acceptodds
Withdrawn from ICLR 2027

scAMod: Generalizing Perturbation Prediction through Adaptive Gene Modules

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In silico perturbation prediction offers a scalable view of how cells respond to interventions. Sample variation, destructive sequencing, and dropout make this prediction hard to generalize to the response distributions of unseen contexts. Existing distributional perturbation models emphasize population-level matching, yet jointly capturing distributional shifts and coordinated gene responses within individual cells remains challenging. We propose scAMod, a distributional perturbation prediction model built on adaptive gene modules. We use prior gene interactions and balanced optimal transport reassignment to construct semantically anchored modules, whose gene composition adapts to cellular expression context while retaining shared module semantics. In our method, the Adaptive Module Tokenizer first converts cell-specific transcriptomes into module tokens, which are compressed through a module-centered bottleneck and reconstructed by a gene query decoder. We then use perturbation-conditioned flow matching to generate perturbed cell distributions. Finally, to model condition-specific responses, we introduce Perturbation Module Readout (PMR), which uses the perturbation condition to guide the readout of generated module states into gene-level predictions. On the PBMC donor leave-out benchmark, scAMod reduces DEG energy distance by 24.2% and improves DEG pseudobulk R² by 8.7% over the strongest evaluated baseline. Together with component ablations, these results support the effectiveness of module-centered modeling for generalizable distributional perturbation prediction. Code is available at https://anonymous.4open.science/r/scAMod-A775.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Adaptive gene modules, Context-dependent gene organization, Biological prior integration, Distributional response modeling, Conditional flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QHiZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

scAMod predicts distributions of post-perturbation gene expression for unseen donors and drugs. It constructs a shared soft gene-to-module prior from pretrained gene representations, a PPI network, graph clustering, and balanced assignment. An adaptive module tokenizer uses each cell's full expression profile to adjust gene participation across modules. Module-level interactions and compression produce a cell representation, and a gene-query decoder reconstructs expression. Optimal transport provides training pseudo-pairs for conditional flow matching in the learned latent space. The autoencoder and flow model are then frozen, and Perturbation Module Readout (PMR) adds a condition-dependent contribution to the gene-level predictions. The paper evaluates the method on PBMC leave-one-donor-out prediction and SciPlex3 held-out-drug prediction, with ablations of the module backbone and PMR objectives, routing analyses, and five-seed results on SciPlex3.

Critical strengths

S1. Adaptive module composition is supported by ablations independent of PMR

The model separates the shared prior SgkS_{gk} from cell-dependent participation weights Agk(x)A_{gk}(x). Decoder-query alignment and batch-averaged anchoring allow cell-specific routing while maintaining a common module reference. Table 3 tests this design without PMR: fixing the routing increases Full ED from 0.7371 to 1.3543 and DEG ED from 0.8364 to 0.9952 on SciPlex3. Shuffling the prior and removing the anchor loss also reduce performance. These results support the predictive value of adaptive module composition and the shared prior.

S2. The experiments cover two complementary forms of generalization

The 12 PBMC leave-one-donor-out folds test transfer across donor backgrounds. SciPlex3 holds out entire drug identities across doses and cell lines, testing generalization to unseen compounds. On PBMC, scAMod improves DEG R2R^2 from 0.7602 for control sampling to 0.8260 and reduces DEG ED from 1.1006 to 0.8339. On SciPlex3, it achieves the best reported result on seven of the eight metrics. The appendix provides donor-level results and five-seed results on SciPlex3, with DEG R2=0.4364±0.0147R^2 = 0.4364 \pm 0.0147 and DEG ED =0.7373±0.0106= 0.7373 \pm 0.0106.

Critical weaknesses

W1. The contribution beyond existing module-based methods is not sufficiently established

scBIG (Ruan et al., ICML 2026) already combines GeneCompass embeddings and PPI priors with Gene-Relation Clustering, a Gene-Cluster-Aware Encoder, and conditional flow matching for perturbation prediction. The relevant contributions of scAMod are therefore its cell-specific gene-to-module routing, the corresponding decoder, and PMR. The paper needs a direct comparison that establishes what these components add beyond existing module-based models.

The PMR results also confound the readout architecture with additional supervision. In Table 4, DEG R2R^2 is 0.1508 without PMR, 0.1760 with ED-only PMR training, 0.4883 with DEG-MSE-only training, and 0.4364 with the combined objective. The gains in DEG mean prediction depend strongly on the training objective. These results do not separate the benefit of PMR's architecture from that of adding a supervised conditional correction stage. A conventional conditional readout trained with the same loss and comparable capacity, or fixed and adaptive module backbones paired with the same PMR, would provide a more informative comparison.

The biological analysis leaves a related question open: do the membership changes learned by adaptive routing add biological information beyond the shared prior or fixed routing? Concrete examples of changed gene groups or cross-module relationships, supported by reproducibility and independent evidence, would strengthen this interpretation.

W2. The evaluation does not directly establish perturbation specificity or DEG recovery

The reported metrics cover mean prediction error and differences between cell-population distributions, but leave perturbation-response patterns, discrimination between perturbations, and recovery of responsive genes insufficiently tested. Pearson delta correlation relative to matched controls, PDS with a specified candidate set and distance, and DES or an equivalent DEG-recovery measure would address these gaps. The existing error and distributional metrics should be retained alongside them.

Questions for authors

Q1. How much does adaptive routing improve on scBIG and other module-based predictors?

Please explain the differences between scAMod and scBIG in cell-specific routing, gene-query decoding, and PMR, and provide a targeted performance comparison using matched data splits, gene sets, conditioning information, and comparable training budgets. A conventional conditional readout trained with the same objective, or fixed and adaptive module backbones paired with the same PMR, would also help separate the contributions of adaptive organization and PMR.

Q2. What biological information is gained by adapting module membership?

The paper reports reproducible routing patterns across donors, but the pathway-enrichment analysis mainly concerns the static prior modules. Does adaptation yield candidate gene groups or membership changes not explicitly specified by the input prior? For one or two cases, describe the changes relative to the prior, their stability, and supporting evidence from data or resources not used in modeling.

Q3. Do the performance gains hold under perturbation-aware evaluation?

Please use the existing predictions to report Pearson delta correlation relative to matched controls, Perturbation Discrimination Score (PDS), and Differential Expression Score (DES) or an equivalent DEG-recovery measure. Specify the Cell-Eval version or other implementation, the PDS distance and candidate perturbation set, and the DEG-scoring procedure.

Q4. How does the noise-regularized training path match the inference initialization?

Algorithm 3 (p. 18) uses

zt=(1−t)z0+tz1+σξ,z_t = (1-t)z_0 + t z_1 + \sigma\xi,

but starts inference from the encoded control state z0z_0 without added noise. Please report the value of σ\sigma and explain how the training endpoints and inference initialization are matched. For nonzero σ\sigma, the training endpoint distributions are noise-smoothed; how is this handled during generation?

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

scAMod predicts a perturbed population from control cells and a perturbation descriptor. It constructs a gene-module prior from GeneCompass embeddings and STRING interactions, allows gene-to-module assignments to vary with cellular context, and learns a module autoencoder followed by conditional flow matching. A separately trained Perturbation Module Readout (PMR) adds a condition-dependent expression correction. PMR uses pseudobulk error on training-condition DEGs together with full-profile energy distance. Experiments test transfer to held-out PBMC donors and nine completely withheld SciPlex3 drugs. The paper also studies pathway enrichment, routing stability and cytokine-dependent module weights.

Critical strengths
  1. The drug holdout tests an identifiable generalization task

In Appendix A.1, the nine test drug identities are excluded at every measured dose and in all three cell lines. This removes the easy route of learning a compound from one dose or cell line and predicting another measurement of the same compound. Table 2 then shows a fairly consistent benefit: scAMod leads seven of eight metrics, with DEG R2 increasing from STATE's 0.2997 to 0.4364 and DEG energy distance decreasing from 0.852 to 0.7373. Its advantage is not confined to one pseudobulk score. The five training seeds are useful evidence that the reported result is not an unstable initialization, even though they do not replace additional drug splits.

  1. The module analysis includes checks that help interpret the representation

The split-half routing analysis on control cells is more informative than a few visually attractive attention maps. Appendix B.3 defines expression, assignment-confidence and stability thresholds, then repeats the analysis across five donors. The prior enrichment analysis also states its gene universe and multiple-testing procedure. I appreciate the explicit qualification in Appendix B: the three highlighted cytokines are selected examples, and 63 of 90 cytokines instead assign their largest PMR weight to M11, without a similarly clear immune annotation. The paper therefore supplies evidence that assignments vary reproducibly, while giving the reader enough information to avoid treating every gate as a recovered pathway mechanism.

Critical weaknesses

Critical Weaknesses

  1. The final performance gain has not been separated from supervised readout correction

On p. 9, Table 3 reports SciPlex3 DEG R2 of 0.1508 for the full backbone without PMR and 0.0849 without adaptive modules. Table 4 starts from that same 0.1508 backbone and reaches 0.4364 after adding PMR with the combined loss. The largest change in the reported comparison occurs when a new conditional correction head and a DEG-focused training objective are introduced. The backbone result is also below STATE's 0.2997 in Table 2, although that cross-table comparison is not a matched component experiment.

The existing loss ablation does not answer whether module structure is needed for the improvement. PMR trained only with DEG MSE obtains DEG R2 of 0.4883, higher than the combined objective, while full-profile R2 falls to 0.8178. With energy distance alone, DEG R2 is 0.1760. Those numbers show that the final DEG result depends strongly on the supervised readout objective. An unstructured latent model with a conditional residual head could plausibly benefit from the same supervision; its performance is not reported.

I would like one controlled comparison in which the backbone representation changes but the readout capacity, perturbation inputs and training losses remain fixed. In particular, the adaptive-versus-static module comparison should be repeated with PMR attached to both. A generic residual head with the same DEG and energy losses is the other necessary reference. If these controls retain the advantage, the paper can attribute the gain to its proposed representation. Without them, the current results establish a successful complete pipeline but leave its central module-specific contribution uncertain. Training-condition DEG supervision is legitimate; the issue is the missing matched control.

  1. The PBMC comparison does not test the claim of prediction beyond population-average responses

Section 4.2 links the PBMC result to predicting individual responses "beyond population-average effects". Yet the strongest DEG comparator in Table 1 is control sampling, which makes no cytokine response correction at all. Its DEG R2 is already 0.7602; scAMod reaches 0.8260. STATE is at 0.3255 and CellFlow at 0.7389. These results show an improvement over untreated cells, but do not tell us whether the model transfers the common cytokine response better than a simple population-average rule.

The missing reference is specific: estimate each cytokine's mean response within each cell type from the training donors, then apply that shift to the held-out donor's matched control cells. This retains the target donor's starting population and uses only responses observed in training. It directly tests whether adaptive cellular context adds value beyond transferring the average response. Swapping the source-control donor at inference, with cytokine and cell type fixed, would further show whether the prediction uses donor-specific starting states in a useful way.

The distributional conclusion also needs qualification. Full-profile energy distance is 1.2665 for scAMod versus 1.0043 for control sampling, although MMD and DEG-space metrics improve. A reader choosing the model to generate a population needs to know which response features improve and which worsen. Appendix Table 7 already reports scAMod's results for all twelve donors; what is missing is the paired gain over the relevant comparators for each donor. Donor 1 informed hyperparameter selection according to p. 17, so the eleven remaining folds should also be summarized separately. For SciPlex3, the independent novelty units are nine drug identities. The 83 dose/cell-line conditions and five fits of one split cannot establish robustness to a different set of drugs. Drug-level paired effects would reveal whether the average improvement is broad or concentrated in a few compounds.

Questions for authors
  1. Can the authors repeat the adaptive-versus-static module comparison with the same PMR head and losses on both, and add a generic conditional residual head to an unstructured backbone? Please keep training-condition DEG masks, parameter budget and checkpoint selection comparable. The decisive result is whether the final module advantage survives this matched comparison.

  2. What is the PBMC gain over training-donor, cell-type/cytokine mean-response transfer applied to the test donor's controls? Please extend the existing Table 7 with paired comparator differences, summarize the eleven folds excluding the tuning donor, and identify the conditions responsible for worse full-profile energy distance.

  3. For the nine SciPlex3 test drugs, how many improve over STATE on both DEG R2 and a distributional metric, and what is the uncertainty when drugs are the resampling unit? Please also state exactly which cells were used to select HVGs, fit each autoencoder and choose checkpoints. The current appendix excludes test responses at generation; the remaining question concerns preprocessing and selection.

Reviewer_GfnfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces scAMod, a distributed framework for predicting single-cell perturbations based on adaptive gene modules. The core concept involves using biological priors to define a set of semantically fixed gene modules, while allowing the assignment of genes to these modules to vary according to the cellular context. The framework generates post-perturbation cellular states at the module level, which are then mapped to gene expression profiles based on the specific perturbation conditions. The method is evaluated using two experimental settings: cytokine responses in PBMCs (with a "leave-one-donor-out" approach) and drug responses in the SciPlex3 dataset (using a "leave-out-entire-drug" approach).

Critical strengths
  • Allowing module semantics to be shared across cells while enabling gene-to-module assignments to adapt to the cellular context represents a sound improvement over static, prior-defined gene programs.

  • The evaluation settings are rigorous and large-scale. In the cross-donor experiments, stimulated cells from the held-out donor are entirely excluded from training (across a 12-fold cross-validation covering 90 different cytokines). In the unseen-drug experiments, all data for a given drug—across all dosages and cell lines—are held out, effectively preventing information leakage.

Critical weaknesses
  • The paper’s motivation is unclear and lacks empirical support. The introduction presents three distinct motivations: 1) existing distributed methods struggle to capture coordinated gene responses within cells; 2) current approaches for incorporating biological priors are insensitive to cellular context; and 3) the importance of generalization across donors. However, the paper provides no experiments or case studies demonstrating that existing methods actually fail due to the first two issues. Furthermore, the third motivation is disconnected from the method’s core mechanism: the "adaptive routing" is designed to address differences between cell types, yet the primary experiments evaluate generalization across donors—the paper fails to explain how the former contributes to the latter.

  • Some results contradict the paper’s claims and remain unaddressed. On the PBMC dataset, the proposed method yields a higher energy distance (1.267) for full-genome expression compared to a control sampling baseline (1.004), which undermines the paper’s emphasis on its capability for distribution-level prediction.

  • The analysis of the routing mechanism remains purely descriptive, failing to establish any link to predictive performance or known biological phenomena. Observed variations in module-switching rates between different cell-type pairs may simply reflect the overall distance between the expression profiles of those cell types.

  • The core problem that the paper claims to address is never actually evaluated. While the abstract and introduction emphasize the modeling of coordinated gene responses within cells, none of the experiments assess whether the predicted intracellular coordination structures—such as inter-gene covariances or co-expression patterns—align with empirical data.

  • The proposed approach lacks a unified mechanistic explanation, and the logical relationships between its various components remain unclear. The paper fails to offer a testable hypothesis as to why adaptive modules improve the prediction of perturbation effects. Furthermore, the construction of module priors involves multiple steps, yet the necessity of each step is neither justified nor evaluated through ablation studies. Overall, the paper resembles a collection of stacked technical components rather than a coherent framework derived logically from the problem statement.

  • It is also questionable whether the "readout heads" actually learn information specific to the perturbations.

Questions for authors

See my weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.