acceptodds
Under review as a conference paper at ICLR 2027

CrispyDrug: A causal framework for learning joint representations of genetic and chemical perturbations in cells

Abstract

Large-scale single-cell RNA-seq perturbation datasets provide detailed measurements of how interventions alter cellular state. However, existing perturbation models typically focus on either genetic or chemical perturbations, while causal methods for jointly modeling both perturbation types are missing. Here, we introduce CrispyDrug, a novel causal framework for learning joint representations of genetic and chemical perturbations from single-cell transcriptomic data. CrispyDrug treats both perturbation types as interventions on a shared cellular outcome and learns representations through their effects on post-treatment gene expression. Our framework proceeds in two stages. (1) First, it learns perturbation representations by predicting post-treatment expression and, at the same time, balances the representations wrt. to the perturbation type. (2) Second, it trains perturbation-type-specific decoders that map the joint representations back to genetic or chemical perturbations. Across large-scale experiments in the K562 leukemia cell line, we find that CrispyDrug captures heterogeneous transcriptional responses and achieves stronger predictive performance than existing perturbation models. By analyzing the learned latent space and cross-perturbation retrieval, we further find that the joint representation links genetic and chemical perturbations through similarities in their transcriptional effects. To the best of our knowledge, CrispyDrug is the first framework to jointly model both genetic and chemical perturbations in single-cell transcriptomics with an explicit causal interpretation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single cell perturbation modeling, chemical and genetic perturbation modeling, representation learning, treatment effect prediction

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 3 reviews, above 27% of the others.
Reviewer_WKdTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CrispyDrug, which jointly models genetic and chemical perturbations and learns shared perturbation representations through their effects on post-treatment expression. CrispyDrug includes two stages: Stage 1 maps both genetic and chemical perturbations to a common representation space with perturbation-type-specific encoders, optimizing post-treatment expression prediction and perturbation-type balancing concurrently. Stage 2 then fixes the learned representations and uses perturbation-type-specific decoders to reconstruct genetic or chemical perturbagen embeddings, such that it can support cross-perturbation retrieval. The paper uses the potential outcomes framework to formalize causal effects from genetic and chemical perturbations, and discusses the multiple-versions-of-treatment problem due to many-to-one treatment representations. The paper further uses natural soft intervention as its causal interpretation for the learned representations by CrispyDrug. Experimental settings include K562 Replogle genetic perturbation data and sci-Plex3 chemical perturbation data. Stage 1 is compared against applicable baselines on held-out perturbations. It also analyzes joint representation and genetic-chemical cross-retrieval for Stage 2.

Critical strengths
  1. The problem is novel and practically meaningful, since CrispyDrug learns joint representations through common cellular outcomes---this is beyond just predicting genetic and chemical perturbations separately. The paper also analyses the multiple-versions-of-treatment problem, and it provides a representation-level causal interpretation via natural soft intervention.

  2. The experimental results for Stage 1 are strong. Under the same held-out perturbation splits, CrispyDrug outperforms chemCPA, PRnet, and BioLORD on the chemical tasks, as well as outperforms CPA and GEARS on the genetic task. The reported figures include 5 random seeds, which is appreciated.

Critical weaknesses
  1. The reported experiments have not demonstrated whether the benefits are from joint training itself.

The paper regards "knowledge sharing between genetic and chemical perturbations" in a joint representation as its main motivation, but Table 2 only compares CrispyDrug with perturbation-type-specific baselines. I believe it lacks at least one important controlled comparison: train the model with the same CrispyDrug architecture with genetic data only or chemical data only, and then compare it with the jointly trained model. Currently, there are many confounders, such as architecture, ZINB objective, perturbation embeddings, or other training choices, that may cause the improvement.

  1. The causal interpretation and empirical validation for Stage 2 are not sufficiently convincing.

Stage 2 is claimed to recover the natural soft intervention mechanism P(S,T∣R=r)P(S, T|R=r), and the paper models it in Eq. (18) using perturbation-type-specific decoders with a fixed-variance Gaussian conditional distribution. Yet, Stage 2 is trained only through perturbagen embedding reconstruction and dose reconstruction with L2 losses, and the paper does not directly evaluate whether these decoders accurately approximate the conditional distribution defined in its theoretical analysis. Also, due to the lack of cross-perturbation retrieval baselines, the same-pathway retrieval reported in Figure 5 only shows meaningful biological evidence, but it is insufficient to establish that CrispyDrug captures stronger cross-perturbation structure than simple alternative retrieval methods.

Questions for authors
  1. Can you provide a direct joint-vs-separate training ablation? For example, train the models using genetic data only and chemical data only, respectively, while keeping the architecture and losses unchanged, and compare them with the jointly trained model on the same held-out perturbations.

  2. Can you explain which evidence shows that the decoders trained with Lrec\mathcal L_{\mathrm{rec}} can be interpreted as an approximation of natural soft intervention? Can you use, for example, random retrieval or nearest-neighbor in the original embedding space as cross-retrieval baselines?

  3. The fused dataset contains genetic and chemical observations drawn from different experimental datasets. Can you clarify why the perturbation type in the fused dataset should be interpreted as randomized treatment assignment, as stated in causal assumption (iii)?

2 comments

Rebuttal [1/2]

We thank reviewer WKdT for their constructive feedback.

Weakness 1: The reported experiments have not demonstrated whether the benefits are from joint training itself. The paper regards "knowledge sharing between genetic and chemical perturbations" in a joint representation as its main motivation, but Table 2 only compares CrispyDrug with perturbation-type-specific baselines. I believe it lacks at least one important controlled comparison: train the model with the same CrispyDrug architecture with genetic data only or chemical data only, and then compare it with the jointly trained model. Currently, there are many confounders, such as architecture, ZINB objective, perturbation embeddings, or other training choices, that may cause the improvement.

Response:

Great suggestion, we thank the reviewer for identifying these important comparisons.

Action To address this concern, we conducted controlled experiments using the same CrispyDrug architecture and the hyperparameters optimized for training on the full joint dataset. We trained CrispyDrug separately on chemical perturbations only and genetic perturbations only, and evaluated each model held-out perturbations of the corresponding type. We also trained CrispyDrug jointly on chemical and genetic perturbations and evaluated it on (i) held-out chemical perturbations, (ii) held-out genetic perturbations, and (iii) held-out perturbations from both types. Furthermore, we also ran an experiment to compare the ZINB vs. the normal objective.

We have added these experiments and their results to Appendix E.2. The results show that despite the clear outperformance of baseline methods in the separate perturbation type tasks - in the joint task CrispyDrug leads in r2 performance. Additionally, our experiment on the ZINB objective does demonstrate that changing to a normal object improves model r2 performance in comparison to ZINB (see Figure 10).


Weakness 2: The causal interpretation and empirical validation for Stage 2 are not sufficiently convincing. Stage 2 is claimed to recover the natural soft intervention mechanism , and the paper models it in Eq. (18) using perturbation-type-specific decoders with a fixed-variance Gaussian conditional distribution. Yet, Stage 2 is trained only through perturbagen embedding reconstruction and dose reconstruction with L2 losses, and the paper does not directly evaluate whether these decoders accurately approximate the conditional distribution defined in its theoretical analysis. Also, due to the lack of cross-perturbation retrieval baselines, the same-pathway retrieval reported in Figure 5 only shows meaningful biological evidence, but it is insufficient to establish that CrispyDrug captures stronger cross-perturbation structure than simple alternative retrieval methods.

Response:

Great point! We thank the reviewer for noting that.

Action We evaluated both the conditional Gaussian decoder and the conditional affine-flow decoder on the same test set of 825 examples (containing both perturbation types). The conditional affine flow achieved a higher mean summed log-likelihood compared with the conditional Gaussian decoder (see table inline). Thus, the flow assigns higher density to the held-out targets under this evaluation. However, we found that the conditional flow decoder performed poorly on the biological task of Stage 2 (see Figure 10)

ModelLog-likelihood (mean ± SE)
Gaussian Probabilistic Decoder2599.203354 ± 6.499594
Conditional Affine-Flow2786.533535 ± 4.499950

Importantly, Stage 2 of our CrispyDrug is generally decoder-agnostic: it conditions a decoder on the Stage 1 representation and trains it to reconstruct the perturbagen representation (and dose). The decoder can therefore be replaced by another model that supports the required conditional prediction, without changing the Stage 1 representation-learning stage. The Gaussian and affine-flow decoders are two examples of this flexibility.

Rebuttal [2/2]

Question 1: Can you provide a direct joint-vs-separate training ablation? For example, train the models using genetic data only and chemical data only, respectively, while keeping the architecture and losses unchanged, and compare them with the jointly trained model on the same held-out perturbations.

Response:

Sure, great idea! .

Action., We have adjusted the benchmark table such that it now includes: Crispydrug trained on only chemical perturbations → evaluated on heldout only chemical perturbations Crispydrug trained on only genetic perturbations → evaluated on heldout only genetic perturbations Crispydrug trained on joint (chemical and genetic) perturbations → evaluated on heldout only chemical perturbations Crispydrug trained on joint (chemical and genetic) perturbations → evaluated on only heldout genetic perturbations Crispydrug trained on joint (chemical and genetic) perturbations → evaluated on heldout joint (chemical and genetic) perturbations These results have helped show the importance of


Question 2: Can you explain which evidence shows that the decoders trained with can be interpreted as an approximation of natural soft intervention? Can you use, for example, random retrieval or nearest-neighbor in the original embedding space as cross-retrieval baselines?

Response:

We thank the reviewer for this suggestion!

Action. To assess whether decoder-based retrieval captures meaningful cross-modal relationships rather than chance or similarity already present in the input embedding space, we added two baselines: (1) random retrieval, which selects a candidate perturbation at random, and (2) nearest-neighbor retrieval in the original perturbation-embedding space, which ranks candidates using their embedding similarity without the learned decoder. We compare these baselines with CrispyDrug using the number of significant pathways overlapping with the drug target on all the data. We report results separately for chemical-to-genetic retrieval. CrispyDrug’s improved retrieval performance over these baselines would indicate that the learned decoder captures cross-modal structure beyond random selection and direct nearest-neighbor matching in the embedding space. We interpret this as evidence consistent with the decoder approximating a natural soft intervention—not as a direct demonstration that the decoder recovers a causal intervention. We have clarified this distinction and added the baseline results to Section 5.


Question 3: The fused dataset contains genetic and chemical observations drawn from different experimental datasets. Can you clarify why the perturbation type in the fused dataset should be interpreted as randomized treatment assignment, as stated in causal assumption (iii)?

Response:

We thank the reviewer for raising this point! Indeed, in general, combining data from separate randomized experiments can only preserve a causal interpretation when the experimental populations are comparable and potential outcomes are not systematically affected by study-specific factors [1], namely, under a cross-study exchangeability assumption. Notably, randomization within each experiment alone does not guarantee this property. Still, our work operates under this cross-study exchangeability assumption: It can be shown mathematically that our assumption (iii) is equivalent to both (a) the within-study randomization and (b) the cross-study exchangeability. Hence, all the fused datasets in our work are treated as good as randomized.

In our experiments specifically, both genetic and chemical perturbation datasets originate from controlled experiments on the same K562 cell line. This shared biological environment supports the plausibility of cross-study exchangeability and, consequently, our randomization assumption (iii) (= we implicitly assume that differences in experimental protocols and batch effects do not systematically affect the potential outcomes).

Action. We have clarified this requirement and acknowledge it as a limitation in the revised manuscript (see a new footnote in Sec. 3 and a paragraph in a new Sec. 6 in the updated PDF).

References:

  • [1] Brantner, Carly Lupton, et al. "Comparison of methods that combine multiple randomized trials to estimate heterogeneous treatment effects." Statistics in medicine 43.7 (2024): 1291-1314.
Reviewer_tWv4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a causal framework for jointly learning representations of genetic and chemical perturbations from single-cell transcriptomic data. The method combines outcome prediction with representation alignment to capture similarities between perturbation types, and provides a potential-outcomes-based causal interpretation through natural soft interventions. The paper is generally well-written and addresses a relevant problem with potential applications in drug discovery and biological understanding. Experiments on two K562 datasets demonstrate improved gene-expression prediction over existing baselines, while latent-space visualizations and cross-modal retrieval analyses provide preliminary evidence of biologically meaningful representations. However, the experimental validation remains somewhat limited, particularly in demonstrating the benefits of joint learning and the reliability of cross-modal retrieval. Overall, the work presents a promising direction, although stronger empirical evidence is needed to fully support its main claims.

Critical strengths
  1. The paper addresses a relevant and challenging problem: jointly modeling genetic and chemical perturbations to identify similarities in their cellular effects. The proposed architecture is relatively straightforward and interpretable, combining modality-specific encoders, a shared predictive objective, and representation alignment.

  2. The experimental setup is clearly explained and includes two large-scale datasets covering both genetic and chemical perturbations, with comparisons against several state-of-the-art methods. The results show improvements in gene-expression prediction over existing baselines for both perturbation types, suggesting that the proposed approach can effectively model cellular responses.

Critical weaknesses
  1. While the method encourages genetic and chemical perturbations to occupy a shared latent space, it remains unclear whether this alignment captures biologically meaningful relationships beyond improving predictive performance. The authors provide PCA visualizations and pathway-based retrieval experiments, but these do not convincingly establish the utility or interpretability of the representations. Stronger quantitative validation, including comparisons against random retrieval and biologically informed baselines, would help demonstrate the practical value of the proposed approach.

  2. Although the proposed method improves MAE and MSE over existing baselines, the baseline models already achieve very high correlation and (R^2) scores, making the practical significance of these improvements unclear. In particular, improvements in overall gene-expression prediction do not necessarily translate into better identification of biologically meaningful perturbation effects. Additional evaluations focusing on differentially expressed genes, perturbation-induced expression changes, or downstream biological tasks would help establish whether the reported improvements provide meaningful practical benefits.

  3. Related to (2), recent studies (Viñas Torné et al., 2026; Ahlmann-Eltze et al., 2025) have shown that simple statistical or linear approaches can achieve comparable or even superior performance to complex deep-learning models in perturbation-response prediction. Given these findings, the absence of simple baselines (e.g., control-mean, perturbed-mean, or linear predictors) makes it difficult to assess whether CrispyDrug learns meaningful perturbation-specific effects or primarily captures general expression patterns. This is particularly relevant given the already very high correlation and (R^2) scores reported for existing methods.

Questions for authors
  1. Have the authors considered evaluating whether the improvements in gene-expression prediction translate into biologically meaningful outcomes, such as more accurate identification of differentially expressed genes or improved prediction of perturbation-induced transcriptional changes?

  2. Is the choice of architecture and the low-dimensional shared latent space motivated by the goal of uncovering interpretable biological factors? If so, have you considered evaluating whether individual latent dimensions or directions correspond to meaningful biological processes, pathways, or mechanisms of action, beyond the observed alignment between perturbation types?

1 comment

We thank Reviewer tWv4 for their constructive feedback.

Weakness 1: Although the proposed method improves MAE and MSE over existing baselines, the baseline models already achieve very high correlation and (R^2) scores, making the practical significance of these improvements unclear. In particular, improvements in overall gene-expression prediction do not necessarily translate into better identification of biologically meaningful perturbation effects. Additional evaluations focusing on differentially expressed genes, perturbation-induced expression changes, or downstream biological tasks would help establish whether the reported improvements provide meaningful practical benefits.

Response:

We thank the reviewer for highlighting the need to assess whether improvements in overall gene-expression prediction translate into more biologically meaningful perturbation predictions.

Action. we added differential-expression-based evaluation metrics, following recent work: cosine similarity of log-fold changes, cosine similarity of ranked expression changes, differential-expression (DE) overlap, differentially expressed gene (DEG) recall, and weighted MSE. These metrics provide a more targeted assessment of perturbation-induced expression changes and the recovery of biologically relevant genes. We have added the corresponding results to the revised manuscript in Appendix C.


Weakness 2: Related to (2), recent studies (Viñas Torné et al., 2026; Ahlmann-Eltze et al., 2025) have shown that simple statistical or linear approaches can achieve comparable or even superior performance to complex deep-learning models in perturbation-response prediction. Given these findings, the absence of simple baselines (e.g., control-mean, perturbed-mean, or linear predictors) makes it difficult to assess whether CrispyDrug learns meaningful perturbation-specific effects or primarily captures general expression patterns. This is particularly relevant given the already very high correlation and (R^2) scores reported for existing methods.

Response:

Great suggestion!

Action. We have adjusted the benchmark table accordingly, so it now includes the new suggested baselines Perturb-mean: essentially the pseudobulked expressions Linear model: we have used the same parameters as those appearing in the original publication (Ahlmann-Eltze et al., 2025) Additionally, we sought to further detail the evaluations for our method in phase 1, therefore, we also add control-mean (Viñas Torné et al., 2026) for evaluating the results based on the met With these new baselines, it is more visible that the added benefit of Crispydrug’s joint training.


Question 1: Have the authors considered evaluating whether the improvements in gene-expression prediction translate into biologically meaningful outcomes, such as more accurate identification of differentially expressed genes or improved prediction of perturbation-induced transcriptional changes?

Response:

Yes. To assess whether improved gene-expression prediction translates into biologically meaningful outcomes, we added differential-expression–related evaluation metrics. These results show that both the biological metrics and the machine-learning metrics indicate an additional benefit from CrispyDrug’s joint training compared with the baseline. In terms of DEG recall (Wu et al., 2025) and weighted MSE(Meiji et al., 2025), CrispyDrug outperforms all baselines, we have added these results in Appendix C. Moreover, we also found that even after adding multiple cell lines we CrispyDrug still preserves good performance on both types of metrics (see Appendix E.2)


Question 2: Is the choice of architecture and the low-dimensional shared latent space motivated by the goal of uncovering interpretable biological factors? If so, have you considered evaluating whether individual latent dimensions or directions correspond to meaningful biological processes, pathways, or mechanisms of action, beyond the observed alignment between perturbation types?

Response:

We thank the reviewer for raising this important question. The low-dimensional shared latent space is primarily designed to align perturbation representations across modalities and learn a representation that is predictive of post-treatment outcome, rather than to guarantee that each individual latent dimension corresponds to a discrete biological factor. Accordingly, the observed alignment between perturbation types should not, by itself, be interpreted as evidence of latent biological interpretability.

Reviewer_w8svOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors introduce a framework for learning joint representations of two different types of perturbations - genetic and chemical - called "CrispyDrug", based on single-cell transcriptomic data. To learn the joint representations, CrispyDrug treats both types of perturbations as interventions on a shared cellular outcome reflected in the single-cell transcriptomic data. This is achieved by learning the perturbation representations by predicting the gene expression changes resulting from the interventions, and then by training perturbation-type-specific decoders that can map the joint representations to either the genetic or chemical perturbations.

Critical strengths

The idea of learning joint representations of different types of perturbations is interesting and timely. There have been, for example, increasing interests in learning joint "unified" representations (e.g., for cell types) from multi-modal and/or multi-omics data, and the proposed work can extend this further for two (or possibly more) perturbation types. Aiming at causal understanding of post-perturbation changes reflected in single-cell transcriptomics might be beneficial in unveiling the underlying mechanisms of changes induced by different types of perturbations, especially, by investigating their similarities and differences. The overall approach and model architecture appear to be well-motivated and reasonable as well.

Critical weaknesses

Despite the interesting idea and reasonable approach proposed in this study, there are unfortunately several major concerns.

  1. Most importantly, it is unclear whether informative and effective joint representations can be learned from the datasets the authors considered in the current work based on the proposed CrispyDrug framework.

For example, the genetic and chemical perturbation data come from different types of experiments, batches, conditions, protocols, etc.. As a result, if the perturbation representations are aligned based on the post-perturbation gene expression, these representations may fit various confounding factors (e.g., batch effects) without reflecting the real biology underlying the changes that arise from a given intervention.

  1. It is mentioned (line 159) that X is the "average expression of the unexposed cells (namely, cells under a baseline condition) from the same parent cell line", which makes one wonder about its value and informativeness as a covariate for the datasets considered in this study. For example, how can one derive any context-dependence claims when the context is constant? This should be clarified.

  2. Although aiming at a causal framework for perturbations and post-perturbation expression is commendable, the method presented in "Causal prediction target", "Causal assumptions", and "Representation-level effects & multiple-versions-of-treatment problem" do not appear to actually lead to interpretable mechanistic understanding of perturbation-induced expression changes. As a result, it is unclear how the proposed CrisyDrug framework can actually add to improving causal understanding.

  3. The results presented in the current manuscript do not convincingly demonstrate (1) whether CrispyDrug effectively learns joint representations of chemical and genetic perturbations, and (2) whether it can decode the representations into genetic or chemical perturbations accurately.

Questions for authors

Please see the comments above in "critical weaknesses". Pleae provide clarifications and further evidence if possible.

7 comments

Rebuttal [1/2]

We thank Reviewer 1 for the careful assessment and constructive suggestions.

Weakness 1: It is unclear whether informative and effective joint representations can be learned from the datasets the authors considered in the current work based on the proposed CrispyDrug framework. For example, the genetic and chemical perturbation data come from different types of experiments, batches, conditions, protocols, etc.. As a result, if the perturbation representations are aligned based on the post-perturbation gene expression, these representations may fit various confounding factors (e.g., batch effects) without reflecting the real biology underlying the changes that arise from a given intervention.

Response:

We agree that differences in experimental design, protocol, and batch between genetic and chemical perturbation datasets could contribute to the learned representations. Importantly, our CrispyDrug does not explicitly model batch identity or enforce batch-invariant latents: In our causal setting, the perturbagen-conditioned latent space alone cannot establish that such technical structure has been removed; and, thus, the need to perform an explicit batch correction is orthogonal to our approach. We therefore interpret the learned representations as predictive representations of perturbation-associated expression, rather than as batch-free or purely biological representations.

Still, our framework can incorporate several measures that may reduce specific sources of batch effect. Specifically, control observations use a zero perturbation embedding, providing a shared control input representation; library size is modeled with an auxiliary prediction loss, which can account for sequencing-depth variation; and recorded covariates can be supplied as model inputs so that comparisons can be made conditional on those covariates. In addition, perturbation-level held-out evaluation tests transfer to unseen perturbations.

Action. We have clarified that, in our CrispyDrug, we do not aim to address unmeasured or confounded batch effects, and our method should not be interpreted as an explicit batch correction (see a new Sec. 6 in the updated PDF).


Weakness 2: It is mentioned (line 159) that X is the "average expression of the unexposed cells (namely, cells under a baseline condition) from the same parent cell line", which makes one wonder about its value and informativeness as a covariate for the datasets considered in this study. For example, how can one derive any context-dependence claims when the context is constant? This should be clarified.

Response:

We thank the reviewer for pointing out this potential ambiguity. Indeed, in our current experiments, the baseline expression (X) is computed from unexposed cells of the corresponding parent cell line. Thus, it is constant across perturbations within a given cell line, and our current results do not establish context-dependent responses within that cell line. However, our framework is more general and can be applied across multiple cell lines or cell types (whose baseline expression profiles can differ, enabling perturbations to be represented in their respective cellular contexts).

Action. We have clarified this distinction in the text Section 5 (Experiments)Also, as a remedy, we conducted new experiments with multiple cell lines (A549, MCF7) (See Appendix E.2 in the manuscript, Tables 7 & 8). Therein,we find that CrispyDrug is consistently performing better than baselines on the joint task on all metrics and that the expression (X) while not useful at one cell line only can help in the general format where more cell lines are introduced.

Rebuttal [2/2]

Weakness 3: Although aiming at a causal framework for perturbations and post-perturbation expression is commendable, the method presented in "Causal prediction target", "Causal assumptions", and "Representation-level effects & multiple-versions-of-treatment problem" do not appear to actually lead to interpretable mechanistic understanding of perturbation-induced expression changes. As a result, it is unclear how the proposed CrisyDrug framework can actually add to improving causal understanding.

Response:

We thank the reviewer for raising this important distinction. We would like to clarify that the term causal encompasses several research directions that differ in their objectives and interpretations. For example, causal discovery aims to recover causal relationships among variables (e.g., gene regulatory networks), whereas causal effect estimation aims to determine how an outcome would change under an intervention (e.g., [1, 2, 3]). Our work follows the latter tradition of causal machine learning, where machine learning methods are used to estimate potential outcomes and treatment effects. Specifically, CrispyDrug models the effects of genetic and chemical interventions on post-treatment gene expression. We do not aim to recover gene regulatory graphs or identify the molecular mechanisms mediating these effects. Rather, we model the intervention-to-outcome relationship (S,T)→Y(S,T)\rightarrow Y (conditional on XX).

Action. We have added the explanation above to a new Sec. 6 “Outline” (see our updated PDF).


Question 1: The results presented in the current manuscript do not convincingly demonstrate (1) whether CrispyDrug effectively learns joint representations of chemical and genetic perturbations, and (2) whether it can decode the representations into genetic or chemical perturbations accurately. While the method encourages genetic and chemical perturbations to occupy a shared latent space, it remains unclear whether this alignment captures biologically meaningful relationships beyond improving predictive performance. The authors provide PCA visualizations and pathway-based retrieval experiments, but these do not convincingly establish the utility or interpretability of the representations. Stronger quantitative validation, including comparisons against random retrieval and biologically informed baselines, would help demonstrate the practical value of the proposed approach.

Response:

Action We added quantitative cross-modal retrieval comparisons to evaluate whether the learned representations support decoding across perturbation types. Specifically, for chemical-to-genetic retrieval, stage 2 aims at identifying genes whose perturbation can be similar to that of drugs’. We have clarified that definition in Section 5. Additionally, we compare CrispyDrug with (1) random retrieval and (2) nearest-neighbor retrieval in the original perturbation-embedding space. We evaluate retrieval by the number of significant pathways overlapping with the retrieved drug target’s associated pathways and report the results in Section 5. These comparisons test whether decoder-based retrieval identifies biologically relevant candidates. We interpret an advantage over these baselines as evidence that CrispyDrug captures useful cross-modal structure and can decode representations for this retrieval task—not as proof that individual latent dimensions are biologically interpretable or that the decoder recovers a causal intervention mechanism. We have clarified this distinction in the revised manuscript.

References:

  • [1] Donald B. Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5):688, 1974.
  • [2] Guido W. Imbens and Donald B. Rubin. Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press, New York, 2015. ISBN 9780521885881.
  • [3] Chernozhukov, V., Hansen, C., Kallus, N., Spindler, M., & Syrgkanis, V. (2024). Applied causal inference powered by ML and AI. arXiv preprint arXiv:2403.02467, 7.

Thank you for the response.

W1: Please note that the concern was not only regarding batch effects but various confounding factors, which include batch effects as well. W2: It is still unclear how the framework's potential "context-dependency" benefits the outcomes when applied across multiple cell lines or cell types as claimed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.