acceptodds
Under review as a conference paper at ICLR 2027

GeneScope: Adaptive Gene Resolution for Chemical Perturbation Prediction

Abstract

Predicting how cells respond to drugs could help researchers decide which experiments to run across a vast space of compounds and cellular contexts. Current methods learn perturbation effects in latent spaces, model changes in cell populations, or draw on pretrained cellular representations. Yet benchmarks of genetic perturbation prediction show that complex deep models do not consistently outperform simple linear baselines. These findings encourage a closer look at what models preserve about cells before treatment. Global compression can blur gene-level differences, but choosing which details to retain is difficult when the responding genes are still unknown. We introduce GeneScope, an adaptive-resolution Transformer built on full coverage with selective refinement. GeneScope keeps a compact view of all input genes, then uses untreated cells to identify where additional detail is needed. Genes that are poorly reconstructed or vary across controls receive finer-resolution representations. The controls thus guide where to preserve detail, while the compound guides how the state changes. Every gene can still respond, whether or not it is selected for refinement. We evaluate GeneScope on Tahoe and two Sci-Plex3-derived benchmarks. Shared-condition comparisons show improvements in response correlation and high-response gene recovery over linear and multiple deep-learning baselines. These gains hold without a pretrained cellular encoder. Adding pretrained representations further improves response rank correlation on both SciPlex benchmarks. GeneScope therefore ties gene resolution to the cellular state we can observe, rather than the response we have yet to predict. Code and preprocessing pipelines are available at an anonymous repository: https://anonymous.4open.science/r/GeneScope/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, chemical perturbation response, adaptive-resolution transformers, gene expression modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_JTrHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

this paper proposed genescope with adaptive resolution transformer for single-cell chemical perturbation prediction. it dynatically determines and select genes that should be kept under higher resolution representation without relying on unknown treatment response, based on features of control cell population. genescope contains following major modules for representation learning, 1) a macro encoder to compress gene represetntation for 2k input genes via three gene partition views, 2) a micro encoder to select 200 genes via control gene variance and reconstruction error and build fine res representation via attention mechanism, and 3) a population encoder to learn context across control cells via transformer. then the response decoder generates post-treatment expression profile using shared affine response, nonlinear control-mean correction and gene-specific residual prediction. main results show genescope get better improment on response Pearson on Tahoe dataset, and improved high-response gene retrieval on cellxgene-sciplex3 dataset.

Critical strengths
  1. clean paper writing explains the method and model clearly, with explaination on choice of architecture, objective and experiment settting, as well as revealing the limitations brought by baseline protocol difference, dataset dependence, etc.
  2. separately handle res allocation and treatment response prediction, to avoid using treated outcome for inference time selection, with dense macro mechanism gauranteeing response can come from any gene.
Critical weaknesses
  1. lack comparison for control-guided selection method to other task-specific/statistic-driven/general feature selection method.
  2. micro-only interventation removes macro attention and gene features, as well changed micro cross-attention memory, population pooling and gene selection score, thus cannot be used to differentiate , thus cannnot prove the lift comes from proposed adapative selection strategy.
  3. too many discrepancies exist between baseline setting and genescope prediction setting, e.g. in normalization, control sampling, profile used, etc., which greatly prevent me to trust the good looking numbers.
  4. high-response gene retrieval targets were calculated from observed response magnitude rather than FDR-controled DEGs which cannot prove compound-specific effects are truely identified from post-treatment population distribution.
Questions for authors
  1. can u provide direct evidence to prove control-guided selection criterion is core to corresponding performance improvement? e.g. under same genescope structure, k (=200), objective and compute resource, directly compare to random/variance-only/reconstruction-error-only/fixed seletion?
  2. can u rerun benchmark for ridge, chemCPA and STATE-ST under unified evaluation criterion?
  3. how did u handle the coverage limitation of adapting chemCPA on unseen treatment prediction?
  4. can you try to explain the better performance over ridge regression on tahoe-100m but degenerating on the tow sciplex data on unseen compound prediction?
1 comment

We appreciate your suggestions for stronger controls and more consistent baseline comparisons. Our responses below present the new experiments and explain how they will inform the revision.


Response to Q1:

Thank you for your constructive suggestion. Control-guided selection improves mean Pearson and AUPRC over random and fixed panels; single-criterion rules remain competitive (Table 1). Following your suggestion, 36 runs compare four alternatives across three datasets and three matched seeds, fixing architecture, K=200, losses, initialization, split, 5,000 updates and four GPUs. Random samples one panel per seed; Fixed uses one training-control variance-ranked panel across conditions and seeds.

For W1-2, selection alone changes; Macro features, cross-attention and pooling remain fixed, unlike Micro-only.

Table 1. Response Pearson correlation on all test conditions, mean (own sample SD).

DatasetGeneScopeRandomVariance-onlyReconstruction-onlyFixed
Tahoe (646)0.4954(0.0065)0.4676(0.0210)0.4914(0.0098)0.4982(0.0060)0.4865(0.0203)
SciPlex-middle (1036)0.1032(0.0018)0.0926(0.0019)0.1045(0.0012)0.1014(0.0030)0.0907(0.0020)
cellxgene-SciPlex3 (1040)0.0785(0.0014)0.0686(0.0018)0.0779(0.0012)0.0792(0.0011)0.0692(0.0023)

Eighteen further runs retain only energy distance, response MSE, reconstruction and selection stability. GeneScope selection still exceeds random in Pearson and AUPRC across datasets (e.g., SciPlex-middle Pearson 0.0804 vs 0.0700). Auditing 54 records showed that control-guided selection favored frequently detected genes.


Response to Q2:

We reran evaluation with common expression units and reference populations. GeneScope leads on Tahoe Pearson and SciPlex RMSE; ridge and STATE-ST have higher SciPlex Pearson (Table 2).

Addressing "normalization, control sampling", models share 2,000 output genes and reference means from all control/treated cells. Population models receive the same 256 controls, processed in two 128-cell groups by STATE-ST. Scores from raw predictions are averaged equally across conditions. We tuned ridge over seven regularization values on validation data, regenerated predictions from six saved STATE-ST models and retrained chemCPA (Q3). Training budgets and molecular/dose inputs remain model-specific; ridge adds training-control means.

Table 2. Means on test conditions with predictions from all models; STATE-ST limits coverage (counts). Space permits summary scores.

DatasetMethodResponse Pearson correlationAUPRCRMSE
Tahoe (529)GeneScope0.5190.4900.173
Ridge0.4360.4510.258
STATE-ST0.4420.4640.242
chemCPA0.1710.3340.353
SciPlex-middle (808)GeneScope0.0960.3620.061
Ridge0.1150.2890.119
STATE-ST0.0990.3270.116
chemCPA0.0800.3460.121
cellxgene-SciPlex3 (812)GeneScope0.0770.1190.134
Ridge0.1240.0770.436
STATE-ST0.1250.0960.415
chemCPA0.1130.1290.338

Your distinction in W4 is helpful: AUPRC evaluates recovery of genes with large observed changes, not FDR-controlled DEGs or causal specificity. A large, noisy change can rank highly without statistical significance. We have not run replicate-aware DEG tests, and the revision will clarify this and include the protocol and full results.


Response to Q3:

The original chemCPA adapter used compound-specific dose parameters and excluded compounds outside its training vocabulary. We addressed this restriction by replacing those parameters with a dose network shared across compounds and representing drugs using frozen Morgan256 fingerprints. An unseen compound can therefore enter through its molecular structure and dose, without fitting new drug-specific parameters. We also recovered 97 missing structures from chemical identities, without using test expression.

We then retrained chemCPA with three seeds per dataset for 100 epochs, selecting checkpoints on validation data. These nine models cover all test conditions, including unseen compounds. On unseen compounds in familiar cell lines, GeneScope outperforms chemCPA on response Pearson, AUPRC and RMSE on Tahoe, while chemCPA performs better on cellxgene-SciPlex3. We will document the adapter changes and include both coverage and transfer results in the revision.


Response to Q4:

Our interpretation is that limited cellular diversity in SciPlex makes state-dependent corrections harder to learn, favoring ridge's regularized chemical-to-response mapping for unseen drugs. Tahoe spans 50 cell lines; both SciPlex versions represent the same three-cell-line screen. For example, untreated A549 cells reveal the starting state but not an unseen compound's mechanism.

With unified scoring, Tahoe's unseen-compound advantage over ridge is in AUPRC (0.401 vs 0.365), while Pearson is slightly lower (0.441 vs 0.460). We have not isolated cell-line diversity from differences in compounds and preprocessing.

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

The authors suggest an adaptive resolution model for drug response prediction. The model is built with three submodules: a macro encoder, which encodes all input genes; a micro encoder, which adaptively selects a subset of genes based on poor reconstruction or large deviations from the population mean of a control population; and an affine branch, which takes the control mean as input. All submodules are conditioned on the compound. The model is compared with ridge and ML baselines on three datasets and shows better performance on one of them and mixed results on the other two. The authors also perform an ablation study, separating each submodule.

Critical strengths
  1. The core idea is clean and well-motivated. Selecting a subset of genes based on control population statistics removes the circularity of choosing them based on predictions
  2. The use of the affine channel is the right design choice given the performance of linear baselines on these tasks. The ablation study confirms the importance of this channel
  3. The results on the Tahoe dataset are strong.
Critical weaknesses
  1. The ablation raises questions about the chosen architecture: Micro-only matches the full model on the Tahoe dataset. Combined with the fact that this is the only dataset where the full model clearly wins, it is not clear that the Macro encoder contributes on the Tahoe dataset, where the model is strongest.
  2. Another point the ablation raises is related to the two metrics that inform the micro encoder. In the Micro-only setting, there is no Macro encoder and therefore no reconstruction term. The subset selection is then based only on the deviation from the mean, which is different from other experiments, and if reconstruction was somehow used without the Macro encoder, this should be clearly stated. The fact that this experiment matches the full model raises the question of whether the reconstruction term is not contributing. Beyond that point, it is important to understand why the authors chose the two criteria and not only one of them.
  3. The resolution sweep shows very weak dependency on K, and is not consistent across the three metrics. By itself, this is a reasonable result, but it leaves the question of whether the micro encoder really selects a meaningful subset of genes. One important missing ablation is with a random subset of genes of the same length.
  4. The fairness of the comparison with the baselines is questionable, mainly because of the objective losses used in this work: two of the five evaluation metrics (Pearson correlation and direction) were used as loss functions during training, which the baseline models didn't use. An ablation study of different objectives can tell if the reported gains are architectural or optimization-based, and it is missing.
  5. The fact that GeneScope generalizes better than the baselines on the Tahoe dataset but not on the SciPlex datasets is noted but not explained. The authors make no hypothesis about the cause of this difference, or further experiments that can point to it (for example, ablating the number of cell lines). The result is a clear win on only one dataset. The other two datasets are two different preprocessing settings on the same dataset. The difference in performance is not marginal, especially in Δr\Delta r, and without further experiments or at least a mechanism that can explain it, the generalization results are not conclusive.
  6. Some of the notations are not explained in the main text. For example, qgq_g, frecf_{rec} and ψb\psi_b in Eq. 1, TmacroT_{macro} and Zi(r)Z_i^{(r)} before Eq. 2, and dcd_c in Eq. 6.
Questions for authors
  1. How does the model perform with a random subset of genes instead of the selected ones?
  2. What genes are selected by the Micro encoder? Do they have any biological significance? Do they expose biases in the selection process?
  3. How do the results change when the model is trained with only LenergyL_{energy}, similar to the baselines? Are there correlations between the objective loss and the corresponding evaluation metric (LpearsonL_{pearson} with Δr\Delta r, LdirectionL_{direction} with Dir, etc.)
  4. Is there any explanation of the difference in generalization performance between the Tahoe and the SciPlex datasets? One possible explanation would be that the number of cell lines is a dominant factor. Can you confirm or refute this hypothesis?
1 comment

Thank you for your thoughtful questions on biological interpretation, training objectives, and generalization. Below, we address them with additional evidence and outline the planned revisions. (For the notions, we will define all symbols clearly at first use in the revised manuscript.)


Response to Q1:

In our additional 36 selection comparisons, random selection gives lower mean Pearson and AUPRC on all three datasets, including nine random controls, with architecture, K=200, losses, initialization and 5,000 updates fixed across three seeds. We fixed a random gene set per seed and compared its model with the matching GeneScope model.

Table 1. Test-set response Pearson correlation (mean ± sample standard deviation).

DatasetGeneScopeRandom
Tahoe (646 conditions)0.4954±0.00650.4676±0.0210
SciPlex-middle (1,036)0.1032±0.00180.0926±0.0019
cellxgene-SciPlex3 (1,040)0.0785±0.00140.0686±0.0018

This fixes K and tests which genes to refine (Weakness 3); the K sweep experiments found no unique optimum. Micro-only changes pooling, attention and selection together (Weakness 1), so it does not isolate Macro's contribution. We will clarify this point in the revised manuscript.


Response to Q2:

Thank you for encouraging us to examine the biological meaning of the selected genes. Frequently selected genes include GAPDH (Tahoe), APBB2 (SciPlex-middle), and NEAT1 (cellxgene-SciPlex3). GAPDH contributes to glycolysis and cellular stress responses, while APBB2 is an APP-binding signaling adaptor, suggesting links to the metabolic and signaling context in which drugs act. NEAT1 offers a more direct connection to drug sensitivity, as its depletion sensitized cancer cells to the PARP inhibitor ABT-888 [1]. These examples provide biological context for the features selected from untreated cells.

We analyzed 54 records across 2,000 genes with equal cell-line weighting. For our rule, only the simplified runs (Q3) saved selections. Selection favors readily detected genes: Spearman correlations with detection rate are 0.894, 0.713 and 0.685 in dataset order. After adjusting for mean expression, detection and variance, correlations with reconstruction error are 0.228, 0.185 and 0.092. Thus, selection has a strong detection bias and a smaller association with reconstruction error beyond these properties.

For Weakness 2, variance finds genes that differ among untreated cells; reconstruction error finds detail lost by Macro. We take 60 genes by variance, 100 by reconstruction error and 40 by their sum, without duplicates. Micro-only uses absolute control deviation, with no Macro reconstruction. In 18 single-criterion runs, individual criteria sometimes yield higher Pearson; combining them gives higher cellxgene-SciPlex3 AUPRC than either alone.

[1] Adriaens C, Standaert L, Barra J, et al. p53 induces formation of NEAT1 lncRNA-containing paraspeckles that modulate replication stress response and chemosensitivity[J]. Nature medicine, 2016, 22(8): 861-868.


Response to Q3:

With combined selection, the extra losses improve correlation and gene recovery, while removing them improves RMSE. We ran 18 controlled comparisons across both selection rules, three datasets and three seeds (Weakness 4). The simplified objective removes Pearson, soft Spearman, direction, MMD and amplitude losses, retaining energy distance, response MSE, reconstruction and stability.

Table 2. Complete test sets without these five losses; three-run means. Joint uses both selection criteria.

DatasetJoint PearsonRandom PearsonJoint AUPRCRandom AUPRC
Tahoe0.48400.47670.46510.4499
SciPlex-middle0.08040.07000.29490.2896
cellxgene-SciPlex30.06820.05890.11090.1054

On SciPlex-middle, simplification lowers RMSE but also correlation and direction accuracy. This measures the losses' combined effect; individual loss–metric associations and exact baseline-loss matching remain untested. AUPRC has no dedicated loss.

Selection still helps without these terms: combined selection exceeds random in mean Pearson and AUPRC on all three datasets. For example, SciPlex-middle Pearson is 0.0804 versus 0.0700; random has slightly lower RMSE. We will add the complete comparisons and run variability to the revision.


Response to Q4:

Thank you for this suggestion. Our interpretation is that Tahoe’s broader cellular diversity, 50 cell lines versus three in SciPlex, provides more opportunities to learn state-dependent drug effects. For example, observing the same compound across different cell lines helps identify which starting-state features matter for its response. With fewer contexts, these corrections are harder to learn, and ridge’s regularized mapping can generalize better to unseen compounds. Cell-line diversity is therefore a plausible contributor, but the current comparisons do not establish it as the dominant factor, since compounds and preprocessing also differ.

Comments on the paper

2 comments

Thank you for your careful assessment of the mechanism and evaluation protocol. Below, we provide paired comparisons, clarify the protocol, and describe the corresponding revisions.


Response to Q1:

Thank you for suggesting a comparison that isolates the selection rule from the fine-resolution prediction path. We conducted 36 additional runs covering four alternative selectors, three datasets and three matched seeds. The full architecture, K=200, losses, initialization and 5,000-update budget were held fixed. Unlike the Micro-only intervention discussed in Weakness 1, these comparisons preserve Macro features, population pooling and cross-attention.

Random selection uses one randomly sampled gene panel per seed, while Fixed uses a single panel ranked by variance in training controls across all conditions and seeds. Table 1 reports absolute performance and the variability of paired differences on Tahoe.

Table 1. All 646 Tahoe test conditions. Scores are each model’s mean ± sample SD across three seeds. Paired SD is the standard deviation of within-seed GeneScope-minus-alternative differences; the same three GeneScope runs are used throughout.

SelectorResponse Pearson correlationPaired SD
GeneScope0.4954 ± 0.0065--
Random0.4676 ± 0.02100.0266
Variance-only0.4914 ± 0.00980.0033
Reconstruction-only0.4982 ± 0.00600.0021
Fixed0.4865 ± 0.02030.0149

The mean paired Pearson gain is 0.0278 over Random and 0.0040 over Variance-only, with paired SDs of 0.0266 and 0.0033, respectively. The clearest improvement is over random selection, while the single-criterion selectors remain competitive. This supports using control-derived information to choose which genes receive additional detail.

The comparison also extends beyond the full objective: in 18 additional runs with reduced losses, control-guided selection retains higher mean Pearson and AUPRC than random selection across all three datasets. For example, SciPlex-middle Pearson is 0.0804 versus 0.0700. We will add these fixed-architecture comparisons and paired variability to the revised ablation study.


Response to Q2:

Thank you for suggesting this controlled comparison. We tuned ridge over seven regularization values on validation, regenerated predictions from six retained STATE-ST checkpoints, and trained nine chemCPA models—three seeds per dataset, 100 epochs, with validation-selected checkpoints.

All methods use the same 2,000 genes and condition-specific reference means computed from all available control and treated cells. We score unclipped predictions and weight conditions equally. Tahoe and cellxgene-SciPlex3 panel counts are normalized to 10,000 per cell before log1p; SciPlex-middle retains released values. Population models share 256 controls, processed as two 128-cell groups by STATE-ST. Budgets and inputs remain model-specific: GeneScope/ridge use chemical descriptors, STATE-ST uses one-hot treatments, and ridge adds training-control means.

For chemCPA, we replaced compound-specific dose parameters with a shared dose network and frozen Morgan256 fingerprints, enabling unseen compounds without fitting new drug-specific parameters. We recovered 97 missing structures without using test expression; the resulting models cover all test conditions.

GeneScope retains higher Tahoe Pearson and SciPlex AUPRC than ridge and STATE-ST (Table 2), although these baselines have higher SciPlex Pearson and chemCPA leads on cellxgene-SciPlex3 AUPRC.

Table 2. Common conditions (STATE-ST coverage): 529 Tahoe, 808 SciPlex-middle, 812 cellxgene-SciPlex3 conditions. Mean ± SD: three GeneScope and chemCPA runs, two STATE-ST runs; one ridge fit.

MethodTahoe response Pearson correlationSciPlex-middle AUPRCcellxgene-SciPlex3 AUPRC
GeneScope0.519 ± 0.0080.362 ± 0.0160.119 ± 0.002
Ridge0.4360.2890.077
STATE-ST0.442 ± 0.0230.327 ± 0.0030.096 ± 0.000
chemCPA0.171 ± 0.0040.346 ± 0.0300.129 ± 0.003

For unseen compounds in seen cell lines, Pearson favors GeneScope over chemCPA on Tahoe (0.441 vs 0.277), but favors chemCPA on cellxgene-SciPlex3 (0.300 vs 0.096). We will replace the historical comparison and document the unified protocol, remaining differences, and adapter changes.


Response to Q3:

Thank you for pointing out this ambiguity. We did follow the standard train/validation/test protocol: model parameters were learned from the training set, validation data guided tuning and checkpoint selection, and the test set was used to evaluate finalized configurations. Test results were not used for model selection.

The wording in Appendix A referred to repeated evaluations of the completed model on the same held-out test split. It did not mean that test results guided model development. The rebuttal comparisons reused this split, with configurations fixed before test scoring. We will revise Appendix A to state the protocol directly and remove the ambiguous wording.

Reviewer_XCfLOpenReview ↗

Updated Recommendation After Rebuttal

Thank you for the additional experiments. At (K=200), control-guided selection improves mean Pearson over random selection, though the single-criterion selectors remain competitive. The common-reference evaluation and revised chemCPA baseline address my main comparison concern. Your clarification that validation, not test data, guided model selection also resolves my question about Appendix A. Please include these results and clarifications in the revision. I now recommend weak acceptance (3), up from weak rejection (2)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.