had other sources to trust that it was a real phenomenon.)

This still describes the way the semiconductor manufacturing world works.

This is sort of an odd remark. Even excluding servers and focusing only on traditional desktop and laptop computers, Windows' dominance is as weak now as it has been at any other point in the last 30 years.

Most desktop and laptop computers? Yeah, probably.

"Almost all"? Surely not.

Reviewer #2 (Public Review):

Summary:

In this work, the authors present a new Python software package, Avian Vocalization Network (AVN) aimed at facilitating the analysis of birdsong, especially the song of the zebra finch, the most common songbird model in neuroscience. The package handles some of the most common (and some more advanced) song analyses, including segmentation, syllable classification, featurization of song, calculation of tutor-pupil similarity, and age prediction, with a view toward making the entire process friendlier to experimentalists working in the field.

For many years, Sound Analysis Pro has served as a standard in the songbird field, the first package to extensively automate songbird analysis and facilitate the computation of acoustic features that have helped define the field. More recently, the increasing popularity of Python as a language, along with the emergence of new machine learning methods, has resulted in a number of new software tools, including the vocalpy ecosystem for audio processing, TweetyNet (for segmentation), t-SNE and UMAP (for visualization), and autoencoder-based approaches for embedding.

Strengths:

The AVN package overlaps several of these earlier efforts, albeit with a focus on more traditional featurization that many experimentalists may find more interpretable than deep learning-based approaches. Among the strengths of the paper are its clarity in explaining the several analyses it facilitates, along with high-quality experiments across multiple public datasets collected from different research groups. As a software package, it is open source, installable via the pip Python package manager, and features high-quality documentation, as well as tutorials. For experimentalists who wish to replicate any of the analyses from the paper, the package is likely to be a useful time saver.

Weaknesses:

I think the potential limitations of the work are predominantly on the software end, with one or two quibbles about the methods.

First, the software: it's important to note that the package is trying to do many things, of which it is likely to do several well and few comprehensively. Rather than a package that presents a number of new analyses or a new analysis framework, it is more a codification of recipes, some of which are reimplementations of existing work (SAP features), some of which are essentially wrappers around other work (interfacing with WhisperSeg segmentations), and some of which are new (similarity scoring). All of this has value, but in my estimation, it has less value as part of a standalone package and potentially much more as part of an ecosystem like vocalpy that is undergoing continuous development and has long-term support. While the code is well-documented, including web-based documentation for both the core package and the GUI, the latter is available only on Windows, which might limit the scope of adoption.

That is to say, whether AVN is adopted by the field in the medium term will have much more to do with the quality of its maintenance and responsiveness to users than any particular feature, but I believe that many of the analysis recipes that the authors have carefully worked out may find their way into other code and workflows.

Second, two notes about new analysis approaches:

(1) The authors propose a new means of measuring tutor-pupil similarity based on first learning a latent space of syllables via a self-supervised learning (SSL) scheme and then using the earth mover's distance (EMD) to calculate transport costs between the distributions of tutors' and pupils' syllables. While to my knowledge this exact method has not previously been proposed in birdsong, I suspect it is unlikely to differ substantially from the approach of autoencoding followed by MMD used in the Goffinet et al. paper. That is, SSL, like the autoencoder, is a latent space learning approach, and EMD, like MMD, is an integral probability metric that measures discrepancies between two distributions. (Indeed, the two are very closely related: https://stats.stackexchange.com/questions/400180/earth-movers-distance-and-maximum-mean-discrepency.) Without further experiments, it is hard to tell whether these two approaches differ meaningfully. Likewise, while the authors have trained on a large corpus of syllables to define their latent space in a way that generalizes to new birds, it is unclear why such an approach would not work with other latent space learning methods.

(2) The authors propose a new method for maturity scoring by training a model (a generalized additive model) to predict the age of the bird based on a selected subset of acoustic features. This is distinct from the "predicted age" approach of Brudner, Pearson, and Mooney, which predicts based on a latent representation rather than specific features, and the GAM nicely segregates the contribution of each. As such, this approach may be preferred by many users who appreciate its interpretability.

In summary, my view is that this is a nice paper detailing a well-executed piece of software whose future impact will be determined by the degree of support and maintenance it receives from others over the near and medium term.

Author response:

Public Reviews:

Reviewer #1 (Public Review):

Summary: 

This paper applies methods for segmentation, annotation, and visualization of acoustic analysis to zebra finch song. The paper shows that these methods can be used to predict the stage of song development and to quantify acoustic similarity. The methods are solid and are likely to provide a useful tool for scientists aiming to label large datasets of zebra finch vocalizations. The paper has two main parts: 1) establishing a pipeline/ package for analyzing zebra finch birdsong and 2) a method for measuring song imitation. 

Strengths: 

It is useful to see existing methods for syllable segmentation compared to new datasets. 

It is useful, but not surprising, that these methods can be used to predict developmental stage, which is strongly associated with syllable temporal structure. 

It is useful to confirm that these methods can identify abnormalities in deafened and isolated songs. 

Weaknesses: 

For the first part, the implementation seems to be a wrapper on existing techniques. For instance, the first section talks about syllable segmentation; they made a comparison between whisperseg (Gu et al, 2024), tweetynet (Cohen et al, 2022), and amplitude thresholding. They found that whisperseg performed the best, and they included it in the pipeline. They then used whisperseg to analyze syllable duration distributions and rhythm of birds of different ages and confirmed past findings on this developmental process (e.g. Aronov et al, 2011). Next, based on the segmentation, they assign labels by performing UMAP and HDBScan on the spectrogram (nothing new; that's what people have been doing). Then, based on the labels, they claimed they developed a 'new' visualization - syntax raster ( line 180 ). That was done by Sainburg et. al. 2020 in Figure 12E and also in Cohen et al, 2020 - so the claim to have developed 'a new song syntax visualization' is confusing. The rest of the paper is about analyzing the finch data based on AVN features (which are essentially acoustic features already in the classic literature). 

First, we would like to thank this reviewer for their kind comments and feedback on this manuscript. It is true that many of the components of this song analysis pipeline are not entirely novel in isolation. Our real contribution here is bringing them together in a way that allows other researchers to seamlessly apply automated syllable segmentation, clustering, and downstream analyses to their data. That said, our approach to training TweetyNet for syllable segmentation is novel. We trained TweetyNet to recognize vocalizations vs. silence across multiple birds, such that it can generalize to new individual birds, whereas Tweetynet had only ever been used to annotate song syllables from birds included in its training set previously. Our validation of TweetyNet and WhisperSeg in combination with UMAP and HDBSCAN clustering is also novel, providing valuable information about how these systems interact, and how reliable the completely automatically generated labels are for downstream analysis. 

Our syntax raster visualization does resemble Figure 12E in Sainburg et al. 2020, however it differs in a few important ways, which we believe warrant its consideration as a novel visualization method. First, Sainburg et al. represent the labels across bouts in real time; their position along the x axis reflects the time at which each syllable is produced relative to the start of the bout. By contrast, our visualization considers only the index of syllables within a bout (ie. First syllable vs. second syllable etc) without consideration of the true durations of each syllable or the silent gaps between them. This makes it much easier to detect syntax patterns across bouts, as the added variability of syllable timing is removed. Considering only the sequence of syllables rather than their timing also allows us to more easily align bouts according to the first syllable of a motif, further emphasizing the presence or absence of repeating syllable sequences without interference from the more variable introductory notes at the start of a motif. Finally, instead of plotting all bouts in the order in which they were produced, our visualization orders bouts such that bouts with the same sequence of syllables will be plotted together, which again serves to emphasize the most common syllable sequences that the bird produces. These additional processing steps mean that our syntax raster plot has much starker contrast between birds with stereotyped syntax and birds with more variable syntax, as compared to the more minimally processed visualization in Sainburg et al. 2020. There doesn’t appear to be any similar visualizations in Cohen et al. 2020. 

The second part may be something new, but there are opportunities to improve the benchmarking. It is about the pupil-tutor imitation analysis. They introduce a convolutional neural network that takes triplets as an input (each tripled is essentially 3 images stacked together such that you have (anchor, positive, negative), Anchor is a reference spectrogram from, say finch A; positive means a different spectrogram with the same label as anchor from finch A, and negative means a spectrogram not related to A or different syllable label from A. The network is then trained to produce a low-dimensional embedding by ensuring the embedding distance between anchor and positive is less than anchor and negative by a certain margin. Based on the embedding, they then made use of earth mover distance to quantify the similarity in the syllable distribution among finches. They then compared their approach performance with that of sound analysis pro (SAP) and a variant of SAP. A more natural comparison, which they didn't include, is with the VAE approach by Goffinet et al. In this paper (https://doi.org/10.7554/eLife.67855, Fig 7), they also attempted to perform an analysis on the tutor pupil song. 

We thank the reviewer for this suggestion, and plan to include a comparison of the triplet loss embedding space to the VAE space for song similarity comparisons in the revised manuscript.

Reviewer #2 (Public Review):

Summary: 

In this work, the authors present a new Python software package, Avian Vocalization Network (AVN) aimed at facilitating the analysis of birdsong, especially the song of the zebra finch, the most common songbird model in neuroscience. The package handles some of the most common (and some more advanced) song analyses, including segmentation, syllable classification, featurization of song, calculation of tutor-pupil similarity, and age prediction, with a view toward making the entire process friendlier to experimentalists working in the field. 

For many years, Sound Analysis Pro has served as a standard in the songbird field, the first package to extensively automate songbird analysis and facilitate the computation of acoustic features that have helped define the field. More recently, the increasing popularity of Python as a language, along with the emergence of new machine learning methods, has resulted in a number of new software tools, including the vocalpy ecosystem for audio processing, TweetyNet (for segmentation), t-SNE and UMAP (for visualization), and autoencoder-based approaches for embedding. 

Strengths: 

The AVN package overlaps several of these earlier efforts, albeit with a focus on more traditional featurization that many experimentalists may find more interpretable than deep learning-based approaches. Among the strengths of the paper are its clarity in explaining the several analyses it facilitates, along with high-quality experiments across multiple public datasets collected from different research groups. As a software package, it is open source, installable via the pip Python package manager, and features high-quality documentation, as well as tutorials. For experimentalists who wish to replicate any of the analyses from the paper, the package is likely to be a useful time saver. 

Weaknesses: 

I think the potential limitations of the work are predominantly on the software end, with one or two quibbles about the methods. 

First, the software: it's important to note that the package is trying to do many things, of which it is likely to do several well and few comprehensively. Rather than a package that presents a number of new analyses or a new analysis framework, it is more a codification of recipes, some of which are reimplementations of existing work (SAP features), some of which are essentially wrappers around other work (interfacing with WhisperSeg segmentations), and some of which are new (similarity scoring). All of this has value, but in my estimation, it has less value as part of a standalone package and potentially much more as part of an ecosystem like vocalpy that is undergoing continuous development and has long-term support. 

We appreciate this reviewer’s comments and concerns about the structure of the AVN package and its long-term maintenance. We have considered incorporating AVN into the VocalPy ecosystem but have chosen not to for a few key reasons. (1) AVN was designed with ease of use for experimenters with limited coding experience top of mind. VocalPy provides excellent resources for researchers with some familiarity with object-oriented programming to manage and analyze their datasets; however, we believe it may be challenging for users without such experience to adopt VocalPy quickly. AVN’s ‘recipe’ approach, as you put it, is very easily accessible to new users, and allows users with intermediate coding experience to easily navigate the source code to gain a deeper understanding of the methodology. AVN also consistently outputs processed data in familiar formats (tables in .csv files which can be opened in excel), in an effort to make it more accessible to new users, something which would be challenging to reconcile with VocalPy’s emphasis on their `dataset`classes. (2) AVN and VocalPy differ in their underlying goals and philosophies when it comes to flexibility vs. standardization of analysis pipelines. VocalPy is designed to facilitate mixing-and-matching of different spectrogram generation, segmentation, annotation etc. approaches, so that researchers can design and implement their own custom analysis pipelines. This flexibility is useful in many cases. For instance, it could allow researchers who have very different noise filtering and annotation needs, like those working with field recordings versus acoustic chamber recordings, analyze their data using this platform. However, when it comes to comparisons across zebra finch research labs, this flexibility comes at the expense of direct comparison and integration of song features across research groups. This is the context in which AVN is most useful. It presents a single approach to song segmentation, labeling, and featurization that has been shown to generalize well across research groups, and which allows direct comparisons of the resulting features. AVN’s single, extensively validated, standard pipeline approach is fundamentally incompatible with VocalPy’s emphasis on flexibility. We are excited to see how VocalPy continues to evolve in the future and recognize the value that both AVN and VocalPy bring to the songbird research community, each with their own distinct strengths, weaknesses, and ideal use cases. 

While the code is well-documented, including web-based documentation for both the core package and the GUI, the latter is available only on Windows, which might limit the scope of adoption. 

We thank the reviewer for their kind words about AVN’s documentation. We recognize that the GUI’s exclusive availability on Windows is a limitation, and we would be happy to collaborate with other researchers and developers in the future to build a Mac compatible version, should the demand present itself. That said, the python package works on all operating systems, so non-Windows users still have the ability to use AVN that way.  

That is to say, whether AVN is adopted by the field in the medium term will have much more to do with the quality of its maintenance and responsiveness to users than any particular feature, but I believe that many of the analysis recipes that the authors have carefully worked out may find their way into other code and workflows. 

Second, two notes about new analysis approaches: 

(1) The authors propose a new means of measuring tutor-pupil similarity based on first learning a latent space of syllables via a self-supervised learning (SSL) scheme and then using the earth mover's distance (EMD) to calculate transport costs between the distributions of tutors' and pupils' syllables. While to my knowledge this exact method has not previously been proposed in birdsong, I suspect it is unlikely to differ substantially from the approach of autoencoding followed by MMD used in the Goffinet et al. paper. That is, SSL, like the autoencoder, is a latent space learning approach, and EMD, like MMD, is an integral probability metric that measures discrepancies between two distributions.

(Indeed, the two are very closely related: https://stats.stackexchange.com/questions/400180/earth-movers-distance-andmaximum-mean-discrepency.) Without further experiments, it is hard to tell whether these two approaches differ meaningfully. Likewise, while the authors have trained on a large corpus of syllables to define their latent space in a way that generalizes to new birds, it is unclear why such an approach would not work with other latent space learning methods. 

We recognize the similarities between these approaches, and plan to include a comparison of triplet loss embeddings compared with MMD and VAE embeddings compared with MMD and EMD in the revised manuscript. Thank you for this suggestion.  

(2) The authors propose a new method for maturity scoring by training a model (a generalized additive model) to predict the age of the bird based on a selected subset of acoustic features. This is distinct from the "predicted age" approach of Brudner, Pearson, and Mooney, which predicts based on a latent representation rather than specific features, and the GAM nicely segregates the contribution of each. As such, this approach may be preferred by many users who appreciate its interpretability. 

In summary, my view is that this is a nice paper detailing a well-executed piece of software whose future impact will be determined by the degree of support and maintenance it receives from others over the near and medium term. 

Reviewer #3 (Public Review):

Summary: 

The authors invent song and syllable discrimination tasks they use to train deep networks. These networks they then use as a basis for routine song analysis and song evaluation tasks. For the analysis, they consider both data from their own colony and from another colony the network has not seen during training. They validate the analysis scores of the network against expert human annotators, achieving a correlation of 80-90%. 

Strengths: 

(1) Robust Validation and Generalizability: The authors demonstrate a good performance of the AVN across various datasets, including individuals exhibiting deviant behavior. This extensive validation underscores the system's usefulness and broad applicability to zebra finch song analysis, establishing it as a potentially valuable tool for researchers in the field. 

(2) Comprehensive and Standardized Feature Analysis: AVN integrates a comprehensive set of interpretable features commonly used in the study of bird songs. By standardizing the feature extraction method, the AVN facilitates comparative research, allowing for consistent interpretation and comparison of vocal behavior across studies. 

(3) Automation and Ease of Use. By being fully automated, the method is straightforward to apply and should introduce barely an adoption threshold to other labs. 

(4) Human experts were recruited to perform extensive annotations (of vocal segments and of song similarity scores). These annotations released as public datasets are potentially very valuable. 

Weaknesses: 

(1) Poorly motivated tasks. The approach is poorly motivated and many assumptions come across as arbitrary. For example, the authors implicitly assume that the task of birdsong comparison is best achieved by a system that optimally discriminates between typical, deaf, and isolated songs. Similarly, the authors assume that song development is best tracked using a system that optimally estimates the age of a bird given its song. My issue is that these are fake tasks since clearly, researchers will know whether a bird is an isolated or a deaf bird, and they will also know the age of a bird, so no machine learning is needed to solve these tasks. Yet, the authors imagine that solving these placeholder tasks will somehow help with measuring important aspects of vocal behavior. 

We appreciate this reviewer’s concerns and apologize for not providing sufficiently clear rationale for the inclusion of our phenotype classifier and age regression models in the original manuscript. These tasks are not intended to be taken as a final, ultimate culmination of the AVN pipeline. Rather, we consider the carefully engineered 55-interpretable feature set to be AVN’s final output, and these analyses serve merely as examples of how that feature set can be applied. That said, each of these models do have valid experimental use cases that we believe are important and would like to bring to the attention of the reviewer.

For one, we showed how the LDA model that can discriminate between typical, deaf, and isolate birds’ songs not only allows us to evaluate which features are most important for discriminating between these groups, but also allows comparison of the FoxP1 knock-down (FP1 KD) birds to each of these phenotypes. Based on previous work (Garcia-Oscos et al. 2021), we hypothesized that FP1 KD in these birds specifically impaired tutor song memory formation while sparing a bird’s ability to refine their own vocalizations through auditory feedback. Thus, we would expect their songs to resemble those of isolate birds, who lack a tutor song memory, but not to resemble deaf birds who lack a tutor song memory and auditory feedback of their own vocalizations to guide learning. The LDA model allowed us to make this comparison quantitatively for the first time and confirm our hypothesis that FP1 KD birds’ songs are indeed most like isolates’. In the future, as more research groups publish their birds’ AVN feature sets, we hope to be able to make even more fine-grained comparisons between different groups of birds, either using LDA or other similar interpretable classifiers. 

The age prediction model also has valid real-world use cases. For instance, one might imagine an experimental manipulation that is hypothesized to accelerate or slow song maturation in juvenile birds. This age prediction model could be applied to the AVN feature sets of birds having undergone such a manipulation to determine whether their predicted ages systematically lead or lag their true biological ages, and which song features are most responsible for this difference. We didn’t have access to data for any such birds for inclusion in this paper, but we hope that others in the future will be able to take inspiration from our methodology and use this or a similar age regression model with AVN features in their research. We will revise the original manuscript to make this clearer. 

Along similar lines, authors assume that a good measure of similarity is one that optimally performs repeated syllable detection (i.e. to discriminate same syllable pairs from different pairs). The authors need to explain why they think these placeholder tasks are good and why no better task can be defined that more closely captures what researchers want to measure. Note: the standard tasks for self-supervised learning are next word or masked word prediction, why are these not used here? 

There appears to be some misunderstanding regarding our similarity scoring embedding model and our rationale for using it. We will explain it in more depth here and provide some additional explanation in the manuscript. First, we are not training a model to discriminate between same and different syllable pairs. The triplet loss network is trained to embed syllables in an 8-dimensional space such that syllables with the same label are closer together than syllables with different labels. The loss function is related to the relative distance between embeddings of syllables with the same or different labels, not the classification of syllables as same or different. This approach was chosen because it has repeatedly been shown to be a useful data compression step (Schorff et al. 2015, Thakur et al. 2019) before further downstream tasks are applied on its output, particularly in contexts where there is little data per class (syllable label). For example, Schorff et al. 2015 trained a deep convolutional neural network with triplet loss to embed images of human faces from the same individual closer together than images of different individuals in a 128-dimensional space. They then used this model to compute 128-dimensional representations of additional face images, not included in training, which were used for individual facial recognition (this is a same vs. different category classifier), and facial clustering, achieving better performance than the previous state of the art. The triplet loss function results in a model that can generate useful embeddings of previously unseen categories, like new individuals’ faces, or new zebra finches’ syllables, which can then be used in downstream analyses. This meaningful, lower dimensional space allows comparisons of distributions of syllables across birds, as in Brainard and Mets 2008, and Goffinet et al. 2021. 

Next word and masked word prediction are indeed common self-supervised learning tasks for models working with text data, or other data with meaningful sequential organization. That is not the case for our zebra finch syllables, where every bird’s syllable sequence depends only on its tutor’s sequence, and there is no evidence for strong universal syllable sequencing rules (James et al. 2020). Rather, our embedding model is an example of a computer vision task, as it deals with sets of twodimensional images (spectrograms), not sequences of categorical variables (like text). It is also not, strictly speaking, a self-supervised learning task, as it does require syllable labels to generate the triplets. A common self-supervised approach for dimensionality reduction in a computer vision task such as this one would be to train an autoencoder to compress images to a lower dimensional space, then faithfully reconstruct them from the compressed representation.  This has been done using a variational autoencoder trained on zebra finch syllables in Goffinet et al. 2021. In keeping with the suggestions from reviewers #1 and #2, we plan to include a comparison of our triplet loss model with the Goffinet et al. VAE approach in the revised manuscript.  

(2) The machine learning methodology lacks rigor. The aims of the machine learning pipeline are extremely vague and keep changing like a moving target. Mainly, the deep networks are trained on some tasks but then authors evaluate their performance on different, disconnected tasks. For example, they train both the birdsong comparison method (L263+) and the song similarity method (L318+) on classification tasks. However, they evaluate the former method (LDA) on classification accuracy, but the latter (8-dim embeddings) using a contrast index. In machine learning, usually, a useful task is first defined, then the system is trained on it and then tested on a held-out dataset. If the sensitivity index is important, why does it not serve as a cost function for training?

Again, there appears to be some misunderstanding of our similarity scoring methodology. Our similarity scoring model is not trained on a classification task, but rather on an embedding task. It learns to embed spectrograms of syllables in an 8dimensional space such that syllables with the same label are closer together than syllables with different labels. We could report the loss values for this embedding task on our training and validation datasets, but these wouldn’t have any clear relevance to the downstream task of syllable distribution comparison where we are using the model’s embeddings. We report the contrast index as this has direct relevance to the actual application of the model and allows comparisons to other similarity scoring methods, something that the triplet loss values wouldn’t allow. 

The triplet loss method was chosen because it has been shown to yield useful lowdimensional representations of data, even in cases where there is limited labeled training data (Thakur et al. 2019). While we have one of the largest manually annotated datasets of zebra finch songs, it is still quite small by industry deep learning standards, which is why we chose a method that would perform well given the size of our dataset. Training a model on a contrast index directly would be extremely computationally intensive and require many more pairs of birds with known relationships than we currently have access to. It could be an interesting approach to take in the future, but one that would be unlikely to perform well with a dataset size typical to songbird research. 

Also, usually, in solid machine learning work, diverse methods are compared against each other to identify their relative strengths. The paper contains almost none of this, e.g. authors examined only one clustering method (HDBSCAN). 

We did compare multiple methods for syllable segmentation (WhisperSeg,  TweetyNet, and Amplitude thresholding) as this hadn’t been done previously. We chose not to perform extensive comparison of different clustering methods as Sainburg et al. 2020 already did so and we felt no need to reduplicate this effort. We encourage this reviewer to refer to Sainburg et al.’s excellent work for comparisons of multiple clustering methods applied to zebra finch song syllables.  

(3) Performance issues. The authors want to 'simplify large-scale behavioral analysis' but it seems they want to do that at a high cost. (Gu et al 2023) achieved syllable scores above 0.99 for adults, which is much larger than the average score of 0.88 achieved here (L121). Similarly, the syllable scores in (Cohen et al 2022) are above 94% (their error rates are below 6%, albeit in Bengalese finches, not zebra finches), which is also better than here. Why is the performance of AVN so low? The low scores of AVN argue in favor of some human labeling and training on each bird. 

Firstly, the syllable error rate scores reported in Cohen et al. 2022 are calculated very differently than the F1 scores we report here and are based on a model trained with data from the same bird as was used in testing, unlike our more general segmentation approach where the model was tested on different birds than were used in testing. Thus, the scores reported in Cohen et al. and the F1 scores that we report cannot be compared. 

The discrepancy between the F1seg scores reported in Gu et al. 2023 and the segmentation F1 scores that we report are likely due to differences in the underlying datasets. Our UTSW recordings tend to have higher levels of both stationary and nonstationary background noise, which make segmentation more challenging. The recordings from Rockefeller were less contaminated by background noise, and they resulted in slightly higher F1 scores. That said, we believe that the primary factor accounting for this difference in scores with Gu et al. 2023 is the granularity of our ‘ground truth’ syllable segments. In our case, if there was ever any ambiguity as to whether vocal elements should be segmented into two short syllables with a very short gap between them or merged into a single longer syllable, we chose to split them. WhisperSeg had a strong tendency to merge the vocal elements in ambiguous cases such as these. This results in a higher rate of false negative syllable onset detections, reflected in the low recall scores achieved by WhisperSeg (see supplemental figure 2b), but still very high precision scores (supplemental figure 2a). While WhisperSeg did frequently merge these syllables in a way that differed from our ground truth segmentation, it did so consistently, meaning it had little impact on downstream measures of syntax entropy (Fig 3c) or syllable duration entropy (supplemental figure 7a). It is for that reason that, despite a lower F1 score, we still consider AVN’s automatically generated annotations to be sufficiently accurate for downstream analyses. 

Should researchers require a higher degree of accuracy and precision with their annotations (for example, to detect very subtle changes in song before and after an acute manipulation) and be willing to dedicate the time and resources to manually labeling a subset of recordings from each of their birds, we suggest they turn toward one of the existing tools for supervised song annotation, such as TweetyNet.  

(4) Texas bias. It is true that comparability across datasets is enhanced when everyone uses the same code. However, the authors' proposal essentially is to replace the bias between labs with a bias towards birds in Texas. The comparison with Rockefeller birds is nice, but it amounts to merely N=1. If birds in Japanese or European labs have evolved different song repertoires, the AVN might not capture the associated song features in these labs well. 

We appreciate the reviewer’s concern about a bias toward birds from the UTSW colony. However, this paper shows that despite training (for the similarity scoring) and hyperparameter fitting (for the HDBSCAN clustering) on the UTSW birds, AVN performs as well if not better on birds from Rockefeller than from UTSW. To our knowledge, there are no publicly available datasets of annotated zebra finch songs from labs in Europe or in Asia but we would be happy to validate AVN on such datasets, should they become available. Furthermore, there is no evidence to suggest that there is dramatic drift in zebra finch vocal repertoire between continents which would necessitate such additional validation. While we didn’t have manual annotations for this dataset (which would allow validation of our segmentation and labeling methods), we did apply AVN to recordings share with us by the Wada lab in Japan, where visual inspection of the resulting annotations suggested comparable accuracy to the UTSW and Rockefeller datasets.  

(5) The paper lacks an analysis of the balance between labor requirement, generalizability, and optimal performance. For tasks such as segmentation and labeling, fine-tuning for each new dataset could potentially enhance the model's accuracy and performance without compromising comparability. E.g. How many hours does it take to annotate hundred song motifs? How much would the performance of AVN increase if the network were to be retrained on these? The paper should be written in more neutral terms, letting researchers reach their own conclusions about how much manual labor they want to put into their data. 

With standardization and ease of use in mind, we designed AVN specifically to perform fully automated syllable annotation and downstream feature calculations. We believe that we have demonstrated in this manuscript that our fully automated approach is sufficiently reliable for downstream analyses across multiple zebra finch colonies. That said, if researchers require an even higher degree of annotation precision and accuracy, they can turn toward one of the existing methods for supervised song annotation, such as TweetyNet. Incorporating human annotations for each bird processed by AVN is likely to improve its performance, but this would require significant changes to AVN’s methodology and is outside the scope of our current efforts.  

(6) Full automation may not be everyone's wish. For example, given the highly stereotyped zebra finch songs, it is conceivable that some syllables are consistently mis-segmented or misclassified. Researchers may want to be able to correct such errors, which essentially amounts to fine-tuning AVN. Conceivably, researchers may want to retrain a network like the AVN on their own birds, to obtain a more fine-grained discriminative method. 

Other methods exist for supervised or human-in-the-loop annotation of zebra finch songs, such as TweetyNet and DAN (Alam et al. 2023). We invite researchers who require a higher degree of accuracy than AVN can provide to explore these alternative approaches for song annotation. Incorporating human annotations for each individual bird being analyzed using AVN was never the goal of our pipeline, would require significant changes to AVN’s design, and is outside the scope of this manuscript.  

(7) The analysis is restricted to song syllables and fails to include calls. No rationale is given for the omission of calls. Also, it is not clear how the analysis deals with repeated syllables in a motif, whether they are treated as two-syllable types or one. 

It is true that we don’t currently have any dedicated features to describe calls. This could be a useful addition to AVN in the future. 

What a human expert inspecting a spectrogram would typically call ‘repeated syllables’ in a bout are almost always assigned the same syllable label by the UMAP+HDBSCAN clustering. The syntax analysis module includes features examining the rate of syllable repetitions across syllable types. See https://avn.readthedocs.io/en/latest/syntax_analysis_demo.html#SyllableRepetitions

(8) It seems not all human annotations have been released and the instruction sets given to experts (how to segment syllables and score songs) are not disclosed. It may well be that the differences in performance between (Gu et al 2023) and (Cohen et al 2022) are due to differences in segmentation tasks, which is why these tasks given to experts need to be clearly spelled out. Also, the downloadable files contain merely labels but no identifier of the expert. The data should be released in such a way that lets other labs adopt their labeling method and cross-check their own labeling accuracy. 

All human annotations used in this manuscript have indeed been released as part of the accompanying dataset. Syllable annotations are not provided for all pupils and tutors used to validate the similarity scoring, as annotations are not necessary for similarity comparisons. We will expand our description of our annotation guidelines in the methods section of the revised manuscript. All the annotations were generated by one of two annotators. The second annotator always consulted with the first annotator in cases of ambiguous syllable segmentation or labeling, to ensure that they had consistent annotation styles. Unfortunately, we haven’t retained records about which birds were annotated by which of the two annotators, so we cannot share this information along with the dataset. The data is currently available in a format that should allow other research groups to use our annotations either to train their own annotation systems or check the performance of their existing systems on our annotations.  

(9) The failure modes are not described. What segmentation errors did they encounter, and what syllable classification errors? It is important to describe the errors to be expected when using the method. 

As we discussed in our response to this reviewer’s point (3), WhisperSeg has a tendency to merge syllables when the gap between them is very short, which explains its lower recall score compared to its precision on our dataset (supplementary figure 2). In rare cases, WhisperSeg also fails to recognize syllables entirely, again impacting its precision score. TweetyNet hardly ever completely ignores syllables, but it does tend to occasionally merge syllables together or over-segment them. Whereas WhisperSeg does this very consistently for the same syllable types within the same bird, TweetyNet merges or splits syllables more inconsistently. This inconsistent merging and splitting has a larger effect on syllable labeling, as manifested in the lower clustering v-measure scores we obtain with TweetyNet compared to WhisperSeg segmentations. TweetyNet also has much lower precision than WhisperSeg, largely because TweetyNet often recognizes background noises (like wing flaps or hopping) as syllables whereas WhisperSeg hardly ever segments nonvocal sounds. 

Many errors in syllable labeling stem from differences in syllable segmentation. For example, if two syllables with labels ‘a’ and ‘b’ in the manual annotation are sometimes segmented as two syllables, but sometimes merged into a single syllable, the clustering is likely to find 3 different syllable types; one corresponding to ‘a’, one corresponding to ‘b’ and one corresponding to ‘ab’ merged. Because of how we align syllables across segmentation schemes for the v-measure calculation, this will look like syllable ‘b’ always has a consistent cluster label, but syllable ‘a’ can carry two different cluster labels, depending on the segmentation. In certain cases, even in the absence of segmentation errors, a group of syllables bearing the same manual annotation label may be split into 2 or 3 clusters (it is extremely rare for a single manual annotation group to be split into more than 3 clusters). In these cases, it is difficult to conclusively say whether the clustering represents an error, or if it actually captured some meaningful systematic difference between syllables that was missed by the annotator. Finally, sometimes rare syllable types with their own distinct labels in the manual annotation are merged into a single cluster. Most labeling errors can be explained by this kind of merging or splitting of groups relative to the manual annotation, not to occasional mis-classifications of one manual label type as another. 

For examples of these types of errors, we encourage this reviewer and readers to refer to the example confusion matrices in figure 2f and supplemental figure 4b&e. We will also expand our discussion of these different types of errors in the revised manuscript. 

(10) Usage of Different Dimensionality Reduction Methods: The pipeline uses two different dimensionality reduction techniques for labeling and similarity comparison - both based on the understanding of the distribution of data in lower-dimensional spaces. However, the reasons for choosing different methods for different tasks are not articulated, nor is there a comparison of their efficacy. 

We apologize for not making this distinction sufficiently clear in the manuscript and will add additional explanation to the main text to make the reasoning more apparent. We chose to use UMAP for syllable labeling because it is a common embedding methodology to precede hierarchical clustering and has been shown to result in reliable syllable labels for birdsong in the past (Sainburg et al. 2020). However, it is not appropriate for similarity scoring, because comparing EMD scores between birds requires that all the birds’ syllable distributions exist within the same shared embedding space. This can be achieved by using the same triplet loss-trained neural network model to embed syllables from all birds. This cannot be achieved with UMAP because all birds whose scores are being compared would need to be embedded in the same UMAP space, as distances between points cannot be compared across UMAPs. In practice, this would mean that every time a new tutor-pupil pair needs to be scored, their syllables would need to be added to a matrix with all previously compared birds’ syllables, a new UMAP would need to be computed, and new EMD scores between all bird pairs would need to be calculated using their new UMAP embeddings. This is very computationally expensive and quickly becomes unfeasible without dedicated high power computing infrastructure. It also means that similarity scores couldn’t be compared across papers without recomputing everything each time, whereas EMD scores obtained with triplet loss embeddings can be compared, provided they use the same trained model (which we provide as part of AVN) to embed their syllables in a common latent space.  

(11) Reproducibility: are the measurements reproducible? Systems like UMAP always find a new embedding given some fixed input, so the output tends to fluctuate. 

There is indeed a stochastic element to UMAP embeddings which will result in different embeddings and therefore different syllable labels across repeated runs with the same input. Anecdotally, we observed that v-measures scores were quite consistent within birds across repeated runs of the UMAP, but we will add an additional supplementary figure to the revised manuscript showing this.

Main Idea of Section: Heller's father highlighted San Francisco's changing urban scene by renting reasonably priced office space at the Ferry Building, which reflected the post-earthquake transition as underutilized places became sought-after real estate.

Reviewer #3 (Public review):

Summary:

The authors aimed to investigate the multifaceted roles of the Inferior Colliculus (IC) in auditory and cognitive processes in monkeys. Through extracellular recordings during a sound duration-based novelty detection task, the authors observed a "climbing effect" in neuronal firing rates, suggesting an enhanced response during sensory prediction. Observations of reward prediction errors within the IC further highlight its complex integration in both auditory and reward processing. Additionally, the study indicated IC neuronal activities could be involved in decision-making processes.

Strengths:

This study has the potential to significantly impact the field by challenging the traditional view of the IC as merely an auditory relay station and proposing a more integrative role in cognitive processing. The results provide valuable insights into the complex roles of the IC, particularly in sensory and cognitive integration, and could inspire further research into the cognitive functions of the IC.

Weaknesses:

Major Comments:

(1) Structural Clarity and Logic Flow:<br /> The manuscript investigates three intriguing functions of IC neurons: sensory prediction, reward prediction, and cognitive decision-making, each of which is a compelling topic. However, the logical flow of the manuscript is not clearly presented and needs to be well recognized. For instance, Figure 3 should be merged into Figure 2 to present population responses to the order of sounds, thereby focusing on sensory prediction. Given the current arrangement of results and figures, the title could be more aptly phrased as "Beyond Auditory Relay: Dissecting the Inferior Colliculus's Role in Sensory Prediction, Reward Prediction, and Cognitive Decision-Making."

(2) Clarification of Data Analysis:<br /> Key information regarding data analysis is dispersed throughout the results section, which can lead to confusion. Providing a more detailed and cohesive explanation of the experimental design would significantly enhance the interpretation of the findings. For instance, including a detailed timeline and reward information for the behavioral paradigms shown in Figures 1C and D would offer crucial context for the study. More importantly, clearly presenting the analysis temporal windows and providing comprehensive statistical analysis details would greatly improve reader comprehension.

(3) Reward Prediction Analysis:<br /> The conclusion regarding the IC's role in reward prediction is underdeveloped. While the manuscript presents evidence that IC neurons can encode reward prediction, this is only demonstrated with two example neurons in Figure 6. A more comprehensive analysis of the relationship between IC neuronal activity and reward prediction is necessary. Providing population-level data would significantly strengthen the findings concerning the IC's complex functionalities. Additionally, the discussion of reward prediction in lines 437-445, which describes IC neuron responses in control experiments, does not sufficiently demonstrate that IC neurons can encode reward expectations. It would be valuable to include the responses of IC neurons during trials with incorrect key presses or no key presses to better illustrate this point.

This work has been published in GigaByte Journal under a CC-BY 4.0 license (https://doi.org/10.46471/gigabyte.132), and has published the reviews under the same license. These are as follows.

Reviewer 1. Michael Hall

Is installation/deployment sufficiently outlined in the paper and documentation, and does it proceed as outlined?

Unable to test. The README states "If you want to test the RiboSnake functions yourself, you can use the same data used for the CI/CD tests." A worked example of how I can do this would be appreciated so I can test the workflow.

Is there enough clear information in the documentation to install, run and test this tool, including information on where to seek help if required?

The Usage instructions say to create a new repository using ribosnake as a template, but ribosnake is not a template repository (see https://docs.github.com/en/repositories/creating-and-managing-repositories/creating-a-repository-from-a-template). The README states "If you want to test the RiboSnake functions yourself, you can use the same data used for the CI/CD tests." A worked example of how I can do this would be appreciated so I can test the workflow.

Have any claims of performance been sufficiently tested and compared to other commonly-used packages?

Not applicable.

Is automated testing used or are there manual steps described so that the functionality of the software can be verified?

Yes, though as mentioned above, the README states "If you want to test the RiboSnake functions yourself, you can use the same data used for the CI/CD tests." A worked example of how I can do this would be appreciated so I can test the workflow.

Additional Comments:

The Introduction could be make far more concise, there's a lot of repetition.

The installation command in figure 1 is three commands, not two as stated in the text (third-last paragraph Introduction), and is slightly misleading from an installation point of view as it assumes conda and snakemake are installed. Though it is mentioned later in the text (p5) that snakemake and conda require manual installation.

The in-text citation for Greengenes2 is just [?] - maybe a latex issue?

The last paragraph of the 'Features and Implementations' section was mostly already stated earlier in the manuscript.

Make the colouring consistent between fig 2a-c and 2d as well as the vertical ordering to make for easier comparison. For example, in figures 2a-c Enterococcus (grey) is on the bottom, whereas in fig 2d it is red and in the middle. Colour legends should also be added to Figures 3-5 to match Fig 2.

A small table should be added showing the comparison of RiboSnake and the original publication for the top 10 most abundant phyla for the Atacama soil dataset and their abundances (see last paragraph of 'Usage and Findings'.

Reviewer 2. Yong-Xin Liu and Salsabeel Yousuf

The manuscript presented by the authors describes a comprehensive study on the “RiboSnake pipeline” for 16S rRNA gene microbiome analysis, which is a user-friendly, robust, and multipurpose. RiboSnake, a validated, automated, reproducible QIIME2-based analysis pipeline implemented in Snakemake, offers parallel processing for efficient analysis of large datasets in both environmental and medical research contexts. Further demonstrating its effectiveness, this pipeline effectively analyzes human-associated microbiomes and environmental samples like wastewater and soil, thus expanding the scope of analysis for 16S rRNA data. The overall computational pipeline is useful and results are sound, validated through rigorous testing on MOCK communities and real-world datasets. However, there are some issues for improvement in the manuscript.

Major comments: 1． In the clinical data section the author mentions rectal swabs were used from a published study [31]. While the source is referenced, it would be helpful to know if any information was provided in the referenced study regarding the collection methods or storage conditions for the rectal swabs. 2． The text mentions using cotton swabs pre-moistened with TE buffer + 0.5% Tween 20. While cotton swabs are common, are there any considerations for using different swab materials depending on the target analytes or sampling surface (e.g., flocked swabs for better epithelial cell collection)? 3． Does RiboSnake require user intervention during any steps, or is it fully automated? 4． The author mentions that contamination filtering parameters should be adjusted based on the sample type. How can users determine the appropriate filtering parameters for their specific samples? Are there guidelines for users to know how much adjustment is needed for specific scenarios? 5． The default abundance threshold for filtering low-frequency reads is chosen based on Nearing et al. [44]. Please discuss the rationale behind using a single threshold for all sample types? Would it be beneficial to allow users to define this threshold based on their data characteristics? 6． Would you like to explain the limitation of RiboSnake, such as specific types of samples it may not be suitable for or potential biases introduced by certain functionalities? 7． The manuscript mentions various visualization tools used throughout the pipeline (QIIME2, qurro). Please clarify which types of data are visualized with each tool, and how users can access or customize these visualizations? 8． To strengthen the manuscript's impact, consider discussing the specific novelty of RiboSnake compared to existing 16S rRNA gene microbiome analysis pipelines. Would you be able to elaborate on the unique features or functionalities of RiboSnake that address limitations of current methods? 9． EasyAmplicon is recently published pipeline and easy using in windows, mac and linux system,

Minor comments: 1. Reference is missing in this sentence. “The default is the SILVA database [47]. Greengenes2 [? ] can be used alternatively”. 2. The author should careful about the lowercase and upper case throughout the manuscript. Please check the following for references:  ..the 2017 published Atacama Soil data set with samples taken fromthe Atacama desert was used [32] as well as samples collected fromsoil under switchgrass published in [33].  based on an Euclidean beta diversitymetric, shows that the positive controls, as well as the samples taken from subjects 1 and 3 (S1 and S3), cluster together.  A wide range of diversity analysis parameters are available in QIIME2 and its associated tools. These include the Shannon diversity index to measure richness, the Pielou index tomeasure evenness, or perform standard correlation analysis using Pearson or Spearman indices, among others. 3. In the introduction part this sentences “However, while these methods enable 16S rRNA analysis with minimal user interaction…” needs attention for clarity. Consider separating it into two sentences to emphasize the limitations of existing pipelines compared to the described methods’. Alternatively, using contrasting words like "in contrast" could highlight these differences. 4. More detail in attached PDF.

https://gigabyte-review.rivervalleytechnologies.comdownload-api-file?ZmlsZV9wYXRoPXVwbG9hZHMvZ3gvVFIvNTM5L2d4LVRSLTE3MTY5Nzk4MTktcmV2aXNlZC5wZGY=

Re-review: The author's response has been fully addressed my concerns. The quality of the paper has apparently improved. I agree with the publication of this article.

Reviewer #2 (Public review):

Deciphering the metabolic alterations characterizing the prediabetes-diabetes spectrum could provide early time windows for targeted preventive measures to extend precision medicine while avoiding disproportionate healthcare costs. The authors identified a panel of 9 circulating metabolites combined with basic clinical variables that significantly improved the prediction from prediabetes to diabetes. These findings provided insights into the integration of these metabolites into clinical and public health practice.

Comments on the revised version:

Congratulations to the authors. I have no more comments.

Author response:

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

Reviewer #1 (Public Review):

Using the UK Biobank, this study assessed the value of nuclear magnetic resonance measured metabolites as predictors of progression to diabetes. The authors identified a panel of 9 circulating metabolites that improved the ability in risk prediction of progression from prediabetes to diabetes. In general, this is a well-performed study, and the findings may provide a new approach to identifying those at high risk of developing diabetes. I have some comments that may improve the importance of this study.

We deeply appreciate the reviewer's invaluable time dedicated to the review of this manuscript and the insightful comments to enhance its overall quality.

(1) It is unclear why the authors only considered the top 20 variables in the metabolite selection and why they did not set a wider threshold.

Thank you for the comment. We set the top 20 variables in the metabolite selection balancing the performance of the final diabetes risk prediction model and the clinical applicability due to measurement costs. We have added this explanation in the “Methods” section.

“We chose the intersection set of the top 20 most important variables selected by the three machine learning models, after balancing the performance of the final diabetes risk prediction model and the clinical applicability associated with measurement costs of metabolites.”

(2) The methods section would benefit from a more detailed exposition of how parameter tuning was conducted and the range of parameters explored during the training of the RSF model.

According to the reviewer’s suggestion, we have added a more detailed description of parameters tunning and the range of parameters explored during the training of the RSF model in the “Method S3” section in the Supplementary material.

“The RSF model was fitted using the “randomForestSRC” package and the grid search method was used for hyperparameter tuning. Specifically, the grid search method was used to tune hyperparameters among the RSF model, through minimizing out-of-sample or out-of-bag error1. Each tree in the RSF is constructed from a random sample of the data, typically a bootstrap sample or 63.2% of the sample size (as in the present study). Consequently, not all observations are used to construct each tree. The observations that are not used in the construction of a tree are referred to as out-of-bag observations. In an RSF model, each tree is built from a different sample of the original data, so each observation is “out-of-bag” for some of the trees. The prediction for an observation can then be obtained using only those trees for which the observation was not used for the construction. A classification for each observation is obtained in this way and the error rate can be estimated from these predictions. The resulting error rate is referred to as the out-of-bag error. Through calculating the out-of-bag error in each iteration, the best hyperparameters were finally determined.

The hyperparameters to be tuned and range of grid search in the present study were below: number of trees (50-1000, by 50), number of variables to possibly split at each node (3-6, by 1), and minimum size of terminal node (1-20, by 1)2.”

(3) It is hard to understand the meaning of the decision curve analysis and the clinical implications behind the net benefit, which are required to clarify the application values of models.

Thank you for the comment. We have added more description and discussion about the decision curve analysis in the “Methods” and “Discussion” sections.

“Furthermore, we used decision curve analysis (DCA) to assess the clinical usefulness of prediction model-based guidance for prediabetes management, which calculates a clinical “net benefit” for one or more prediction models in comparison to default strategies of treating all or no patients3.”

“Most importantly, a model with good discrimination does not necessarily have high clinical value. Hence, DCA was used to compare the clinical utility of the model before and after adding the metabolites, and this showed a higher net benefit for the latter than the basic model, suggesting the addition of the metabolites increased the clinical value of prediction, i.e., the potential benefit of guiding management in individuals with prediabetes3,4. These results provided novel evidence supporting the value of metabolic biomarkers in risk prediction and stratification for the progression from prediabetes to diabetes.”

(4) Notably, the NMR platform utilized within the UK Biobank primarily focused on lipid species. This limitation should be discussed in the manuscript to provide context for interpreting the results and acknowledge the potential bias from the measuring platform.

Thank you for the comment. We acknowledged this limitation that NMR platform within the UK Biobank primarily focused on lipid species and the potential bias from the measuring platform and have added this in “Discussion” section.

“Third, the Nightingale metabolomics platform primarily focused on lipids and lipoprotein sub-fractions, and thus the predictive value of other metabolites in the progression from prediabetes to diabetes warranted further research using an untargeted metabolomics approach.”

(5) The manuscript should explain the potential influence of non-fasting status on the findings, particularly concerning lipoprotein particles and composition. There should be a detailed discussion of how non-fasting status may impact the measurement and the findings.

According to the reviewer’s suggestion, we have added more details to explain the potential influence of non-fasting status on our findings in the “Discussion” section.

“Additionally, the use of non-fasting blood samples might increase inter-individual variation in metabolic biomarker concentrations, however, fasting duration has been reported to account for only a small proportion of variation in plasma metabolic biomarker concentrations5. Therefore, we believe the impact of non-fasting samples on our findings would be minor.”

(6) Cross-platform standardization is an issue in metabolism, and further descriptions of quality control are recommended.

Thank you for the comment. We have added more description of quality control in the “Method S1” section in the Supplementary material.

“Metabolic biomarker profiling by Nightingale Health’s NMR platform provides consistent results over time and across spectrometers. Furthermore, the sample preparation is minimal in the Nightingale Health’s metabolic biomarker platform, circumventing all extraction steps. These aspects result in highly repeatable biomarker measurements. Pre-specified quality metrics were agreed between UK Biobank and Nightingale Health to ensure consistent results across the samples, and pilot measurements were conducted. Nightingale Health performed real-time monitoring of the measurement consistency within and between spectrometers throughout the UK Biobank samples. Two control samples provided by Nightingale Health were included in each 96-well plate for tracking the consistency across multiple spectrometers. Furthermore, two blind duplicate samples provided by the UK Biobank were included in each well plate, with the position information unlocked only after results delivery. Coefficient of variation (CV) targets across the metabolic biomarker profile were pre-specified for both Nightingale Health’s internal control samples and UK Biobank’s blind duplicates. The targets were met for each consecutively measured batch of ~25,000 samples. For the majority of the metabolic biomarkers, the CVs were below 5% (https://biobank.ndph.ox.ac.uk/showcase/refer.cgi?id=3000). Further, the distributions of measured biomarkers from 5 sample batches indicated absence of batch effects (https://biobank.ctsu.ox.ac.uk/ukb/ukb/docs/nmrm_app1).”

Reviewer #2 (Public Review):<br /> Deciphering the metabolic alterations characterizing the prediabetes-diabetes spectrum could provide early time windows for targeted preventive measures to extend precision medicine while avoiding disproportionate healthcare costs. The authors identified a panel of 9 circulating metabolites combined with basic clinical variables that significantly improved the prediction from prediabetes to diabetes. These findings provided insights into the integration of these metabolites into clinical and public health practice. However, the interpretation of these findings should take account of the following limitations.

We appreciate the reviewer’s positive comments and encouragement.

(1) First, the causal relationship between identified metabolites and diabetes or prediabetes deserves to be further examined particularly when the prediabetic status was partially defined. Some metabolites might be the results of prediabetes rather than the casual factors for progression to diabetes.

Thank you for your insightful comments. We agree with you that the panel of metabolites in this study might not be the causal factor for progression from prediabetes to diabetes, which needs further validation in experimental studies. We have added this limitation in the “Discussion” section.

“Fifth, we could not draw any conclusion about the causality between the identified metabolites and the risk for progression to diabetes due to the observational nature, which remained to be validated in further experimental studies.”

(2) The blood samples were taken at random (not all in a non-fasting state) and so the findings were subjected to greater variability. This should be discussed in the limitations.

According to the reviewer’s suggestion, we have added more details to explain the potential influence of non-fasting status on our findings in the “Discussion” section.

“Additionally, the use of non-fasting blood samples might increase inter-individual variation in metabolic biomarker concentrations, however, fasting duration has been reported to account for only a small proportion of variation in plasma metabolic biomarker concentrations5. Therefore, we believe the impact of non-fasting samples on our findings would be minor.”

(3) The strength of NMR in metabolic profiling compared to other techniques (i.e., mass spectrometry [MS], another commonly used metabolic profiling method) could be added in the Discussion section.

According to the reviewer’s suggestion, we have added the strength of NMR in metabolic profiling compared to other techniques in the “Discussion” section.

“Circulating metabolites were quantified via NMR-based metabolome profiling within the UK Biobank, which offers metabolite qualification with relatively lower costs and better reproducibility6.”

(4) Fourth, the applied platform focuses mostly on lipid species which may be a limitation as well.

Thank you for the comment. We acknowledged this limitation that NMR platform within the UK Biobank primarily focused on lipid species and the potential bias from the measuring platform and have added this in the “Discussion” section.

“Third, the Nightingale metabolomics platform primarily focused on lipids and lipoprotein sub-fractions, and thus the predictive value of other metabolites in the progression from prediabetes to diabetes warranted further research using an untargeted metabolomics approach.”

(5) It is a very large group with pre-diabetes, but the results only apply to prediabetes and not to the general population. This should be clear, although the authors have also validated the predictive value of these metabolites in the general population.

Thank you for the comment. We agree with you that the results only apply to prediabetes and not to the general population, though they also showed potential predictive value among participants with normoglycemia. We have accordingly modified the relevant expressions in the “Conclusion” section to restrict these findings to participants with prediabetes.

“In this large prospective study among individuals with prediabetes, we detected a panel of circulating metabolites that were associated with an increased risk of progressing to diabetes.”

Recommendations for the Authors:

Thank you for providing the valuable feedback and the time you have dedicated to our work.

(1) In the first paragraph of the Discussion section, please include the specific names of the metabolites selected from machine learning methods.

Thank you for your comment and we have added accordingly in the first paragraph of the “Discussion” section.

“More importantly, our findings suggested that adding the selected metabolites (i.e., cholesteryl esters in large HDL, cholesteryl esters in medium VLDL, triglycerides in very large VLDL, average diameter for LDL particles, triglycerides in IDL, glycine, tyrosine, glucose, and docosahexaenoic acid) could significantly improve the risk prediction of progression from prediabetes to diabetes beyond the conventional clinical variables.”

(2) To enhance the readability and simplicity of the paper, the description of covariate collection in the methods section should be streamlined, with detailed information provided in the supplementary materials.

Thank you for your suggestion and we have moved details about covariates collection to the “Supplementary method S2” to enhance the readability and simplicity of the paper.

“Information on covariates was collected through a self-completed touchscreen questionnaire or verbal interview at baseline, including age, sex, ethnicity, Townsend deprivation index, household income, education, employment status, smoking status, moderate alcohol, physical activity, healthy diet score, healthy sleep score, family history of diabetes, history of cardiovascular disease (CVD), history of hypertension, history of dyslipidemia, history of chronic lung diseases (CLD), and history of cancer.

Physical measurements included systolic (SBP) and diastolic blood pressure (DBP), height, weight, waist circumference (WC), and hip circumference (HC). Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters (kg/m²). Missing covariates were imputed by the median value for continuous variables and a missing indicator for categorical variables. More details about covariates collection can be found in Method S2.”

3. Title for Table 2, using Cox proportional hazards prediction models is not common. You may consider the title "Performance of Cox proportional hazards regression models in prediction of progression of prediabetes to diabetes".

Thank you for your suggestion and we have revised it accordingly.

4. Figure 3, did the authors consider competing risk to compute cumulative incidence function?

Thank you for your comment. We did not consider competing risk from death when plotting the cumulative hazard curves. However, following your suggestion, we have included an additional cumulative hazard plot after considering the competing

References

(1) Janitza S, Hornung R. On the overestimation of random forest's out-of-bag error. PLoS One. 2018;13(8):e0201904.

(2) Tian D, Yan HJ, Huang H, et al. Machine Learning-Based Prognostic Model for Patients After Lung Transplantation. JAMA Netw Open. 2023;6(5):e2312022.

(3) Vickers AJ, van Calster B, Steyerberg EW. A simple, step-by-step guide to interpreting decision curve analysis. Diagn Progn Res. 2019;3:18.

(4) Li J, Xi F, Yu W, Sun C, Wang X. Real-Time Prediction of Sepsis in Critical Trauma Patients: Machine Learning-Based Modeling Study. JMIR Form Res. 2023;7:e42452.

(5) Li-Gao R, Hughes DA, le Cessie S, et al. Assessment of reproducibility and biological variability of fasting and postprandial plasma metabolite concentrations using 1H NMR spectroscopy. PLoS One. 2019;14(6):e0218549.

(6) Geng T-T, Chen J-X, Lu Q, et al. Nuclear Magnetic Resonance–Based Metabolomics and Risk of CKD. American Journal of Kidney Diseases. 2023.

Get that Linux feeling - on Windows

Reviewer #3 (Public Review):

Summary:<br /> The authors introduce a tool that employs thermal cameras to automatically detect urine and feces deposits in rodents. The detection process involves a heuristic to identify potential thermal regions of interest, followed by a transformer network-based classifier to differentiate between urine, feces, and background noise. The tool's effectiveness is demonstrated through experiments analyzing social preference, stress response, and temporal dynamics of deposits, revealing differences between male and female mice.

Strengths:<br /> The method effectively automates the identification of deposits<br /> The application of the tool in various behavioral tests demonstrates its robustness and versatility.<br /> The results highlight notable differences in behavior between male and female mice

Weaknesses:<br /> The definition of 'start' and 'end' periods for statistical analysis is arbitrary. A robustness check with varying time windows would strengthen the conclusions.<br /> The paper could better address the generalizability of the tool to different experimental setups, environments, and potentially other species.<br /> The results are based on tests of individual animals, and there is no discussion of how this method could be generalized to experiments tracking multiple animals simultaneously in the same arena (e.g., pair or collective behavior tests, where multiple animals may deposit urine or feces).

Author response:

We want to thank the reviewers for their constructive feedback.

General

The recall values of our method range between 78.6% for all urine cases to 83.3% for feces (and not between 70-80%, as stated by reviewer #2), with a mean precision of 85.6%. This is rather similar to other machine learning-based methods commonly used for the analysis of complicated behavioral readouts. For example, in the paper presenting DeepSqueak for analysis of mouse ultrasonic vocalizations (Coffey et al. DeepSqueak: a deep learning-based system for detection and analysis of ultrasonic vocalizations. Neuropsychopharmacol. 44, 859–868 (2019). https://doi.org/10.1038/s41386-018-0303-6), the recall values reported for both DeepSqueak, Mupet and Ultravox (Fig. 2c, f) are very similar to our method.

We have analyzed and reported all the types of errors made by our methods, which are mostly technical. For example, depositions that overlap the mouse blob for too long till getting cold will be associated with the mouse and therefore will not be detected (“miss” events). These technical errors are not supposed to create a bias for a specific biological condition and, hence, shouldn’t interfere with the use of our method. A video showing all of the mistakes made by our algorithm on the test set was submitted (Figure 2-video 1).

Below we will to relate to specific points and describe our plan to revise the manuscript accordingly.

Detection accuracy

a. It should be noted that when large urine spots are considered, our algorithm got 100% correct classification (Figure 2, supplement 1, panel b). However, small urine deposits are very similar to feces in their appearance in the thermal picture. In fact,  if the feces are not shifted, discrimination can be quite challenging even for human annotators. To demonstrate the accuracy of the proposed method relative to human annotators, we plan to compare its results with the accuracy of a second human annotator.

b. As part of the revision, we plan to test general machine learning-based object detectors such as faster-RCNN or YOLO (as suggested by Reviewer 2) and compare them with our method.

c. To check if our method may introduce bias to the results, we plan to check if the errors are distributed evenly across time, space, and genders.

Design choices

(A) The preliminary detection algorithm has several significant parameters. These are:

a. Minimal temperature rise for detection: 1.1°C rise during 5 sec.

b. Size limits of the detection: 2 - 900 pixels.

c. Minimal cooldown during 40 sec: 1.1°C and at least half the rise.

d. Minimal time between detections in the same location: 30 sec.

We chose to use low thresholds for the preliminary detection to allow detection of very small urinations and to minimize the number of “miss” events, relying on the classifier to robustly reject false alarms. Indeed, we achieved a low rate of miss events: 5 miss events for the entire test set (1 miss event per ~90 minutes of video). We attribute these 5 “miss” events to partial occlusion of the detection by the mouse.

To adjust the preliminary detection parameters to a new environment, one will need to calibrate these parameters in their own setup. Mainly, the size of the detection depends on the resolution of the video, and the cooldown rate might be affected by the material of the floor, as well as the room temperature.

We plan to explore the robustness of these parameters in our setup and report the influence on the accuracy of the preliminary algorithm.

(B) We chose to feed the classifier with 71 seconds of videos (11 seconds before the event and 60 seconds after it) as we wanted the classifier to be able to capture the moment of the deposition, the cooldown process, as well as urine smearing or feces shifting which might give an additional clue for the classification. In the revised paper we plan to report accuracy when using a shorter video for classification.

Generability

a. In the revised version, we plan to report the accuracy of the method used on a different strain of mice (C57), with a different arena color (white arena instead of black).

Statistics

a. In the revised paper, we will explain why we chose each time window for analysis. Also, we will report statistics for different time windows, as suggested by Reviewer 3.

b. Unlike reviewer #2, we don’t think that the small difference in recall rate between urine and feces (78.6% vs. 83.3%, respectively) creates a bias between them. Moreover, we don’t compare the urine rate to the feces rate.

c. In the revised manuscript we will explicitly report the precision scores, although they also appear in our manuscript in Fig. 2- Supplement 1b.

Reviewer #1 (Public Review):

The authors are attempting to use the internal workings of a language hierarchy model, comprising phonemes, syllables, words, phrases, and sentences, as regressors to predict EEG recorded during listening to speech. They also use standard acoustic features as regressors, such as the overall envelope and the envelopes in log-spaced frequency bands. This is valuable and timely research, including the attempt to show differences between normal-hearing and hearing-impaired people in these regards.

I will start with a couple of broader questions/points, and then focus my comments on three aspects of this study: The HM-LSTM language model and its usage, the time windows of relevant EEG analysis, and the usage of ridge regression.

Firstly, as far as I can tell, the OSF repository of code, data, and stimuli is not accessible without requesting access. This needs to be changed so that reviewers and anybody who wants or needs to can access these materials.

What is the quantification of model fit? Does it mean that you generate predicted EEG time series from deconvolved TRFs, and then give the R2 coefficient of determination between the actual EEG and predicted EEG constructed from the convolution of TRFs and regressors? Whether or not this is exactly right, it should be made more explicit.

About the HM-LSTM:

• In the Methods paragraph about the HM-LSTM, a lot more detail is necessary to understand how you are using this model. Firstly, what do you mean that you "extended" it, and what was that procedure? And generally, this is the model that produces most of the "features", or regressors, whichever word we like, for the TRF deconvolution and EEG prediction, correct? A lot more detail is necessary then, about what form these regressors take, and some example plots of the regressors alongside the sentences.<br /> • Generally, it is necessary to know what these regressors look like compared to other similar language-related TRF and EEG/MEG prediction studies. Usually, in the case of e.g. Lalor lab papers or Simon lab papers, these regressors take the form of single-sample event markers, surrounded by zeros elsewhere. For example, a phoneme regressor might have a sample up at the onset of each phoneme, and a word onset regressor might have a sample up at the onset of each word, with zeros elsewhere in the regressor. A phoneme surprisal regressor might have a sample up at each phoneme onset, with the value of that sample corresponding to the rarity of that phoneme in common speech. Etc. Are these regressors like that? Or do they code for these 5 linguistic levels in some other way? Either way, much more description and plotting is necessary in order to compare the results here to others in the literature.<br /> • You say that the 5 regressors that are taken from the trained model's hidden layers do not have much correlation with each other. However, the highest correlations are between syllable and sentence (0.22), and syllable and word (0.17). It is necessary to give some reason and interpretation of these numbers. One would think the highest correlation might be between syllable and phoneme, but this one is almost zero. Why would the syllable and sentence regressors have such a relatively high correlation with each other, and what form do those regressors take such that this is the case?<br /> • If these regressors are something like the time series of zeros along with single sample event markers as described above, with the event marker samples indicating the onset of the relevant thing, then one would think e.g. the syllable regressor would be a subset of the phoneme regressor because the onset of every syllable is a phoneme. And the onset of every word is a syllable, etc.

For the time windows of analysis:

• I am very confused, because sometimes the times are relative to "sentence onset", which would mean the beginning of sentences, and sometimes they are relative to "sentence offset", which would mean the end of sentences. It seems to vary which is mentioned. Did you use sentence onsets, offsets, or both, and what is the motivation?<br /> • If you used onsets, then the results at negative times would not seem to mean anything, because that would be during silence unless the stimulus sentences were all back to back with no gaps, which would also make that difficult to interpret.<br /> • If you used offsets, then the results at positive times would not seem to mean anything, because that would be during silence after the sentence is done. Unless you want to interpret those as important brain activity after the stimuli are done, in which case a detailed discussion of this is warranted.<br /> • For the plots in the figures where the time windows and their regression outcomes are shown, it needs to be explicitly stated every time whether those time windows are relative to sentence onset, offset, or something else.<br /> • Whether the running correlations are relative to sentence onset or offset, the fact that you can have numbers outside of the time of the sentence (negative times for onset, or positive times for offset) is highly confusing. Why would the regressors have values outside of the sentence, meaning before or after the sentence/utterance? In order to get the running correlations, you presumably had the regressor convolved with the TRF/impulse response to get the predicted EEG first. In order to get running correlation values outside the sentence to correlate with the EEG, you would have to have regressor values at those time points, correct? How does this work?<br /> • In general, it seems arbitrary to choose sentence onset or offset, especially if the comparison is the correlation between predicted and actual EEG over the course of a sentence, with each regressor. What is going on with these correlations during the middle of the sentences, for example? In ridge regression TRF techniques for EEG/MEG, the relevant measure is often the overall correlation between the predicted and actual, calculated over a longer period of time, maybe the entire experiment. Here, you have calculated a running comparison between predicted and actual, and thus the time windows you choose to actually analyze can seem highly cherry-picked, because this means that most of the data is not actually analyzed.<br /> • In figures 5 and 6, some of the time window portions that are highlighted as significant between the two lines have the lines intersecting. This looks like, even though you have found that the two lines are significantly different during that period of time, the difference between those lines is not of a constant sign, even during that short period. For instance, in figure 5, for the syllable feature, the period of 0 - 200 ms is significantly different between the two populations, correct? But between 0 and 50, normal-hearing are higher, between 50 and 150, hearing-impaired are higher, and between 150 and 200, normal-hearing are higher again, correct? But somehow they still end up significantly different overall between 0 and 200 ms. More explanation of occurrences like these is needed.

Using ridge regression:

• What software package(s) and procedure(s) were specifically done to accomplish this? If this is ridge regression and not just ordinary least squares, then there was at least one non-zero regularization parameter in the process. What was it, how did it figure in the modeling and analysis, etc.?<br /> • It sounds like the regressors are the hidden layer activations, which you reduced from 2,048 to 150 non-acoustic, or linguistic, regressors, per linguistic level, correct? So you have 150 regressors, for each of 5 linguistic levels. These regressors collectively contribute to the deconvolution and EEG prediction from the resulting TRFs, correct? This sounds like a lot of overfitting. How much correlation is there from one of these 150 regressors to the next? Elsewhere, it sounds like you end up with only one regressor for each of the 5 linguistic levels. So these aspects need to be clarified.<br /> • For these regressors, you are comparing the "regression outcomes" for different conditions; "regression outcomes" are the R2 between predicted and actual EEG, which is the coefficient of determination, correct? If this is R2, how is it that you have some negative numbers in some of the plots? R2 should be only positive, between 0 and 1.

Ni la pregunta ni la respuesta capturan el contexto.

有空试试 Windows电脑好久不用了，重新梳理一下使用场景

Reviewer #3 (Public Review):

Summary:

This paper aims to investigate how the human brain represents different forms of value and uncertainty that participate in active inference within a free-energy framework, in a two-stage decision task involving contextual information sampling, and choices between safe and risky rewards, which promotes shifting between exploration and exploitation. They examine neural correlates by recording EEG and comparing activity in the first vs second half of trials and between trials in which subjects did and did not sample contextual information, and perform a regression with free-energy-related regressors against data "mapped to source space."

Strengths:

This two-stage paradigm is cleverly designed to incorporate several important processes of learning, exploration/exploitation and information sampling that pertain to active inference. Although scalp/brain regions showing sensitivity to the active-inference related quantities do not necessary suggest what role they play, they are illuminating and useful as candidate regions for further investigation. The aims are ambitious, and the methodologies impressive. The paper lays out an extensive introduction to the free energy principle and active inference to make the findings accessible to a broad readership.

Weaknesses:<br /> In its revised form the paper is complete in providing the important details. Though not a serious weakness, it is important to note that the high lower-cutoff of 1 Hz in the bandpass filter, included to reduce the impact of EEG noise, would remove from the EEG any sustained, iteratively updated representation that evolves with learning across trials, or choice-related processes that unfold slowly over the course of the 2-second task windows.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

Summary:<br /> I really enjoyed this manuscript from Torsekar et al on "Contrasting responses to aridity by

different-sized decomposers cause similar decomposition rates across a precipitation gradient". The authors aimed to examine how climate interacts with decomposers of different size categories to influence litter decomposition. They proposed a new hypothesis: "The opposing climatic dependencies of macrofauna and that of microorganisms and mesofauna should lead to similar overall decomposition rates across precipitation gradients".

This study emphasizes the importance as well as the contribution of different groups of organisms (micro, meso, macro, and whole community) across different seasons (summer with the following characteristics: hot with no precipitation, and winter with the following characteristics: cooler and wetter winter) along a precipitation gradient. The authors made use of 1050 litter baskets with different mesh sizes to capture decomposers contribution. They proposed a new hypothesis that was aiming to understand the "dryland decomposition conundrum". They combined their decomposition experiment with the sampling of decomposers by using pittfall traps across both experiment seasons. This study was carried out in Israel and based on a single litter species that is native to all seven sites. The authors found that microorganism contribution dominated in winter while macrofauna decomposition dominated the overall decomposition in summer. These seasonality differences combined with the differences in different decomposers groups fluctuation along precipitation resulted in similar overall decomposition rates across sites.<br /> I believe this manuscript has a potential to advance our knowledge on litter decomposition.

Strengths:

Well design study with combination of different approaches (methods) and consideration of seasonality to generalize pattern.

The study expands to current understanding of litter decomposition and interaction between factors affecting the process (here climate and decomposers).

Weaknesses:

The study was only based on a single litter species.

We now discuss the advantages and limitations of this approach in the methods and devote a completely new paragraph to this important point in the discussion (lines 394-401).

Reviewer #2 (Public Review):

Summary: Torsekar et al. use a leaf litter decomposition experiment across seasons, and in an aridity gradient, to provide a careful test of the role of different-sized soil invertebrates in shaping the rates of leaf litter decomposition. The authors found that large-sized invertebrates are more active in the summer and small-sized invertebrates in the winter. The summed effects of all invets then translated into similar levels of decomposition across seasons. The system breaks down in hyper-arid sites.

Strengths: This is a well-written manuscript that provides a complete statistical analysis of a nice dataset. The authors provide a complete discussion of their results in the current literature.

Weaknesses:

I have only three minor comments. Please standardize the color across ALL figures (use the same color always for the same thing, and be friendly to color-blind people).

Thank you for this important suggestion. We have now changed all figures to standardize all colors and chose a more color-blind friendly pallete.

Fig 1 may benefit from separating the orange line (micro and meso) into two lines that reflect your experimental setup and results. I would mention the dryland decomposition conundrum earlier in the Introduction.

We based our novel hypotheses on a thorough literature search. Accordingly, decomposition is expected to be positively associated with moisture, regardless of the decomposer body size. Our contribution to theory was to suggest that macro-detritivores may respond very differently to climatic conditions and dominate litter decomposition in warm arid-lands (we listed the reasons in the text). Consequently, we did not distinguish between microorganisms and mesofauna. We assumed that both groups inhabit the litter substrate and have limited adaptation to dry conditions. Our results provide strong evidence that this presumption is likely wrong and that mesofauna respond to climate very differently from micro-decomposers. Yet, we cannot use hindsight understanding to improve our original hypothesis. We now emphasize this important point at the discussion as important future direction. 

Although we are very appreciative and pleased with the reviewer enthusiasm to highlight the importance of our work as a possible solution to the longstanding dryland decomposition conundrum, we decided not to move it to the introduction. This is because we think that our work is not centred on resolving the DDC but provides more general principles that may lead to a paradigm shift in the way ecologists study nutrient cycling across ecosystems.

And the manuscript is full of minor grammatical errors. Some careful reading and fixing of all these minor mistakes here and there would be needed.

We apologize and did our best to find and fix those mistakes

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

I really enjoyed this manuscript from Torsekar et al on "Contrasting responses to aridity by different-sized decomposers cause similar decomposition rates across a precipitation gradient". The authors aimed to examine how climate interacts with decomposers of different size categories to influence litter decomposition. They proposed a new hypothesis: "The opposing climatic dependencies of macrofauna and that of microorganisms and mesofauna should lead to similar overall decomposition rates across precipitation gradients".

This study emphasizes the importance as well as the contribution of different groups of organisms (micro, meso, macro, and whole community) across different seasons (summer with the following characteristics: hot with no precipitation, and winter with the following characteristics: cooler and wetter winter) along a precipitation gradient. The authors made use of 1050 litter baskets with different mesh sizes to capture decomposers contribution. They proposed a new hypothesis that was aiming to understand the "dryland decomposition conundrum". They combined their decomposition experiment with the sampling of decomposers by using pitfall traps across both experiment seasons. This study was carried out in Israel and based on a single litter species that is native to all seven sites. The authors found that microorganism contribution dominated in winter while macrofauna decomposition dominated the overall decomposition in summer. These seasonality differences combined with the differences in different decomposers groups fluctuation along precipitation resulted in similar overall decomposition rates across sites.

I believe this manuscript has the potential to advance our knowledge on litter decomposition. Below i provide my general and specific comments.

General comments:

(1) Study in general is well designed and well thought beforehand,

(2) Study aims to expand the current understanding of the dryland decomposition conundrum

(3) The should put a caveat to the fact they only use one litter species and call for examining litter mixture in the same gradient.

(4) Please check the way you reduce the random effects from your initial model, I have provided a better way to do so in my specific comments

(5) For Figure 1, authors can check my comment on this and see if they could revise the figure.

Thank you for the positive feedback and your valuable comments. We have tried to best address all comments and suggestions for improvement and clarification

Specific comments

Line # 57 Please write "Theory suggests" instead of "Theory suggest"

We changed the text as suggested

Line # 70, please write "Indeed, handful evidence shows" instead of "Indeed, handful evidence show"

We changed the text as suggested

Figure 1: I like this conceptual framework. I have a silly question, why is it that the slopes of the whole community at the beginning (between Hyperarid and Arid) is the same as the Macro fauna, I would think the slope should be higher as this is adding up right? and also the same goes for the decomposition of whole community later on. For me this should reflect the adding or summing up (if i am right) then the authors should think about how this could be reflected in the figure.

We agree with your interpretation that the whole community decomposition reflects the addition by constituent decomposers. The slope of the whole community decomposition between hyper-arid and arid is slightly higher than the one of macro decomposition to reflect the additive effect of macro with meso+micro decomposition. We have now changed the figure slightly to make this point more visible (Line 106).

Line # 111 Please make "Methods" bold as well to be consistent with others headings.

We changed the formatting as suggested

Line #125 and in other lines as well please replace "X" by "x" to denote multiplication.

We changed the formatting as suggested

Table 1 Please add "*" to climate like this "Climate*" so that the end note of the table could make sense

Thank you for this suggestion. We have now added the asterisk referring to the note below the Table.

Figure 2, please consider putting at line #133, mean annual precipitation (MAP), as such for line # 135 You can directly says The precipitation map ....

We made both changes as suggested.

Line # 138 I would not use the different units for the same values. I do understand that you want to emphasize the accuracy but i would write instead 3 +- 0.001 g

We changed the units as suggested.

Line # 145, how is the litter basket customized to rest at 1 cm above ground level?

We have now clarified –that we cut-open windows one centimeter above the cage floor. The cages were positioned on the soil (line 144).

Lines # 181-183, I like the approach of checking the necessity of having the random effects. However, it has been reported that likelihood ratio test (LRT) are not really reliable to test for random effects. I will suggest you rather use permutations instead. I think the function is confint(MODEL) you need to specify the number of permutation the higher the better but you should start with 99 first and see how the results look like if promising then you can even go to 9999. But it will need computation power and and time.

Thank you for the suggestion. We now used a simulation-based exact test, instead of a LRT, to examine the random effect, as recommended by the authors from the “lme4” package. As recommended, we used 9999 simulations. The simulation test yielded a similar result to those originally reported (see lines 181-183).

Line # 187, 188, 188, please do not use capital letter to start mesofauna, macrofauna and whole-community

We changed the formatting as suggested

Line # 205 Please add the version number of R in the text.

We now included the version number as suggested.

Line # 209-211, could you please check whether "then" is the word you want to use or "than"

Our bad- we indeed meant “than” and have made the appropriate changes.

Line # 227 and in other places as well please provide the second degree of freedom of the F test.

Thank you for this important comment. We have now added the second degree of freedom to the relevant results (lines 229, 232).

Figure 3 and Figure 4 show some results that are negative, can you please explain what might be the reasons behind this?

We now explain this important point in the figures’ captions.

Figure 5 Please add label to the x-axis.

Thank you-we have now included a label.

Line # 357, the sentence "... meso-decomposition, like microbial decomposition,...", I don't understand which criteria authors used to classify microbial decomposition as "meso-decomposition"?

We now remove this potential cause of confusion by using the term ‘meso-decomposition’ to distinguish from microbial decomposition (Line 366).

Line # 380 Kindly put "per se" in italic.

We changed the formatting as suggested

References

The references format are not consistent. For example for the same journal (say Trends in Ecology and Evolution) the authors sometimes wrote the full name like at line # 36 (and also realize that "vol" should not be written as such) but wrote the abbreviations at line #42

Our bad- we apologize and carefully checked all references to make sure the style is consistent.

-[] good #share

Author response:

We are grateful to the reviewers for recognizing the importance of our work and for their helpful suggestions. We will revise our manuscript in the revised version. However, we’d like to provide provisional responses now to answer the key questions and comments from the reviewers.

(1) Both reviewers asked why we chose 24-120 hpf to measure the apoptotic rates. We chose this time window based on the following two reasons: 1) Previous studies showed that although the motor neuron death time windows vary in chick (E5-E10), mouse (E11.5-E15.5), rat (E15-E18) and human (11-25 weeks of gestation), the common feature of these time windows is that they are all the developmental periods when motor neurons contact with muscle cells. The contact between zebrafish motor neurons and muscle cells occurs before 72 hpf, which is included in our observation time window. 2) Zebrafish complete hatching during 48-72 hpf, and most organs form before 72 hpf. More importantly, zebrafish start swimming around 72 hpf, indicating that motor neurons are fully functional.

Thus, we are confident that this 24-120 hpf time window covers the time window during which motor neurons undergo programmed cell death during zebrafish early development. We frequently used “early development” in this manuscript to describe our observation. However, we missed “early” in our title. We will add “early” in the title in the revised version.

(2) Both reviewers also asked about the neurogenesis of motor neurons. Previous studies have shown that the production of spinal cord motor neurons largely ceases before 48 hpf and then the motor neurons remain largely constant until adulthood. Our observation time window covers the major motor neuron production process. Therefore, we believe that neurogenesis will not affect our data and conclusions.

(3) Both reviewers questioned the specificity of using the mnx1 promoter to label motor neurons. The mnx1 promoter has been widely used to label motor neurons in transgenic zebrafish. Previous studies have shown that most of the cells labeled in the mnx1 transgenic zebrafish are motor neurons. In this study, we observed that the neuronal cells in our sensor zebrafish formed green cell bodies inside of the spinal cord and extended to the muscle region, which is an important morphological feature of the motor neurons. Furthermore, a few of those green cell bodies turned into blue apoptotic bodies inside the spinal cord and changed to blue axons in the muscle regions at the same time, which strongly suggests that those apoptotic neurons are not interneurons. Although the mnx1 promoter might have labeled some interneurons, this will not affect our major finding that only a small portion of motor neurons died during zebrafish early development.

(4) Reviewer 2 is concerned that the estimated 50% of motor neuron death was in limb-innervating motor neurons but not in body wall-innervating motor neurons. The death of motor neurons in limb-innervating motor neurons has been extensively studied in chicks and rodents, as it is easy to undergo operations such as amputation. However, previous studies have shown this dramatic motor neuron death does not only occur in limb-innervating motor neurons but also occurs in other spinal cord motor neurons. In our manuscript, we studied the naturally occurring motor neuron death in the whole spinal cord during the early stage of zebrafish development.

(5) Reviewer 2 mentioned that we ignored the death of an identified motor neuron. Our study was to examine the overall motor neuron apoptosis rather than a specific type of motor neuron death, so we did not emphasize the death of VaP motor neurons. We agree that the dead motor neurons observed in our manuscript contain VaP motor neurons. However, there were also other types of dead motor neurons observed in our study. The reasons are as follows: 1) VaP primary motor neurons die before 36 hpf, but our study found motor neuron cells died after 36 hpf and even at 84 hpf. 2) The position of the VaP motor neuron is together with that of the CaP motor neuron, that is, at the caudal region of the motor neuron cluster. Although it’s rare, we did observe the death of motor neurons in the rostral region of the motor neuron cluster. 3) There is only one or zero VaP motor neuron in each hemisegment. Although our data showed that usually one motor neuron died in each hemisegment, we did observe that sometimes more than one motor neuron died in the motor neuron cluster. We will include this information in the revised manuscript.

(6) For the morpholinos, we did not confirm the downregulation of the target genes. These morpholino-related data are a minor part of our manuscript and shall not affect our major findings. Thus, we didn’t think we missed “important” controls. We will perform experiments to confirm the efficiency of the morpholinos or remove these morpholino-related data from the revised version.

The image of someone having three different windows that each portray a misleading landscape made me reconsider the different conditions that one is put under and how it could influence decisions. I see it being that if you see the bright greenery versus the image of the broken fire hydrant, it would create vastly contrasting mental modes that would contribute to your thinking.

Reviewer #1 (Public Review):

Summary:

In their paper, Hou and co-workers explored the use of a FRET sensor for endogenous g-sec activity in vivo in the mouse brain. They used AAV to deliver the sensor to the brain for neuron specific expression and applied NIR in cranial windows to assess FRET activity; optimizing as well an imaging and segmentation protocol. In brief they observe clustered g-sec activity in neighboring cells arguing for a cell non-autonomous regulation of endogenous g-sec activity in vivo.

Strengths:

Mone.

Weaknesses:

Overall the authors provide a very limited data set and in fact only a proof of concept that their sensor can be applied in vivo. This is not really a research paper, but a technical note. With respect to their observation of clustered activity, they now provide an overview image, next to zoomed details. However, from these images one cannot conclude 'by eye' any clustering event. This aligns with the very low r values. All neurons in the field show variable activity and a clustering is not really evident from these examples. Even within a cluster, there is variability. The authors now confirm that expression levels are indeed variable but are independent from the ratio measurements. Further, they controlled for specificity by including DAPT treatments, but opposite to their own in vitro data (in primary neurons) the ratios increased. The authors argue that both distance and orientation can either decrease or increase ratios and that the use of this biosensor should be explored model-by-model. This doesn't really confer high confidence and may hinder other groups in using this sensor reliably.

Secondly, there is still no physiological relevance for this observation. The experiments are performed in wild-type mice, but it would be more relevant to compare this with a fadPSEN1 KI or a PSEN1cKO model to investigate the contribution of a gain of toxic function or LOF to the claimed cell non-autonomous activations. The authors acknowledge this shortcoming but argue that this is for a follow-up study.

For instance, they only monitor activity in cell bodies, and miss all info on g-sec activity in neurites and synapses: what is the relevance of the cell body associated g-sec and can it be used as a proxy for neuronal g-sec activity? If cells 'communicate' g-sec activities, I would expect to see hot spots of activity at synapses between neurons.

Without some more validation and physiologically relevant studies, it remains a single observation and rather a technical note paper, instead of a true research paper.

After release from G1, subtelomeric regions were more under-replicated than regions further away from telomeric regions.

As you move further from the telomere, the underreplicated DNA decreases.

Mount, rather than rsync for offline use when traveling.

Reviewer #3 (Public Review):

This study by Guan and co-workers focuses on a model neuronal lineage in the developing Drosophila nervous system, revealing interesting aspects about: a) the generation of supernumerary cells, later destined for apoptosis; and, b) new insights into the mechanisms that regulate this process. The two RNA-binding proteins, Imp and Syp, are shown to be expressed in temporally largely complementary patterns, their expression defining early vs later born neurons in this lineage, and thus also regulating the apoptotic elimination. Moreover, neuronal 'fate' transcription factors that are downstream of Imp and signatures of early-born neurons, can also be sufficient to convert later born cells to an earlier 'fate', including survival. The authors provide solid evidence for most of their statements, including the temporal windows during which the early and the later-born motoneurons are generated by this model lineage, how this relates to patterns of cell death by apoptosis and that mis-expression of early-born transcription factors in later-born cells can be sufficient to block apoptosis (part of, and perhaps indicative of the late-born identity). Other studies have previously outlined analogous, mutually antagonistic roles for Imp and Syp during nervous system development in Drosophila, in different parts and at different stages, with which the working model of this study aligns. Overall, this study adds to and extends current working models and evidence on the developmental mechanisms that underlie temporal cell fate decisions.

Author response:

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

Reviewer #1 (Public Review):

This study addresses the temporal patterning of a specific Drosophila CNS neuroblast lineage, focusing on its larval development. They find that a temporal cascade, involving the Imp and Syb genes changes the fate of one daughter cell/branch, from glioblast (GB) to programmed cell death (PCD), as well as gates the decommissioning of the NB at the end of neurogenesis.

I believe there are some inaccuracies in this summary. We address temporal patterning during larval and pupal stages until the adult stage. The Imp and Syp genes change the fate of one daughter cell/branch from survival to programmed cell death (PCD). The change from glioblast (GB) to PCD, which occurs at an early time point, is not addressed here. The main point of the paper is missing:

• Last-born MNs undergo apoptosis due to their failure to express a functional TF code, and this code is post-transcriptionally regulated by the opposite expression of Imp and Syp in immature MNs.

Reviewer #2 (Public Review):

Summary:

Guan and colleagues address the question of how a single neuroblast produces a defined number of progeny, and what influences its decommissioning. The focus of the experiments are two well-studied RNA-binding proteins: Imp and Syp. The Authors find that these factors play an important role in determining the number of neurons in their preferred model system of VNC motor neurons coming from a single lineage (LinA/15) by separate functions taking place at specific stages of development of this lineage: influencing the life-span of the LinA neuroblast to control its timely decommissioning and functioning in the Late-born post-mitotic neurons to influence cell death after the appropriate number of progeny is generated. The post-mitotic role of Imp/Syp in regulating programmed-cell death (PCD) is also correlated with a specific code of key transcription factors that are suspected to influence neuronal identity, linking the fate of neuronal survival with its specification. This paper addresses a wide scope of phenotypes related to the same factors, thus providing an intriguing demonstration of how the nervous system is constructed by context-specific changes in key developmental regulators.

The bulk of conclusions drawn by the authors are supported by careful experimental evidence, and the findings are a useful addition to an important topic in developmental neuroscience.

I cannot summarize better the paper.

Strengths:

A major strength is the use of a genetic labeling tool that allows the authors to specifically analyze and manipulate one neuronal lineage. This allows for simultaneous study of both the progenitors and post-mitotic progeny. As a result the paper conveys a lot of useful information for this particular neuronal lineage. Furthermore addressing the association of cell fate specification, taking advantage of this lab's extensive prior work in the system, with developmentally-regulated programmed celldeath is an important contribution to the field.

Beyond Imp/Syp, additional characterization of this model system is provided in characterizing a previously unrecognized death of a hemilineage in early-born neurons.

Thanks!

Weaknesses:

The main observations that distinguish this study from others that have investigated Imp/Syp in the fly nervous system is the role played in late-born post-mitotic neurons to regulate programmed cell death. This is an important and plausible (based on the presented findings) newly discovered role for these proteins. However the precision of experiments is not particularly strong, which limits the authors claims. The genetic strategy used to manipulate Imp/Syp or the TF code appears to be done throughout the entire lineage, or all neuronal progeny, and not restricted to only the late born cells. Can the authors rule out survival of the early born hemi-lineage normally fated to die? Therefore statements such as this: 

To further investigate this possibility, we used the MARCM technique to change the TF code of lastborn MNs without affecting the expression of Imp and Syp should be qualified to specify that the result is obtained by misexpressing these factors throughout the entire lineage.

We agree that our genetic manipulations affect the entire lineage or all neuronal progeny. We do not have genetic tools to gain such precision. We have changed our descriptions to specify the entire lineage or all neuronal progeny. As the reviewer raised, we were also concerned about the possibility that the overexpression of Imp or knockdown of Syp could induce the survival of the early-born hemilineage. We have two experiments that rule out this possibility:

(1) In late LL3 larvae, Imp OE or syp MARCM clones do not change the number of cells in LL3 larvae (see Guan et al., 2022), indicating that the hemilineage that died by PCD is not affected. If Imp or Syp played a role in the survival of the hemilineage, we would see at least a 50% increase in the number of MNs at this stage.

(2) The MARCM experiment using the VGlut driver to overexpress P35 or Imp allows us to manipulate only elav+ VGlut+ neurons. The hemilineage removed by PCD is elav- VGlut- and is not affected by this experiment. Consequently, the increase in MNs in adults with genetic manipulation can only be the result of the survival of the other hemilineage (elav+, VGlut+). Moreover, this experiment shows an increase in the number of neurons in the adult but not in LL3, demonstrating that the hemilineage (elav- VGlut-) is still removed by PCD with this genetic manipulation.

The authors make an observation that differs from other systems in which Imp/Syp have been studied: that the expression of the two proteins appears to be independent and not influenced by cross-regulation. However there is a lack of investigation as to what effect this may have on how Imp/Syp regulate temporal identity. A key implication of the previously observed cross-regulation in the fly mushroom body is that the ratio of Imp/Syp could change over the life of the NB which would permit different neuronal identities. Without cross-regulation, do the authors still observe a gradient in the expression pattern of time? Because the data is presented with Imp and Syp stained in different brain samples, and without quantification across different stages, this is unclear. The authors use the term 'gradient' but changes in levels of these factors are not evident from the presented data.

We have now quantified the transcriptional activity of Imp and Syp in the NB over time using smFISH. We have also quantified the relative expression of Imp and Syp protein in the NB over time by co-immunostaining. Additionally, we quantified the relative expression of Imp and Syp protein in postmitotic neurons as a function of their birth order in late LL3 larvae. All these data show an opposite temporal gradient of Imp and Syp in the NB and an opposite spatial gradient in immature neurons according to their birth order (Figure. 4). How these gradients are established in our system remains to be elucidated. 

Reviewer #3 (Public Review):

This study by Guan and co-workers focuses on a model neuronal lineage in the developing Drosophila nervous system, revealing interesting aspects about: a) the generation of supernumerary cells, later destined for apoptosis; and, b) new insights into the mechanisms that regulate this process. The two RNA-binding proteins, Imp and Syp, are shown to be expressed in temporally largely complementary patterns, their expression defining early vs later born neurons in this lineage, and thus also regulating the apoptotic elimination. Moreover, neuronal 'fate' transcription factors that are downstream of Imp and signatures of early-born neurons, can also be sufficient to convert later born cells to an earlier 'fate', including survival.

The authors provide solid evidence for most of their statements, including the temporal windows during which the early and the later-born motoneurons are generated by this model lineage, how this relates to patterns of cell death by apoptosis and that mis-expression of early-born transcription factors in later-born cells can be sufficient to block apoptosis (part of, and perhaps indicative of the late-born identity).

Other studies have previously outlined analogous, mutually antagonistic roles for Imp and Syp during nervous system development in Drosophila, in different parts and at different stages, with which the working model of this study aligns.

Overall, this study adds to and extends current working models and evidence on the developmental mechanisms that underlie temporal cell fate decisions.

I cannot summarize better the paper.

Reviewer #1 (Recommendations For The Authors):

While this is an interesting topic, I raised two issues in my original review.

(1) Against the backdrop of numerous previous studies linking many developmental regulators, including tTFs, to programmed cell death in the developing CNS, which in several cases have involved identifying key PCD genes and decoding the molecular regulatory interplay between regulators and PCD genes, this study does not provide any new insight into the regulation of developmental PCD in the CNS.

The authors have not added any new data to address this shortcoming.

I agree with the reviewer that we did not attempt to link Imp/Syp with the temporal transcription factor (tTF) cascade or spatial selectors such as Hox genes. However, this decision was intentional as our primary focus was on studying immature MNs. It is worth noting that the decommissioning of NBs by autophagic cell death or terminal differentiation, which is mediated by Imp/Syp in other lineages, has not been correlated with tTFs or spatial selectors. Although we have not directly examined the involvement of the hb + sv > kr > pdm > cas > cas-svp > Grh cascade in the decommissioning of the Lin A neuroblast, our preliminary data indicate that Hb, Sv, Pdm, and Cas are not expressed in the Lin A NB, while Grh is consistently expressed in the NB (Wenyue et al., 2022). Thus, it is less likely that this particular tTF cascade is not implicated in Lin A neuroblast decommissioning. In contrast, spatial selectors, such as the Hox gene Antp, play an opposing role compared to HOX transcription factors in abdominal NBs. In the Lin A lineage, Antp promotes survival (Baek, Enriquez, & Mann, 2013). Here, to avoid repeating what has already been described in the literature, we focused on the role of Imp/Syp in postmitotic neurons and revealed that the precise elimination of MNs is linked to the control of TFs expressed in the MNs.

(2) I raised the issue that it is unclear if Imp/Syp acts in the NB, and/or in IMC/GMC, and/or in the daughter cells generated from these.

I agree with the reviewer's concern regarding the unclear function of Imp/Syp, i.e., whether it acts in the NB, IMC/GMC, or daughter cells. To address this, one possible approach would be to attempt rescuing Imp and Syp mutants by transgenic expression in specific cell types, such as NBs, IMC/GMC, or GB/daughter cells. However, we have not conducted such experiments as we were skeptical about the outcome. Previous published work has used drivers expressed in NBs, IMC/GMC, or postmitotic neurons to decipher the function of a gene in a specific cell type. But the results of these experiments must be taken with caution. Using NB/GMC drivers to study gene function can lead to effects not only in the NB but also in its progeny, including GMC or postmitotic neurons, due to the perdurance and stability of the Gal4 and UAS-gene expression system. For instance, dpn-Gal4 UASGFP not only labels the NB but also many of its progeny, even if Dpn is only expressed in NBs. And elav-Gal4 is expressed in the NB and GMCs.

However, our overexpression of Imp in immature neurons using Vglut demonstrates that Imp promotes cell survival through an autonomous function in these neurons. This driver is only expressed in postmitotic neurons (elav+) and not in the NB, IMC/GMC, or in the hemilineage eliminated by cell death (elav-vglut-).

Reviewer #2 (Recommendations For The Authors):

Oddly knockdown of Imp in the neuroblast (Fig. 5D) only led to death at 8h APF, when Imp is no longer expressed. Do the authors have an explanation as to how the stem cell can survive until this point? A discussion would be helpful.

The simple explanation is the efficiency of RNAi. The imp-/- MARCM clones (Guan et al., 2022) lead to a stronger reduction of MNs in LL3.

A simple experiment I would recommend is to repeat the antibody stainings of staged larvae/pupae (Fig. 4) having the anti-Imp/Syp antibodies in the same brain sample, and perhaps a quantification of the ratio in the NB. Given the species in which the ABs were raised seem compatible, this should be feasible. As it stands now, there is no indication of whether the ratio of Imp vs Syp change over time.

We have now quantified the transcriptional activity of Imp and Syp in the NB over time. We have also quantified the relative expression of Imp and Syp proteins in the NB over time and quantified the relative expression of Imp and Syp proteins in postmitotic neurons as a function of their birth in late LL3 larvae. How these gradients are established in our system still remains to be 

Minor errors/suggestions:

Fig 4. Time legend at the top goes A, B, C, E, F (no D). So it doesn't match the panels below

Yes, we have made the corrections.

Sentence repeated in Intro:

The process of terminating NB neurogenesis through autophagic cell death or terminal differentiation is commonly referred to as decommissioning.

Yes, corrections have been made.

IN FIGURE 1 THEY SAY 'TYPE IB' AND IN FIGURE 2 THEY SAY 'TYPE 1B'

We have changed it to type 1b.

In Fig2A-It's hard to see lack of Elav and Fig2G-It's hard to see presence of Dcp1. Panels could be adjusted to emphasize these results

We have increased the size of the panels and made two separate panels where only the elav and Dcp1 signals are present.

Observations that the result is equivalent in all thoracic segments is expected, since all legs need the same number of neurons. This is nice to have but can be in the supplement.

Overall the figure number seems excessive, especially considering much of the results included(particularly the NB results) are findings consistent with previous papers and some is characterization of the system that does not fit well with the main focus regarding Imp/Syp (i.e death of one hemi-lineage:

Figure 5 and 6 can be joined as one.

We have combined Figures 5 and 6, showing only the T1 segments.

There is some discrepancy between graphs Fig7F and K: At LL3 the number of neurons is different for the control in 7F and the count in K

Yes, because the genetic backgrounds are not the same and we are not counting the same type of cells. In 7F, we are counting the elav+ and VGlut+ cells, whereas in Figure 7K, we are counting all the elav+ in Lin A, including those elav+ VGlut-. VGlut expression arrives a bit later after elav+, which is why we have fewer elav+ cells in 7F. In other words, VGlut MARCM clones do not label all Lin A elav+ cells. I have clarified this in the figure.

Reviewer #3 (Recommendations For The Authors):

Main comment: on the notion of Imp and Syp gradients:

p. 5, related to figure 4 - there are clearly distinct windows for predominantly (if not exclusively) Imp, and later, Syp expression in lineage 15, with a phase of co-expression.

However, based on the data shown, it is unclear whether these windows represent gradients, as repeatedly stated. If the notion of gradients is derived from other studies, on other lineages, then this would be good to clarify. Alternatively, the idea of temporally opposing gradients of Imp and Syp would need to be demonstrated for this lineage.

For example, a more accurate way to describe this study's data is given on p.7 "In conclusion, our findings demonstrate that the opposite expression pattern of Imp and Syp in postmitotic neurons precisely shapes the size of Lin A/15 lineage by controlling the pattern of PCD in immature MNs (Fig. 8)."

We have now quantified the transcriptional activity of Imp and Syp in the NB over time. We have also quantified the relative expression of Imp and Syp proteins in the NB over time. We have also quantified the relative expression of Imp and Syp proteins in postmitotic neurons as a function of their birth in late LL3 larvae. How these gradients are established in our system still remains to be identified.

Minor points:

p.6, related to figure 7: Are numbers of EDU- early born and EDU+, late born, MNs expressed as means in the main text? As written, it suggests absence of any variability, which one would expect and which is shown in Fig.7 data.

Yes, we have added averages in the text.

Methods: the author name 'Lacin' has been mis-spelled

Sorry about that, it's been corrected.

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

The authors investigate the role of chirping in a species of weakly electric fish. They subject the fish to various scenarios and correlate the production of chirps with many different factors. They find major correlations between the background beat signals (continuously present during any social interactions) or some aspects of social and environmental conditions with the propensity to produce different types of chirps. By analyzing more specifically different aspects of these correlations they conclude that chirping patterns are related to navigation purposes and the need to localize the source of the beat signal (i.e. the location of the conspecific).

The study provides a wealth of interesting observations of behavior and much of this data constitutes a useful dataset to document the patterns of social interactions in these fish. Some data, in particular the high propensity to chirp in cluttered environments, raises interesting questions. Their main hypothesis is a useful addition to the debate on the function of these chirps and is worth considering and exploring further.

After the initial reviewers' comments, the authors performed a welcome revision of the way the results are presented. Overall the study has been improved by the revision. However, one piece of new data is perplexing to me. The new Figure 7 presents the results of a model analysis of the strength of the EI caused by a second fish to localize when the focal fish is chirping. From my understanding of this type of model, EOD frequency is not a parameter in the model since it evaluates the strength of the field at a given point in time. Therefore the only thing that matters is the phase relationship and strength of the EOD. Assuming that the second fish's EOD is kept constant and the phases relationship is also the same, the only difference during a chirp that could affect the result of the calculation is the potential decrease in EOD amplitude during the chirp. It is indeed logical that if the focal fish decreased its EOD amplitude the target fish's EOD becomes relatively stronger. Where things are harder to understand is why the different types of chirps (e.g. type 1 vs type 2) lead to the same increase in signal even though they are typically associated with different levels of amplitude modulations. Also, it is hard to imagine that a type 2 chirps that is barely associated with any decrease in EOD amplitude (0-10% maybe), would cause doubling of the EI strength. There might be something I don't understand but the authors should provide a lot more details on how this result is obtained and convince us that it makes sense.

We thank the author for the comments and we agree that the approach could have been better detailed. As anticipated by the Reviewer, the Boundary Element Method (BEM) model can be used simply to calculate the electric field and electric image at a specific point in time (instantaneously), regardless of EOD frequency. However, our model allows for the concatenation of consecutive instants and thus is able to render an entire sequence of electric fields - and resulting electric images - incorporating realistic EOD characteristics such as shape, duration, and frequencies (see Pedraja et al., 2014).

Chirp-triggered EIs were modeled using real chirps produced by interacting fish. Each chirp was thus associated to its duration and peak parameters, as well as the fish positional information (distance and angle). 

However, since we did not know the beat phase at which chirps were produced, we computed electric images for each fish position and chirp scenario by simulating various phases (here referred to the initial offset of the two EODs, set at 4 phases, equally spaced). These are intended as phases of the sender EOD and simply refer to the initial OFFSET between the two interacting EODs. However, since our simulations were run over a time window of 500 msec, all phases are likely to be covered, with a different temporal order relative to the chirp (always centered within the 500 msec).

The simulation was run maintaining consistent timing for both chirp and non-chirp conditions, across approximately 800 body nodes. At each node, the current flow was calculated from the peak-to-peak of the EOD sum (i.e. the point-to-point of the difference between the beat positive and negative envelopes). Analyzing the EIs over this fixed time window enables us to assess the unitary changes of current flow induced by chirps over units of time (ΔI/Δt). From this, we can calculate a cumulative sum of current flow changes - expressed as delta(EI) and use it to show the effect of the chirps on the spatiotemporal EI (Figure 7C).

One can express this cumulative change mapped onto the fish body (keeping the 800 points separated, as in Figure 7C) or further sum the current changes to obtain a single total (as shown in Figure 7D).

One can check this by considering that a sum for example of a set of 500/800 points - judging from the size of the blue areas in C not all 800 points have a detectable change - each valued 0.1-to-0.3 mA/s, one could get circa 100 mA/s, which is what is shown in D. (is this what is happening ?)

We do not know why chirps of different types triggered similar effects. It is possible that, since EI measurements are pooled over several chirps produced at different angles and distances, in case of a lower amount of chirps considered for a given type (as in the case of rises, very low) these measurements may not highlight more marked differences among types. In a publication we are currently working on, we are considering a larger dataset to better assess these results.

The methods section has been edited to clarify the approach (not yet).

Reviewer #2 (Public Review):

Studying Apteronotus leptorhynchus (the weakly electric brown ghost knifefish), the authors provide evidence that 'chirps' (brief modulations in the frequency and amplitude of the ongoing electric signal) function in active sensing (specifically homeoactive sensing) rather than communication. Chirping is a behavior that has been well studied, including numerous studies on the sensory coding of chirps and the neural mechanisms for chirp generation.

Chirps are largely thought to function in communication behavior, so this alternative function is a very exciting possibility that could have a great impact on the field.

We thank the Reviewer for the extensive and constructive comments. We would like to add that, while it is true that many detailed studies have been published on the anatomy and physiology of the circuits implicated in the production and modulation of “electric chirps”, most of this  research assumed, and focused exclusively on, their possible role in communication.  In addition, most behavioral studies did the same and a meta-analysis of the existing literature on chirping allows to trace back the communication idea mainly to two studies: Hagedorn and Heiligenberg, 1985 (“Court and spark: electric signals in the courtship and mating of gymnotoid fish”) and Hopkins, 1974 (“Electric Communication: Functions in the Social Behavior of Eigenmannia Virescens”), among the main sources. Importantly, in these studies only contextual observations have been made (no playback experiment or other attempts to analyze more quantitatively the correlation of chirping with other behaviors).

The authors do provide convincing evidence that chirps may function in homeoactive sensing. However, their evidence arguing against a role for chirps in communication is not as strong, and fails to sufficiently consider the evidence from a large body of existing research. Ultimately, the manuscript presents very interesting data that is sure to stimulate discussion and follow-up studies, but it suffers from dismissing evidence in support of, or consistent with, a communicative function for chirps.

Although the tone of some statements present in our earlier draft may suggest otherwise, through our revisions, we have made an effort to clarify that we do not intend to dismiss a function of chirps in communication, we only intend to debate and discuss valid alternative hypothesis, advanced from reasonable considerations.

Before writing this manuscript, we have attempted to survey  literally all the existing literature on chirps (including studies focused on behavior, peripheral sensory physiology as well as brain physiology). Although it is not unlikely that some studies have eluded our attention, an effort for a comprehensive review was made. Based on this survey we realized that none of the studies provided a clear  and  unambiguous piece of evidence to support the communication hypothesis (we refer here to the weak points highlighted in the discussion and mentioned in the previous comment). Which in fact does not come without its weak points and contradictions (see later comments).

It follows a summary of the mentions made to the communication theory in the different section of the manuscript including several edits we have applied in response to the Reviewer’s concern:

In the abstract we clearly state that we are considering an alternative that is only hypothetically complementary, not for sure.  Nonetheless, we have identified a couple of instances that could sound dismissive of the “communication hypothesis” in the following section.

In the introduction we write in fact about the possibility of interference between communication signals and conspecific electrolocation cues, as they are both detected as beat perturbations. We did not mean to use “Interference” here as “reciprocal canceling”, rather we intended it as “partial or more or less conspicuous overlap” in the responses triggered in electroreceptors.

Hoping to convey a clearer message, we have edited the related statement and changed it to “both types of information are likely to overlap and interact in highly variable ways”.

We have also removed the statement: “According to this idea, beats and chirps are not only detected through the same input channel, but also used for the same purpose.” as at this point in the manuscript it may be too strong.

In the results section we do not include statements that might be seen as dismissive of the communication hypothesis but only statements in support of the “probing with chirps” idea (which is the central hypothesis of the study).

In the discussion paragraphs we elaborate on why the current functional view is either flawed or incomplete (first paragraph “existing functional hypotheses''). Namely: 1)  multiple triggering factors implied in chirp responses covary and need to be disentangled (example DF/ sex), 2) findings on brown ghosts and a few other gymnotiforms have been used to advance the hypothesis of “communication through chirps'' in all weakly electric fish (including pulse species). 3) social encounters - in which chirps are recorded - imply also other behaviors (such as probing) which have not been considered so far. This point is related to the first one on covariates. 4) most studies referring to big chirps as courtship chirps were not done in reproductive animals (added now)  and 5) no causal evidence has been provided so far to justify a role of chirps in social communication.

We are discussing these points as challenges to the communication hypothesis, not to dismiss the hypothesis, but rather to motivate future studies addressing these challenges.

We do not want to appear dismissive of the communication hypothesis and had therefore previously edited the manuscript to avoid the impression of exclusivity of the probing hypothesis. We have now gone over the manuscript once more and edited several sentences. Nevertheless, we want to point out again that - despite the large consensus - the communication hypothesis has, until now, never been investigated with the kind of rigor applied here.

The authors do acknowledge that chirps could function as both a communication and homeactive sensing signal, but it seems clear they wish to argue against the former and for the latter, and the evidence is not yet there to support this.

In both rounds of revision we have made an effort to convey a more inclusive interpretation of our findings. We tried our best to express our ideas as hypothetical, not as proof that communication through chirps does not exist. The aim of this study is to propose an alternative view, and this cannot be done without underlining the weak points of an existing hypothesis while providing and supporting reasonable arguments in favor of the alternative we advance. The actual evidence for a role of chirping in communication is much less strong than appears from the pure number of articles that have discussed chirps in this context.

Regarding the weak evidence against communication, here we can list a few additional important points related to the proposed interpretations of chirp function (more specific than those made earlier):

(1) A formally sound assessment of signal value/meaning - as typically done in animal communication studies should involve: 

a) the isolation of a naturally occurring signal and determination of the context in which it is produced 

b) the artificial replication of the signal

c) the observation that such mimic is capable of triggering reliable and stereotyped responses in a group of individuals (identified by sex and/or species) under the same conditions (conditioned, unconditioned, state-dependent, etc.). As discussed for instance in Bradbury and Vehrencamp, 2011; Laidre and Johnstone, 2013; Wyatt, 2015; Rutz et al., 2023.

This approach has so far not been applied to weakly electric fish. The initial purpose of the present study was in fact to conduct this type of validation.

(2) The hypothesis of chirps used for DF-sign discrimination - for “social purposes” - although plausible in the face of theoretical considerations,  does not seem to be reasonable in practice, when one considers emission rates of 150 chirps per minute. We do find a strong correlation of chirp type with DF, which is often very abrupt and sudden (as if the fish were tracking beat frequency to guess its value) but the consideration made above on chirp rates seems to discourage this interpretation.

(3) The hypothesis of chirp-patterning (i.e. chirping may have meaning based on the sequence of chirps of different types, a bit like syllables in birdsongs) - assessed by only one study conducted in our group - has not been enough substantiated by replication. We have surveyed all possible combinations of chirps produced by interacting pairs in different behavioral conditions using different value for chirp sequence size: 2, 3,... ,8 chirps (both considering the sender alone as well as sender+receiver together). In all cases we found no evidence for  a context dependent “modulation” of chirp types (i.e. no specific chirp type sequence in specific contexts).

(4) The hypothesized role of “large chirps” as courtship signals could be easily criticized by noting the symmetrical distribution of these events around  a DF of 0 Hz . Although one could argue about a failure to discriminate DF-sign, to explain this well known pattern. However, we know from Walter Heiligenberg’s work and physiological considerations that such task can be solved easily through t-units and … in principle even just by motion (which would change the EOD phase in frequency dependent ways, thus potentially revealing the DF sign).

Overall, these considerations made us think that certainly chirping occurs in a social context, but it is the meaning of this behavior that remains elusive.  We noticed that environmental factors are also strongly implied … we then formulate an alternative hypothesis to explain chirping but we do so  without dismissing the communication idea.

All this seems to us just a careful way to critically discuss our results and those of other studies, without considering the issue resolved.

In the introduction, the authors state, "Since both chirps and positional parameters (such as size, orientation or motion) can only be detected as perturbations of the beat, and via the same electroreceptors, the inputs relaying both types of information are inevitably interfering." I disagree with this statement, which seems to be a key assumption. Both of these features certainly modulate the activity of electroreceptors, but that does not mean those modulations are ambiguous as to their source. You do not know whether the two types of modulations can be unambiguously decoded from electroreceptor afferent population activity.

We thank the Reviewer for noting this imprecision. We have addressed the Reviewer’s concern in another reply (see above).

My biggest issue with this manuscript is that it is much too strong in dismissing evidence that chirping correlates with context. In your behavioral observations, you found sex differences in chirping as well as differences between freely interacting and physically separated fish. Chirps tended to occur in close proximity to another fish. Your model of chirp variability found that environmental experience, social experience, and beat frequency (DF) are the most important factors explaining chirp variability. Are these not all considered behavioral or social context? Beat frequency (DF) in particular is heavily downplayed as being a part of "context" but it is a crucial part of the context, as it provides information about the identity of the fish you're interacting with. The authors show quite convincingly that the types of chirps produced do not vary with these contexts, but chirp rates do.

We believe the “perceived claim” may be an issue of unclear writing. We have now tried to better clarify that “context” affects chirp rates, but it does not affect chirp types as much (except when beat frequency is high).  

We have edited two statements possibly susceptible to misinterpretation: 

(1) In the results: “It also indicates that chirp parameters such as duration and FM do not seem to be associated with any particular context in a meaningful way, other than being affected by beat frequency.”

(2) In the discussion: the statement

“Recordings from interacting fish pairs confirmed the absence of any significant correlation between chirp type choice and behavioral context (Figure S2) although the variance of chirp parameters appears to be significantly affected by this factor (Figure 2). This may suggest that the effect of behavioral context is mainly detectable in the number of chirps produced (Figure S1), rather than the type (Figure S2).”

has been changed to:

“Recordings from interacting fish pairs confirmed the absence of any significant correlation between chirp type choice and behavioral context, except for those cases characterized by higher beat frequencies  (Figure S2). This suggests that the effect of behavioral context highlighted in our factor analysis (Figure 2) is mainly due to the number of chirps produced (Figure S1), rather than their type (Figure S2).”

Eventually, in the results we emphasize the relatively higher impact of previously unexplored factors on chirp variance: “The plot of individual chirps (Figure 2C) shows the presence of clustering around different categorical variables and it reveals that experience levels or swimming conditions are important factors affecting chirp distribution (note for instance the large central “breeding” cluster in which fish are divided and the smaller ones in which fish are free). Sender or receiver identity does not individuate any clear clustering relative to either sex (see the overlap of male_s/male_r and female_s/female_r) or social status (dominant/subordinate). Chirps labeled based on tank experience (i.e. resident vs intruder) are instead clearly separated.”.

Further, in your playback experiments, fish responded differently to small vs. large DFs, males chirped more than females, type 2 chirps became more frequent throughout a playback, and rises tended to occur at the end of a playback. These are all examples of context-dependent behavior.

We do note that male brown ghosts chirp more than females. But we do also say - and show in figure 8 - that males move more in proximity to and around conspecifics. We do acknowledge that chirp time-course may be different during playbacks in a type-dependent manner. But how this can support the communication hypothesis - or other alternatives - is unclear. This result could equally imply the use of different chirp types for different probing needs. Since we cannot be sure about either, we do not want to put too much emphasis to it. Eventually, the fact that “context” (here meant broadly to define different experimental situations in which social but also physical and environmental parameters are altered) affects chirping is undeniable: cluttered and non-cluttered environments do represent different contexts which differently affect chirping in conspicuous ways.

In the results, the authors state, "Overall, the majority of chirps were produced by male subjects, in comparable amounts regardless of environmental experience (resident, intruder or equal; Figure S1A,C), social status (dominant or subordinate; Figure S1B) or social experience (novel or experienced; Figure S1D)." This is not what is shown in Figure S1. S1A shows clear differences between resident vs. intruder males, S1B shows clear differences between dominant vs. subordinate males, and S1D shows clear differences between naïve and experienced males. The analysis shown in Figure 2 would seem to support this. Indeed, the authors state, "Overall, this analysis indicated that environmental and social experience, together with beat frequency (DF) are the most important factors explaining chirp variability."

The Reviewer is right in pointing at this imprecise reference and we are grateful for spotting this incongruence. The writing refers probably to an earlier version of the figure in which data were grouped and analyzed differently. We now edited the text and changed it to: “Overall, the majority of chirps were produced by male subjects, at rates that seemed  affected by environmental experience (resident, intruder or equal; Figure S1A,C), social status (dominant or subordinate; Figure S1B) and social experience (novel or experienced; Figure S1D).”

The choice of chirp type varied widely between individuals but was relatively consistent within individuals across trials of the same experiment. The authors interpret this to mean that chirping does not vary with internal state, but is it not likely that the internal states of individuals are stable under stable conditions, and that individuals may differ in these internal states across the same conditions? Stable differences in communication signals between individuals are frequently interpreted as reflecting differences between those individuals in certain characteristics, which are being communicated by these signals.

It seems here we have been unclear in the writing: while it is true that behavioral states are stable and can imply stable chirp patterning (if the two are related), since chirp types vary abruptly and in a reliable DF-dependent manner, different types of chirps are unlikely to be matched to different internal states following the same temporal order in such a reliable way (similarly repeated through consecutive trials).

This would imply the occurrence of different internal states in rapid sequence, reliably triggered by repeated EOD ramps, regardless of whether the playback is 20 sec long or 180 sec long.

We have edited this paragraph to better explain this: “The reliability by which the chirping response adapts to both the rate and direction of beat frequency is variable across individuals but rather stable across trials (relative to a given subject), further suggesting that chirp type variations may not reflect changes in internal states or in the animal motivation to specific behavioral displays (which are presumably subject to less abrupt variations and stereotypical patterning based on DF).”

I am not convinced of the conclusion drawn by the analysis of chirp transitions. The transition matrices show plenty of 1-2 and 2-1 transitions occurring.

The only groups in which 1-2 and 2-1 transitions are as frequent as 1-1 and 2-2 (being 1 and 2 the numerical IDs of the two interacting fish) are F-F pairs. This is a result of the fact that in females chirp rates are so low that within-fish-correlations end up being as low as between-fish-correlations. We believe the impression of the Reviewer could be due to the fact that these are normalized maps (see legend of Figure 5A-B).

Further, the cross-correlation analysis only shows that chirp timing between individuals is not phase-locked at these small timescales. It is entirely possible that chirp rates are correlated between interacting individuals, even if their precise timing is not.

We agree with the Reviewer, this is a possibility. To address this point, we did edit the results section to acknowledge that what we see may be related to the time window chosen (i.e. 4 sec):

“More importantly, they show that - at least in the social conditions analyzed here and within small-sized time windows - chirp time series produced by different fish during paired interactions are consistently independent of each other.”

Further, it is not clear to me how "transitions" were defined. The methods do not make this clear, and it is not clear to me how you can have zero chirp transitions between two individuals when those two individuals are both generating chirps throughout an interaction.

We thank the Reviewer for bringing up this unclear point. We have now clarified how transitions were calculated in the method section: “The number of chirp transitions present in each recording (dataset used for Figures 1, 2, 5) was measured by searching in a string array containing the 4 chirp types per fish pair, all their possible pairwise permutations (i.e. all possible permutations of 4+4=8 elements are: 1-1, 1-2, 1-3 … 7-6, 7-7, 7-8; considering the following legend 1 = fish1 type 1, 2 = fish 1 type 2, 3 = fish1 type 3 … 6 = fish2 type 2, 7 = fish2 type 3 and 8 = fish2 rise).”.

Zero transitions are possible if two fish (or groups of fish) do not produce chirps of all types. Only transitions of produced types can be counted.

In the results, "Although all chirp types were used during aggressive interactions, these seemed to be rather less frequent in the immediate surround of the chirps (Figure 6A)." A lack of precise temporal correlation on short timescales does not mean there is no association between the two behaviors. An increased rate of chirping during aggression is still a correlation between the two behaviors, even if chirps and specific aggressive behaviors are not tightly time-locked.

The Reviewer is right in pointing out the limited temporal scaling of our observations/analysis. We have now edited the last paragraph of the results related to figure 6 to include the possibility mentioned by the Reviewer: “The significantly higher extent of chirping during swimming and locomotion, consistently confirmed by 4 different approaches (PSTH, TM, CN, MDS), suggests that - although chirp-behavior correlations may exist at time-scales larger than those here considered - chirping may be linked more strongly with scanning and environmental exploration than with a particular motivational state, thus confirming findings from our playback experiments.”

The Reviewer here remarks an important point, yet, due to space limitations, we have considered only a sub-second scale. Most playback experiments in weakly electric fish implied the use of EOD mimics for a few tens of seconds - to avoid habituation in the fish behavioral responses -  while inter-chirp intervals usually range between a few hundreds of milliseconds to seconds (depending on how often a fish would chirp). This suggested to us that a 4 second time window may not be a bad choice to start with.

In summary, it is simply too strong to say that chirping does not correlate with context, or to claim that there is convincing evidence arguing against a communication function of chirps. Importantly, however, this does not detract from your exciting and well-supported hypothesis that chirping functions in homeoactive sensing. A given EOD behavior could serve both communication and homeoactive sensing. I actually suspect this is quite common in electric fish (both gymnotiforms and mormyrids), and perhaps in other actively sensing species such as echolocating animals. The two are not mutually exclusive.

We agree with the Reviewer that context - broadly speaking - does affect chirping (as we mentioned above). We hope we have improved the writing and clarified that we do not dismiss communication functions of chirping, but we do lean towards electrolocation based on the considerations above made and our results.

We do conclude the manuscript remarking that communication and electrolocation are not mutually exclusive: ”probing cues could function simultaneously as proximity signals to signal presence, deter approaches, or coordinate behaviors like spawning, if properly timed (Henninger et al., 2018).” (see the conclusion paragraph of the discussion) .

Therein, we further add “These findings aim to stir the pot and initiate a discussion on possible alternative functions of chirps beyond their presumed communication role.”.

With this, we hope we’ve made it clear how we intend our manuscript to be read.

Reviewer #3 (Public Review):

Summary:

This important paper provides the best-to-date characterization of chirping in weakly electric fish using a large number of variables. These include environment (free vs divided fish, with or without clutter), breeding state, gender, intruder vs resident, social status, locomotion state and social and environmental experience, without and with playback experiments. It applies state-of-the-art methods for reducing the dimensionality of the data and finding patterns of correlation between different kinds of variables (factor analysis, K-means). The strength of the evidence, collated from a large number of trials with many controls, leads to the conclusion that the traditionally assumed communication function of chirps may be secondary to its role in environmental assessment and exploration that takes social context into account. Based on their extensive analyses, the authors suggest that chirps are mainly used as probes that help detect beats caused by other fish and as well as objects.

Strengths:

The work is based on completely novel recordings using interaction chambers. The amount of new data and associated analyses is simply staggering, and yet, well organized in presentation. The study further evaluates the electric field strength around a fish (via modelling with the boundary element method) and how its decay parallels the chirp rate, thereby relating the above variables to electric field geometry.

The main conclusions are that the lack of any significant behavioural correlates for chirping, and the lack of temporal patterning in chirp time series, cast doubt on a primary communication goal for most chirps. Rather, the key determinants of chirping are the difference frequency between two interacting conspecifics as well as individual subjects' environmental and social experience. The paper concludes that there is a lack of evidence for stereotyped temporal patterning of chirp time series, as well as of sender-receiver chirp transitions beyond the known increase in chirp frequency during an interaction.

These conclusions by themselves will be very useful to the field. They will also allow scientists working on other "communication" systems to perhaps reconsider and expand the goals of the probes used in those senses. A lot of data are summarized in this paper, with thorough referencing to past work.

The alternative hypotheses that arise from the work are that chirps are mainly used as environmental probes for better beat detection and processing and object localization, and in this sense are self-directed signals. This led to their prediction that environmental complexity ("clutter") should increase chirp rate, which is fact was revealed by their new experiments. The authors also argue that waveform EODs have less power across high spatial frequencies compared to pulse-type fish, with a resulting relatively impoverished power of resolution. Chirping in wave-type fish could temporarily compensate for the lower frequency resolution while still being able to resolve EOD perturbations with a good temporal definition (which pulse-type fish lack due to low pulse rates).

The authors also advance the interesting idea that the sinusoidal frequency modulations caused by chirps are the electric fish's solution to the minute (and undetectable by neural wetware) echo-delays available to it, due to the propagation of electric fields at the speed of light in water. The paper provides a number of experimental avenues to pursue in order to validate the non-communication role of chirps.

We thank the reviewer for the kind assessment.

Weaknesses:

My main criticism is that the alternative putative role for chirps as probe signals that optimize beat detection could be better developed. The paper could be clearer as to what that means precisely, especially since beating - and therefore detection of some aspects of beating due to the proximity of a conspecific - most often precedes chirping. One meaning the authors suggest, tentatively, is that the chirps could enhance electrosensory responses to the beat, for example by causing beat phase shifts that remediate blind spots in the electric field of view.

We agree with the Reviewer that a better and more detailed explanation of how beat processing for conspecific electrolocation may be positively affected by chirps would be important to provide. We are currently working on a follow-up manuscript in which we intend to include these aspects. For space limitations and readability we had to discard from the current manuscript a lot of results that could further clarify these issues.

A second criticism is that the study links the beat detection to underwater object localization. The paper does not significantly develop that line of thought given their data - the authors tread carefully here given the speculative aspect of this link. It is certainly possible that the image on the fish's body of an object in the environment will be slightly modified by introducing a chirp on the waveform, as this may enhance certain heterogeneities of the object in relation to its environment. The thrust of this argument derives mainly from the notion of Fourier analysis with pulse type fish EOD waveforms (see above, and radar theory more generally), where higher temporal frequencies in the beat waveform induced by the chirp will enable a better spatial resolution of objects. It remains to be seen whether experiments can show this to be significant.

Perhaps the Reviewer refers to the last discussion paragraph before the conclusions in which we mention the performance of pulse or wave-type EODs in electrolocation (referring here to ideas illustrated in a recent review by Crampton, 2019). We added to this paragraph a statement which could better clarify that we do not propose that chirping could enhance object electrolocation. What we mean is that, in a context in which object electrolocation occurs through wave-type EODs - given the generally lower performance of such narrow-band signals in resolving the spatial features of any object, even a 3D electric field  - chirping could improve beat detection during social encounters by increasing the amount of information obtained by the fish.

The edited paragraph now reads: “While broadband pulse signals may be useful to capture highly complex environments rich in foliage, roots and other structures common in vegetation featuring the more superficial habitats in which pulse-type fish live, wave-type EODs may be a better choice in the relatively simpler river-bed environments in which many wave-type fish live (e.g., the benthic zone of deep river channels; Crampton, 2019). In this case, achieving a good spatial resolution is critical during social encounters, especially considering the limited utility of visual cues in these low-light conditions. In such habitats, social encounters may “electrically” be less “abrupt”, but spatially less “conspicuous” or blurred (as a 3D electric field may be). In such a scenario, chirps could serve as a means to supplement the spatial information acquired via the beat, accentuating these cues during periods of reduced resolution.”

Recommendations for the authors:

Reviewer #3 (Recommendations For The Authors):

None, my points in the original review have been properly addressed in this resubmission.

Author response:

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

Reviewer #1 (Public Review):  

Summary: 

In this study, Nandy and colleagues examine neural and behavioral correlates of perceptual variability in monkeys performing a visual change detection task. They used a laminar probe to record from area V4 while two macaque monkeys detected a small change in stimulus orientation that occurred at a random time in one of two locations, focusing their analysis on stimulus conditions where the animal was equally likely to detect (hit) or not-detect (miss) a briefly presented orientation change (target). They discovered two behavioral measures that are significantly different between hit and miss trials - pupil size tends to be slightly larger on hits vs. misses, and monkeys are more likely to miss the target on trials in which they made a microsaccade shortly before target onset. They also examined multiple measures of neural activity across the cortical layers and found some measures that are significantly different between hits and misses. 

Strengths: 

Overall the study is well executed and the analyses are appropriate (though multiple issues do need to be addressed). 

We thank the reviewer for their enthusiasm and their constructive comments which we address below.

Weaknesses: 

My main concern with this study is that with the exception of the pre-target microsaccades, the physiological and behavioral correlates of perceptual variability (differences between hits and misses) appear to be very weak and disconnected. Some of these measures rely on complex analyses that are not hypothesis-driven and where statistical significance is difficult to assess. The more intuitive analysis of the predictive power of trial outcomes based on the behavioral and neural measures is only discussed at the end of the paper. This analysis shows that some of the significant measures have no predictive power, while others cannot be examined using the predictive power analysis because these measures cannot be estimated in single trials. Given these weak and disconnected effects, my overall sense is that the current results do not significantly advance our understanding of the neural basis of perceptual variability. 

Reviewer #1 (Recommendations For The Authors): 

(1) Most of the effects are very small. For example, the difference in pupil size between hits and misses is ~0.08 z-score units. The differences in firing rates between hits and misses are in the order of 1-2% of normalized firing rates. While these effects may be significant, their contribution to perceptual variability could be negligible, as suggested by the analysis of predictive power at the end of the result section. On a related note, it would be useful to mention the analysis of predictive power earlier in the paper. The finding that some of the measures do not have significant predictive power w/r to behavioral outcome raises questions regarding their importance. Finally, it would strengthen the paper if the authors could come up with methods to assess the predictive power of the PPC and interlaminar SSC. Without such analyses, it is difficult to assess the importance of these measures. 

We expect that relatively small differences in early to intermediate sensory areas could cumulatively result in large differences in higher areas and contribute to the binary distinction between hits and misses. We certainly do not claim that these results completely explain state-dependent differences that determine the outcome of these trials. Instead, we have focused on neural signatures at the level of the V4 columnar microcircuit that might ultimately contribute to the variability in perception.

We would like to emphasize that, based on the reviewer’s recommendation, we have now analyzed our results separately for each animal (see below). The consistency and significance of our findings across both animals give us confidence that what we have reported here are important neural signatures underlying perceptual variability at threshold.

We would also like to note that SSC and PPC are now part of the standard toolkit of systems neuroscience and have been employed in numerous studies to our knowledge. While all measures come with their set of caveats and limitations, these two measures provide a frequency-resolved metric of the relationship between two temporal processes (point or continuous), which we believe provide insights into the interlaminar flow of information that we report here.

Unfortunately, limitations in the GLM method and the reliability of these analyses with limited data make it impossible for these two measures to be included. The GLM requires all variables to be defined for each trial in the input. SSC and PPC can be undefined at low firing rates and require a substantial amount of data to be reliably calculated. While we did consider imputing data or estimating SSC and PPC using multiple trials, we ultimately did not pursue this idea as the purpose of the GLM was to use simultaneous measurements from single trials. 

(2) What is the actual predictive power of the GLM model (i.e., what is the accuracy of predicting whether a given held-out trial will lead to a hit or a miss)? How much of this predictive power is accounted for by the effect of microsaccades? 

As the GLM is not a decoder, it does not classify whether a given left out trial will be a hit or a miss. However, the GLM was highly predictive compared to a constant model. This information has been added to Table 3. The deviance of the GLM with and without microsaccades as a variable was not significantly different (p >0.9).  

(3) The role of stimulus contrast is not explained clearly. Are all the analyses and figures restricted to a single contrast level? Was the contrast the same on both sides? If multiple contrasts are used, could contrast account for some of the observed neural-behavioral covariations? 

All of the analyses include stimuli of all tested contrast levels. Stimulus contrasts were the same at both locations (attended and unattended). We have added a more detailed description of the contrast in hit and miss trials (Lines 289-296 and reproduced that here: 

“Non-target stimulus contrasts were slightly different between hits and misses (mean:

33.1% in hits, 34.0% in misses, permutation test, 𝑝 = 0.02), but the contrast of the target was higher in hits compared to misses (mean: 38.7% in hits, 27.7% in misses, permutation test, 𝑝 = 1.6 𝑒 − 31). Firing rates were normalized by contrast in Figure 3. In all other figures, we considered only non-target stimuli, which had very minor differences in contrast (<1%) across hits and misses. While we cannot completely rule out any other effects of stimulus contrast, the normalization in Figure 3 and minor differences for non-target stimuli should minimize them.”

(4) Do the animals make false alarms (i.e., report seeing a target in non-target epochs)?

If not, then it is not clear that the animals are performing near their perceptual threshold. If the false-alarm rate is non-zero, it should be reported and analyzed for neural/behavioral correlates. Does the logistic regression fit allow for a false alarm rate? More generally, it would be useful to see a summary of behavioral performance, such as distribution of thresholds, lower and upper asymptotes, and detection rates on foil trials vs. matched target trials. 

The logistic regression does allow for a false alarm rate. We have reported additional behavioral parameters in Figure 1-figure supplement 3A-G.  

(5) As far as I can tell, all the analyses in the paper are done on data combined across the two animals. Given that these effects are weak and that the analyses are complex, it is important to demonstrate for each analysis/figure that the results hold for each animal separately before combining the data across animals. This can be done in supplementary figures. 

We have updated the paper to include all main results plotted separately for each animal as supplementary figures. 

- Figure 2-figure supplement 2

- Figure 3-figure supplement 1

- Figure 3-figure supplement 2

- Figure 4-figure supplement 1

- Figure 5-figure supplement 2

- Figure 7-figure supplement 1

All the results except for the canonical correlation analysis were present, consistent, and significant when we analyzed them in each monkey independently.

(6) The selection of the temporal interval used for the various analyses appears somewhat post hoc and is not explained clearly. Some analyses are restricted to the period immediately before or during target onset (e.g., 400 ms before target onset for analysis of the effect of microsaccade, 60 ms before stimulus onset for the analysis of the effect of neural variability). Other analyses are done on non-target rather than target stimuli. What is the justification for selecting these particular periods for these analyses? The differences in firing rates between hits and misses are restricted to the target epoch and are not present in the non-target epochs. Given these results, it seems important to compare the effects in target and non-target epochs in other analyses as well.

Restricting the analysis of the Fano Factor to 60 ms before non-target onset seems odd. Given that the duration of the interval between stimulus presentations is random, how could this pre-stimulus effect be time-locked to target onset? 

We selected a 200ms time window during the pre-stimulus or stimulus-evoked period for almost all our analyses. The results relating to microsaccade occurrence were robust to narrower time windows more consistent with the other pre-stimulus windows we used, but we chose to use the 400ms window to capture a larger fraction of trials with microsaccades. 

Only the Fano factor time window was selected post-hoc based on the traces in Figure 4A, and the result is robust across animals (new Figure 4-figure supplement 1). The inter-stimulus intervals are random, and we do not believe the neural variability is timelocked to upcoming stimuli, but that lower variability in this pre-stimulus window is characteristic of hits. 

We believe that the consistency of our results across both animals provides further evidence that our time window selection was appropriate. 

We are interested in the extent to which these effects would remain consistent when applied only to target stimuli. However, restricting our analyses to only target stimuli substantially reduces the amount of neural data available for analysis. We plan to explore target stimulus representation more thoroughly in future studies.   

(7) Can the measured neural response be used to discriminate between target and nontarget stimuli? If so, is the discriminability between target and non-target higher in hits vs. misses? 

Thank you for raising this interesting point. We performed this analysis and find that target stimuli are more discriminable from non-targets in hits compared to misses. This has been added as a new Figure 3A.  

(8) How many trials were performed per session? Did miss probability tend to increase over time over the session? If so, could this slow change in hit probability account for some of the observed neural and behavioral correlations with perceptual decisions? 

Monkeys initiated a median of 905 trials (range of 651 to 1086). This has been added to the manuscript (Line 106). Approximately 1/8 of those trials were at perceptual threshold. Hit probability at threshold does not change substantially over the course of the session. We now report this in new Figure 1- figure Supplement 3I (error bars show standard deviation). 

(9) Did miss probability depend on the time of the change within the trial? If so, do any of the behavioral/neural metrics share a similar within-trial time course? 

Change times were not significantly different across hit and miss trials (p=0.15, Wilcoxon rank sum test). We now report this in new Figure 1-figure supplement 3H.

(10) "Deep layer neurons exhibit reduced low-frequency phase-locking in hit trials than in misses (Figure 5B), suggesting an improvement in pooled signal-to-noise among this neural population." - why does this metric suggest improved SNR? Is there any evidence for improved SNR in the data? Why just in deep layers? 

Thank you for raising this question. We agree this statement is not fully supported by the data and have removed it.  

(11) I may have missed this but what were the sizes of the Gabor stimuli? 

This has been added to the methods section (Line 454). The Gaussian halfwidth was 2 degrees.  

Reviewer #2 (Public Review):  

In this manuscript, the authors conducted a study in which they measured eye movements, pupil diameter, and neural activity in V4 in monkeys engaged in a visual attention task. The task required the monkeys to report changes in the orientation of Gabors' visual stimuli. The authors manipulated the difficulty of the trials by varying the degree of orientation change and focused their analysis on trials of intermediate difficulty where the monkeys' hit rate was approximately 50%. Their key findings include the following: 1) Hit trials were preceded by larger pupil diameter, reflecting higher arousal, and by more stable eye positions; 2) V4 neurons exhibit larger visual responses in hit trials; 3) Superficial and deep layers exhibited greater coherence in hit trials during both the pre-target stimulus period and the non-target stimulus presentation period. These findings have useful implications for the field, and the experiments and analyses presented in this manuscript validly support the authors' claims. 

Strengths: 

The experiments were well-designed and executed with meticulous control. The analyses of both behavioural and electrophysiological data align with the standards in the field. 

We thank the reviewer for their enthusiasm about our study and their constructive comments which we address below.

Weaknesses: 

Many of the findings appear to be incremental compared to previous literature, including the authors' own work. While incremental findings are not necessarily a problem, the manuscript lacks clear statements about the extent to which the dataset, analysis, and findings overlap with the authors' prior research. For example, one of the main findings, which suggests that V4 neurons exhibit larger visual responses in hit trials (as shown in Fig. 3), appears to have been previously reported in their 2017 paper. Additionally, it seems that the entire Fig1-S1 may have been reused from the 2017 paper. These overlaps should have been explicitly acknowledged and correctly referenced. 

While the raw data used in this paper overlaps entirely with Nandy et al. (2017), all the analyses and findings in this manuscript are new and have not been previously reported. Figure 1-figure supplement 1 is modified and reproduced from that paper only to allow readers to understand the recording methods used to collect the data without needing to go back to the previous paper. We have added an explicit acknowledgment of this to the figure caption.

Previous studies have demonstrated that attention leads to decorrelation in V4 population activity. The authors should have discussed how and why the high coherence across layers observed in the current study can coexist with this decorrelation. 

We have updated the discussion section (Lines 347-351) to further elaborate on this interpretation. 

Furthermore, the manuscript does not explore potentially interesting aspects of the dataset. For instance, the authors could have investigated instances where monkeys made 'false' reports, such as executing saccades towards visual stimuli when no orientation change occurred. It would be valuable to provide the fraction of the monkeys' responses in a session, including false reports and correct rejections in catch trials, to allow for a broader analysis that considers the perceptual component of neural activity over pure sensory responses. 

We appreciate this feedback. While we agree these are interesting directions, we decided to limit the scope of this study to only focus on trials at threshold with an orientation change, and are considering these directions for future studies. 

Reviewer #2 (Recommendations For The Authors): 

• Figure Design: Since eLife does not impose space limitations, it is advisable for the authors to avoid using very small font sizes. Consistency in font size throughout the figures is recommended. Some figures are challenging to discern, for example, the mean+-sem in Fig. 2B, and the alpha values of green and purple colours for superficial/deep layers are too high, making them too transparent or pale. 

We have increased the size of some small fonts and improved font size consistency throughout the figures. We have changed the layer colors to improve legibility. 

• Line 119: trail, 

This has been fixed.

Exactly, it is illegal to trespass in the capitol. I've been in the capitol, and you're only allowed there under very controlled circumstances, when you're with a tour, or when you have official business and an appointment with congressional staff. You have to go through security, etc. So, the protestors who broke in were trespassing and breaking the law. To argue that after the doors and windows were beaten down that they were "invited in" by security guards that were scared shitless at that point is absolute BS.

何をやってもダメだったのにこれで回復した。 自分は一般的にはPCの知識がある方だと思っていたけど、これは知らなかった。 本当に感謝。

improve

vi behöver även lägga till (som attachment eller som en del av resultatet) available discharged power" for all cases, at the end of the discharge season

det saknas minimum discharge temperature

det saknas minimum discharge temperature

figure 9 is missing

har TRNSYS beräknat Rb eller har vi angett ett lambdavärde för grouting lika med 3?

see earlier comment. This needs to be corrected. The minimum temperature is more important than the maximum

MAKE THESE BE in the same bullet as the text "Number of bhs"

SOFTWARE USED? TRNSYS?

vad mer säger Swecos rapport om detta? lägg gärna till om det anges mer info

se tidigare kommentar. Om vi vill lämna kvar denna text så behöver vi en gång till påminna läsare om källan för detta påstående

se tidigare kommentar. Är det samma resultat än Ruden? Kommentera gärna.

lägg gärna till en rad som berättar att simuleringarna tar hänsyn till olika layers med olika egenskaper och tjocklek

unclear, please clarify for the reader

this is unclear. What do we mean with previous? Specify!

se tidigare resultat. If we want to keep the text, make sure to reformulate it so that it is clear that this is Rudens recommendation, but even better, just remain the reader about the source of this text.

replace with the source

Se tidigare kommentar

move this to an own bullet point

kan vi förtydliga sekvensen mellan dessa? det är lite otydligt.

Vi kan annars upprepa här att sekvensen mellan dessa är oklar.

stor skillnad mellan dessa två case. Är vi säkra på att det inte anges? Vänligen dubbelkolla. Below we write 392. I guess thats the one?

Om man tittar på siffrorna i output, efterliknader siffrorna ett stort lager med 392? eller med 91? vi kan isf lägga till en kommentar där vi skriver vad vi tror de har simulerat.

add source

har vi lambdavärdet för alla andra borrhål? redovisa gärna dem i så fall. Detsamma gäller temperatur och Rb

se tidigare kommentar angående detta. Behöver förtydligas

dollartecken

det finns några konstiga tecken här

(or part of it)

add a source

lägg till en referens till denna text.

Det saknas en mellanrubrik med dokumentnummer enligt tidigare instruktion, dvs projektnummer-X.X-Y.Y. Viktigt!

det saknas mellanrubrik för UA (project responsible) enligt tidigare instruktion. Ni behöver inte fixa det denna gång men mallen måste innehålla det framöver.

I like this part a lot because it gives the reader a new perspective of this character that is in the narrator's life. We also get to see how she comes to realize the deeper part of his life.

Welcome back. In this lesson, I want to talk about CloudWatch, a core product inside AWS used for operational management and monitoring. CloudWatch performs three main jobs: it collects and manages operational data, monitors metrics, and performs actions based on these metrics.

CloudWatch collects and manages operational data generated by an environment, including performance details, nominal operations, and logging data. It can be considered three products in one: CloudWatch, CloudWatch Logs, and CloudWatch Events.

Firstly, CloudWatch allows the collection, monitoring, and actions based on metrics related to AWS products, applications, or on-premises systems. Metrics include data such as CPU utilization, disk space usage, or website traffic. CloudWatch can gather metrics from AWS, on-premises environments, or other cloud platforms using a public internet connection. Some metrics are gathered natively by CloudWatch, while others require the CloudWatch Agent, especially for monitoring non-AWS environments or specific processes on AWS instances.

CloudWatch provides a user interface, command line interface, or API to access and manage this data. The second part of CloudWatch, CloudWatch Logs, handles the collection, monitoring, and actions based on logging data from various sources like Windows event logs, web server logs, and more. For custom logs or non-AWS systems, the CloudWatch Agent is also needed.

The third part is CloudWatch Events, which functions as an event hub. It generates events based on AWS service actions (e.g., starting or stopping an EC2 instance) and can also create scheduled events for specific times or days.

The core concepts of CloudWatch include namespaces, metrics, datapoints, and dimensions. A namespace is a container for monitoring data, helping to organize and separate different areas of data. AWS uses a reserved namespace format (e.g., AWS/EC2 for EC2 metrics), while you can create custom namespaces for your data. Metrics are collections of related data points in a time-ordered structure, such as CPU utilization. Each datapoint includes a timestamp and value. Dimensions, which are name-value pairs, help separate and identify data within a metric, like distinguishing datapoints from different EC2 instances.

CloudWatch also uses alarms to take actions based on metrics. Alarms can be in an OK state (indicating no issues), an ALARM state (indicating a problem), or an INSUFFICIENT_DATA state (indicating not enough data to assess). Actions could include notifications or more complex responses. You’ve already seen an example of this with the billing alarm created at the start of the course.

In the next demo lesson, we’ll provision an EC2 instance, let it run, and then create an alarm to monitor CPU usage, providing practical exposure to how CloudWatch works.

Thanks for watching. Complete this video and join me in the demo when you’re ready.

add admin user using windows pro +

Securing Windows 10 With A Separate Administrator Account

Reviewer #1 (Public Review):

Summary:

This study addresses the question of how task-relevant sensory information affects activity in the motor cortex. The authors use various approaches to address this question, looking at single units and population activity. They find that there are three subtypes of modulation by sensory information at the single unit level. Population analyses reveal that sensory information affects the neural activity orthogonally to motor output. The authors then compare both single unit and population activity to computational models to investigate how encoding of sensory information at the single unit level is coordinated in a network. They find that an RNN that displays similar orbital dynamics and sensory modulation to the motor cortex also contains nodes that are modulated similarly to the three subtypes identified by the single unit analysis.

Strengths:

The strengths of this study lie in the population analyses and the approach of comparing single-unit encoding to population dynamics. In particular, the analysis in Figure 3 is very elegant and informative about the effect of sensory information on motor cortical activity. The task is also well designed to suit the questions being asked and well controlled.

It is commendable that the authors compare single units to population modulation. The addition of the RNN model and perturbations strengthen the conclusion that the subtypes of individual units all contribute to the population dynamics. However, the subtypes (PD shift, gain, and addition) are not sufficiently justified. The authors also do not address that single units exhibit mixed modulation, but RNN units are not treated as such.

Weaknesses:

The main weaknesses of the study lie in the categorization of the single units into PD shift, gain, and addition types. The single units exhibit clear mixed selectivity, as the authors highlight. Therefore, the subsequent analyses looking only at the individual classes in the RNN are a little limited. Another weakness of the paper is that the choice of windows for analyses is not properly justified and the dependence of the results on the time windows chosen for single-unit analyses is not assessed. This is particularly pertinent because tuning curves are known to rotate during movements (Sergio et al. 2005 Journal of Neurophysiology).

This paper shows sensory information can affect motor cortical activity whilst not affecting motor output. However, it is not the first to do so and fails to cite other papers that have investigated sensory modulation of the motor cortex (Stavinksy et al. 2017 Neuron, Pruszynski et al. 2011 Nature, Omrani et al. 2016 eLife). These studies should be mentioned in the Introduction to capture better the context around the present study. It would also be beneficial to add a discussion of how the results compare to the findings from these other works.

This study also uses insights from single-unit analysis to inform mechanistic models of these population dynamics, which is a powerful approach, but is dependent on the validity of the single-cell analysis, which I have expanded on below.

I have clarified some of the areas that would benefit from further analysis below:

(1) Task:<br /> The task is well designed, although it would have benefited from perhaps one more target speed (for each direction). One monkey appears to have experienced one more target speed than the others (seen in Figure 3C). It would have been nice to have this data for all monkeys.

(2) Single unit analyses:<br /> In some analyses, the effects of target speed look more driven by target movement direction (e.g. Figures 1D and E). To confirm target speed is the main modulator, it would be good to compare how much more variance is explained by models including speed rather than just direction. More target speeds may have been helpful here too.

The choice of the three categories (PD shift, gain addition) is not completely justified in a satisfactory way. It would be nice to see whether these three main categories are confirmed by unsupervised methods.

The decoder analyses in Figure 2 provide evidence that target speed modulation may change over the trial. Therefore, it is important to see how the window considered for the firing rate in Figure 1 (currently 100ms pre - 100ms post movement onset) affects the results.

(3) Decoder:<br /> One feature of the task is that the reach endpoints tile the entire perimeter of the target circle (Figure 1B). However, this feature is not exploited for much of the single-unit analyses. This is most notable in Figure 2, where the use of a SVM limits the decoding to discrete values (the endpoints are divided into 8 categories). Using continuous decoding of hand kinematics would be more appropriate for this task.

(4) RNN:<br /> Mixed selectivity is not analysed in the RNN, which would help to compare the model to the real data where mixed selectivity is common. Furthermore, it would be informative to compare the neural data to the RNN activity using canonical correlation or Procrustes analyses. These would help validate the claim of similarity between RNN and neural dynamics, rather than allowing comparisons to be dominated by geometric similarities that may be features of the task. There is also an absence of alternate models to compare the perturbation model results to.

Author response:

eLife assessment

This useful study examines the neural activity in the motor cortex as a monkey reaches to intercept moving targets, focusing on how tuned single neurons contribute to an interesting overall population geometry. The presented results and analyses are solid, though the investigation of this novel task could be strengthened by clarifying the assumptions behind the single neuron analyses, and further analyses of the neural population activity and its relation to different features of behaviour.

Thanks for recognizing the content of our research, and please stay tuned for our follow-up studies on neural dynamics during interception.

Public Reviews:

Reviewer #1 (Public Review):

Summary:

This study addresses the question of how task-relevant sensory information affects activity in the motor cortex. The authors use various approaches to address this question, looking at single units and population activity. They find that there are three subtypes of modulation by sensory information at the single unit level. Population analyses reveal that sensory information affects the neural activity orthogonally to motor output. The authors then compare both single unit and population activity to computational models to investigate how encoding of sensory information at the single unit level is coordinated in a network. They find that an RNN that displays similar orbital dynamics and sensory modulation to the motor cortex also contains nodes that are modulated similarly to the three subtypes identified by the single unit analysis.

Strengths:

The strengths of this study lie in the population analyses and the approach of comparing single-unit encoding to population dynamics. In particular, the analysis in Figure 3 is very elegant and informative about the effect of sensory information on motor cortical activity. The task is also well designed to suit the questions being asked and well controlled.

We appreciate these kind comments.

It is commendable that the authors compare single units to population modulation. The addition of the RNN model and perturbations strengthen the conclusion that the subtypes of individual units all contribute to the population dynamics. However, the subtypes (PD shift, gain, and addition) are not sufficiently justified. The authors also do not address that single units exhibit mixed modulation, but RNN units are not treated as such.

We’re sorry for not providing sufficient grounds to introduce the subtypes. We determined the PD shift, gain, and addition as pertinent subtypes based on classical cosine tuning model (Georgopoulos et al., 1982) and referred to some gain modulation studies (e.g. Pesaran et al. 2010, Bremner and Andersen, 2012). Here, we applied this subtype analysis as a criteria to identify the modulation in neuronal population rather than to sort neuron into distinct cell types. We will update Methods in the revised version of manuscript.

Weaknesses:

The main weaknesses of the study lie in the categorization of the single units into PD shift, gain, and addition types. The single units exhibit clear mixed selectivity, as the authors highlight. Therefore, the subsequent analyses looking only at the individual classes in the RNN are a little limited. Another weakness of the paper is that the choice of windows for analyses is not properly justified and the dependence of the results on the time windows chosen for single-unit analyses is not assessed. This is particularly pertinent because tuning curves are known to rotate during movements (Sergio et al. 2005 Journal of Neurophysiology).

The mixed selectivity or precisely the mixed modulation is indeed a significant feature of neuronal population in the present study. The purpose of the subtype analysis was to serve as a criterion for the potential modulation mechanisms. However, the results appear to be a spectrum than clusters. It still through some insights to understand the modulation distribution and we will refine the description in the next version. In the current version, we observed single-unit tuning and population neural state with sliding windows, focusing on the period around movement onset (MO) due to the emergence of a ring-like structure. We will clarify the choice of windows and the dependence assessment in the next version. It’s a great suggestion to consider the role of rotating tuning curves in neural dynamics during interception.

This paper shows sensory information can affect motor cortical activity whilst not affecting motor output. However, it is not the first to do so and fails to cite other papers that have investigated sensory modulation of the motor cortex (Stavinksy et al. 2017 Neuron, Pruszynski et al. 2011 Nature, Omrani et al. 2016 eLife). These studies should be mentioned in the Introduction to capture better the context around the present study. It would also be beneficial to add a discussion of how the results compare to the findings from these other works.

Thanks for the reminder. We will introduce the relevant research in the next version of manuscript.

This study also uses insights from single-unit analysis to inform mechanistic models of these population dynamics, which is a powerful approach, but is dependent on the validity of the single-cell analysis, which I have expanded on below.

I have clarified some of the areas that would benefit from further analysis below:

(1) Task:

The task is well designed, although it would have benefited from perhaps one more target speed (for each direction). One monkey appears to have experienced one more target speed than the others (seen in Figure 3C). It would have been nice to have this data for all monkeys.

Great suggestion! However, it’s hard to implement as the implanted arrays have been removed.

(2) Single unit analyses:

In some analyses, the effects of target speed look more driven by target movement direction (e.g. Figures 1D and E). To confirm target speed is the main modulator, it would be good to compare how much more variance is explained by models including speed rather than just direction. More target speeds may have been helpful here too.

Nice suggestion! The fitting goodness of the simple model (just motor direction) is much less than the complex model (including target speed). We will update the results in the next version.

The choice of the three categories (PD shift, gain addition) is not completely justified in a satisfactory way. It would be nice to see whether these three main categories are confirmed by unsupervised methods.

A good point. We will have a try with unsupervised methods. 

The decoder analyses in Figure 2 provide evidence that target speed modulation may change over the trial. Therefore, it is important to see how the window considered for the firing rate in Figure 1 (currently 100ms pre - 100ms post movement onset) affects the results.

Thanks for the suggestion and close reading. We will test the decoder in other epochs.

(3) Decoder:

One feature of the task is that the reach endpoints tile the entire perimeter of the target circle (Figure 1B). However, this feature is not exploited for much of the single-unit analyses. This is most notable in Figure 2, where the use of a SVM limits the decoding to discrete values (the endpoints are divided into 8 categories). Using continuous decoding of hand kinematics would be more appropriate for this task.

This is a very reasonable suggestion. In this study, we discrete the reach-direction as the previous studies (Li et al., 2018&2022) and thought that the discrete decoding was already enough to show the interaction of sensory and motor variables. In future studies, we will try continuous decoding of hand kinematics.

(4) RNN:

Mixed selectivity is not analysed in the RNN, which would help to compare the model to the real data where mixed selectivity is common. Furthermore, it would be informative to compare the neural data to the RNN activity using canonical correlation or Procrustes analyses. These would help validate the claim of similarity between RNN and neural dynamics, rather than allowing comparisons to be dominated by geometric similarities that may be features of the task. There is also an absence of alternate models to compare the perturbation model results to.

Thank you for these helpful suggestions. We will perform decoding analysis on RNN units to verify if there is interaction of sensory and motor variables as in real data, as well as the canonical correlation or Procrustes analysis.

Reviewer #2 (Public Review):

Summary:

In this manuscript, Zhang et al. examine neural activity in the motor cortex as monkeys make reaches in a novel target interception task. Zhang et al. begin by examining the single neuron tuning properties across different moving target conditions, finding several classes of neurons: those that shift their preferred direction, those that change their modulation gain, and those that shift their baseline firing rates. The authors go on to find an interesting, tilted ring structure of the neural population activity, depending on the target speed, and find that (1) the reach direction has consistent positioning around the ring, and (2) the tilt of the ring is highly predictive of the target movement speed. The authors then model the neural activity with a single neuron representational model and a recurrent neural network model, concluding that this population structure requires a mixture of the three types of single neurons described at the beginning of the manuscript.

Strengths:

I find the task the authors present here to be novel and exciting. It slots nicely into an overall trend to break away from a simple reach-to-static-target task to better characterize the breadth of how the motor cortex generates movements. I also appreciate the movement from single neuron characterization to population activity exploration, which generally serves to anchor the results and make them concrete. Further, the orbital ring structure of population activity is fascinating, and the modeling work at the end serves as a useful baseline control to see how it might arise.

Thank you for recognizing our work.

Weaknesses:

While I find the behavioral task presented here to be excitingly novel, I find the presented analyses and results to be far less interesting than they could be. Key to this, I think, is that the authors are examining this task and related neural activity primarily with a single-neuron representational lens. This would be fine as an initial analysis since the population activity is of course composed of individual neurons, but the field seems to have largely moved towards a more abstract "computation through dynamics" framework that has, in the last several years, provided much more understanding of motor control than the representational framework has. As the manuscript stands now, I'm not entirely sure what interpretation to take away from the representational conclusions the authors made (i.e. the fact that the orbital population geometry arises from a mixture of different tuning types). As such, by the end of the manuscript, I'm not sure I understand any better how the motor cortex or its neural geometry might be contributing to the execution of this novel task.

The present study shows the sensory modulation on motor tuning in single units and neural state during motor execution period. It’s a pity that the findings were constrained in certain time windows. We are still working this topic, and hopefully will address related questions in our follow-up studies.

Main Comments:

My main suggestions to the authors revolve around bringing in the computation through a dynamics framework to strengthen their population results. The authors cite the Vyas et al. review paper on the subject, so I believe they are aware of this framework. I have three suggestions for improving or adding to the population results:

(1) Examination of delay period activity: one of the most interesting aspects of the task was the fact that the monkey had a random-length delay period before he could move to intercept the target. Presumably, the monkey had to prepare to intercept at any time between 400 and 800 ms, which means that there may be some interesting preparatory activity dynamics during this period. For example, after 400ms, does the preparatory activity rotate with the target such that once the go cue happens, the correct interception can be executed? There is some analysis of the delay period population activity in the supplement, but it doesn't quite get at the question of how the interception movement is prepared. This is perhaps the most interesting question that can be asked with this experiment, and it's one that I think may be quite novel for the field--it is a shame that it isn't discussed.

Great idea! We are on the way, and close to complete the puzzle.

(2) Supervised examination of population structure via potent and null spaces: simply examining the first three principal components revealed an orbital structure, with a seemingly conserved motor output space and a dimension orthogonal to it that relates to the visual input. However, the authors don't push this insight any further. One way to do that would be to find the "potent space" of motor cortical activity by regression to the arm movement and examine how the tilted rings look in that space (this is actually fairly easy to see in the reach direction components of the dPCA plot in the supplement--the rings will be highly aligned in this space). Presumably, then, the null space should contain information about the target movement. dPCA shows that there's not a single dimension that clearly delineates target speed, but the ring tilt is likely evident if the authors look at the highest variance neural dimension orthogonal to the potent space (the "null space")--this is akin to PC3 in the current figures, but it would be nice to see what comes out when you look in the data for it.

Nice suggestion. Target-speed modulation mainly influences PC3, which is consistent with ‘null space’ hypothesis. We will try other methods of dimensionality reduction (e.g. dPCA, Manopt) to determine the potent and null space.

(3) RNN perturbations: as it's currently written, the RNN modeling has promise, but the perturbations performed don't provide me with much insight. I think this is because the authors are trying to use the RNN to interpret the single neuron tuning, but it's unclear to me what was learned from perturbing the connectivity between what seems to me almost arbitrary groups of neurons (especially considering that 43% of nodes were unclassifiable). It seems to me that a better perturbation might be to move the neural state before the movement onset to see how it changes the output. For example, the authors could move the neural state from one tilted ring to another to see if the virtual hand then reaches a completely different (yet predictable) target. Moreover, if the authors can more clearly characterize the preparatory movement, perhaps perturbations in the delay period would provide even more insight into how the interception might be prepared.

We are sorry that we didn’t clarify the definition of “none” type, which can be misleading. The 43% unclassified nodes include those inactive ones, when only activate (task-related) nodes included, the ratio of unclassified nodes would be much lower. By perturbing the connectivity, we intended to explore the interaction between different modulations.

Thank you for the great advice. We tried moving neural states from one ring to another without changing the directional cluster, but this perturbation didn’t have a significant influence on network performance as expected. We will check this result again and try perturbations in the delay period.

Reviewer #3 (Public Review):

Summary:

This experimental study investigates the influence of sensory information on neural population activity in M1 during a delayed reaching task. In the experiment, monkeys are trained to perform a delayed interception reach task, in which the goal is to intercept a potentially moving target.

This paradigm allows the authors to investigate how, given a fixed reach endpoint (which is assumed to correspond to a fixed motor output), the sensory information regarding the target motion is encoded in neural activity.

At the level of single neurons, the authors found that target motion modulates the activity in three main ways: gain modulation (scaling of the neural activity depending on the target direction), shift (shift of the preferred direction of neurons tuned to reach direction), or addition (offset to the neural activity).

At the level of the neural population, target motion information was largely encoded along the 3rd PC of the neural activity, leading to a tilt of the manifold along which reach direction was encoded that was proportional to the target speed. The tilt of the neural manifold was found to be largely driven by the variation of activity of the population of gain-modulated neurons.

Finally, the authors studied the behaviour of an RNN trained to generate the correct hand velocity given the sensory input and reach direction. The RNN units were found to similarly exhibit mixed selectivity to the sensory information, and the geometry of the « neural population » resembled that observed in the monkeys.

Strengths:

- The experiment is well set up to address the question of how sensory information that is directly relevant to the behaviour but does not lead to a direct change in behavioural output modulates motor cortical activity.

- The finding that sensory information modulates the neural activity in M1 during motor preparation and execution is non trivial, given that this modulation of the activity must occur in the nullspace of the movement.

- The paper gives a complete picture of the effect of the target motion on neural activity, by including analyses at the single neuron level as well as at the population level. Additionally, the authors link those two levels of representation by highlighting how gain modulation contributes to shaping the population representation.

Thanks for your recognition.

Weaknesses:

- One of the main premises of the paper is the fact that the motor output for a given reach point is preserved across different target motions. However, as the authors briefly mention in the conclusion, they did not record muscle activity during the task, but only hand velocity, making it impossible to directly verify how preserved muscle patterns were across movements. While the authors highlight that they did not see any difference in their results when resampling the data to control for similar hand velocities across conditions, this seems like an important potential caveat of the paper whose implications should be discussed further or highlighted earlier in the paper.

Thanks for the suggestion. We will highlight the resampling results as important control in the next version of manuscript.

- The main takeaway of the RNN analysis is not fully clear. The authors find that an RNN trained given a sensory input representing a moving target displays modulation to target motion that resembles what is seen in real data. This is interesting, but the authors do not dissect why this representation arises, and how robust it is to various task design choices. For instance, it appears that the network should be able to solve the task using only the motion intention input, which contains the reach endpoint information. If the target motion input is not used for the task, it is not obvious why the RNN units would be modulated by this input (especially as this modulation must lie in the nullspace of the movement hand velocity if the velocity depends only on the reach endpoint). It would thus be important to see alternative models compared to true neural activity, in addition to the model currently included in the paper. Besides, for the model in the paper, it would therefore be interesting to study further how the details of the network setup (eg initial spectral radius of the connectivity, weight regularization, or using only the target position input) affect the modulation by the motion input, as well as the trained population geometry and the relative ratios of modulated cells after training.

Great suggestions. It’s a considerable pity that we didn’t dissect the formation reason and influence factor of the representation in the current version. We’ve tried several combinations of inputs before: in the network which received only motor intention and GO inputs, there were rings but not tilting related to target-speed; in the network which received only target location and GO inputs, there were ring-like structures but not clear directional clusters. We will check these results and try alternative models in the next version. In future studies, we will examine the influence of network setup details.

- Additionally, it is unclear what insights are gained from the perturbations to the network connectivity the authors perform, as it is generally expected that modulating the connectivity will degrade task performance and the geometry of the responses. If the authors wish the make claims about the role of the subpopulations, it could be interesting to test whether similar connectivity patterns develop in networks that are not initialized with an all-to-all random connectivity or to use ablation experiments to investigate whether the presence of multiple types of modulations confers any sort of robustness to the network.

Thank you for the great suggestions. By perturbations, we intended to explore the contribution of interaction between certain subpopulations. We tried ablation experiments, but the result was not significant. Probably because the most units were of mixed selectivity, the units of only modulations were not enough for bootstrapping, or the random sampling from single subpopulation (bearing mixed selectivity) could be repeated. We will consider these suggestions carefully in the revised version.

- The results suggest that the observed changes in motor cortical activity with target velocity result from M1 activity receiving an input that encodes the velocity information. This also appears to be the assumption in the RNN model. However, even though the input shown to the animal during preparation is indeed a continuously moving target, it appears that the only relevant quantity to the actual movement is the final endpoint of the reach. While this would have to be a function of the target velocity, one could imagine that the computation of where the monkeys should reach might be performed upstream of the motor cortex, in which case the actual target velocity would become irrelevant to the final motor output. This makes the results of the paper very interesting, but it would be nice if the authors could discuss further when one might expect to see modulation by sensory information that does not directly affect motor output in M1, and where those inputs may come from. It may also be interesting to discuss how the findings relate to previous work that has found behaviourally irrelevant information is being filtered out from M1 (for instance, Russo et al, Neuron 2020 found that in monkeys performing a cycling task, context can be decoded from SMA but not from M1, and Wang et al, Nature Communications 2019 found that perceptual information could not be decoded from PMd)?

How and where sensory information modulates M1 are very interesting and open questions. We will discuss further about this topic in the next version.

Reviewer #1 (Public Review):

Summary:

Previous research from the Margoliash laboratory has demonstrated that the intrinsic electrophysiological properties of one class of projection neurons in the song nucleus HVC, HVCX neurons, are similar within birds and differ between birds in a manner that relates to the bird's song. The current study builds on this research by addressing how intrinsic properties may relate to the temporal structure of the bird's song and by developing a computational model for how this can influence sequence propagation of activity within HVC during singing.

First, the authors identify that the duration of the song motif is correlated with the duration of song syllables and particularly the length of harmonic stacks within the song. They next found positive correlations between some of the intrinsic properties, including firing frequency, sag ratio, and rebound excitation area with the duration of the birds' longest harmonic syllable and some other measure of motif duration. These results were extended by examining measures of firing frequency and sag ratio between two groups of birds that were experimentally raised to learn songs that only differed by the addition of a long terminal harmonic stack in one of the groups. Lastly, the authors present an HH-based model elucidating how the timing and magnitude of rebound excitation of HVCX neurons can function to support previously reported physiological network properties of these neurons during singing.

Strengths:

By trying to describe how intrinsic properties (IPs) may relate to the structure of learned behavior and providing a potentially plausible model (see below for more on this) for how differences in IPs can relate to sequence propagation in this neural network, this research is addressing an important and challenging issue. An understanding of how cell types develop IPs and how those IPs relate to the function and output of a network is a fundamental issue. Tackling this in the zebra finch HVC is an elegant approach because it provides a quantifiable and reliable behavior that is explicitly tied to the neurons that the authors are studying. Nonetheless, this is a difficult problem, and kudos to the authors for trying to unravel this.

Correlations between harmonic stack durations and song durations are well-supported and interesting. This provides a new insight that can and will likely be used by other research groups in correlating neuronal activity patterns to song behavior and motif duration. Additionally, correlations between IPs associated with rebound excitation are also well supported in this study.

The HH-model presented is important because it meaningfully relates how high or low rebound excitation can set the integration time window for HVCX neurons. Further, the synaptic connectivity of this model provides at least one plausible way in how this functions to permit the bursting activity of HVCX neurons during singing (and potentially during song playback experiments in sleeping birds). Thus, this model will be useful to the field for understanding how this network activity intersects with 'learned' IPs in an important class of neurons in this circuit.

Weaknesses:

The main weakness of the study is that there is somewhat of a disconnect between the physiological measurements described and the key components of the circuit model presented at the end of the paper. Thus, better support could be provided to link the magnitude of rebound excitation with song temporal structure. The rebound excitation area is shown to be positively correlated with the longest harmonic stack. Does this correlation hold when the four birds with unusually long stacks (>150ms) are excluded? Is rebound excitation area positively correlated with motif duration? Additionally, rebound excitation was not considered when examining experimentally tutored birds. Further analysis of these correlations can better link this research to the model presented.

The HH model is of general interest, but I am concerned about the plausibility of some of this circuitry, particularly because synaptic connectivity underlying information flow is a central component of the model. At several steps in the model, excitatory drive onto HVCX neurons is coming from another HVCX neuron. Although disynaptic inhibition between HVCX neurons and between HVCRA and HVCX neurons is well established, I am not aware of any data indicating direct synaptic connections between HVCX neurons.

Thus, how does the model change if all excitatory drive onto HVCX neurons are coming from HVCRA neurons? Currently, the model is realized through neurons being active at syllable or gesture transitions. What does the model predict about the distribution of HVCRA neurons activity across songs if they are the exclusive excitatory input to HVCX neurons? A better consideration of these issues can improve the suitability of the model in the context of known connectivity.

If I understand the model and ideas correctly, birds with longer motifs should exhibit longer pauses in the activity of tonically active HVC interneurons during singing and they should exhibit longer post-rebound integration windows. Experimental evidence supporting either of these ideas is not provided and would strengthen this research.

Reviewer #2 (Public Review):

Intrinsic properties of a neuron refer to the ion channels that a neuron expresses. These ion channels determine how a neuron responds to its inputs. How intrinsic properties link to behavior remains poorly understood. Medina and Margoliash address this question using the zebra finch, a well-studied songbird. Previous studies from their lab and other labs have shown that the intrinsic properties of adult songbird basal-ganglia projecting premotor neurons, are more similar within a bird than across birds. Across birds, this similarity is related to the extent of similarity in the songs; the more similar the song between two birds, the more similar the intrinsic properties between the neurons of these two birds. Finally, the intrinsic properties of these neurons change over the course of development and are sensitive to intact auditory feedback. However, the song features that relate to these intrinsic properties and the function of the within-bird homogeneity of intrinsic properties are unclear.

In this manuscript, the authors address these two questions by examining the intrinsic properties of basal-ganglia projecting premotor neurons in zebra finch brain slices. Specifically, they focus on the Ih current (as this is related to rhythmic activity in many pattern-generating circuits) and correlate the properties of the Ih current with song features. They find that the sag ratio (a measure of the driving force of the Ih current) and the rebound area (a measure of the post-inhibitory depolarisation) are both correlated with the temporal features of the song. First, they show the presence of correlations between the length of the song motif and the length of the longest syllable (most often a harmonic stack syllable). Based on this, they conclude that longer song motifs are composed of longer syllables. Second, they show that HVCX neurons within a bird have more similar sag ratios and rebound areas than across birds. Third, the mean sag ratio and mean rebound areas across birds were correlated with the duration of the longest harmonic stack within the song. These two results suggest that IPs are correlated with the temporal structure of the song. To further test this, the authors used natural and experimental tutoring procedures to have birds that learned two different types of songs that only differed in length; the longer song had an extra harmonic stack at the end. Using these two sets of birds, the authors find larger sag ratios and higher firing frequencies in birds with longer songs. Fifth, they show that the post-inhibitory rebound area allows neurons to respond to excitatory inputs and produce spikes. Neurons with a larger rebound area have a larger time window for responding to excitatory inputs. Based on this, they speculate that HVCX neurons with larger rebound areas integrate over larger time windows. Finally, they make a network model of HVC and show that one specific model could explain sequence-specific bursting of HVCX neurons.

Strengths

The question being addressed is an interesting question and the authors use appropriate techniques. The authors find a new temporal structure within the song, specifically, they find that longer songs typically have more syllables and longer syllables. As far as I know, this has not been shown earlier. The authors build on existing literature to suggest that IPs of HVCX neurons are correlated with the temporal structure of songs.

Weaknesses

I have a number of concerns with the statistics and interpretation of the results, insufficient controls for one experiment, and the specifics of the model that affect the implications of these results. These concerns are listed in the recommendations for the authors.

Possible solutions for high CPU usage by vmmem

「pane」を「領域」と約しているようですが、Windowsの用語として「ペイン」はあるようなので、そのまま片仮名にしていいと思います。

https://learn.microsoft.com/ja-jp/windows/terminal/panes https://learn.microsoft.com/ja-jp/windows/apps/design/controls/two-pane-view

Welcome back. In this lesson, I want to introduce another core AWS service, the simple storage service known as S3. If you use AWS in production, you need to understand S3. This lesson will give you the very basics because I'll be deep diving into a specific S3 section later in the course, and the product will feature constantly as we go. Pretty much every other AWS service has some kind of interaction with S3. So let's jump in and get started.

S3 is a global storage platform. It's global because it runs from all of the AWS regions and can be accessed from anywhere with an internet connection. It's a public service. It's regional based because your data is stored in a specific AWS region at rest. So when it's not being used, it's stored in a specific region. And it never leaves that region unless you explicitly configure it to. S3 is regionally resilient, meaning the data is replicated across availability zones in that region. S3 can tolerate the failure of an AZ, and it also has some ability to replicate data between regions, but more on that in the S3 section of the course.

Now S3 might initially appear confusing. If you utilize it from the UI, you appear not to have to select a region. Instead, you select the region when you create things inside S3, which I'll talk about soon. S3 is a public service, so it can be accessed from anywhere as long as you have an internet connection. The service itself runs from the AWS public zone. It can cope with unlimited data amounts and it's designed for multi-user usage of that data. So millions of users could be accessing cute cat pictures added by the Animals for Life Rescue Officers. S3 is perfect for hosting large amounts of data. So think movies or audio distribution, large scale photo storage like stock images, large textual data or big data sets. It could be just as easily used for millions or billions of IOT devices or to store images for a blog. It scales from nothing to near unlimited levels.

Now S3 is economical, it's a great value service for storing and allowing access to data. And it can be accessed using a variety of methods. There's the GUI, you can use the command line, the AWS APIs or even standard methods such as HTTP. I want you to think of S3 as the default storage service in AWS. It should be your default starting point unless your requirement isn't delivered by S3. And I'll talk more about the limitations and use cases later in this lesson.

S3 has two main things that it delivers: Objects and Buckets. Objects are the data the S3 stores, your cat picture, the latest episode of Game of Thrones, which you have stored legally, of course, or it could be large scale datasets showing the migration of the koala population in Australia after a major bushfire. Buckets are containers for objects. It's buckets and objects that I want to cover in this lesson as an introduction to the service.

So first, let's talk about objects. An object in S3 is made up of two main components and some associated metadata. First, there is the object key. And for now you can think of the object key, similar to a file name. The key identifies the object in a bucket. So if you know the object key and the bucket, then you can uniquely access the object, assuming that you have permissions. Remember by default, even for public services, there is no access in AWS initially, except for the account root user.

Now, the other main component of an object is its value. And the value is the data or the contents of the object. In this case, a sequence of binary data, which represents a koala in his house. In this course, I want to avoid suggesting that you remember pointless values. Sometimes though, there are things that you do need to commit to memory. And this is one of those times. The value of an object, in essence, how large the object is, can range from zero bytes up to five terabytes in size. So you can have an empty object or you can have one that is a huge five TB. This is one of the reasons why S3 is so scalable and so useful in a wide range of situations because of this range of sizes for objects.

Now, the other components of an object, aren't that important to know at this stage, but just so you hear the terms that I'll use later, objects also have a version ID, metadata, some access control, as well as sub resources. Now don't worry about what they do for now, I'll be covering them all later. For this lesson, just try to commit to memory what an object is, what components it has and the size range for an object.

So now that we've talked about objects, let's move on and look at buckets. Buckets are created in a specific AWS region. And let's use Sydney or ap-southeast-2 as an example. This has two main impacts. Firstly, your data that's inside a bucket has a primary home region. And it never leaves that region, unless you as an architect or one of your system admins configures that data to leave that region. That means that S3 has stable and controlled data sovereignty. By creating a bucket in a region, you can control what laws and rules apply to that data. What it also means is that the blast radius of a failure is that region.

Now this might be a new term. What I mean by blast radius is that if a major failure occurs, say a natural disaster or a large scale data corruption, the effect of that will be contained within the region. Now a bucket is identified by its name, the bucket name in this case, koala data. A bucket name needs to be globally unique. So that's across all regions and all accounts of AWS. If I pick a bucket name, in this case, koala data, nobody else can use it in any AWS account. Now making a point of stressing this as it often comes up in the exam. Most AWS things are often unique in a region or unique in your account. For example, I can have an IAM user called Fred and you can also have an IAM user called Fred. Buckets though are different, with buckets, the name has to be totally unique, and that's across all regions and all AWS accounts. I've seen it come up in the exam a few times. So this is definitely a point to remember.

Now buckets can hold an unlimited number of objects. And because objects can range from zero to five TB in size, that essentially means that a bucket can hold from zero to unlimited bytes of data. It's an infinitely scalable storage system. Now one of the most important things that I want to say in this lesson is that as an object storage system, an S3 bucket has no complex structure. It's flat, it has flat structure. All objects stored within the bucket at the same level. So this isn't like a file system where you can truly have files within folders, within folders. Everything is stored in the bucket at the root level.

But, if you do a listing on an S3 bucket, you will see what you think are folders. Even the UI presents this as folders. But it is important for you to know at this stage that that's not how it actually is. Imagine a bucket where you see three image files, koala one, two and three dot JPEG. The first thing is that inside S3, there's no concept of file type based on the name. These are just three objects where the object key is koala1.JPEG, koala2.JPEG and koala3.JPEG. Now folders in S3 are represented when we have object names that are structured like these. So the objects have a key, a forward slash old forward slash koala one, two and three dot JPEG. When we create object names like this, then S3 presents them in the UI as a folder called old. So because we've got object names that begin with slash old, then S3 presents this as a folder called old. And then inside that folder, we've got koala one, two, and three dot JPEG.

Now nine out of 10 times, this detail doesn't matter, but I want to make sure that you understand how it actually works. Folders are often referred to as prefixes in S3 because they're part of the object names. They prefix the object names. As you move through the various stages of your AWS learnings, this is gonna make more and more sense. And I'm gonna demonstrate this in the next lesson, which is a demo lesson.

Now to summarize buckets are just containers, they're stored in a region, and for S3, they're generally where a lot of permissions and options are set. So remember that buckets are generally the default place where you should go to, to configure the way the S3 works.

Now, I want to cover a few summary items and then step through some patterns and anti-patterns for S3, before we move to the demo. But first an exam powerup. These are things that you should try to remember and they will really help in the exam. First bucket names are globally unique. Remember that one because it will really help in the exam. I've seen a lot of times where AWS have included trick questions, which test your knowledge of this one. If you get any errors, you can't create a bucket a lot of the time it's because somebody else already has that bucket name.

Now bucket names do have some restrictions. They need to be between 3 and 63 characters, all lower case and no underscores. They have to start with a lowercase letter or a number, and they can't be formatted like IP addresses. So you can't have 1.1.1.1 as your bucket name. Now there are some limits in terms of buckets. Now limits are often things that you don't need to remember for the exam, but this is one of the things that you do. There is a limit of a hundred buckets that you can have in an AWS account. So this is not per region, it's for the entire account. There's a soft limit of 100 and a hard limit so you can increase all the way up to this hard limit using support requests, and this hard limit is a thousand.

Now this matters for architectural reasons. It's not just an arbitrary number. If you're designing a system which uses S3 and users of that system store data inside S3, you can implement a solution that has one bucket per user if you have anywhere near this number of users. So if you have anywhere from a hundred to a thousand users or more of a system, then you're not gonna be able to have one bucket per user because you'll hit this hard limit. You tend to find this in the exam quite often, it'll ask you how to structure a system, which has potentially thousands of users. What you can do is take a single bucket and divide it up using prefixes, so those folders that aren't really folders, and then in that way, you can have multiple users using one bucket. Remember the 100/1000, it's a 100 soft limit and a 1000 hard limit.

You aren't limited in terms of objects in a bucket, you can have zero to an infinite number of objects in a bucket. And each object can range in size from zero bytes to 5 TB in size. And then finally, in terms of the object structure, an object consists of a key, which is its name and then the value, which is the data. And there are other elements to an object which I'll discuss later in the course, but for now, just remember the two main components, the key and the value. Now, all of these points are worth noting down, maybe make them into a set of flashcards and you can use them later on during your studies.

S3 is pretty straightforward and that there tend to be a number of things that it's really good at and a fairly small set of things that it's not suitable for. So let's take a look. S3 is an object storage system. It's not a file storage system, and it's not a block storage system, which are the other main types. What this means is that if you have a requirement where you're accessing the whole of these entities, so the whole of an object, so an image, an audio file, and you're doing all of that at once, then it's a candidate for object storage. If you have a Window server which needs to access a network file system, then it's not S3 that needs to be file-based storage. S3 has no file system, it's flat. So you can't browse to an S3 bucket like you would a file share in Windows. Likewise, it's not block storage, which means you can't mount it as a mount point or a volume on the Linux or Windows. When you're dealing with virtual machines or instances, you mount block storage to them. Block storage is basically virtual hard disks. In EC2, you have EBS, which is block storage. Block storage is generally limited to one thing accessing it at a time, one instance in the case of EBS. S3 doesn't have that single user limitation and it's not block storage, but that means you can't mount it as a drive.

S3 is great for large scale data storage or distribution. Many examples I'll show you throughout the course will fit into that category. And it's also good for offloading things. If you have a blog with lots of posts and lots of images or audio or movies, instead of storing that data on an expensive compute instance, you can move it to an S3 bucket and configure your blog software to point your users at S3 directly. You can often shrink your instance by offloading data onto S3. And don't worry, I'll be demoing this later in the course. Finally, S3 should be your default thought for any input to AWS services or output from AWS services. Most services which consume data and or output data can have S3 as an option to take data from or put data to when it's finished. So if you're faced with any exam questions and there's a number of options on where to store data, S3 should be your default. There are plenty of AWS services which can output large quantities of data or ingest large quantities of data. And most of the time, it's S3, which is an ideal storage platform for that service.

Okay time for a quick demo. And in this demo, we're just gonna run through the process of creating a simple S3 bucket, uploading some objects to that bucket, and demonstrating exactly how the folder functionality works inside S3. And I'm also gonna demonstrate a number of elements of how access and permissions work with S3. So go ahead and complete this video, and when you're ready join me in the next, which is gonna be a demo of S3.

Welcome back. In this lesson, we will introduce an essential AWS service, the Elastic Compute Cloud (EC2). EC2 is a critical service you need to understand for the AWS exam and for designing real-world AWS solutions, as it is the default compute service within AWS. EC2 provides access to virtual machines known as instances. When deploying compute resources that require an operating system, runtime environment, database dependencies, applications, and application interfaces, EC2 is the service to use. It should be the starting point for any compute requirement in AWS.

Let's review some important features of EC2 before discussing architectural points. EC2 is an Infrastructure as a Service (IaaS), providing access to virtual machines called EC2 instances. As an IaaS, the unit of consumption is the instance, which is an operating system configured with allocated resources. EC2 is a private AWS service, meaning it runs in the private AWS zone by default. An EC2 instance is configured to launch into a single VPC subnet, set during instance launch, and requires configuration for public access if needed.

With the default VPC, public access configuration is handled automatically. For custom VPCs, you need to manage this configuration. An EC2 instance's availability zone (AZ) resilience means that if the AZ fails, the instance will likely fail too. Understanding this is crucial for the course and the exam. When launching an instance, you can choose from various sizes and capabilities, influencing the resources and additional features like GPUs, advanced storage, or networking. Some configurations can be changed post-launch.

As a consumer of EC2, you manage the operating system and upward in the infrastructure stack, while AWS handles virtualization, physical hardware, networking, storage, and facilities. EC2 offers on-demand billing, either by the second or hour, depending on the instance's software, charging only for consumed resources. Instance charges include CPU and memory consumption, storage usage, and extras for any commercial software.

EC2 instances can use different storage types, including local host storage and Elastic Block Store (EBS), a network storage service. An EC2 instance's state, such as running, stopped, or terminated, indicates its condition. Running instances consume resources and generate charges, while stopped instances do not incur CPU, memory, or networking costs but still generate storage charges. Terminated instances stop all resource usage and delete allocated storage, making this action irreversible.

An Amazon Machine Image (AMI) is an image of an EC2 instance used to create or derived from an instance. AMIs contain attached permissions, determining which accounts can use them. They can be public, owner-accessible, or explicitly shared with specific AWS accounts. AMIs include the boot volume and a block device mapping, linking volumes to the device IDs expected by the operating system.

In the next lesson, we will connect to EC2 instances. Windows instances use the Remote Desktop Protocol (RDP) on port 3389, while Linux instances use the SSH protocol on port 22, authenticated with an SSH key pair. When creating an EC2 instance, you select a key pair, download the private key, and AWS retains the public key. For Windows instances, the private key retrieves the local administrator password for RDP access. For Linux instances, the private and public key match enables SSH connection.

Next, we will demonstrate creating an EC2 instance in the default VPC, assigning a public IPv4 address, and connecting using the discussed methods. Finish this video and join me in the demo lesson when ready.

Author response:

“Overall, the paper has several strengths, including leveraging large-scale, multi-modal datasets, using computational reasonable tools, and having an in-depth discussion of the significant results.”

We thank the reviewer for the very supportive comments.

Based on the comments and questions, we have grouped the concerns and corresponding responses into three categories.

(1) The scope and data selection

“The results are somewhat inconclusive or not validated.

The overall results are carefully designed, but most of the results are descriptive. While the authors are able to find additional evidence either from the literature or explain the results with their existing knowledge, none of the results have been biologically validated. Especially, the last three result sections (signaling pathways, eQTLs, and TF binding) further extended their findings, but the authors did not put the major results into any of the figures in the main text.”

The goal of this manuscript is to provide a list of putative childhood obesity target genes to yield new insights and help drive further experimentation. Moreover, the outputs from signaling pathways, eQTLs, and TF binding, although noteworthy and supportive of our method, were not particularly novel. In our manuscript we placed our focus on the novel findings from the analyses. We did, however, report the part of the eQTLs analysis concerning ADCY3, which brought new insight to the pathology of obesity, in Figure 4C.

“The manuscript would benefit from an explanation regarding the rationale behind the selection of the 57 human cell types analyzed. it is essential to clarify whether these cell types have unique functions or relevance to childhood development and obesity.”

We elected to comprehensively investigate the GWAS-informed cellular underpinnings of childhood development and obesity. By including a diverse range of cell types from different tissues and organs, we sought to capture the multifaceted nature of cellular contributions to obesity-related mechanisms, and open new avenues for targeted therapeutic interventions.

There are clearly cell types that are already established as being key to the pathogenesis of obesity when dysregulated: adipocytes for energy storage, immune cell types regulating inflammation and metabolic homeostasis, hepatocytes regulating lipid metabolism, pancreatic cell types intricately involved in glucose and lipid metabolism, skeletal muscle for glucose uptake and metabolism, and brain cell types in the regulation of appetite, energy expenditure, and metabolic homeostasis.

While it is practical to focus on cell types already proven to be associated with or relevant to obesity, this approach has its limitations. It confines our understanding to established knowledge and rules out the potential for discovering novel insights from new cellular mechanisms or pathways that could play significant roles in the pathogenesis if obesity. Therefore, it is was essential to reflect known biology against the unexplored cell types to expand our overall understanding and potentially identify innovative targets for treatment or prevention.

“I wonder whether the used epigenome datasets are all from children. Although the authors use literature to support that body weight and obesity remain stable from infancy to adulthood, it remains uncertain whether epigenomic data from other life stages might overlook significant genetic variants that uniquely contribute to childhood obesity.”

The datasets utilized in our study were derived from a combination of sources, both pediatric and adult. We recognize that epigenetic profiles can vary across different life stages but our principal effort was to characterize susceptibility BEFORE disease onset.

“Given that the GTEx tissue samples are derived from adult donors, there appears to be a mismatch with the study's focus on childhood obesity. If possible, identifying alternative validation strategies or datasets more closely related to the pediatric population could strengthen the study's findings.” 

We thank the reviewer for raising this important point. We acknowledge that the GTEx tissue samples are derived from adult donors, which might not perfectly align with the study's focus on childhood obesity. The ideal strategy would be a longitudinal design that follows individuals from childhood into adulthood to bridge the gap between pediatric and adult data, offering systematic insights into how early-life epigenetic markers influencing obesity later in life. In future work, we aim to carry out such efforts, which will represent substantial time and financial commitment.

Along the same lines, the Developmental Genotype-Tissue Expression (dGTEx) Project is a new effort to study development-specific genetic effects on gene expression at 4 developmental windows spanning from infant to post-puberty (0-18 years). Donor recruitment began in August 2023 and remains ongoing. Tissue characterization and data production are underway. We hope that with the establishment of this resource, our future research in the field of pediatric health will be further enhanced.

“Figure 1B: in subplots c and d, the results are either from Hi-C or capture-C. Although the authors use different colors to denote them, I cannot help wondering how much difference between Hi-C and capture-C brings in. Did the authors explore the difference between the Hi-C and capture-C?”.

Thank you for your comment. It is not within the scope of our paper to explore the differences between the Hi-C and Capture-C methods. In the context of our study, both methods serve the same purpose of detecting chromatin loops that bring putative enhancers to sometimes genomically distant gene promoters. Consequently, our focus was on utilizing these methods to identify relevant chromatin interactions rather than comparing their technical differences.

(2) Details on defining different categories of the regions of interest

“Some technical details are missing.

While the authors described all of their analysis steps, a lot of the time, they did not mention the motivation. Sometimes, the details were also omitted.”

We will add a section to the revision to address the rationale behind different OCRs categories.

“Line 129: should "-1,500/+500bp" be "-500/+500bp"? 

A gene promoter was defined as a region 1,500 bases upstream to 500 bases downstream of the TSS. Most transcription factor binding sites are distributes upstream (5’) from TSS, and the assembly of transcription machinery occurs up to 1000 bases 5’ from TSS. Given our interest in SNPs that can potentially disrupt transcription factor binding, this defined promoter length allowed us to capture such SNPs in our analyses.

“How did the authors define a contact region?”

Chromatin contact regions identified by Hi-C or Capture-C assays are always reported as pairs of chromatin regions. The Supplementary eMethods provide details on the method of processing and interaction calling from the Hi-C and Capture-C data.

“The manuscript would benefit from a detailed explanation of the methods used to define cREs, particularly the process of intersecting OCRs with chromatin conformation data. The current description does not fully clarify how the cREs are defined.”

“In the result section titled "Consistency and diversity of childhood obesity proxy variants mapped to cREs", the authors introduced the different types of cREs in the context of open chromatin regions and chromatin contact regions, and TSS. Figure 2A is helpful in some way, but more explanation is definitely needed. For example, it seems that the authors introduced three chromatin contacts on purpose, but I did not quite get the overall motivation.”

We apologize for the confusion. Our definition of cREs is consistent throughout the study. Figure 2A will be the first Figure 1A in the revision in order to aid the reader.

The 3 representative chromatin loops illustrate different ways the chromatin contact regions (pairs of blue regions under blue arcs) can overlap with OCRs (yellow regions under yellow triangles – ATAC peaks) and gene promoters.

[1] The first chromatin loop has one contact region that overlaps with OCRs at one end and with the gene promoter at the other. This satisfies the formation of cREs; thus, the area under the yellow ATAC-peak triangle is green.

[2] The second loop only overlapped with OCR at one end, and there was no gene promoter nearby, so it is unqualified as cREs formation.

[3] The third chromatin loop has OCR and promoter overlapping at one end. We defined this as a special cRE formation; thus, the area under the yellow ATAC-peak triangle is green.

To avoid further confusion for the reader, we will eliminate this variation in the new illustration for the revised manuscript.

“Figure 2A: The authors used triangles filled differently to denote different types of cREs but I wonder what the height of the triangles implies. Please specify.”

The triangles are illustrations for ATAC-seq peaks, and the yellow chromatin regions under them are OCRs. The different heights of ATAC-seq peaks are usually quantified as intensity values for OCRs. However, in our study, when an ATAC-seq peak passed the significance threshold from the data pipeline, we only considered their locations, regardless of their intensities. To avoid further confusion for the reader, we will eliminate this variation in the new illustration for the revised manuscript.

“Figure 1B-c. the title should be "OCRs at putative cREs". Similarly in Figure 1B-d.”

cREs are a subset of OCRs.

- In the section "Cell type specific partitioned heritability", the authors used "4 defined sets of input genomic regions". Are you corresponding to the four types of regions in Figure 2A? 

Figure 2A will be the first Figure 1A in the revision and will be modified to showcase how we define OCRs and cREs.

“It seems that the authors described the 771 proxies in "Genetic loci included in variant-to-genes mapping" (ln 154), and then somehow narrowed down from 771 to 94 (according to ln 199) because they are cREs. It would be great if the authors could describe the selection procedure together, rather than isolated, which made it quite difficult to understand.”

In the Methods section entitled “Genetic loci included in variant-to-genes mapping," we described the process of LD expansion to include 771 proxies from 19 sentinel obesity-significantly associated signals. Not all of these proxies are located within our defined cREs. Figure 2B, now Figure 2A in the revision, illustrates different proportions of these proxies located within different types of regions, reducing the proxy list to 94 located within our defined cREs.

“Figure 2. What's the difference between the 771 and 758 proxies? “

13 out of 771 proxies did not fall within any defined regions. The remaining 758 were located within contact regions of at least one cell type regardless of chromatin state.

(3) Typos

“In the paragraph "Childhood obesity GWAS summary statistics", the authors may want to describe the case/control numbers in two stages differently. "in stage 1" and "921 cases" together made me think "1,921" is one number.”

This will be amended in the revision.

“Hi-C technology should be spelled as Hi-C. There are many places, it is miss-spelled as "hi-C". In Figure 1, the author used "hiC" in the legend. Similarly, Capture-C sometime was spelled as "capture-C" in the manuscript.”

“At the end of the fifth row in the second paragraph of the Introduction section: "exisit" should be "exist".

“In Figure 2A: "Within open chromatin contract region" should be "Within open chromatin contact region". 

These typos and terminology inconsistencies will be amended in the revision.

DOI: 10.3791/62757

Resource: (MMRRC Cat# 012039-MU,RRID:MMRRC_012039-MU)

Curator: @Naa003

SciCrunch record: RRID:MMRRC_012039-MU

What is this?

Author response:

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

Public Reviews:

Reviewer #1 (Public Review):

The development of effective computational methods for protein-ligand binding remains an outstanding challenge to the field of drug design. This impressive computational study combines a variety of structure prediction (AlphaFold2) and sampling (RAVE) tools to generate holo-like protein structures of three kinases (DDR1, Abl1, and Src kinases) for binding to type I and type II inhibitors. Of central importance to the work is the conformational state of the Asp-Phy-Gly "DFG motif" where the Asp points inward (DFG-in) in the active state and outward (DFG-out) in the inactive state. The kinases bind to type I or type II inhibitors when in the DFG-in or DFG-out states, respectively.

It is noted that while AlphaFold2 can be effective in generating ligand-free apo protein structures, it is ineffective at generating holo-structures appropriate for ligand binding. Starting from the native apo structure, structural fluctuations are necessary to access holo-like structures appropriate for ligand binding. A variety of methods, including reduced multiple sequence alignment (rMSA), AF2-cluster, and AlphaFlow may be used to create decoy structures. However, those methods can be limited in the diversity of structures generated and lack a physics-based analysis of Boltzmann weight critical to their relative evaluation.

To address this need, the authors combine AlphaFold2 with the Reweighted Autoencoded Variational Bayes for Enhanced Sampling (RAVE) method, to explore metastable states and create a Boltzmann ranking. With that variety of structures in hand, grid-based docking methods Glide and Induced-Fit Docking (IFD) were used to generate protein-ligand (kinase-inhibitor) complexes.

The authors demonstrate that using AlphaFold2 alone, there is a failure to generate DFG-out structures needed for binding to type II inhibitors. By applying the AlphaFold2 with rMSA followed by RAVE (using short MD trajectories, SPIB-based collective variable analysis, and enhanced sampling using umbrella sampling), metastable DFG-out structures with Boltzmann weighting are generated enabling protein-ligand binding. Moreover, the authors found that the successful sampling of DFG-out states for one kinase (DDR1) could be used to model similar states for other proteins (Abl1 and Src kinase). The AF2RAVE approach is shown to result in a set of holo-like protein structures with a 50% rate of docking type II inhibitors.

Overall, this is excellent work and a valuable contribution to the field that demonstrates the strengths and weaknesses of state-of-the-art computational methods for protein-ligand binding. The authors also suggest promising directions for future study, noting that potential enhancements in the workflow may result from the use of binding site prediction models and free energy perturbation calculations.

Reviewer #2 (Public Review):

Summary:

This manuscript explores the utility of AlphaFold2 (AF2) and the author's own AF2-RAVE method for drug discovery. As has been observed elsewhere, the predictive power of docking against AF2 structures is quite limited, particularly for proteins like kinases that have non-trivial conformational dynamics. However, using enhanced sampling methods like RAVE to explore beyond AF2 starting structures leads to a significant improvement.

Strengths:

This is a nice demonstration of the utility of the authors' previously published RAVE method.

Weaknesses:

My only concern is the authors' discussion of induced fit. I'm quite confident the structures discussed are present in the absence of ligand binding, consistent with conformational selection. It seems the author's own data also argues for an important role in conformational selection. It would be nice to acknowledge this instead of going along with the common practice in drug discovery of attributing any conformational changes to induced fit without thoughtful consideration of conformational selection.

The reviewer is correct. We aim to highlight the significant role of conformational selection. To clarify this, we have expanded the discussion on conformational selection in the introduction.

Reviewer #3 (Public Review):

In this manuscript, the authors aim to enhance AlphaFold2 for protein conformation-selective drug discovery through the integration of AlphaFold2 and physics-based methods, focusing on improving the accuracy of predicting protein structures ensemble and small molecule binding of metastable protein conformations to facilitate targeted drug design.

The major strength of the paper lies in the methodology, which includes the innovative integration of AlphaFold2 with all-atom enhanced sampling molecular dynamics and induced fit docking to produce protein ensembles with structural diversity. Moreover, the generated structures can be used as reliable crystal-like decoys to enrich metastable conformations of holo-like structures. The authors demonstrate the effectiveness of the proposed approach in producing metastable structures of three different protein kinases and perform docking with their type I and II inhibitors. The paper provides strong evidence supporting the potential impact of this technology in drug discovery. However, limitations may exist in the generalizability of the approach across other structures, especially complex structures such as protein-protein or DNA-protein complexes.

Proteins undergo thermodynamic fluctuations and can occasionally reach metastable configurations. It can be assumed that other biomolecules, such as proteins and DNA, stabilize these metastable states when forming protein-protein or protein-DNA complexes. Since our method has the potential to identify these metastable states, it shows promise for designing drugs targeting proteins in allosteric configurations induced by other biomolecules.

The authors largely achieved their aims by demonstrating that the AF2RAVE-Glide workflow can generate holo-like structure candidates with a 50% successful docking rate for known type II inhibitors. This work is likely to have a significant impact on the field by offering a more precise and efficient method for predicting protein structure ensemble, which is essential for designing targeted drugs. The utility of the integrated AF2RAVE-Glide approach may streamline the drug discovery process, potentially leading to the development of more effective and specific medications for various diseases.

Recommendations for the authors:

Reviewer #1 (Recommendations For The Authors):

Suggestions

(1) The computational protocol is found to be insufficient to generate precise values of the relative free energies between structures generated. The authors note in the Conclusion that an enhancement in the workflow might result from the addition of free energy calculations. Can the authors comment on the prospects for generating more accurate estimates of the free energy that might be used to qualitatively evaluate poses and the free energy landscape surrounding putative metastable states? What are the principal challenges and what might help overcome them? What would the most effective computational protocol be?

More accurate estimates of the free energy can theoretically be achieved by increasing the number of umbrella sampling windows and extending the simulation length until the PMF converges. However, there is always a trade-off between PMF accuracy and computational costs, so we have chosen to stick with the current setup. Metadynamics is another method to obtain a more accurate free energy profile, which we have used in previous versions of AlphaFold2-RAVE, but for the specific systems we investigated, it had issues in achieving back and forth movement given the high entropic nature of the activation loop. Research in enhanced sampling methods and dimensionality reduction techniques for reaction coordinates is continually evolving and will play a critical role in alleviating this problem.

(2) I was surprised that there was not more correlation of a funnel-like shape in Figures S16 and S18, showing a stronger correlation between low RMSD and better docking score. This is true for both the ponatinib and imatinib applications in DDR1 and Abl1. That also seems true for the trimmed results for Src kinase in Figure S19. I was also surprised that there are structures with very large RMSD but docking scores comparable to the best structures of the lowest RMSD. Might something be done to make the docking score a more effective discriminator?

The docking algorithm and docking score are used to filter out highly improbable docking poses. False positives in predicted docking poses are a common issue across all docking methods as described for instance in:

Fan, Jiyu, Ailing Fu, and Le Zhang. "Progress in molecular docking." Quantitative Biology 7 (2019): 83-89.

Ferreira, R.S., Simeonov, A., Jadhav, A., Eidam, O., Mott, B.T., Keiser, M.J., McKerrow, J.H., Maloney, D.J., Irwin, J.J. and Shoichet, B.K., 2010. "Complementarity between a docking and a high-throughput screen in discovering new cruzain inhibitors." Journal of medicinal chemistry, 53(13), pp.4891-4905.

Moreover, there is always a trade-off between docking accuracy and computational cost. While employing more accurate docking methods may decrease false positives, it can also be resource-intensive. In such scenarios, our approach to enriching holo-structures can be impactful by reducing the number of pocket structures in the input ensembles and significantly enhancing docking efficiency.

(3) I think that it is fine to identify one structure as "IFD winner" but also feel that its significance is overstressed, especially given that it can be identified only in a retrospective analysis rather than through de novo prediction.

We agree with the reviewer. We did not intend to emphasize the specific structure "IFD winner". Rather, we aimed to demonstrate that our method can enrich promising candidates for holo-structures. We verified this by showing that our holo-structure candidates performed well in retrospective docking using IFD, which we previously referred to as "IFD winner". We have now revised this term to "holo-model".

Minor Points

p. 3 "DymanicBind" should be "DynamicBind"

p. 3 Change "We chosen" to "We have chosen" or "we chose."

p. 3 In identifying the Schrödinger software Glide and IFD, I recommend removing the subjective modifier "industry-leading."

Modifications done.

Reviewer #2 (Recommendations For The Authors):

In the view of this reviewer, the writing is 'choppy'.

We have tried to improve the writing.

Reviewer #3 (Recommendations For The Authors):

(1) In Figure 1, the workflow labels (i) to (iv) are not shown on the figures, making it difficult for readers to follow. Consider adding these labels to the figures.

Modifications done.

(2) Explain how Boltzmann ranks were calculated based on unbiased MD simulations to guide the enrichment of holo-like structures in metastable states.

The Methods section is now updated for clarification.

(3) The authors could clarify how the classical DFG-out decoys in the DDR1 rMSA AF2 ensemble are transferred to Abl1 kinase in the Methods section.

The Methods section is now updated for clarification.

(4) The authors can clarify the methodology section by providing more detailed explanations about how the unbiased MD simulations are performed, including which MD simulation software was used and whether energy minimization and equilibrium steps were needed as in conventional MD simulations, and other setup details.

The Methods section is now updated for clarification.

(5) The validation of the proposed approach in this work used three kinase proteins. The authors can enhance the discussion section by addressing other types of protein structure prediction that can use the proposed approach in drug discovery, beyond the three kinase proteins tested.

The proposed approach is theoretically applicable to other types of proteins, such as GPCRs, where both conformational selection and the induced-fit effect are crucial. We have expanded the discussion on the generalization of our protocol in the Conclusion section.

(6) The authors should add appropriate citations for the software and tools used in the manuscript. For example, a reference should be added for the Glide XP docking experiments that utilized the Maestro software. Double-check all related software citations.

We have now updated the citations for docking experiments based on the instruction of the Maestro Glide User manual and IFD User manual.

(7) The authors should consider offering a comprehensive list of software tools and databases utilized in the study to assist in replicating the experiments and further validating the results.

We have now added a summary of tools used in the Methods section.

It might be helpful to have a list of the everyday programs the user will need to install, along with links to their corresponding pages in the help portal.