acceptodds
Under review as a conference paper at ICLR 2027

Beyond Gene Counts: Self-Supervised Cell Representations from Individual Molecules

Abstract

Image-based spatial transcriptomics records the position of every mRNA molecule inside each cell, yet representation learning for these data still reduces each cell to a vector of gene counts, discarding the subcellular organization of these molecules. Where transcripts sit within a cell, near the nucleus, along the membrane or in protrusions, varies between cells in ways a count vector cannot record. We formulate subcellular representation learning as self-supervised learning over point clouds of molecules, each labeled with one of thousands of genes, observed in an arbitrary orientation and with incomplete detection. We present Molm, the first self-supervised model for this setting, which represents a cell as a hierarchy of molecules, the localization programs they form, and the cell as a whole. At every layer, a transformer groups molecules into programs, and the geometry of each program is computed exactly, respecting rotations and reflections by construction. We also introduce a benchmark with known subcellular ground truth, built on real cells whose molecule positions are resampled from known localization programs, and use it, with three real datasets from different imaging platforms, to evaluate nine existing methods. Molm recovers a cell subtype defined only by molecular arrangement at 0.780 balanced accuracy against 0.426 for the best existing method, and localization patterns at 0.448 against 0.259. On real data it leads all label-free methods in cell-type prediction while additionally capturing cell shape, which count-based and foundation-model embeddings do not encode.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Spatial Transcriptomics, Subcellular Biology, Self-supervised representation learning, O(2)-equivariance

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_E3ADOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Molm treats a cell as a point cloud of transcript identities and positions. Context-dependent molecular weights define groups whose local geometric moments feed back into the cell representation. Self-supervised training is evaluated with cell-label and morphology prediction on real spatial transcriptomics data, and with subtype and localization recovery on synthetic data.

Critical strengths

The model gives molecular geometry a clear role in representation learning. The rebuttal addresses the main concerns: Molm retains its advantage over FOCUS and pyTrance at comparable GPU-hour budgets, and the additional synthetic datasets support the result across different signal strengths. The real-data validation and same-input ablations also strengthen the evidence for the learned representation.

Critical weaknesses

The new synthetic experiments change the generation seed and signal strength together, so they do not separately measure variation across generated datasets at a fixed setting. This is a remaining limitation, although the additional experiments substantially improve the original evaluation. The EGNN comparison is still preliminary, so the comparison with another geometric encoder remains incomplete.

Questions for authors

If feasible, please repeat one fixed generator setting with a different generation seed to check variation across generated datasets. Please include the complete EGNN results once the remaining runs are finished.

3 comments

Follow-up: the revision is complete (1/2)

Thank you again for your review. The equivariant graph network runs that our reply (W2b) promised are complete and are now in the paper. Below are the new results and the one statement of our reply that they refine.

In brief. EGNN, trained on Molm's input with the shared terms of our objective, shows that distances in the encoder are needed but not sufficient to reach Molm's level. On a graph network as on the transformer, a capable encoder recovers the subtype on every dataset with subtype signal, and the measurement adds the localization pattern on all four datasets.

W2b. Attribution to the geometric architecture

A generic geometric encoder given the same input does not reach Molm's subtype score. A class token and pre-norm blocks recover the subtype, and Molm's measurement recovers the pattern. We train EGNN (Satorras et al., 2021) on each molecule's gene and position, with a 32-nearest-neighbor graph and the shared terms of our objective, for 220 epochs (Molm: 215). It comes in three forms: as published; with a class token and pre-norm blocks (EGNN + cls + preLN); and with Molm's measurement added after each node update, which is Molm with an EGNN encoder, Molm (EGNN). We report three training seeds at κ₂ = 6 and one seed on each further dataset.

Subtypeκ₂ = 61848null
FOCUS / pyTrance0.484 / 0.6960.417 / 0.4930.405 / 0.4720.333 / 0.334
EGNN0.6650.4850.4570.335
EGNN + cls + preLN0.8660.6600.5890.338
Molm (EGNN)0.8860.7060.5980.336
Molm (transformer)0.8640.6560.5670.331
Patternκ₂ = 61848null
FOCUS / pyTrance0.089 / 0.1130.085 / 0.1120.085 / 0.1140.079 / 0.119
EGNN (h⁰)0.2240.2200.2090.217
EGNN + cls + preLN (h⁰)0.2350.2330.2080.228
Molm (EGNN) (measures)0.2980.2950.2880.298
Molm (transformer) (measures)0.2950.2940.2760.288
  • Distances alone are not enough. EGNN takes distances into account by construction. It exceeds FOCUS but, at Molm's number of epochs (220 against 215, about half its compute), stays below pyTrance on every dataset with subtype signal (0.665 against 0.696 at κ₂ = 6); with twice the epochs it reaches 0.720 (one seed), still well below Molm.
  • A capable encoder recovers the subtype. A class token and pre-norm blocks bring EGNN level with Molm (transformer) on subtype (0.866 against 0.864). Its pattern score stays at the level of Molm's h⁰ (0.235 against 0.236).
  • The measurement carries the pattern on either encoder. Adding it raises the pattern lift of EGNN + cls + preLN from 0.235 to 0.298, against 0.295 for Molm (transformer). It also raises subtype to 0.886. Molm (EGNN) matches Molm (transformer) on cell type and pattern and exceeds it on subtype, at about 1.8 times its compute (110 against 60 GPU-hours).

This refines one statement in our reply. We wrote that exceeding pyTrance "takes geometry in the encoder". Geometry in the encoder is needed but not sufficient. With input, objective and epochs fixed, recovering the subtype needs a capable encoder and the localization pattern the measurement; neither depends on the transformer in particular. The paper adds EGNN to Section 5 and Figure 2, with details in Appendices G and N and its compute in Appendix J.

(2/2)

W1a, Q1. Four datasets, each from its own generation seed

On every dataset with subtype signal, the three EGNN forms keep their subtype order seen at κ₂ = 6, and Molm (EGNN) stays ahead of FOCUS and pyTrance; its measures keep their pattern lift on all four datasets, and in the null every form is at chance (tables in our first comment; Appendix M).

W1b. "Do not yet validate recovery of real subcellular programs"

Two new appendices show the measures on single real DCIS cells. With depth, the measures move from encoding gene identity toward spatial patterns such as the nucleus and the cell outline, and single measures, each the best match for its target chosen on the training section, track the nucleus, the cell's center, the cell outline and DAPI within gene (r 0.51 to 0.73 on the test section; Appendix Q, Figure 6). Where the probe on the measures calls nucleus beyond the segmented nucleus, the DAPI stain is brighter (AUROC 0.67 against 0.50, chance level, for the baseline; Appendix R, Figures 8 and 9), so the measures follow the nucleus in the DAPI stain, which the model does not take as input.

W2b (continued). Same molecular inputs: the hand-crafted baseline

On the 20,000 cells of that comparison, Molm's subtype score is now the mean of three seeds, 0.807 (our reply gave one seed, 0.801), against 0.507 for the hand-crafted features (Appendix O, Table 26).

Pointers. The appendices of the final revision are lettered differently from those our reply cites. The synthetic data are in Appendix F, robustness is in Appendix M (the 57-class label is in Table 22), training compute is in Appendix J, and localization in real cells is in Appendix P. Appendix O, Table 2 and Figure 2 keep their numbers.

We would be glad to discuss any remaining questions.

Dear Reviewer E3AD,

Thank you for reassessing the paper and acknowledging the stronger evaluation. We would like to highlight that the EGNN comparison identified as incomplete in your updated review is now finished and included in the manuscript (Section 5; Appendices M and N).

Across the four synthetic datasets, the comparison clarifies the contributions of encoder design and the learned measurement. A strengthened EGNN recovers the subtype, while adding Molm’s measurement raises localization-pattern lift from 0.235 to 0.298 at the main setting, comparable to 0.295 for transformer-based Molm. This supports the measurement’s benefit across both encoder architectures. We also added individual real-cell visualizations and quantitative validation against nuclear geometry and DAPI imaging (Appendices Q and R).

Your point about generation-seed variation at a fixed signal strength remains valid: the existing sweep changes both together and does not isolate that source of variation.

With the EGNN comparison now complete and the additional real-cell evidence included, would you consider whether the revised paper warrants a higher score? We appreciate your consideration and would be happy to address any final questions.

Best regards,
The Authors

Reviewer_gLdQOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Molm, a self-supervised framework for learning cell representations directly from individual RNA molecules in image-based spatial transcriptomics. Each cell is represented as a set of molecules with gene identities and spatial coordinates. Molm uses a transformer together with learned measures and rotation/reflection-aware geometric moments to represent molecules, localization programs, and whole cells. The method is evaluated on MERFISH, CosMx, and Xenium datasets using cell-type and morphology prediction, and on a synthetic benchmark with known subcellular localization ground truth. The results show strong recovery of synthetic localization patterns and competitive performance on cell-level prediction tasks.

Critical strengths
  1. The paper addresses an important limitation of current spatial transcriptomics representations: aggregation of individual molecules into gene-count vectors discards intracellular spatial organization. Treating molecules as the basic representation unit is well motivated and potentially useful.
  2. The combination of learned localization measures with explicitly rotation- and reflection-aware geometric moments is technically interesting. The synthetic benchmark also provides a useful setting in which subcellular organization can be evaluated with known ground truth.
Critical weaknesses

1.The strongest evidence that Molm learns subcellular organization comes from the synthetic benchmark. On real datasets, the main evaluations are cell-type prediction and morphology regression, while arrangement-only and molecule-level localization are evaluated only on synthetic data. Since the main contribution of the paper is the representation of subcellular molecular organization, I would like to see direct validation that the learned programs correspond to meaningful localization patterns in real cells.

2.The current experiments do not convincingly establish generalization or provide fully matched comparisons. The Lung and DCIS train/test sections come from the same donor or tissue block, and the Colon split also does not separate animals. In addition, Molm is trained substantially longer than the main subcellular baselines, which makes it difficult to separate the contribution of the proposed representation from differences in model capacity and training.

Questions for authors

1.Can the authors provide direct real-data validation of the learned localization programs, for example by comparing learned measures with known compartment-associated genes, experimentally supported RNA localization patterns, or cell polarity features?

2.Since the Lung dataset contains multiple donors, can the authors report a donor-held-out experiment? This would substantially affect my assessment of the generalization claim.

3.Can the authors provide a stronger matched baseline, such as a generic point-cloud or equivariant model trained with the same self-supervised objective and a comparable training budget? Alternatively, can the training budgets of the existing subcellular baselines be better matched?

3 comments

Follow-up: the equivariant model you asked for, and the measures on real cells (1/2)

Thank you again for suggesting the equivariant baseline. The revision is complete: the EGNN runs promised in our reply to Q3 are in the paper (Section 5, Figure 2, Appendix N), and two new appendices show the measures on real cells. Below is only what is new since our reply.

Q3 "generic … equivariant model": the result

Trained on Molm's input with the shared terms of our objective, for Molm's number of epochs (about half its compute), EGNN as published stays below pyTrance on subtype. A class token and pre-norm blocks bring it to Molm's subtype score, and adding Molm's measurement, which gives Molm (EGNN), also brings it to Molm's localization pattern. Synthetic DCIS, κ₂ = 6, three training seeds; every nearest-neighbor EGNN design trains for 220 epochs, Molm for 215 (Table 24):

ModelParameters (M)GPU-hoursCell typeSubtypePattern, h⁰ / measures
EGNN13.3300.8650.665(9)0.224(1) / –
EGNN + cls + preLN21.2350.879(1)0.866(7)0.235(1) / –
Molm (EGNN)36.71100.897(2)0.886(6)0.233(1) / 0.298(1)
Molm (transformer)40.9600.896(1)0.864(1)0.236(1) / 0.295(1)

cls: a class token; preLN: pre-norm residual blocks; Molm (EGNN): EGNN + cls + preLN with Molm's measurement. Standard error over seeds in units of the last digit. The same order on subtype holds on the three further datasets (Figure 4; Table 21).

W2c / Q3 "geometry in the encoder": needed but not sufficient

Our reply said that exceeding pyTrance takes geometry in the encoder. EGNN shows that geometry is needed but not sufficient to reach Molm's level. EGNN uses distances by construction, yet at Molm's number of epochs (220 against 215, about half its compute) it stays below pyTrance on subtype (0.665(9) against 0.696(1)); with twice the epochs, at about Molm's compute, it reaches 0.720 (one seed, Appendix N), still well below Molm. A class token and pre-norm blocks, both of which Molm's transformer has, bring it to 0.866(7). The 0.05–0.08 on subtype that our Q3 answer credited to Molm's invariant design and measures holds among the transformers we compared; EGNN shows that another capable encoder reaches that level without the measures. Without the measures, no design exceeds 0.236 on the pattern, the level of Molm's scalar stream h⁰ (table above). So recovering the subtype needs a capable encoder, and the localization pattern needs the measurement.

(2/2)

W2c "model capacity": size does not order these models

EGNN + cls + preLN reaches Molm's subtype score with 21.2 million parameters, while the generic point-cloud transformer of our reply (Transformer, xy; 25.6 million) stays at 0.598(23) (Table 9). The smaller Molm of our W2c answer is now quoted from the reference row of Table 35: 0.762(12) on subtype at about 25 GPU-hours. This replaces the 0.72–0.82 range in our reply, which came from a length-scale table that the revision reruns at d = 512 and eight layers (Table 34). So parameter count does not explain the differences on subtype; the encoder's design does.

Q3 "the configuration they use for molecular property prediction": refined

EGNN keeps the node update and readout of Satorras et al.'s molecular-property experiments but connects each molecule to its 32 nearest neighbors (Appendix G); their fully connected graph, which we also ran, scores lower and trains about 3.5 times more slowly (Appendix N). It is lower on all three scores (subtype 0.553(17) against 0.665(9)), with messages at half the width to fit the same batch in memory.

W1 / Q1 "meaningful localization patterns in real cells": the measures on single cells

Individual measures follow cell-scale localization patterns in real cells, and they add most where position alone is a poor guide. Both appendices use the DCIS test section; all real-data results are those of Molm (transformer).

  • Depth (Appendix Q, Figure 6). Gene identity explains a median 85% of a measure's variance in the first layer and 7% in the last; deeper measures form patterns at the scale of the cell, such as the cell outline, the cell's center or the nucleus. Single measures, selected on the training section, track depth inside the nucleus, closeness to the cell's center, the cell outline and DAPI, at r 0.51–0.73 within gene on the test section. The measures predict DAPI better than the baseline in 85% of test cells (Figure 7).
  • Compartments (Appendix R, Figures 8 and 9). The measures' margin over the baseline from gene identity, position and local context grows with the distance of the nucleus from the cell's center, from 0.06 to 0.11 balanced accuracy (0.75 against 0.64 in the fifth of cells with the most off-center nucleus). Molecules the segmentation places more than 2 µm into the cytoplasm but the measures call nucleus lie on brighter DAPI (AUROC 0.67 over three seeds; 0.50, chance level, for the baseline's calls).

Pointers in our reply after the final reordering of the appendices: held-out individuals, Appendix K (was J); compute, Appendix J (was M); localization in real cells, Appendix P (was N); synthetic robustness, Appendix M with Tables 21 and 22 (was L, Tables 20 and 21). The third row of our Q3 table is now named "Transformer, invariant + proximity".

We would be glad to discuss any of these results further.

Dear Reviewer gLdQ,

Thank you again for your thoughtful review. As the discussion closes, we would like to briefly summarize the completed additions addressing your three main concerns:

  • Direct validation in real cells: The learned measures agree with independent RNA fractionation data and predict molecular localization beyond a baseline using gene identity, position and local context. On the held-out DCIS section, compartment balanced accuracy is 0.766 versus 0.697, and DAPI prediction reaches R² = 0.56 versus 0.36. We also added visualizations and quantitative analyses of individual measures in real cells (Appendices P–R).
  • Generalization across individuals: We evaluated four held-out lung donors and two held-out colon mice. Molm performs comparably to scVI on lung donors and leads the evaluated baselines at all three cell-type label resolutions in colon (Appendix K).
  • Stronger, controlled comparisons: We matched the GPU-hour budgets of FOCUS and pyTrance and completed the requested EGNN comparison. A strengthened EGNN matches Molm on subtype prediction; adding Molm’s measurement increases its localization-pattern lift from 0.235 to 0.298, supporting the measurement’s contribution across encoder architectures (Appendices J and N).

All these results are included in the revised manuscript. If these additions address the concerns underlying your assessment, would you consider raising your score to reflect the strengthened evidence? We would be happy to clarify any remaining point.

Best regards,
The Authors

Reviewer_zKdiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors point out that standard analysis of image-based spatial transcriptomics reduces each cell to a list of gene counts, which throws away where inside the cell each RNA molecule sits. They build a model, called "moments of a learned measure", that learns from the raw molecule positions without labels. It groups molecules into "localization programs", such as genes that line the cell edge, and computes the shape of each group with exact geometry that does not change when a cell is rotated or flipped. They also create a benchmark from real cells whose molecule positions are resampled from known patterns, so the true answer is known. On it, the model recovers a hidden cell subtype defined only by molecule arrangement at 0.780 balanced accuracy versus 0.426 for the best existing method, and on real data it leads the label-free methods at predicting cell type while also capturing cell shape.

Critical strengths
  1. Clear problem and design.
  2. The paper states its caveats (unequal training budgets, a baseline dominated by cell size, a small cell-type margin) and supports its design with ablations.
  3. Innovative label-free training signal. Two random subsamples of a cell are valid views of the same cell, so requiring their geometry to agree gives a natural self-supervised objective without annotations.
  4. Rigorous evaluation.
Critical weaknesses
  1. The main subcellular evidence is synthetic and built by the authors. The hidden subtype and the localization patterns come from a generator the authors designed, with seven compartments and a fixed set of programs.
  2. Real data give no proof that the learned programs are biologically meaningful. On real data the model is only tested on cell type and on cell size and shape.
  3. The cell shape win is close to built in. Size and shape targets are computed from the same molecule positions the model reads, so the model has direct access to them while count-based methods do not.
  4. The baseline comparisons are not fully controlled. Training budgets differ (215 epochs for the model versus 50 for pyTrance and 10 for FOCUS), and the length scale was tuned at a smaller setting.
Questions for authors

N/A

3 comments

Follow-up: the full-size length-scale sweep is in the paper (W4b) (1/2)

Thank you again for your review. The revision is complete, and three results bear on your points about controlled comparisons.

W4b. Length scale "tuned at a smaller setting"

The full-size sweep, now with the learned scales and the localization pattern, confirms the trade-off: small scales favor subtype and the pattern, larger scales favor shape, and at every scale Molm stays far ahead of FOCUS and pyTrance. It runs entirely at Molm's main configuration (d = 512, eight layers, three seeds) and is in the paper (Table 34, Appendix S, summarized in Section 5); the table at the smaller configuration is gone.

Length scale σSubtypeShape (h²)Pattern
1.25 µm, fixed0.8840.4570.314
2.5 µm, fixed (Molm)0.8640.5500.295
5 µm, fixed0.8490.6180.272
10 µm, fixed0.8410.6130.257
Eight fixed, 1.25–10 µm0.8540.5870.287
One learned, from 2.5 µm0.8580.6170.286
One learned, from 5 µm0.8490.6250.267
Eight learned, from 1.25–10 µm0.8500.5920.283

Synthetic DCIS (κ₂ = 6). Subtype: balanced accuracy within the true cell type (chance 1/3). Shape: ridge R² from the type-2 stream h², as in the paper (from h⁰, as in our earlier reply, 1.25 µm is slightly better: 0.486 against 0.476). Pattern: per-gene lift over chance from the layer-4 measures, averaged over all 313 genes. Table 34 also gives shape from the scalar stream h⁰, with cell type and size. FOCUS and pyTrance: 0.484 and 0.696 on subtype, 0.089 and 0.113 on pattern.

  • The trade-off. 1.25 µm is best on subtype (0.884 against 0.864) and on the pattern (0.314 against 0.295), but worst on shape from h² (0.457 against 0.550). Larger, multi-scale and learned scales reverse this. They raise shape to 0.587–0.625 and cost 0.006–0.023 on subtype and 0.008–0.038 on the pattern, against 2.5 µm.
  • Learning the scale brings no gain on subtype or pattern. The learned scales behave like the larger fixed ones.
  • We keep 2.5 µm, chosen before this sweep, because it sits between the two ends: within 0.020 of the best subtype and 0.019 of the best pattern, and 0.093 above 1.25 µm on shape.
  • No scale changes the comparison with the baselines. At every scale, fixed or learned, Molm stays at or above 0.841 on subtype and 0.257 on the pattern.

So the comparison with FOCUS and pyTrance does not depend on the length scale, and no part of it now rests on the smaller configuration.

W4a. Controlled comparisons: "geometry in the encoder"

A generic geometric network with the same input, objective and training length refines our statement that exceeding pyTrance takes geometry in the encoder: such geometry is needed but not sufficient. We trained EGNN (Satorras et al., 2021) with the shared terms of Molm's objective for 220 epochs (Molm: 215) and added capabilities one at a time (Section 5; Appendix N, Table 24):

  • EGNN, which uses distances by construction, exceeds FOCUS on subtype but, at Molm's number of epochs (220 against 215, about half its compute), stays below pyTrance (0.665 against 0.696); with twice the epochs it reaches 0.720 (one seed), still well below Molm.
  • EGNN + cls + preLN (a class token and pre-norm blocks) is level with Molm (transformer) on subtype (0.866 against 0.864), but not on the pattern (0.235 from h⁰, against 0.295 from Molm's measures).
  • Molm (EGNN), with Molm's measurement added, matches Molm (transformer) on cell type and on the pattern from the measures (0.298), and exceeds it on subtype (0.886), at about 1.8 times its compute.

So with input, objective and training length held fixed, recovering the subtype needs a capable encoder, and recovering the localization pattern needs the measurement. Molm's lead over FOCUS and pyTrance holds with a second encoder. Appendix J (Table 9) gives parameters, epochs and GPU-hours for every model in this comparison.

(2/2)

W2. Real data, "no proof that the learned programs are biologically meaningful"

Two new appendices show the measures on single real DCIS cells, in figures that also make them easier to follow (Figures 6 to 9). With depth, the measures move from encoding gene identity, which explains 85% of their variance in the first layer and 7% in the last, toward spatial patterns such as the nucleus and the cell outline; single measures, each the best match for its target chosen on the training section, track the depth inside the nucleus, the cell's center, the cell outline and DAPI within gene (r 0.51 to 0.73 on the test section; Appendix Q, Figure 6). Where the probe on the measures calls nucleus beyond the segmented nucleus, the DAPI stain is brighter (AUROC 0.67 against 0.50, chance level, for the baseline; Appendix R, Figures 8 and 9). So the measures follow the nucleus in the DAPI stain, which the model does not take as input.

Appendix letters in our earlier reply

The appendices were reordered in the final revision. In our reply, Appendix E is now F (synthetic data), L is M (robustness; the 57-class table is Table 22), K is L (nucleus targets), M is J (compute) and N is P (localization in real cells). The length-scale table is now Table 34 (Appendix S), not Table 3.

Thank you again; we would be glad to discuss any remaining questions.

Dear Reviewer zKdi,

Thank you again for your constructive review. As the discussion closes, we would like to highlight the completed revisions addressing your main concerns:

  • Evidence for localization in real cells: We added validation against independent fractionation data, nuclear annotations and DAPI imaging. Individual measures selected on the training section track nuclear and cellular geometry and DAPI on the held-out section, with within-gene correlations of r = 0.51–0.73 (Appendices P–R).
  • Controlled comparisons and full-size tuning: FOCUS and pyTrance were rerun at comparable GPU-hour budgets. The complete length-scale sweep—including fixed, multiple and learned scales—now uses the main model configuration and three training seeds. Every configuration remains ahead of FOCUS and pyTrance on subtype and localization-pattern recovery (Appendix S, Table 34).
  • Broader synthetic evaluation: We added three generated datasets spanning weaker subtype signals and a null control. The advantage over FOCUS and pyTrance persists where subtype signal is present, and subtype prediction falls to chance in the null (Appendix M).
  • Clearer interpretation of shape: We added geometric-feature baselines and DAPI-derived nuclear targets, and explicitly frame transcript-derived shape prediction as a check of retained information (Appendix L).

These results are now in the revised manuscript, with the remaining limitations stated explicitly. If they address your principal concerns, would you consider raising your score to reflect the revised evidence? We remain available for any final clarification.

Best regards,
The Authors

Reviewer_k7Z5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Molm, a self-supervised transformer that learns cell representations from individual gene-tagged molecules instead of count vectors. Each layer learns soft molecule groupings and computes their local moments in closed form. Training combines iBOT with a consistency loss between random subsamples of a cell. The authors also build a synthetic benchmark on real DCIS cells with resampled molecule positions. Molm slightly exceeds count-based and foundation-model baselines on cell-type probing and cell shape prediction.

Critical strengths
  1. The problem is well motivated, and the method is principled. Using detection thinning as a self-supervised view and computing equivariant moments of learned measures are reasonable ideas.
  2. The synthetic benchmark makes the subtype independent of expression by construction, which gives a clean test of subcellular arrangement.
Critical weaknesses
  1. The real-data evidence for subcellular organization is not enough. Cell-type gains are small, e.g., 0.758 vs. 0.753 for scVI on Lung. The shape targets come from the same transcript cloud that Molm takes as input, while count-based baselines never see positions, so the shape results are close to a sanity check. A simple baseline built from coordinate moments or hull features would be benefitial.

  2. The synthetic result mostly reflects weak baselines. FOCUS is label-supervised, pyTrance is a shallow GNN, and both train far fewer epochs than Molm. In Table 3, the measurement line already reaches 0.724, so the proposed geometry adds only about 0.06. Stronger baselines are required for a fair comparison.

  3. The benchmark covers one dataset and one generator setting, and its compartments resemble what local Gaussian moments are built to detect. A sweep over signal strength and a null control would show how robust the advantage is.

Questions for authors
  1. How do count-based baselines score on the synthetic subtype task?
  2. What is the training cost per dataset?
3 comments

Follow-up: the EGNN baseline and other results of the revision

Thank you again for your review. The revision is complete, including the equivariant graph network (EGNN) we promised in W2b. Below is only what is new since our reply, by your point.

W2b, "stronger baselines" (and Q2, cost). EGNN as published, trained on the same input with the shared terms of our objective, stays below Molm on subtype. A class token and pre-norm blocks bring it level with Molm, and Molm's measurement then brings it level on the pattern too. We trained EGNN (Satorras et al., 2021) for 220 epochs (Molm: 215). Synthetic DCIS, κ₂ = 6, three seeds each:

ModelParameters (M)GPU-hours per runSubtypePattern score (h⁰ / measures)
EGNN13.3300.6650.224 / –
EGNN + cls + preLN21.2350.8660.235 / –
Molm (EGNN)36.71100.8860.233 / 0.298
Molm (transformer)40.9600.8640.236 / 0.295

cls: class token; preLN: pre-norm residual blocks. Molm (EGNN) is EGNN + cls + preLN with Molm's measurement; Molm (transformer) is the Molm of our reply. pyTrance 0.696 / 0.113, FOCUS 0.484 / 0.089.

This refines one statement of our reply. We wrote that exceeding pyTrance "takes geometry in the encoder". Geometry is needed but not sufficient to reach Molm's level: EGNN has distances by construction, yet at Molm's number of epochs (220 against 215, about half its compute) it stays below pyTrance on subtype (0.665 against 0.696); with twice the epochs, at about Molm's compute, it reaches 0.720 (one seed, Appendix N), still well below Molm (0.864). The class token and pre-norm blocks add 0.20. Section 5 and Figure 2 now include EGNN; Appendix N gives every design (Table 24) and the training curves (Figure 5), and Appendix J (Table 9) gives the compute. So the margin over FOCUS and pyTrance needs a capable encoder as well as geometry.

W2c, "the proposed geometry adds only about 0.06". The measurement's main contribution is the localization pattern; most of the subtype comes from a capable encoder. On subtype it adds 0.041 on Molm's transformer and 0.020 on EGNN + cls + preLN. On the pattern it adds about 0.06 on both encoders (0.236 to 0.295, and 0.235 to 0.298), and no model without it exceeds 0.24. This refines a second statement of our reply: we wrote that Molm's measures raise the pattern score "beyond every model in the table". Molm (EGNN), at about 1.8 times Molm's compute (110 against 60 GPU-hours), matches or exceeds Molm (transformer), so the precise statement is that the measures stay ahead of every model without them, on either encoder.

W3b, "compartments that suit Gaussian moments". A model with no measurement, and so no local moments at all, matches Molm on subtype: EGNN + cls + preLN reaches 0.866 at κ₂ = 6, and 0.660 and 0.589 at κ₂ = 18 and 48 (Molm (transformer): 0.864, 0.656, 0.567; Appendix M, Table 21). So Molm's subtype lead over FOCUS and pyTrance is not an artifact of compartments shaped for its moments.

W1a, "real-data evidence for subcellular organization". Two new appendices show the measures on single real DCIS cells, in figures that also make them easier to follow (Figures 6 to 9).

  • Depth (Appendix Q). In the first layer, gene identity explains a median of 85% of the measures' variance, and in the last only 7%; deeper measures form patterns at the scale of the cell, such as the nucleus and the cell outline (Appendix Q). The measures are thus not gene lists: deeper layers organize molecules by their place in the cell.
  • Compartments (Appendix R). The compartment probe's margin over the baseline from gene identity, position and local context grows from 0.06 to 0.11 balanced accuracy as the nucleus moves off the cell's center, and where the probe on the measures calls nucleus more than 2 µm outside the segmented nucleus, the DAPI stain is brighter (AUROC 0.67 over three seeds, against 0.50, chance level, for the baseline's calls; Appendix R).

So on real cells the measures carry localization beyond what gene identity and position give.

Pointers. The appendices are reordered so that the results follow Section 5. The letters in our reply map as follows: synthetic data E → F; compute M → J; size and shape K → L; robustness L → M (its Tables 20 and 21 are now 21 and 22); localization in real cells N → P. On the 20,000-cell set of Appendix O, Molm's subtype is now the mean of three seeds, 0.807 (our reply gave one seed, 0.801).

Thank you again for your careful review; we would be glad to discuss any of these points further.

Thank you for taking the time to provide the very detailed response. It addressed most of my concerns, so I've raised my score to 3. Best of luck.

Dear Reviewer k7Z5,

Thank you for revisiting our response and raising your score. We appreciate your constructive feedback, which helped us strengthen the evaluation and clarify the contribution of the learned measures. Thank you again for your time and consideration.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.