acceptodds
Under review as a conference paper at ICLR 2027

Generalization to Unseen Perturbations via Multiple Biological Priors

Abstract

Predicting cellular responses to genetic and chemical perturbations can reduce the cost of experimental screening. A key challenge is generalization to unseen perturbations, since perturbation identities alone do not specify how their responses relate to those observed during training. However, biological knowledge from existing research provides multiple priors that can connect observed and unseen perturbations through shared molecular and functional properties. We thus formulate unseen perturbation prediction as a generalization problem over multiple biological priors, and introduce a corresponding hypothesis space that accommodates different priors and predictor architectures. Using this methodology, we design a two-stage learning algorithm, connected to late fusion, stacked generalization, and boosting, that first trains a response predictor for each prior and then integrates their predictions in response space through a shared integration function. Across two chemical and two genetic perturbation datasets spanning three cell lines, our approach outperforms single-prior predictors and representation-level fusion methods across various metrics, providing an effective way to combine multiple priors.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation, fusion, computational biology, generalization

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_kJZ6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks when to combine several external "priors" (compound or gene embeddings) when predicting single-cell responses to unseen perturbations. Instead of representation-level fusion, it proposes "response-level fusion": train one predictor per prior, then combine the M predicted values of each gene with a rule γ\gamma shared across genes and perturbations (Eq. 3-5). γ\gamma is a mean, median, trimmed mean, learned weighted mean, or a small "Shared MLP". Evaluation uses SciPlex3 and Replogle with cross-validation over perturbations and top-50 DEG metrics.

Critical strengths
  • The comparison is set up the way I'd want it to be. All arms share backbone, folds and prior coverage (Sec 5.2, 5.3) so for the param-free rules only the stage of fusion changes.
  • Tab. 2 has the rows needed for calibration, also where they're unflattering like on random control and perturbed cells, one-hot, a per-perturbation oracle, a Systema-style centered Pearson, and fold stds. Credit to the authors!
Critical weaknesses
  • The biggest numbers (37.4% MMD gain, 11 of 12 cells, L348-350) come from the Shared MLP, and it reads to me like repair of an error shared by all base models, not fusion of priors? γ\gamma only sees M numbers per gene (Eq. 5). On SciPlex3 K562 its six inputs have MMD 0.329-0.347, worse than random perturbed cells (0.284), and sign agreement 0.535-0.549, below random control. The MLP turns these into 0.206 and 0.971 (Tab. 2). Picking the best prior per test perturbation (the oracle) only gets 0.299, and learned global weights equal the plain mean (0.299 vs 0.300). γ\gamma could still use how the six values agree but nothing shows that that is what it's doing. The authors do call the combiner error correction, but then the arms aren't treated equally. No row applies the same MLP to one prior or to the concatenation model. In their favour on RPE1 the inputs beat both trivial baselines and the MLP still winsbut the control is missing there too
  • Without the Shared MLP, the evidence for response-level over representation-level fusion is mixed. The 12-of-12 and 10-of-12 counts compare the best method per class in each cell, picked on the test folds (L338-341). The median alone gets 7 of 12 (L343-344). Their largest gain (sign agreement) I can't really interpret on MCF7 the mean gives 0.581 and the median 0.928, with no input above 0.584, while MMD barely moves.

Clean question, careful table. I just can't attribute the main result to combining priors.

Questions for authors
  1. (experiment) What does the Shared MLP, with the same loss and fitting data, give with M=1, on the concatenation model, and on M seeds of one prior, in all three settings?
  2. (clarification) Why does the median lift sign agreement from 0.58 to 0.93 on MCF7? If it removes single-prior outliers, that's real fusion. What fraction of the top-50 DEGs is down-regulated?
  3. (experiment) Top-50 DEGs are standard, but does the ranking in Tab. 2 hold on all 5,000 genes? That would also move me up.
5 comments
AuthorsOpenReview ↗

Response to Reviewer kJZ6

We thank the reviewer for their careful examination of our results and constructive feedback.

Critical weakness 1 and Question 1: Shared MLP and error correction

Reviewer comment — Critical weakness 1

“The biggest numbers (37.4% MMD gain, 11 of 12 cells, L348-350) come from the Shared MLP, and it reads to me like repair of an error shared by all base models, not fusion of priors? only sees M numbers per gene (Eq. 5). On SciPlex3 K562 its six inputs have MMD 0.329-0.347, worse than random perturbed cells (0.284), and sign agreement 0.535-0.549, below random control. The MLP turns these into 0.206 and 0.971 (Tab. 2). Picking the best prior per test perturbation (the oracle) only gets 0.299, and learned global weights equal the plain mean (0.299 vs 0.300). could still use how the six values agree but nothing shows that that is what it's doing. The authors do call the combiner error correction, but then the arms aren't treated equally. No row applies the same MLP to one prior or to the concatenation model. In their favour on RPE1 the inputs beat both trivial baselines and the MLP still wins but the control is missing there too”

Reviewer question — Question 1

“(experiment) What does the Shared MLP, with the same loss and fitting data, give with M=1, on the concatenation model, and on M seeds of one prior, in all three settings?”

Response

Please see our general response, where we provide the ablations to compare fusing the same prior with multiple seeds and fusing multiple priors. This disentangles the contributions of learned error correction and complementary biological priors, providing empirical evidence explaining when and why each contributes to the gains of response-level fusion.

To test the concatenation model, we ensemble multiple concatenation models using our response-level fusion techniques. Here, in order to keep comparison fair, we use PCA to reduce the dimensionality of the concatenation to normalize for model parameter count. Here we note that fusing a concatenation model is worse than both fusing multiple seeds of the best prior and fusing multiple priors.

Critical weakness 2: Best-of-class comparisons and aggregation

Reviewer comment

“Without the Shared MLP, the evidence for response-level over representation-level fusion is mixed. The 12-of-12 and 10-of-12 counts compare the best method per class in each cell, picked on the test folds (L338-341). The median alone gets 7 of 12 (L343-344). Their largest gain (sign agreement) I can't really interpret on MCF7 the mean gives 0.581 and the median 0.928, with no input above 0.584, while MMD barely moves.”

Response

Simple response-level aggregation methods, such as mean and median, introduce no additional learnable parameters, whereas sophisticated representation-level fusion methods could involve greater model complexity. Their ability to outperform any representation-level approaches on several metrics is therefore a meaningful result. Moreover, the Shared MLP achieves the strongest overall performance and is also favored by validation-based model selection, so excluding it would not reflect the performance of response-level fusion under practical model selection.

Additional reviewer comment: Attribution to priors

Reviewer comment

“Clean question, careful table. I just can't attribute the main result to combining priors.”

Response

Our new ablations (Tables 1–3) explicitly try to separate the contributions of learned error correction from complementary information provided across biological priors. Although we found that the original genetic prior set provided limited gains over same-prior ensembling, this does not necessarily mean that combining heterogeneous priors offers no additional benefit. To show this, we expand the set to seven biological priors and found that multi-prior fusion outperformed a seven-seed best single-prior ensemble on four of six metrics across the two genetic datasets (Table 3). These results suggest that when priors provide complementary biological information, the multi-prior ensembling still plays a major role.

Question 2: Median aggregation on MCF7

Reviewer question

“(clarification) Why does the median lift sign agreement from 0.58 to 0.93 on MCF7? If it removes single-prior outliers, that's real fusion. What fraction of the top-50 DEGs is down-regulated?”

Response

On MCF7, 89.1% of the top 50 DEGs are down-regulated, while individual predictors tend to underestimate perturbation effects. Median aggregation is less sensitive to unusually high predictions than mean aggregation, potentially shifting small predicted effects toward downregulation, which explains the improvement we observed in sign agreement.

AuthorsOpenReview ↗

Response to Reviewer kJZ6 (Part 2)

Question 3: Evaluation across all 5,000 genes

Reviewer question

“(experiment) Top-50 DEGs are standard, but does the ranking in Tab. 2 hold on all 5,000 genes? That would also move me up.”

Response

We have additionally evaluated all methods across all 5,000 genes in both pseudo-bulk and single-cell settings. Response-level fusion remains competitive in pseudo-bulk prediction and in single-cell genetic perturbation prediction, although its performance is more mixed for single-cell chemical perturbations.

However, recent studies have concluded that all-gene evaluation is fundamentally limited because of the noisy nature of the large number of non-differentially expressed genes. Because these metrics and models weight genes uniformly, the majority of genes that are not differentially expressed can mask biologically meaningfully differences in gene expression and lead to misleading model rankings and metrics (Miller et al., 2026; Vollenweider and Bühlmann, 2026).

[1] Henry E. Miller, Gabriel M. Mejia, Francis J. A. Leblanc, Brendan Swain, Bo Wang, and Lucas Paulo de Lima Camillo. Deep learning perturbation models can outperform baselines on calibrated metrics. Nature Biotechnology, October 2026. ISSN 1546-1696. doi: 10.1038/s41587-026-03307-w. URL https://doi.org/10.1038/s41587-026-03307-w.

[2] Michael Vollenweider and Peter B¨uhlmann. Signal, Bounds, and Baselines: Principles for Evaluating Virtual Cell Perturbation Models. bioRxiv, 2026. doi: 10.64898/2026.04.20.719650. URL https://www.biorxiv.org/content/early/2026/05/27/2026.04.20.719650.

AuthorsOpenReview ↗

Response to Reviewer kJZ6 (Part 3)

Table 6. All-5,000-gene single-cell results

Single-cell prediction on all 5,000 genes, with the lowest MMD biological prior shown for each dataset. Entries are mean ± SD across cross-validation folds. Bold marks the best displayed mean for each dataset and metric, including ties.

DatasetFusion typeMethodMMD ↓Sign ↑Pearson ↑
SP3 K562Learning-freeRandom control cell0.528 ± 0.0030.785 ± 0.0020.082 ± 0.022
SP3 K562Learning-freeRandom perturbed cell0.526 ± 0.0020.785 ± 0.0020.030 ± 0.007
SP3 K562Best single priorChemBERTa0.529 ± 0.0020.309 ± 0.0020.120 ± 0.025
SP3 K562RepresentationConcatenation0.529 ± 0.0020.312 ± 0.0040.124 ± 0.021
SP3 K562RepresentationSoft-argmax0.530 ± 0.0020.307 ± 0.0020.091 ± 0.013
SP3 K562RepresentationPerceiver-style0.530 ± 0.0020.312 ± 0.0040.121 ± 0.016
SP3 K562RepresentationPRESAGE-style0.529 ± 0.0020.309 ± 0.0030.109 ± 0.021
SP3 K562ResponseMean0.536 ± 0.0070.309 ± 0.0020.116 ± 0.022
SP3 K562ResponseMedian0.536 ± 0.0080.356 ± 0.0220.117 ± 0.021
SP3 K562ResponseTrimmed mean0.536 ± 0.0080.344 ± 0.0170.117 ± 0.021
SP3 K562ResponseLearned weighted mean0.536 ± 0.0070.309 ± 0.0020.116 ± 0.021
SP3 K562ResponseShared MLP0.586 ± 0.0050.363 ± 0.0240.108 ± 0.018
SP3 MCF7Learning-freeRandom control cell0.494 ± 0.0110.850 ± 0.0010.231 ± 0.046
SP3 MCF7Learning-freeRandom perturbed cell0.492 ± 0.0100.847 ± 0.0010.045 ± 0.013
SP3 MCF7Best single priorECFP0.488 ± 0.0100.263 ± 0.0040.275 ± 0.035
SP3 MCF7RepresentationConcatenation0.487 ± 0.0090.252 ± 0.0020.298 ± 0.038
SP3 MCF7RepresentationSoft-argmax0.496 ± 0.0070.246 ± 0.0050.266 ± 0.032
SP3 MCF7RepresentationPerceiver-style0.488 ± 0.0090.255 ± 0.0080.297 ± 0.048
SP3 MCF7RepresentationPRESAGE-style0.489 ± 0.0090.260 ± 0.0060.291 ± 0.043
SP3 MCF7ResponseMean0.519 ± 0.0110.261 ± 0.0010.287 ± 0.029
SP3 MCF7ResponseMedian0.516 ± 0.0120.371 ± 0.0030.295 ± 0.030
SP3 MCF7ResponseTrimmed mean0.518 ± 0.0120.365 ± 0.0050.295 ± 0.030
SP3 MCF7ResponseLearned weighted mean0.519 ± 0.0110.261 ± 0.0010.287 ± 0.029
SP3 MCF7ResponseShared MLP0.505 ± 0.0110.285 ± 0.0440.256 ± 0.050
Rep. K562Learning-freeRandom control cell0.564 ± 0.0020.499 ± 0.0010.045 ± 0.009
Rep. K562Learning-freeRandom perturbed cell0.562 ± 0.0010.528 ± 0.001-0.009 ± 0.002
Rep. K562Best single priorDepMap0.560 ± 0.0010.552 ± 0.0040.216 ± 0.006
Rep. K562RepresentationConcatenation0.559 ± 0.0010.556 ± 0.0030.231 ± 0.006
Rep. K562RepresentationSoft-argmax0.562 ± 0.0010.550 ± 0.0040.197 ± 0.010
Rep. K562RepresentationPerceiver-style0.561 ± 0.0010.550 ± 0.0010.206 ± 0.007
Rep. K562RepresentationPRESAGE-style0.560 ± 0.0010.554 ± 0.0040.222 ± 0.006
Rep. K562ResponseMean0.601 ± 0.0030.559 ± 0.0040.233 ± 0.005
Rep. K562ResponseMedian0.586 ± 0.0020.546 ± 0.0040.228 ± 0.006
Rep. K562ResponseTrimmed mean0.593 ± 0.0020.550 ± 0.0040.231 ± 0.005
Rep. K562ResponseLearned weighted mean0.600 ± 0.0030.559 ± 0.0040.234 ± 0.004
Rep. K562ResponseShared MLP0.590 ± 0.0130.525 ± 0.0090.197 ± 0.021
Rep. RPE1Learning-freeRandom control cell0.645 ± 0.0030.501 ± 0.0010.115 ± 0.015
Rep. RPE1Learning-freeRandom perturbed cell0.615 ± 0.0020.570 ± 0.0020.091 ± 0.007
Rep. RPE1Best single priorDepMap0.601 ± 0.0020.596 ± 0.0030.349 ± 0.026
Rep. RPE1RepresentationConcatenation0.597 ± 0.0030.604 ± 0.0040.386 ± 0.015
Rep. RPE1RepresentationSoft-argmax0.609 ± 0.0040.591 ± 0.0030.264 ± 0.019
Rep. RPE1RepresentationPerceiver-style0.603 ± 0.0030.596 ± 0.0040.334 ± 0.013
Rep. RPE1RepresentationPRESAGE-style0.600 ± 0.0030.598 ± 0.0040.355 ± 0.012
Rep. RPE1ResponseMean0.630 ± 0.0050.606 ± 0.0020.338 ± 0.010
Rep. RPE1ResponseMedian0.618 ± 0.0040.604 ± 0.0020.332 ± 0.010
Rep. RPE1ResponseTrimmed mean0.624 ± 0.0040.605 ± 0.0020.335 ± 0.010
Rep. RPE1ResponseLearned weighted mean0.630 ± 0.0040.607 ± 0.0020.347 ± 0.010
Rep. RPE1ResponseShared MLP0.650 ± 0.0210.585 ± 0.0060.319 ± 0.026
Reviewer_JEzfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents a study on perturbation response prediction models, analyzing the impact of different modeling choices. In particular, the study focuses on approaches for integrating multiple biological priors and their impact on prediction performance, with an emphasis on generalization to unseen perturbations. Validation spans both genetic and chemical perturbations across multiple biological priors, model architectures and evaluation metrics. Findings suggest that response-level fusion generally provides superior performance in the tested settings.

Critical strengths
  • The problem setting is very relevant, as perturbation prediction is an important task with applications in drug discovery, and generalization to unseen perturbations is an open challenge in computational biology. Given recent works focused on leveraging prior knowledge for this task, this work addresses a relevant question: how to combine heterogeneous biological priors, especially in settings where data is limited.
  • Strong empirical results, with improvements over single priors and representation-level fusion.
  • Extensive application settings, spanning both genetic and chemical perturbations, multiple representation types, and architectural decisions.
Critical weaknesses
  • Even though the study includes a large number of experimental settings, it is hard to fully disentangle the contribution of combining specific biological priors versus those deriving from ensembling independent predictors. To better disentangle these aspects, authors should investigate (1) ensembling independently trained models with the same prior and (2) applying a comparable learned calibration MLP to a single-prior predictor.
  • Novelty. As the authors also note, the proposed approach is very close to established late-fusion frameworks. The authors should better clarify the key novel contributions relative to prior work. Similarly, the theoretical framework seems to summarize an established formalization and bias-variance trade-off.
  • Currently, the study only focuses on generalization within the same dataset. Given the importance of generalizing to new biological settings, it would be interesting to understand whether the same trends apply generalizing to new cellular contexts or different experimental settings.
Questions for authors

See Weaknesses, especially the first two points.

2 comments
AuthorsOpenReview ↗

Response to Reviewer JEzf

We thank the reviewer for the constructive feedback.

Critical weakness 1: Prior complementarity versus ensembling

Reviewer comment

“Even though the study includes a large number of experimental settings, it is hard to fully disentangle the contribution of combining specific biological priors versus those deriving from ensembling independent predictors. To better disentangle these aspects, authors should investigate (1) ensembling independently trained models with the same prior and (2) applying a comparable learned calibration MLP to a single-prior predictor.”

Response

We thank the reviewer for this suggestion. Please see our general response, where we provide the ablations to disentangle the contributions of learned error correction and complementary biological priors, along with empirical evidence explaining when and why each contributes to the gains of response-level fusion.

Critical weakness 2: Novelty

Reviewer comment

“Novelty. As the authors also note, the proposed approach is very close to established late-fusion frameworks. The authors should better clarify the key novel contributions relative to prior work. Similarly, the theoretical framework seems to summarize an established formalization and bias-variance trade-off.”

Response

We clarify our key contributions relative to prior work in the general response. As stated in the paper, Appendix A is intended to provide theoretical intuition through a degenerate case of our nonlinear framework, while our primary contribution is a systematic investigation of response-level fusion as an underexplored approach to perturbation prediction, supported by extensive empirical evaluations.

Critical weakness 3: Generalization across contexts

Reviewer comment

“Currently, the study only focuses on generalization within the same dataset. Given the importance of generalizing to new biological settings, it would be interesting to understand whether the same trends apply generalizing to new cellular contexts or different experimental settings.”

Response

We agree that generalization across biological contexts is an important problem. However, our primary goal is to systematically investigate whether response-level fusion provides an effective alternative for unseen perturbation prediction. This should then provide the basis for adapting such approach for cross-context generalization, which we view as beyond the scope of this study but an important future direction. The same framework can be extended to such settings, provided that base predictors produce responses in a compatible output space.

Questions for authors

Reviewer comment

“See Weaknesses, especially the first two points.”

The relevant responses are under Critical weakness 1 and Critical weakness 2, above.

Thank you for the additional analyses and clarifications. The results are informative, although it seems the benefits of combining heterogeneous priors are sometimes modest copared to same-prior ensembling. Regarding novelty, the additional clarifications help position the empirical contributions, however, the novel methodological contributions remain limited compared to existing late-fusion approaches, and the theoretical work primarily explanatory. Finally, I believe that additional analyses studying generalization across contexts would help strengthen and generalize the findings.

Overall, I therefore maintain my initial score.

Reviewer_ybEYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies cellular-response prediction for unseen genetic and chemical perturbations using multiple biological priors. It trains a separate response predictor for each prior and then combines their predictions using coordinate-wise aggregation, global learned weights, or an MLP shared across response genes and perturbations. Evaluation covers four perturbation screens, two distribution-prediction backbones, and a mean-response setting.

Critical strengths

The method is simple and modular, allowing different biological priors and response predictors to be used. Experiments cover chemical and genetic perturbations, different backbones, and different prediction targets. This breadth provides evidence that the approach is usable across several settings.

Critical weaknesses

1.Insufficient motivation and theoretical justification. The paper does not convincingly explain why response-level fusion should be preferred over representation-level fusion from a first-principles perspective. Beyond avoiding the alignment of heterogeneous representations, what makes this design particularly suitable for biological perturbation prediction? Does it offer distinctive advantages in modeling biological mechanisms, temporal dynamics, or generalization to unseen perturbations? These questions remain unanswered. Appendix A provides only a simplified bias–variance argument under restrictive assumptions, which does not sufficiently justify the proposed design in realistic settings.

2.Limited methodological novelty and inadequate experimental comparisons. The proposed method largely follows established late-fusion and stacking paradigms, with limited methodological novelty. More importantly, the experimental comparisons are insufficient. Recent perturbation prediction methods, for example MAP and TxPert(and a handful of other methods), have explored sophisticated representation-level modeling and fusion strategies. However, the paper primarily compares against relatively generic fusion implementations and a limited set of existing methods, without sufficiently benchmarking against recent competitive perturbation prediction approaches. This makes it difficult to establish whether response-level fusion offers a meaningful advantage over representation-level methods. Furthermore, the experiments do not adequately disentangle the proposed design's benefits from conventional ensembling effects. Consequently, the provided evidence is insufficient to support the methodological claims.

Questions for authors
  1. Motivation and novelty: What is the fundamental justification for preferring response-level fusion over representation-level fusion in perturbation prediction? Beyond conventional late fusion and stacking, what distinctive technical insight or capability does the proposed approach offer?

  2. Experimental comparisons: Can the authors provide comparisons with recent competitive perturbation prediction methods, such as MAP and TxPert, that employ representation-level modeling or fusion? Under fair experimental settings, does the proposed approach still outperform these methods and strong ensemble baselines?

  3. Theoretical justification: How does the simplified analysis in Appendix A relate to the actual nonlinear, two-stage learning procedure? Can the authors clarify its limitations and provide stronger evidence explaining when and why response-level fusion should be preferable?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ybEY

We thank the reviewer for their feedback and particularly appreciate the opportunity to clarify (1) the goal of our paper, (2) comparisons against recent representation-level fusion methods, and (3) the interpretation and limitations of our analysis.

Critical weakness 1 and Question 1: Motivation and theoretical justification

Reviewer comment — Critical weakness 1

“Insufficient motivation and theoretical justification. The paper does not convincingly explain why response-level fusion should be preferred over representation-level fusion from a first-principles perspective. Beyond avoiding the alignment of heterogeneous representations, what makes this design particularly suitable for biological perturbation prediction? Does it offer distinctive advantages in modeling biological mechanisms, temporal dynamics, or generalization to unseen perturbations? These questions remain unanswered. Appendix A provides only a simplified bias--variance argument under restrictive assumptions, which does not sufficiently justify the proposed design in realistic settings.”

Reviewer question — Question 1

“Motivation and novelty: What is the fundamental justification for preferring response-level fusion over representation-level fusion in perturbation prediction? Beyond conventional late fusion and stacking, what distinctive technical insight or capability does the proposed approach offer?”

Response

We do not claim a general first-principles justification for preferring response-level fusion. Rather, our contribution is to systematically investigate it as an underexplored alternative for unseen-perturbation prediction. The biological problem carries its own domain-specific questions, which we address and provide technical insights into: whether response-level fusion offers predictive benefits and, if so, what accounts for them. Our experiments provide a positive answer to the first question. For the second, we hypothesize that its potential advantages arise from two aspects: integrating heterogeneous biological priors in a common response space and learning to correct prediction errors from individual prior-specific predictors. Appendix A examines a simplified setting in which the bias–variance trade-off can be analyzed explicitly, illustrating one possible source of benefit from response-level fusion. Our empirical results further demonstrate the effectiveness of this design in the evaluated biological settings. As discussed in the general response, additional ablations provide further evidence on the respective contributions of learned error correction and complementary information across biological priors.

AuthorsOpenReview ↗

Response to Reviewer ybEY (Part 2)

Critical weakness 2 and Question 2: Experimental comparisons

Reviewer comment — Critical weakness 2

“Limited methodological novelty and inadequate experimental comparisons. The proposed method largely follows established late-fusion and stacking paradigms, with limited methodological novelty. More importantly, the experimental comparisons are insufficient. Recent perturbation prediction methods, for example MAP and TxPert(and a handful of other methods), have explored sophisticated representation-level modeling and fusion strategies. However, the paper primarily compares against relatively generic fusion implementations and a limited set of existing methods, without sufficiently benchmarking against recent competitive perturbation prediction approaches. This makes it difficult to establish whether response-level fusion offers a meaningful advantage over representation-level methods. Furthermore, the experiments do not adequately disentangle the proposed design's benefits from conventional ensembling effects. Consequently, the provided evidence is insufficient to support the methodological claims.”

Reviewer question — Question 2

“Experimental comparisons: Can the authors provide comparisons with recent competitive perturbation prediction methods, such as MAP and TxPert, that employ representation-level modeling or fusion? Under fair experimental settings, does the proposed approach still outperform these methods and strong ensemble baselines?”

Response

We would like to note that our original submission already includes a comparison with PRESAGE, a state-of-the-art sophisticated representation-level fusion method. Our approach outperforms PRESAGE across all six metrics on the two genetic perturbation datasets, even when PRESAGE uses 8 times more biological priors (Table 4).

As suggested, we now further add comparisons with TxPert using its publicly released data splits and checkpoints, as its training code is unavailable (GitHub issue). We retrain our method on the same data and split using our original five priors. Our approach outperforms the strongest TxPert variant across all three metrics (Table 5).

Regarding MAP, we attempted to reproduce the method using its publicly available implementation but encountered issues such as dose not being represented in the code and a missing model dictionary, preventing a reliable comparison (GitHub issue).

Table 4. PRESAGE comparison

Pseudo-bulk comparison with PRESAGE on genetic perturbations. Our method uses a shared MLP. Metrics use each perturbation's top 50 DEGs. Entries are mean ± SD across five cross-validation folds; bold marks the best mean per dataset and metric.

DatasetMethodRMSE ↓Sign ↑Pearson ↑
Replogle K562PRESAGE (our 5 priors)0.171 ± 0.0120.687 ± 0.0160.401 ± 0.040
Replogle K562PRESAGE (40 sources)0.167 ± 0.0050.691 ± 0.0100.431 ± 0.019
Replogle K562Ours (5 priors)0.154 ± 0.0050.821 ± 0.0260.461 ± 0.009
Replogle RPE1PRESAGE (our 5 priors)0.212 ± 0.0050.760 ± 0.0140.324 ± 0.042
Replogle RPE1PRESAGE (40 sources)0.202 ± 0.0080.772 ± 0.0110.385 ± 0.021
Replogle RPE1Ours (5 priors)0.188 ± 0.0060.864 ± 0.0340.430 ± 0.026

Table 5. TxPert comparison

Pseudo-bulk comparison with TxPert on Replogle K562. We retrain our five-prior shared-MLP model on TxPert's data and split and evaluate TxPert's released checkpoints. Metrics use each perturbation's top 50 DEGs; bold marks the best result per metric.

MethodRMSE ↓Sign ↑Pearson ↑
TxPert GAT (STRING)0.1540.8130.565
TxPert Exphormer0.1650.8070.540
TxPert Exphormer-MG (STRING + GO)0.1620.8100.553
Ours (5 priors, shared MLP)0.1520.8690.568
AuthorsOpenReview ↗

Response to Reviewer ybEY (Part 3)

Reviewer question — Question 3

“Theoretical justification: How does the simplified analysis in Appendix A relate to the actual nonlinear, two-stage learning procedure? Can the authors clarify its limitations and provide stronger evidence explaining when and why response-level fusion should be preferable?”

Response

We clarify that the analysis in Appendix A is not intended to provide a theoretical guarantee for response-level fusion. As stated in the paper, it is a stylized analysis of our framework, representing a degenerate case of the nonlinear two-stage procedure to provide intuition about how combining predictors can reduce prediction error. Extending the analysis to the full nonlinear framework is challenging, as both the base predictors and the fusion function are learned nonlinear mappings, requiring additional assumptions to establish tractable theoretical guarantees.

Our justification is therefore primarily empirical. The ablations (see general response, Tables 1–3) clarify when improvements arise from learned error correction versus complementary biological priors. Additionally, comparisons with SOTA methods, including PRESAGE and TxPert (Tables 4–5), justify the effectiveness of response-level fusion.

Thank you for the detailed rebuttal and the additional experiments, but I do not think they fully resolve my concerns. I would prefer to maintain my original rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (Part 1)

Thank you to the reviewers for their thoughtful feedback. Here we respond to common concerns raised by the reviewers.

Interpreting response-level fusion benefits

Response

To understand the gains of response-level fusion (i.e., whether it's from fusion of complementary information across biological priors or a learned correction of prediction errors), we performed ablations by fixing the integration function to a shared MLP and comparing: (1) predictions from a single model using the best prior for each dataset; (2) predictions by fusing multiple models using that same prior with different random seeds; and (3) predictions by fusing multiple models, each using a different prior (Tables 1 and 2).

These ablations reveal two patterns. When one prior dominates (e.g., DepMap is the best prior across all metrics for both of the genetic datasets), its seed ensemble performs comparably to or better than multi-prior fusion, while both outperform the best single prior. This suggests that learned correction accounts for much of the gain in this setting. When no single prior consistently dominates (e.g., in SciPlex3 MCF7, no single prior is the best across all metrics), combining different priors provides additional benefits. Across the two chemical datasets, multi-prior fusion outperforms the best-prior seed ensemble on five of six pseudo-bulk metrics, supporting contributions from both learned correction and complementary prior information.

When the original prior set is dominated by a single strong prior (as in the genetic datasets), heterogeneous prior fusion may still provide additional value if we are able to expand the prior set. To further investigate this, we expanded the genetic prior set to seven priors, including five newly identified priors, and compared multi-prior fusion against an ensemble of seven independently trained models using the strongest original prior (DepMap). With the number of base models and the second-stage integration function matched, multi-prior fusion outperforms the same-prior ensemble on four of six metrics across the two genetic datasets (Table 3).

Together, these experiments suggest that when one prior dominates, error correction contributes a lot to the gains of response-level fusion, whereas complementary information across priors can provide additional benefits when it's not. However, even when one prior initially dominates, identifying additional priors with complementary information can further improve performance. This validates that both parts contribute to the advantages of response-level fusion.

Table 1. Pseudo-bulk ablation

All configurations use a shared MLP. Best priors, selected by RMSE: ChemBERTa (SciPlex3 K562), ECFP (SciPlex3 MCF7), and DepMap (both Replogle datasets). Entries are mean ± SD across cross-validation folds; bold marks the best mean per dataset and metric.

DatasetConfigurationRMSE ↓Sign ↑Pearson ↑
SciPlex3 K562Best prior (1 model)0.077 ± 0.0150.881 ± 0.0260.229 ± 0.074
SciPlex3 K562Best prior (6 seeds)0.075 ± 0.0150.867 ± 0.0160.210 ± 0.054
SciPlex3 K5626 priors (1 model each)0.076 ± 0.0110.885 ± 0.0200.217 ± 0.017
SciPlex3 MCF7Best prior (1 model)0.128 ± 0.0250.756 ± 0.2540.234 ± 0.036
SciPlex3 MCF7Best prior (6 seeds)0.123 ± 0.0250.856 ± 0.1120.340 ± 0.050
SciPlex3 MCF76 priors (1 model each)0.121 ± 0.0270.923 ± 0.0110.351 ± 0.053
Replogle K562Best prior (1 model)0.162 ± 0.0030.760 ± 0.1150.416 ± 0.025
Replogle K562Best prior (5 seeds)0.155 ± 0.0020.822 ± 0.0320.468 ± 0.014
Replogle K5625 priors (1 model each)0.154 ± 0.0050.821 ± 0.0260.461 ± 0.009
Replogle RPE1Best prior (1 model)0.190 ± 0.0100.805 ± 0.0860.428 ± 0.018
Replogle RPE1Best prior (5 seeds)0.185 ± 0.0050.873 ± 0.0190.454 ± 0.016
Replogle RPE15 priors (1 model each)0.188 ± 0.0060.864 ± 0.0340.430 ± 0.026
AuthorsOpenReview ↗

General Response (Part 2)

Table 2. Single-cell ablation

All configurations use a shared MLP. Best priors, selected by MMD: ChemBERTa (SciPlex3 K562), ECFP (SciPlex3 MCF7), and DepMap (both Replogle datasets). Entries are mean ± SD across cross-validation folds; bold marks the best mean per dataset and metric.

DatasetConfigurationMMD ↓Sign ↑Pearson ↑
SciPlex3 K562Best prior (1 model)0.212 ± 0.0450.922 ± 0.0200.403 ± 0.135
SciPlex3 K562Best prior (6 seeds)0.260 ± 0.0540.963 ± 0.0140.496 ± 0.090
SciPlex3 K5626 priors (1 model each)0.228 ± 0.0350.943 ± 0.0240.431 ± 0.110
SciPlex3 MCF7Best prior (1 model)0.278 ± 0.0480.731 ± 0.1710.280 ± 0.059
SciPlex3 MCF7Best prior (6 seeds)0.290 ± 0.0550.683 ± 0.0410.281 ± 0.100
SciPlex3 MCF76 priors (1 model each)0.274 ± 0.0520.808 ± 0.0500.310 ± 0.081
Replogle K562Best prior (1 model)0.478 ± 0.0150.704 ± 0.0450.302 ± 0.017
Replogle K562Best prior (5 seeds)0.457 ± 0.0170.757 ± 0.0380.328 ± 0.007
Replogle K5625 priors (1 model each)0.474 ± 0.0110.748 ± 0.0350.301 ± 0.026
Replogle RPE1Best prior (1 model)0.594 ± 0.0350.812 ± 0.0410.346 ± 0.035
Replogle RPE1Best prior (5 seeds)0.557 ± 0.0360.876 ± 0.0150.372 ± 0.033
Replogle RPE15 priors (1 model each)0.579 ± 0.0290.831 ± 0.0190.363 ± 0.030

Table 3. Expanded genetic prior set

Pseudo-bulk prediction with seven priors versus seven DepMap seeds. Both use a shared MLP and are evaluated on the same held-out perturbation genes covered by all seven priors. Entries are mean ± SD across five cross-validation folds; bold marks the best mean per dataset and metric.

DatasetConfigurationRMSE ↓Sign ↑Pearson ↑
Replogle K562Best prior (7 seeds)0.155 ± 0.0030.832 ± 0.0220.482 ± 0.016
Replogle K5627 priors (1 model each)0.151 ± 0.0040.827 ± 0.0180.493 ± 0.020
Replogle RPE1Best prior (7 seeds)0.201 ± 0.0090.870 ± 0.0090.461 ± 0.015
Replogle RPE17 priors (1 model each)0.198 ± 0.0110.881 ± 0.0130.445 ± 0.049

Novelty

Response

Our contribution is a systematic investigation of response-level fusion as an alternative to the prevailing representation-level approaches for integrating biological priors in unseen-perturbation prediction. Rather than combining prior-derived representations before prediction, we first use each biological prior to construct an independent response predictor and then integrate their predictions in a common gene-expression space. This formulation offers a fundamentally different way to exploit heterogeneous biological knowledge: each prior contributes through its own predictive model, allowing the integration to capture both prior-specific predictive signals and their complementary information.

We investigate the predictive performance of this approach for both genetic and chemical perturbations and analyze the sources of its improvements. With the integration function fixed, our ablations demonstrate that response-level fusion can benefit from both learned error correction and complementary information across biological priors, when available. Moreover, even with a simple second-stage integration function, our approach outperforms current state-of-the-art methods for predicting unseen genetic perturbation effects (Tables 4 and 5).

Together, these results establish response-level fusion as a competitive and effective alternative to representation-level fusion, and provide empirical evidence for when and why integrating multiple biological priors can improve unseen-perturbation prediction over single-prior approaches.

We will revise the manuscript to incorporate these additional results and analyses and clarify our claims and contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.