The classification results(Figure 9) show that transfer learning can successfully classify spectra across instruments
Since the classification task fails without instrument correction, I think that what this shows is that your batch correction approach successfully eliminates instrument-level effects (which, given how aggressive and elaborate it is, is not surprising)
First, we calculated the intensityratio between the spectra collected from the two instruments. Then, we applied a threshold,considering only peaks where the ratio exceeded a specific value for further correction. Thisapproach helped us to isolate and highlight the most relevant spectral peaks. Based on thesecriteria, we filtered the significant peaks for intensity correction. We then used a 3rd-orderpolynomial to fit the selected peaks (to avoid overfitting), generating a polynomial function forintensity correction
I may be missing something, but this feels like an elaborate and potentially over-engineered approach to what is fundamentally a batch-correction problem. I would suggest also trying simpler approaches to batch correction like linear mixed models or any of the many third-party algorithms developed for batch correction in transcriptomics and cell painting communities.
The cellquant pipeline is distributed as a single Python script alongside an environment specification (environment.yml)
I'm not sure there is any benefit to organizing this software as a single script rather than as a modern python package with a modular structure and a single CLI entrypoint. Single large scripts (this one is 2300 lines) are hard to work with for both humans and agents (they are hard to navigate for humans, and it is impossible for two agents to concurrently work on unrelated parts of the code, as they would be trying to edit the same file).
A complete command line interface (CLI) reference (CLI_REFERENCE.md) documents every argument with its default value, valid range, and interaction with cell-type presets. This reference is structured to be readable by both humans and chatbots, so that a user can paste it into a conversation and ask the AI to find the relevant parameter.
This kind of documentation is a double-edged sword, as it creates two sources of truth and is hard to keep in sync with the code (this is true for both humans and agents, who each tend to forget to update the docs). I would suggest eliminating this document in favor of writing and maintaining thorough docstrings in the code itself, which humans can introspect in their IDE and which agents will "know" to read without any prompting.
Distributing the pipeline as a single Python script within a minimal repository eliminates the installation complexity that often derails non-programmers before they reach the analysis itself.
In my experience, the installation complexity referenced here is usually due to creating and managing virtual environments, and distributing this software as a single script in a github repo will not eliminate that (indeed it arguably makes it worse). I would suggest that structuring this script as a python package and distributing it via the usual channels (PyPI and conda-forge) would likely facilitate, rather than hinder, its adoption.
Together, these results indicate that aging in lung is generally associated with elevated Raman intensity at the tissue level but reduced intensity in specific immune population, whereas in skin, cell type specific spectral profiles highlight distinct metabolic adaptations and functional states that differentiate aging from senescence.
It would be helpful to explain the possible biological significance or interpretation of this difference, as it feels a little inscrutable at first glance.
using 70% of the cells for training and 30% for testing to compare the performance in classifying senescence cells
Does this include cells from all tissue and cell types? This seems like a major result, but the differences in performance are modest, so it would be informative to cross-validate on one of these dimensions (e.g. train on lung and predict for skin, and vice versa)
The backscattered Raman light from the sample passes through two dichroic mirrors (DM1: Semrock LPD01 785RU 25, DM2: Semrock LPD01 785RU 25×36×1.1) and was collected by a multi-mode fiber (Thorlabs M14L 01). The collected signal was delivered to the imaging spectrograph (Holospec f/1.8i, Kaiser Optical Systems) and detected by a thermoelectric cooled, back illuminated and deep depleted CCD (PIXIS: 100BR_eXcelon, Princeton Instruments).
What is the spectral resolution and sampling rate of this system? The datasheet for this spectrograph lists a resolution of 3-6 cm^-1, and if the spectra have 873 dimensions and cover 600-1800 cm^-1, then the spectral sampling rate is around 1.4 cm^-1. Assuming these numbers are roughly correct, this makes it hard to interpret figures that highlight differences between Raman intensities at wavenumbers closer than either of these values (e.g. 1643, 1644, and 1645 cm^-1 in Extended data Fig 11).
Consistent with our multimodal analysis, senescent cells in wounded skin displayed increased Raman intensities at 1130/1134/1135 cm⁻¹, corresponding to 12-Methyltetradecanoic acid (15Aiso), 13-Methylmyristic acid (15iso), 15-Methylpalmitic acid (17iso),
How were these molecular identities obtained? I would be surprised if Raman intensities alone could support molecular assignments this specific.
We also observed Raman-only model in skin exhibited superior predictive performance relative to Raman-only model in lung across all evaluated metrics (e.g., 0.6000 vs. 0.4868 for Accuracy, 0.672 vs. 0.5367 for AUC). This enhanced predictive power for skin
Calling this "enhanced predictive power for skin" feels a bit misleading, as the AUC of 0.53 in lung corresponds to no predictive power.
To evaluate the performance of the combined multimodal features in identifying senescent cells, we built a random forest classifier
I might be missing something, but it seems like this is a bit circular, as the features (DEGs and DPs) are selected using the same marker (p21+/-) that the model is predicting, so the feature space is "pre-aligned" to the prediction task.
A total of 360 decreased and 422 increased DPs
Small comment: this is the first time the abbreviation "DP" is used, but what it stands for is not explained. Is it equivalent to the abbreviation "DPP" used in the methods section?
This allowed us to compare Raman intensities between different cells at single-cell resolution.At the tissue level, old lung samples exhibited higher Raman intensity than young samples (46.7208±28.5472 vs. 43.2734±11.3974, p= 2.10E-13), whereas Raman intensity was lower in old T cells (42.4769±20.1176 vs. 43.0874±11.1290, p= 4.60E-07)
It seems like it would be important to normalize the mean Raman intensities by cell area or volume, otherwise changes in Raman intensity could be due to changes in cell area rather than a true change in the overall abundance of raman-active molecules or bonds.
Rather than being “fooled”by the occasional errors of Annotator1, Cellpose-SAM reverts to its inductive biases to selectively learnthe generalizable structure in the data.
It'd be cool to show some examples of this!
hus, Cellpose-SAM can run out-of-the-box on images that have been acquired with varyinglevels of image degradation, at different pixel sizes orin arbitrary channel order, substantially simplifying thelogistics typically associated with setting up an image
What is your intuition for the origin of this capability? e.g. to what extent is it likely due to the size of the SAM encoder, its architecture, the size of the pretraining dataset, etc.
We reducedthis to 256x256 and 8x8, which required us to adaptthe position embeddings and patch embedding filtersvia appropriate subsampling.
Does this later the number of parameters in the pre-trained SAM model? And would it be expected to degrade performance? (of course the tradeoff could still be worthwhile, but it would be nice to know if this downsampling comes with a cost)
The most important aspect of biological software is thatit works well in the hands of biologists.
This is a bit of a nit, but I think it would be worth clarifying what is meant by "works well" as it could be interpreted to mean "easy to use", which feels much less important than that biological software produce correct and reproducible results.
Since we are especially interested in generalizationperformance, we wanted to choose a test datasetin which images are relatively different from thosein the training set.
This could be prohibitively expensive, but a more robust test of generalization would be some kind of leave-one-dataset-out cross-validation (e.g. calculate a dataset-dataset similarity and hold out the dataset least similar to all other datasets)
We also reverted the local attention layers of thecustom ViT transformer from SAM back to the defaultglobal attention of ViT-L, with almost no runtime penalty
what was the motivation for making this change?
The first important observation is that state-of-the-art approaches,except CPM, fail to produce an embedding for the complete dataset (containing 100,000 cells),due to their reliance on pairwise distances for the computation of embeddings, which scalesquadratically in the number of cells
This doesn't feel quite fair, as UMAP and tSNE were designed to handle datasets of this size and have been widely used to generate embeddings for single-cell datasets of this size and larger. Also, I believe at least UMAP is sub-quadratic in the number of samples, as it uses an approximate kNN algorithm that is n log n.
Figure 3: Space and time complexity analysis.
Minor comment: using a log-log scale for these plots would be helpful, as it would prevent the reference methods (UMAP, tSNE, PHATE) from appearing as a flat line.
On synthetic trees with up to 5 generations and 34,000individuals, CPM cuts distortion by > 99%
It would be helpful to clarify what this claim is based on, as I can't see anything in Figure 2 that indicates a 99% change in any of the metrics between CPM and PM.
Overall, these re-sults show that the RamanMAE embeddings capture biolog-ically relevant information and have utility for downstreamclassification tasks in challenging biological datasets
I agree that this analysis sounds promising, but I think it would be more convincing with a comparison to more traditional methods (e.g. random forest after PCA or NMF)
e shaped400-dimensional Raman spectra in the fingerprint wavenum-ber region
Since several different datasets were used here, what preprocessing was necessary to obtain spectra with this consistent range and resolution? Was it necessary to upsample or downsample the spectra from some datasets?
As expected, the UMAP projection showsthat the P231, CTC, and LM cells are progressively clustered
I would be very cautious about drawing this (or any) conclusions on the basis of a UMAP, as UMAP embeddings can be strongly dependent on hyperparameter choices.
Since the circulating tumor cells will have a population ofcells that will become lung metastatic cells in the future, asignificant degree of overlap between CTC and LM cells isreasonably expected
I'm not sure I follow this: just because some CTCs will eventually become metastatic doesn't necessarily mean that they will look similar or "overlap" in the present.
We split the entire high SNR dataset into training,validation, and test dataset
I think more details abour how the data were split into train and test sets would be important to include here. From reading the supmat of [20], it sounds like the 172k spectra were obtained from 11 hyperspectral images, implying that many of the individual spectra are likely to be highly correlated, so I think care would be needed when splitting into train/test sets to avoid data leakage (e.g. perhaps by splitting at the level of the original hyperspectral images)
We leveraged a large publicly available Raman spec-tral datase
This is missing a citation - presumably this dataset was from [20]?
we useda masking ratio of 0.5 to randomly mask half the region ofthe spectra in each step
Could you comment on this was value was chosen? It seems like it could be a really important hyperparameter...
For each spectrum,we used a patch size of 16 in 1D spectra, which translatedto a patch size of 4 in 2D image, yielding 16 patches foreach spectrum
If the spectra are 400-dimensional and the patch size is 16, there would be 20 patches per spectrum, not 16.
Figure axes are Cartesian coordinates of 2-dimensionalvectors representing each nucleotide and then summed up to get the 2D line as by Yau et al. [21].
I think it would be helpful to briefly describe how these 2D lines are constructed, as they may be an unfamiliar way of visualizing sequence data for many biologists.
As model size increases, outputaccuracy seems to improve, with the biggest model showing a generated sequence without repeats. This behavior arisesfrom the abundance of training samples for Enterovirus C (4141) compared with Gallivirus A (15).
Could it instead be that larger models are simply less prone to getting stuck in repeats?
Incontrast, in panel B, the two largest models learn the Enterovirus C function very well, suggestingmemorisation rather than generalisation,
From the figure, it doesn't look like the models are simply recapitulating the true sequence. I think a more direct comparison like a sequence alignment between the true and generated sequence would be more informative here.
Said differently,although mutations at any replication point might be Brownian and viral-cell interactions chaotic,viral genomes adapted to a host (i.e. a particular niche) approximate a temporary Lyapunovstable point to minimize system-wide energy expenditure
This feels pretty vague and hand-wave-y. The term "Lyapunov stable point" is mentioned without establishing or discussing the mathematical conditions or assumptions that would justify using it. Also, phrases like "system-wide energy expenditure" are not defined; what is meant by "energy", and why would viruses adapt in order to minimize it?
Two, its constraints arefunctional not sequential, backward not forward looking, developmental not constructive
This is so succinct that it may be hard for readers to understand what is being referred to here. If these constraints are important, I think it would be worth elaborating on them a little bit.
From this foundation, we can define Ω to be a summed characteristicacross 𝑁 entities
I appreciate that this presentation is intentionally abstract, but I think it would be helpful to explain more explicitly what omega represents here. In other words, what is the nature of the "characteristics" z_i? I think readers may assume that these characteristics refer to phenotypes, but later on, it becomes clear that these "characteristics" actually refer to something like "genomic state" (if not simply a genotype)
We suggestinadequate data science education as a key factor behind these issues
In the absence of any evidence favoring one factor or another, I think it would be fair to include other possible factors here (especially as, imo, a lack of data science education is not the most plausible one)
minimize data misinterpretation
This is a bit of a nitpick, but I'm not sure it's fair to equate "visualization errors" with "data misinterpretation", either on the part of authors or readers, because visualization errors can be made for other reasons by authors and won't necessarily lead to data misinterpretation by readers (even if they do make it more likely)
The prevalence of misused bar graphs is potentially caused by a systematic lack of data science training inscience, technology, engineering, and mathematics (STEM) education, research training, and the academic
fwiw, another possible explanation that strikes me as more plausible is that there are very strong incentives for authors to publish in prestigious journals, which in turn incentives them to present their results in the most compelling possible light.
consistent with our hypothesis thathaving more authors increases the probability of having coauthors who contribute visualization mistakes
Perhaps a simpler explanation is that papers with more authors tend to be longer and to have more figures, increasing the chances that at least one figure contains a mistake.
The ground truth in this dataset is known a priori215because there are 6 chromosomal crossovers in almost every cel
I don't think it's quite accurate to say that the ground truth is known here unless the meaning of "in almost every cell" can be quantified. I agree that the fact that SpotMAX finds six foci in 97% of nuclei is qualitatively consistent with the biology of COSA-1 foci, but there's no way to quantify this performance without first quantifying the true prevalence of nuclei with more or less than six foci (and I think we do know it can't be 100%).
It's also important to specify the meiotic stage (that is, the position of the nuclei within the gonad) at which foci are counted, since (if I recall correctly) COSA-1 foci appear gradually, so I think in mid-pachytene there will be many more nuclei with fewer than six foci than at late pachytene.
revealed an unexpected mechanism that selectively targets nuclei282with errors in crossover regulation for apoptosis independently of the DNA damage checkpoint
Is it known that the presence of RAD-51 foci alone is a reliable proxy for the activation of the DNA damage checkpoint in late meiotic prophase? (I'm not sure, but I wouldn't be surprised if it is not)
SpotMAX performs local peak detection with a footprint determined from the expected spot size
Does SpotMAX support a range of expected spot sizes, or does it require one specific size?
Crucially, we provide an intuitive and user-friendly GUI integrated into the Cell-ACDC software9
I think it'd be helpful to briefly explain what Cell-ACDC is and why it is important that SpotMAX is integrated into it, as some readers may not be familiar with it.
The representative images show two instances where the human made a counting218mistake
If I recall correctly, COSA-1 foci aren't always diffraction-limited, so I'm not sure that it's really possible to determine the true number of foci in the two examples shown, at least from the 2D images shown in the figure (though I agree that the example on the left does look a lot like two barely-separated distinct foci)
. In addition, fluorescence images showed that the324body length of nematodes after exposure to L511A was significantly shorter than325that of the control group,
Does "significantly shorter" mean that the difference was statistically significant? If not statistically significant, I'm not sure it's worth mentioning. If it is, could this data be added to Figure 5?
the342total body relative fluorescence intensit
The phrase "total body" here implies the intensity is a sum, which would mean it depends on worm length. I think it'd be important to clarify this.
the overall fluorescence intensity of all GFP-labeled neurons
Is this the average intensity? (so that the intensity is independent of worm length)
Combining the above results, we can draw the following conclusions:
If the differences were not statistically significant, I don't think these conclusions can be drawn.
This suggests that these two flavoring substances and298tobacco component may produce neuroexcitatory effects at low doses in nematodes
Is an increase in bend frequency known to indicate neuroexcitation? It seems like it could also result from metabolic changes, perturbation of feeding behavior, or direct effects on muscle cells.
For X6145A and Fla-1, the bend296frequency of worms also showed a trend of slightly increasing in low-dose group and297decreasing in high-dose group
If the trend wasn't statistically significant, I'm not sure this is worth commenting on.
its bend frequency per minute was basically consistent with that of the291control grou
Is it possible to make a more more precise statement than "basically consistent"? This is a little confusing, especially since there was a statistically significant difference between the control and the higher dose.
ig. 3A showed that treatment of MCP266and BDE-47 induced significant general toxicity in worms on this platform
How was worm death measured?
ig. 3C&F showed270that the bend frequency of worms decreased significantly (p < 0.001) with the271increase in dose after 24 h treatment
For the 1mM MCP condition, it looks like most worms were still alive after 24 hours, but the bend frequency of all worms was zero. Does this make sense?
Fig. 3. Toxicity evaluation of MCP and BDE-47 on the platform.
It looks like this figure is missing the number of worms analyzed for each condition. Also, the survival curves in Panel A are in discrete steps of the same size, suggesting a smaller number of worms than there are dots in Figures B and C.
Asphenotypic changes during keratinocyte differentiation span across both space and time, this applicationperfectly showcases the power of ESPRESSO spatiotemporal omics in identifying not only the presenceof distinct phenotypes, but also providing insights about their spatiotemporal evolution.
Again, I think a baseline here would make this claim more convincing. In other words, what aspects of the differentiation dynamics described here could only be captured by ESPRESSO?
As shown in Figures 1c and 1d, GMM clustering easily identified the cell type-specific phenotypes andallowed the quantification of properties of interest in their organelle networ
It would be helpful to compare this result to some baseline obtained from an established method like cell painting. In other words, can existing techniques also readily distinguish these cell types?
to increase the acquisition speed 16-fold
it would be helpful to also provide some absolute measures of throughput here, such as how many FOVs of a given size and resolution can be imaged per unit time.
organelle properties are normalized, selected and reduced in dimensionality byPacMAP35, generating low-dimensional embeddings that encode the high-dimensional organelleproperties of each cell. A Gaussian Mixture Model (GMM36) clustering algorithm is then applied
It sounds like the clustering was done after the embedding step; that is, using the low-dimensional embeddings from PacMAP, rather than the original feature matrix. If so, I'm worried that this will result in inaccurate clusters, as PacMAP (like all such methods) does not perfectly preserve the relationships between the original high-dimensional feature vectors.
This suggests that genes not annotated by eggNOG-mapper are probablyproteins that either catalyze some protein, RNA, or DNA chemical modification, or bind to othermolecules, form macromolecular complexes, and are involved in the regulation of essentialprocesses for animals
This is a bit confusing; it's so vague and general that it sounds like it could describe almost any protein.
We therefore considered this evidence assupportive for not filtering
I'm not sure that two examples can constitute evidence for or against filtering. Is it possible to use a ground-truth dataset to make this kind of filtering/no-filtering decision with more confidence?
We show thatprotein language model-based annotations outperformed deep learning-based ones
This is a bit confusing, because protein language models are a kind of deep learning model. It would help to clarify what "deep-learning-based models" refers to in this context.
with a reliability index of 1
What does a reliability index value of "1" mean?
The search starts with acytochrome from corn (Zea Mays), and within the first 50 hits,we find similar structures originating from various animals(fish, eagle, mouse, cat, horse, etc.)
The phrase "within the first 50 hits" feels tantalizing. What else appeared among the top hits? Were there hits that were surprising or potentially false positives? And were there proteins that should have appeared among the top hits, but didn't?
Here, AlphaFind shows us (Figure 2)that highly similar hemoglobin structures can also be found inother species.
Again, it would be really great to quantify what "highly similar" means here.
in an average of 7 seconds withnegligible back-end load
It would be helpful to mention details about the hardware here, as the time cost is hard to interpret without that information.
Therefore, high occurrence of unstructuredregions in the input structure can bias the search. Thisphenomenon is more prevalent in coiled-coil structures but canbe also observed in some small structures
Again, it would be great to quantify this and/or to discuss some examples of proteins for which this is a real problem.
We tested AlphaFind on a diverse set of proteins varying insize, complexity, and quality. AlphaFind provided biologicallyrelevant results even for small, large and lower qualitystructures. When AlphaFind did not offer structures withhigh TM-scores, the results remained biologically relevant.
I think these claims would be more convincing if they could be quantified and if the performance of AlphaFind could be compared to other existing tools, if possible.
he latter two methodsin conjunction with (10) establish the basis of the indexingsolution presented in here
What is the relationship between this approach and approaches to indexing or similarity-based lookup used by common vector databases?
In the offline phase, we first extract semantic information fromraw cif files into vector embeddings,
It would be helpful to explain in more detail how this is done, since it seems like a crucial step.
Every376chromosome group is combined into a single sequence, with chromosome order randomly deter-377mined.
It's surprising to me that chromosomes are randomly ordered; this feels a bit like the equivalent of randomly shuffling the clauses of a sentence. It would be helpful to explain this choice or discuss reasons why it might or might not be a concern.
Start tokens are unique to each chromosome and species
This feels confusing: if start tokens are unique to species, how is UCE able to generate embeddings for datasets from species it was not trained on?
However, beyond that,238the effect levels off (Supplementary Fig. 6). This is expected due to the curse of dimensionality239in high-dimensional spaces and the variability in the level of ontological refinement in different240branches of the ontology
This feels awfully hand-wavy. I can understand that a leveling off is expected at some distance, by why at 5 hops in particular?
or all three species we observed204very high agreement between independent annotations of the novel species’ data and the nearest205cell type centroids in the IMA
It would be helpful to mention here what these three species were and how distantly related they are to the eight species on which UCE was trained.
We train185a simple logistic classifier on the UCE embeddings of the Immune Cell Atlas [38], and then apply186the classifier to B cell embeddings from Tabula Sapiens v2. This classifier accurately classifies the187Tabula Sapiens v2 cells as memory and naive B cell
This result feels hard to interpret without a comparison to other approaches or models. In other words, are embeddings from UCE uniquely able to capture the information required for this classification task?
UCE embeddings174distinctly separate cell types more effectively than other methods tested in zero-shot
This feels a bit subjective; I think this claim would be more convincing if it were grounded in a quantitative measure of clustering accuracy.
We compared several methods and found that UCE substantially out-167performs the next best method Geneformer by 9.0% on overall score, 10.6% on biological conser-168vation score, and 7.4% on batch correction scor
If possible, it would be helpful to contextualize these relative increases in performance, particular given that none of the models listed in Supp Table 1 appear to significantly outperform using the log-normalized raw data. (the "overall score" is 0.74 for UCE and 0.72 for "log-normalized expression"). Without more context, it's hard to know what this means, whether it should be surprising, whether it reflects limitations of the metrics or of the models, etc.
Also, I think it would be more transparent to mention here that there are two metrics for which UCE does not outperform other models (the ARI score and the "ASW (batch) score").
Genes belonging to the same chromosome are grouped111together by placing them in between special tokens and are then sorted by genomic location
It would be helpful to understand the context and motivation for this design decision. In other words, what aspects of UCE's performance depend (or are suspected to depend) on including information about genomic position?
This allows UCE to meaningfully99represent any gene, from any species, regardless of whether the species had appeared in the training100data
It would be good to clarify here if "training data" refers to the data used to train the protein language model or UCE itself.
When calculating doubling times based on mitotic events in the remaining cells that were470not undergoing apoptosis (Figure 6D), the doubling times are similar to those for unexposed cells
Again, it's great to see something like this quantified so carefully!
Higher intensities of excitation light exposure led to significant cell death that was apparent by manual447inspection of images, and by the reduced relative cell numbers as shown by the green lines
It seems surprising that there is such a big difference from 1x to 1.4x. Is this by design? (was the 1x intensity chosen from prior experience or experiments to be as high as possible without inhibiting cell division?)
s shown in Figure 6A,438exposure of cells to the minimal intensity of fluorescence excitation light (56 mJ/cm2 referred to as 1x)
It's super helpful that an absolute measure of intensity is provided here, but it would be great to also include the wavelength (or range of wavelengths) of the excitation light.
Individual cells in the center of the colony tend to move less than cells near the
Is it possible to correct this for the fact that, as the colony itself expands, cells near the edge necessarily must move more than cells in the center (which will not move at all, if the colony as a whole is stationary)?
Average mitotic rates do not appear to depend on431distance from the colony edge (Figure 5D) and do not correlate with the increased cell motion
It's great to see a subtle detail like this quantified so carefully! Is this consistent with prior work (if there is any)?
The233manual data was paired to the 3D U-Net inferenced results using a linear sum assignment routine with234the cost function being proportional to the distance between mitosis events in space with an empirically235determined spatial cutoff of 15 pixels and a time cutoff of 6 frames.
This is a bit hard to understand. How is distance in time measured? (i.e. the difference between the time of mitosis onset in the manual annotations and the segmentation results)
are 1 to 2 to 20 to 20.
how were these weights chosen? And is it correct to think of these weights as a kind of correction for the class imbalance between non-mitotic and mitotic nuclei?
in each of the 5202frames before division, and as class 3 (one or two daughter cells) in each of the 3 frames after division.
how were these numbers of frames chosen?
The binary masks are created by151inferencing with 3 instances of the same model and thresholded by 2 (as explained in more detail in152Supplemental Figure 1B and 1C.
This is a little confusing, especially the "thresholded by 2" part, and I didn't find the caption in Supp Fig 1 to be that much clearer. It would help to explain the origin of the variability in the predictions (in other words, what is an "instance of the same model"?)
We trained a 2D U-Net to segment single-cell nuclei from phase contrast images starting with a pre-136trained U-Net (14) as our initial network
It would be great to mention what kind of images the pre-trained model trained on. Do you have a sense for how important it is that pre-training be done on similar images? (and what kinds of similarity are most important: cell type, imaging modality, magnification, etc)
Wereport the number of gene KOs with an AU-ROC > 0.55
Why 0.55 and not 0.5?
We trained a ViT-small model with patch size = 8, number of global crops = 2, number of local crops = 8on 4 nodes x 8 NVIDIA-V100 GPUs per node (32 GPUs) for 100 epochs
would it be possible (and meaningful) to mention how many GPU hours this required? Also, some more details would be helpful for non-ML experts; e.g., why the choice of 100 epochs, was a stopping criterion used, which epoch was used for the final analysis/results, etc.
we re-parameterized the first layer ofthe model as:
This equation is a bit opaque; it would be helpful to explain what the superscripts and subscripts of theta mean.
(both ~1-1.5million cell tile images)
Does the 1-1.5m figure mean single-cell images? or FOVs? It would also be super helpful to comment on how this dataset size was chosen. Was it the minimum amount of data required for this level of performance? More generally, did you do any experiments varying the quantity or diversity of the training data?
The superior performance of CP-DINO 1640 is unlikely a result oftrivial memorization, as the 1640-genes druggable genome library and 300-genes MoA library sharesimilar numbers of overlapping genes with the 124 PoC library (30 and 26 genes respectively).
I think to make this claim more convincing, it would be important to show how many genes in the 1640 library are very similar to (rather than merely identical to) genes in the 124 PoC library ("very similar" is obviously subjective but I'm thinking of homologs/paralogs or genes that are components of the same complex or pathway)
nti-phospho-S6 (pS6) antibodywith AlexaFluor 750-conjugated secondary antibody was used in the 6th channel as an establishedbiomarker
it would be helpful to mention here what cellular structures of features the pS6 antibody labels, and also (for the non-biologists among us) what mTORC1 is
Nevertheless, CP-DINO 300 trained on bioimaging data yielded a moreinformative embedding that has higher median prediction accuracy than the other two models (Fig.S4a-b), and correctly classified more perturbations with better accuracy (Fig. 4c). CP-DINO 300 alsorecovered more known biological relationships from StringDB as measured by cosine similarity of theaggregate gene KO embeddings (Methods) than the other two models (Fig. 4d)
It's awesome to see such an explicit and direct comparison of classic feature engineering with modern unsupervised ML models!
If possible it would be great to quantify how much better the DINO-based approach is; Figures 4a-d are a bit hard to understand at first and obscure the relative differences; Fig 4d in particular doesn't give the impression that DINO is that much better than the CellStats approach (even though the 0.12 of DINO vs the 0.09 of CellStats is actually a 30% improvement!). Also, some measure of statistical significance would be helpful; in particular, how likely is it that the 0.09 vs 0.12 in Fig 4d is reproducible?
phenotypic clustering of genes by their annotated mechanism of action,
It feels like there's a typo here somewhere, since genes don't really have a "mechanism of action" and the screen here does not involve compounds but rather gene KOs. Is the idea to use the phenotype of the KOs to cluster genes by the MoA of the compounds that target them? In any case, the reference to MoAs here is doubly confusing because the clustering shown in Fig 4E appears to capture cellular localization (and also pathway membership?), but I couldn't see any discussion of the clustering relative to the MoAs of the compounds used to select the 300 genes