acceptodds
Under review as a conference paper at ICLR 2027

SubcellPert: Benchmarking Subcellular Perturbation Prediction

Abstract

Perturbations change both RNA abundance and intracellular organization, but gene-count predictors omit the positions needed to model spatial responses. To address this limitation, we introduce SubcellPert, a benchmark for perturbation-response prediction at single-RNA-molecule resolution. It comprises 14 imaging datasets with 13.88 million cells and 3.65 billion RNA molecules. Given source observations and a directed condition change, models generate target RNA populations in normalized two-dimensional coordinates. Two independently trained tasks separate abundance from localization: Oracle predicts coordinates given exact target counts; Joint predicts both. Eight baselines span source resampling, classical transport, and frozen foundation-model adapters. Fixed perturbation and cellular-context holdouts evaluate abundance error, target-gene coverage, and conditional spatial error. Through within-dataset analyses, we find context-dependent gains from spatial features in condition decoding. Interpreting spatial responses requires distinguishing changes within spatial clusters from population composition, and RNA fractions from enrichment relative to compartment geometry. Through comprehensive experiments, we find that Gaussian transport achieves the lowest Oracle spatial error on eight of eleven cellular-context OOD datasets. Audits show that better centroid accuracy can accompany worse radial distributions, and that 19 of 190 Joint adapter fits retain source counts after validation. SubcellPert opens a spatial frontier for virtual cell research, laying the groundwork for predictive models that capture both the molecular composition and intracellular organization of cellular responses.

Primary area:
datasets and benchmarks
Keywords:
spatial transcriptomics, virtual cell, perturbation prediction, subcellular RNA organization, datasets and benchmarks, out-of-distribution generalization.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_x86qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce SubcellPert, a benchmark for predicting how RNA abundance and its spatial distribution within cells change across perturbations or conditions. They bring together 14 imaging datasets containing about 13.9 million cells and 3.65 billion RNA molecules. They test their models on held-out cellular contexts or perturbations. In the Oracle setting, models receive target RNA counts and predict locations; in the Joint setting, they predict both. The authors evaluate eight baselines, including five frozen foundation encoders with a shared prediction head. Simple transport methods are often competitive, and the results show why spatial error needs to be interpreted alongside abundance accuracy and gene coverage.

Critical strengths
  1. The problem is well motivated: perturbations can affect both RNA abundance and where RNA is located within a cell. Bringing 14 imaging datasets into a common benchmark is a useful contribution.

  2. The Oracle and Joint settings help separate errors in count prediction from errors in localization. The authors also clearly show how low spatial error can coexist with poor abundance prediction or missing genes.

Critical weaknesses
  1. The spatial score can miss important errors within individual cells. Equation 15 pools each gene’s coordinates across cells and compares these distributions separately. For example, two genes could be together in every target cell but separated in every predicted cell, while their pooled distributions remain unchanged. The spatial score would not detect this error. As a result, low spatial error does not by itself establish that a model has recovered how RNA is organized within individual cells. This limits the conclusions that can be drawn from the current evaluation.

  2. The neural comparison does not isolate the value of the pretrained encoders. All five models use frozen encoders, the same spatial prediction head, and five epochs of training. The head permits only limited affine changes to the source RNA coordinates (Equation 8; Appendix E.4). The common contraction reported in Section 4.2 could therefore reflect the shared prediction setup. Without an otherwise identical model that omits pretrained features, it remains unclear how much the encoders contribute. The authors appropriately restrict their claims to this protocol, but an additional control is needed to interpret these results.

Questions for authors
  1. Evaluation validation: Can the authors provide the fixed-count spatial-sensitivity check mentioned in Appendix F, including a case where pooled gene-wise distributions are preserved but colocalization within cells changes? Which aspects of within-cell organization is the benchmark intended to validate?

  2. Baseline validation: How does the shared adapter perform without pretrained encoder features? How much of the reported spatial improvement can be reproduced by simply contracting the source coordinates by a fixed factor?

  3. Reproducibility clarification: Are the processed data, exact train/validation/test splits, evaluator, and baseline code available to reviewers? I could not locate a link in the PDF. Please also point to the supporting details for the contraction and checkpoint-selection audits reported in Section 4.2.

6 comments
AuthorsOpenReview ↗

Response to Reviewer x86q (1/4)

Thank you for the concrete requests for metric validation, model controls, and reproduction resources. We have completed the additional controls on PF-Astro, Adipose, and ATM-RT under their frozen seed-1 protocols. Neural models completed five epochs at batch size 256; PF-Astro/Adipose use the best validation checkpoint including initialization, and ATM-RT uses the final epoch under its original no-validation protocol. Results below are descriptive, use the all-gene panel, and weight responses equally within each dataset.

W1. The spatial score can miss important within-cell errors

We agree. Equation 15 measures gene-wise marginal localization accuracy after unit/cell-balanced pooling. Cell balancing addresses unequal contributions to that marginal measure; it does not preserve cross-gene dependence or fully describe cell-to-cell heterogeneity. We have narrowed the metric claims in Section 3.5 and the conclusion, distinguishing complete-cell generation from validation of complete intracellular organization.

We have added two complementary diagnostics to the saved predictions: sliced W2² between distributions of complete-cell spatial feature vectors, and W1 between distributions of within-cell gene-centroid separations. The first concatenates 29 moment/Fourier features per gene across the full vocabulary, excludes the explicit count feature, and uses native feature scales with 64 fixed normalized random directions (seed 42). The second uses 496 pairs among 32 genes selected by vocabulary-name hashing, retaining units with at least five coexpressing cells. Both compare unpaired populations, balancing units and cells; pair scores average supported pairs and leave unsupported pairs missing.

DatasetMethodOracle primary W1 ↓Whole-cell W2² ↓Pair W1 ↓
PF-AstroGaussian transport0.2563970.01401670.245528
PF-AstroscGPT0.2497100.01359930.225668
AdiposeGaussian transport0.2873560.001042720.334938
AdiposescGPT0.3031590.001151210.341364
ATM-RTGaussian transport0.0827670.0004690340.153169
ATM-RTscGPT0.0959040.0004065870.174614

ATM-RT illustrates a ranking disagreement: scGPT has lower whole-cell feature W2² but higher marginal and pair errors. These diagnostics capture additional aspects of organization, not a complete validation of it. Oracle holds exact cell-by-gene counts and pair support fixed: supported pairs average 212/496 on PF-Astro (minimum 136), 78/496 on Adipose, and 476/496 on ATM-RT. Joint diagnostics also depend on predicted gene presence, because it affects absent-gene feature filling and coexpression support; we therefore use Oracle for this controlled spatial comparison.

AuthorsOpenReview ↗

Response to Reviewer x86q (2/4)

W2. The neural comparison does not isolate pretrained encoders from the shared prediction setup

We agree that contraction cannot be attributed to the frozen encoders alone. The shared head is initialized at identity and uses A = I + 0.1 tanh(R); diagonal entries approach 0.90 at their lower limit. Both this restricted parameterization and random-pair spatial regression can contribute to the observed behavior.

We have now retrained an adapter without pretrained features in both tasks, preserving the other inputs, architecture, optimization budget, and selection rules. It zeros both the 512 pretrained-state channels and the availability channel, retains the 513-channel interface, and does not load pretrained weights or vocabulary. Oracle results appear under Q2; the Joint abundance comparison is:

DatasetModelJoint MSE ↓PCC-delta ↑Coverage ↑
PF-AstroscGPT0.0063660.0368720.998048
PF-AstroAdapter-only0.0108960.0403830.993343
AdiposescGPT0.0090920.7826100.896774
AdiposeAdapter-only0.0081310.7803730.892473
ATM-RTscGPT0.0896110.2496810.986577
ATM-RTAdapter-only0.0959010.2228771.000000

The effects are mixed. PF-Astro's full scGPT and SpatialFormer models select the epoch-zero checkpoint, reproducing source-preserving resampling; adapter-only selects epoch 2. Thus, their lower PF-Astro MSE is not evidence of a learned pretrained-feature advantage. The control establishes the performance recoverable without pretrained inputs under this adapter protocol. It does not isolate pretraining from encoder architecture, alignment, or the availability indicator; no random-encoder control was run.

We separately replaced random-pair spatial regression with condition/context-grouped MMD, keeping the scGPT encoder, affine head, budget, and checkpoint-selection rule fixed. Oracle W1 changes from 0.249710 to 0.249465 on PF-Astro, 0.303159 to 0.297894 on Adipose, and 0.095904 to 0.085372 on ATM-RT. On ATM-RT, radial and pair errors also improve, but whole-cell W2² worsens from 0.000406587 to 0.000471035 and centroid RMSE from 0.071900 to 0.075034. Radial error worsens on PF-Astro and Adipose. This minibatch MMD uses a biased estimator including singleton groups, so it tests one distributional objective implementation. We have not tested a wider head or longer training, and these comparisons do not identify the cause of contraction.

AuthorsOpenReview ↗

Response to Reviewer x86q (3/4)

Q1. Fixed-count sensitivity validation, including preserved pooled marginals

We have executed both the fixed-count gene-location swap and the preserved-marginals counterexample with the evaluator used for this experiment wave. For these two-cell diagnostic constructions, the pair-support threshold is one cell; benchmark comparisons retain the five-cell threshold. Appendix F.4 and Table 10 now include the measured validation and both constructions.

Executed synthetic checkCount MSEPrimary W1Whole-cell W2²Mean pair W1
Identity0000
Gene-location swap, counts fixed01.2349760.3504350
Cross-gene arrangement changed, marginals fixed000.1458221.333333

The gene-location swap preserves per-cell counts and the unmarked RNA cloud while changing gene marginals. Its pair-distance score remains zero because that construction preserves pair separations; it is a positive sensitivity check for marginal localization, not a universal check for all spatial diagnostics.

For the preserved-marginals construction, each cell contains one molecule of each gene, all at y = 0:

PopulationCell 1 x coordinates: (A, B, C, D)Cell 2 x coordinates: (A, B, C, D)
Target(+1, +1, −1, −1)(−1, −1, +1, +1)
Prediction(+1, −1, +1, −1)(−1, +1, −1, +1)

Every cell has centroid zero and RMS radius one, and every gene's pooled coordinate distribution is preserved. The primary metric correctly returns zero marginal error, while A–B separation W1 is 2, mean separation W1 over all six pairs is 1.333333, and whole-cell feature W2² is 0.145822. Thus, the construction survives coordinate normalization and explicitly demonstrates the primary metric's blind spot. These are executed synthetic checks, not fitted-model results. The revised manuscript describes the validated scope as marginal localization, supplemented by finite-feature and gene-pair checks of dependence and heterogeneity.

Q2. Adapter without pretrained features and a fixed-contraction baseline

Both controls are complete. The adapter-only model was retrained, as described under W2. The fixed-contraction control applies r_pred = 0.90 × r_resampling to the same sealed Oracle count-matched resampling coordinates. Exact counts, source and coordinate draws, seeds, and evaluation frames are unchanged. Predictions are neither recentered nor renormalized after scaling. The factor 0.90 was fixed from the reported audit, not tuned on test outcomes; scaling by 1 reproduces the reference scores.

DatasetResampling W1 ↓Fixed 0.90 W1 ↓scGPT W1 ↓SpatialFormer W1 ↓Adapter-only W1 ↓
PF-Astro0.2538890.2489660.2497100.2498550.249966
Adipose0.3113410.2958430.3031590.3010870.297296
ATM-RT0.0855860.0961380.0959040.0959540.095982

Adapter-only changes Oracle W1 relative to scGPT by +0.10%, −1.93%, and +0.08%; there is no consistent pretrained-feature advantage here. For PF-Astro, fixed contraction achieves 117.8% and 122.1% of the scGPT and SpatialFormer W1 improvements over resampling; for Adipose, the corresponding ratios are 189.4% and 151.1%. We calculate these as (W1_resampling − W1_fixed) / (W1_resampling − W1_neural). Values above 100% mean the fixed control improves the metric more than the neural model; these are descriptive error-gap comparisons, not a causal decomposition. On ATM-RT both neural references and fixed contraction are worse than resampling, so a fraction of positive neural gain is undefined.

The additional error measures show why contraction is not uniformly beneficial:

DatasetRadial W1: resampling → fixed 0.90 ↓Centroid RMSE: resampling → fixed 0.90 ↓
PF-Astro0.141238 → 0.1610310.275876 → 0.262182
Adipose0.251959 → 0.2390980.340681 → 0.320979
ATM-RT0.074034 → 0.1001430.075154 → 0.071539

On ATM-RT, fixed contraction also worsens pair W1 from 0.154131 to 0.175531, despite improving whole-cell feature W2² from 0.000477174 to 0.000406257. The control can match the primary-metric improvement on some datasets, but neither a centroid improvement nor one favorable distributional score establishes recovery of intracellular organization.

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

This paper introduces SubcellPert, a benchmark for predicting changes in both RNA abundance and intracellular localization following perturbations. SubcellPert combines 14 imaging datasets (~14m cells, ~3.65b RNA molecules). The authors define localization-only and joitn abundance-localizaiton prediction tasks, and evaluate eight baselines.

Critical strengths
  1. Important problem. As the authors state, existing perturbation benchmarks largely focus on gene abundance. Predicting spatial localization is a meaningful exercise.
  2. Substantial benchmark contribution. The authors have assembled heterogeneous datasets, which is quite useful.
  3. Interesting evaluation. Separating abundance from spatial localization and reporting gene coverage alongside spatial error are useful design choices.
Critical weaknesses
  1. It's not clear that the benchmark rewards the intended capabilities. The primary spatial metric evaluates gene-wise marginal coordinate distributions, not joint RNA organization within cells. A model could perform well without capturing dependencies among RNA species or cell-to-cell variation.
  2. It's not clear how much perturbation identity matters. If RNA count is strongly linked to organization, then the Oracle supplying exact target RNA counts may already explain much of the spatial organization.
  3. The neural training objective seems mis-aligned. If I'm understanding this correctly, the neural baselines use randomly paired source and target cells with a per-cell regression loss, despite being evaluated on population distributions. Such objectives can encourage predictions toward conditional central tendencies rather than preserving the target distribution. This may explain the reported spatial contraction.
  4. The foundation-model contribution is unclera. All neural baselines use a shared adapter that receives abundance, spatial features, and condition descriptions. An adapter without pretrained representations would help determine whether foundation-model features contribute useful information.
Questions for authors
  1. How should we interpret the strong performance of Gaussian transport? Do the authors believe this primarily reflects properties of the spatial responses, limitations of the neural training procedure, or both?
  2. Have you tested whether perturbation identity provides predictive information beyond the exact target counts supplied in Oracle? If so, how much does removing the perturbation descriptor affect spatial prediction accuracy?
5 comments
AuthorsOpenReview ↗

Response to Reviewer dPaw (1/4)

Thank you for recognizing the benchmark contribution and for identifying the controls needed to interpret its results. We have added cell-preserving spatial diagnostics and completed matched retraining on PF-Astro, Adipose, and ATM-RT under frozen seed 1. The revised manuscript now includes metric validation in Appendices F.3–F.4 (Table 10) and the control methods and results in Appendix H.2 (Tables 37–39); revisions are marked in blue. These representative comparisons use the all-gene panel and the original five-epoch training and checkpoint-selection protocols, with equal response weights within each dataset.

W1. Does the spatial metric reward the intended within-cell capabilities?

The added diagnostics detect the specific failure that gene-wise marginal W1 misses. We executed a four-gene, two-cell construction that changes within-cell gene associations while preserving every cell's counts, centroid, and RMS radius, and every gene's pooled coordinate distribution. Primary W1 is zero, but whole-cell feature W2² is 0.145822 and mean gene-pair separation W1 is 1.333333. The A–B pair changes from colocalized to separated by two normalized units. Appendix F.4 and Table 10 now give the construction and measured validation:

Executed synthetic checkCount MSEPrimary W1Whole-cell W2²Pair W1
Identity0000
Gene-location swap, counts fixed01.2349760.3504350
Within-cell associations changed, marginals fixed000.1458221.333333

The complementary checks compare distributions of (i) complete-cell vectors of spatial moment/Fourier features and (ii) within-cell gene-centroid separations. They preserve aspects of cell-level dependence lost by pooling genes separately. Neither is a complete test of intracellular organization; the gene-location swap, for example, preserves pair distances and therefore has zero pair error.

We also applied these diagnostics to saved benchmark predictions, including resampling as a reference. Oracle fixes exact cell-by-gene counts and hence gene-presence and pair support across methods. All errors below are lower-is-better.

DatasetMethodPrimary W1Whole-cell W2²Pair W1
PF-AstroResampling0.2538890.01423310.250330
PF-AstroGaussian transport0.2563970.01401670.245528
PF-AstroscGPT0.2497100.01359930.225668
AdiposeResampling0.3113410.001045350.341604
AdiposeGaussian transport0.2873560.001042720.334938
AdiposescGPT0.3031590.001151210.341364
ATM-RTResampling0.0855860.0004771740.154131
ATM-RTGaussian transport0.0827670.0004690340.153169
ATM-RTscGPT0.0959040.0004065870.174614

These checks change the interpretation of performance: on ATM-RT, scGPT improves whole-cell feature error over resampling but worsens marginal and pair errors, whereas Gaussian improves all three. We have therefore revised the main-text claims to distinguish marginal localization accuracy from the selected dependencies assessed by the supplementary diagnostics.

Appendix F specifies the finite features and support rules. Benchmark pair scores use 496 candidate pairs among 32 deterministically selected genes and require at least five coexpressing cells per unit; the two-cell synthetic checks use threshold one. Supported pairs average 212/496 on PF-Astro, 78/496 on Adipose, and 476/496 on ATM-RT. Unsupported pairs remain missing. In Joint, predicted gene presence also changes feature filling and pair support, so these scores are not pure geometry comparisons.

AuthorsOpenReview ↗

Response to Reviewer dPaw (2/4)

W2. Could exact target counts explain much of Oracle spatial performance?

Removing explicit perturbation identity produces no consistent loss of Oracle accuracy. We retrained scGPT while removing all condition-descriptor, embedding, and graph-lookup routes, retaining source observations, context, exact target counts, model width, and training/selection rules.

DatasetFull Oracle W1 ↓Without descriptor W1 ↓Relative error change
PF-Astro0.2497100.250046+0.134%
Adipose0.3031590.299163−1.318%
ATM-RT0.0959040.094577−1.383%

This directly limits the claim: Oracle evaluates count-conditioned localization, and these results do not establish incremental predictive value from explicit condition identity given the other inputs. Exact counts can convey condition information; the ablation does not separate that information from source observations and context. We have made this distinction explicit in Section 3.2 and Appendix H.2.

Joint provides the complementary task without target counts. A separate source-geometry ablation increases Joint abundance MSE by 17.95% on Adipose and 3.38% on ATM-RT and lowers PCC-delta on both. Thus the benchmark can measure a useful contribution from spatial inputs when counts must be predicted, although this does not establish a contribution from condition descriptors. The full Joint table and PF-Astro checkpoint qualification are in Appendix H.2.

W3. Is random-pair regression misaligned with population-distribution evaluation?

A matched population-loss control improves primary W1, most substantially on ATM-RT. We replaced the random-pair spatial regression loss with condition/context-grouped MMD over spatial features, retaining the encoder, affine head, five-epoch budget, and checkpoint-selection rule.

DatasetOriginal Oracle W1 ↓Grouped-MMD W1 ↓Error reductionRadial W1: original → MMD ↓
PF-Astro0.2497100.2494650.10%0.151826 → 0.156778
Adipose0.3031590.2978941.74%0.237418 → 0.239579
ATM-RT0.0959040.08537210.98%0.099142 → 0.073635

On ATM-RT, pair W1 also improves from 0.174614 to 0.153859, while whole-cell feature W2² worsens from 0.000406587 to 0.000471035 and centroid RMSE from 0.071900 to 0.075034. Radial error increases slightly on PF-Astro and Adipose. The original training loss therefore affects the measured performance, and the improvement is not uniform across spatial properties.

Appendix H.2 gives the grouped RBF-MMD definition and singleton-group accounting. This is a minibatch, biased V-statistic, and the restricted affine head remains unchanged; we have not performed a new affine-map audit. Consequently, the comparison does not establish the cause of contraction or show that MMD eliminates it. The revised interpretation concerns the combined encoder–adapter–training protocol, rather than attributing the observed behavior to pretrained encoders alone.

AuthorsOpenReview ↗

Response to Reviewer dPaw (3/4)

W4. What do pretrained representations contribute beyond the shared adapter?

The requested retrained adapter-only control recovers similar Oracle performance without pretrained inputs. It retains source counts, learned geometry, gene identities, condition inputs, architecture, training budget, and selection rules. Both the 512 pretrained-state channels and their availability indicator are zeroed; no pretrained weights or vocabulary are loaded. This preserves the 513-channel interface of the scGPT and SpatialFormer references.

DatasetscGPT Oracle W1 ↓SpatialFormer Oracle W1 ↓Adapter-only Oracle W1 ↓
PF-Astro0.2497100.2498550.249966
Adipose0.3031590.3010870.297296
ATM-RT0.0959040.0959540.095982

Removing pretrained inputs changes Oracle W1 relative to scGPT by +0.10%, −1.93%, and +0.08%. There is no consistent pretrained-feature advantage here. Joint results are mixed:

DatasetModelJoint MSE ↓PCC-delta ↑Coverage ↑
PF-AstroscGPT0.0063660.0368720.998048
PF-AstroAdapter-only0.0108960.0403830.993343
AdiposescGPT0.0090920.7826100.896774
AdiposeAdapter-only0.0081310.7803730.892473
ATM-RTscGPT0.0896110.2496810.986577
ATM-RTAdapter-only0.0959010.2228771.000000

PF-Astro's full scGPT and SpatialFormer models select epoch zero and equal source-preserving resampling, whereas adapter-only selects epoch 2; their lower MSE is not evidence of learned pretrained-feature utility. On Adipose, adapter-only has lower MSE but slightly lower PCC-delta; on ATM-RT, scGPT has better MSE and PCC-delta but lower coverage.

Reviewer_MQm8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SubcellPert, a perturbation prediction benchmark at subcellular spatial and single-RNA-molecule resolution. The dataset contains collections across multiple tissues and perturbation types.

Critical strengths

Thorough benchmarking and ablation studies. The model is built on creative integration of RNA localization and RNA counts. The dataset is unique.

Critical weaknesses

The imaging data is a key part of this benchmark. But there is no figure or examples of this imaging data modality used for RNA localization. The model prediction of RNA counts and localization should be additionally tested to see if supplementing initial coordinates conditionally enables better prediction of RNA counts.

Questions for authors
  1. How is the difference between unpaired populations dealt with in the benchmarking task?
  2. Do the authors plan to use the imaging data to run other tasks apart from RNA localization?
3 comments
AuthorsOpenReview ↗

Reply to Reviewer MQm8 (1/2)

Thank you for recognizing the value of the dataset collection and the integration of RNA abundance and localization. We have completed the requested source-geometry control on PF-Astro, Adipose, and ATM-RT under their frozen seed-1 protocols. We address your comments individually below.

W1(a). Examples of the imaging data used for RNA localization

We have added a native fluorescence microscopy example to Appendix D (Figure 8), showing an Adipose Xenium DAPI field and its magnified view from the original deposited image. The RNA-coordinate examples remain in Figures 5–7.

W1(b). Does supplying initial RNA coordinates improve RNA-count prediction?

We retrained scGPT-based Joint adapters with and without source-coordinate features. The control retains source counts, condition inputs, pretrained features, source assignments, output budgets, and the optimization and checkpoint-selection rules. It zeros the point-encoder features entering the shared response branch, including the abundance predictor. Donor coordinates remain available for generating RNA positions. Thus, this tests the contribution of geometry to the shared predictor; it is not an intervention restricted to the abundance head alone.

The table reports all-gene abundance MSE (lower is better), PCC-delta (higher is better), and target-gene coverage. Responses are weighted equally within each dataset.

DatasetSource geometryJoint MSE ↓PCC-delta ↑Coverage ↑
PF-AstroIncluded0.0063660.0368720.998048
PF-AstroRemoved0.0107430.0385310.993342
AdiposeIncluded0.0090920.7826100.896774
AdiposeRemoved0.0107240.7619190.898925
ATM-RTIncluded0.0896110.2496810.986577
ATM-RTRemoved0.0926400.2045560.996644

Removing geometry increases MSE by 68.75%, 17.95%, and 3.38%, respectively; PCC-delta also decreases on Adipose and ATM-RT. This supports an incremental abundance-prediction benefit from source geometry on Adipose and ATM-RT under the tested protocol. PF-Astro needs a more qualified interpretation: validation selects the initialized epoch-zero checkpoint for the full model, whose predictions equal source-preserving resampling, whereas the ablated model selects epoch 2. Its MSE gap therefore does not establish a benefit from learned geometry features, and its PCC-delta slightly improves after ablation. Appendix H.2 and Table 38 now include these controls.

AuthorsOpenReview ↗

Reply to Reviewer MQm8 (2/2)

Q1. How are differences between unpaired populations handled?

We evaluate predictions at the population level because source and target cells are not biologically paired. Each response is defined within the same annotated cellular context and anatomical reference for a specified condition transition. Abundance metrics compare population mean log-counts and changes from the source, while spatial W1 compares each gene's coordinate distributions. Units and cells are balanced when constructing these summaries and distributions (Appendices A and F).

Random source–target assignments during training are computational supervision, not observed cell trajectories. Likewise, Oracle's supplied target count vectors do not imply biological source–target matches. Context restriction and population-level evaluation make comparisons possible without cell pairing, but do not remove all donor, batch, or within-context composition differences.

Q2. Do we plan to use the imaging data for tasks beyond RNA localization?

In this paper, our current work focuses on subcellular perturbation prediction, including changes in both RNA abundance and localization. We do not currently have plans to use these data for tasks beyond this scope.

Dear Reviewer MQm8,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, 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.