acceptodds
Withdrawn from ICLR 2027

RePert: Effect-Grounded Identity Learning for Generalizable Inverse Perturbation Retrieval

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inverse perturbation retrieval seeks to identify the intervention responsible for an observed cellular response, yet current retrievers generalize poorly beyond the cellular settings seen during training. Two barriers underlie this failure: (i) response distributions for the same perturbation can differ substantially across cell lines, so patterns learned in one cell line may not transfer to another. (ii) perturbation identity embeddings learned from training cell lines lack direct functional evidence and transfer poorly to unseen ones. To address these barriers, we introduce RePert, a novel Response-to-Perturbation framework with functionally-grounded two-stage retrieval. RePert introduces two key innovations: (i) Effect encoder pretraining learns transferable effect representations from multiple responses sharing the same perturbation or target. (ii) Effect-grounded identity retrieval builds a non-parametric effect memory and uses its prototypes to ground identity embeddings. RePert combines identity and effect matching views through a reliability-calibrated score. We evaluate RePert in six genetic and chemical retrieval settings spanning in-domain evaluation, held-out cell-line generalization, and zero-shot cross-platform transfer to external single-cell platforms. For gene-level retrieval, RePert improves recall@10 over strongest baseline by factors of , , and in single-gene, multiple-gene, and held-out cell-line retrieval, respectively. For compound-level cross-platform transfer without target-platform training, it improves top-1% macro recall against strongest baseline by factors of , , and on Cellarity-Cancer, Cellarity-Primary, and sciPlex3, respectively. These results establish transferable effect learning and identity grounding as a robust and promising design for generalizable inverse perturbation retrieval. Code is available at https://anonymous.4open.science/r/RePert-DEC3/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Single Cell, Perturbation Prediction

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_EYDSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This study introduces RePert, a framework for inverse perturbation retrieval from transcriptional responses. More specifically, the goal is to generalise across cellular contexts and experimental platforms.

RePert combines contrastive pretraining of an effect encoder with a non-parametric memory of perturbation-specific effect prototypes, and a discriminative retrieval model whose identity embeddings are aligned with these prototypes. Then, at inference, identity-based and effect-based retrieval scores are combined using a validation-selected weighting parameter.

Evaluation is performed across six settings, reporting improvements over existing approaches.

Critical strengths

RePert addresses a relevant research direction: inverse perturbation retrieval is an interesting, contemporary problem with enormous potential in drug discovery. The experimental framework is well-designed and covers a satisfactory range of tasks and settings. I particularly valued the cross-platform experiment (although it would be good to have statistical tests on the tables' results and pay more attention to numerical precision and significant figures). Code is released and matches the architecture described in the paper. There are no preprocessing steps or splits available, so proper reproducibility of results cannot be established. RePert has promising empirical performance compared to other models. Most interestingly, against DrugReflector.
The framework is composed of a relatively simple and efficient architecture, providing a potentially valid alternative to heftier solutions.

Critical weaknesses

The Manuscript

The manuscript is -- in my opinion -- generally poorly structured. The relatively simple framework is unnecessarily complicated by the unintuitive presentation. The different experimental settings are mentioned as C1-6 on page 3, while Table 1 is introduced on page 5 and located on page 6. Table 4, top1% gain should be ~1.2 if I am not wrong. Additionally, citations should be in parentheses unless the reference itself is part of the sentence. Arrows on Figure 1 are badly aligned, and "our proposed" sounds like incomplete labelling. I struggle to believe that anyone reviewing this paper internally would not notice how hard it is to follow.

Scientific Claims

I think that baseline comparisons would benefit from further clarification. In C1-C3, several baselines were originally designed for a different task (forward perturbation prediction), and it is not clear to me whether any attempt to adapt them has been performed or not. Although these settings seem to follow those from PDGrapher, I find the comparison principally flawed.

The contribution of the individual components could be explored further. It would be particularly informative to extend the ablation analysis to the held-out cell-line and cross-platform settings, where the proposed effect representations and identity grounding are expected to be most useful. I would also appreciate clarification of the term "reliability-calibrated", as the current procedure appears to select a global combination weight based on validation performance, rather than explicitly modelling the reliability of individual predictions or effect prototypes.

Questions for authors

Q1. Novelty and simpler baselines. Could the authors provide a more detailed discussion of the relationship between RePert and existing supervised contrastive or prototype-based retrieval methods? In particular, it would be interesting to compare against a supervised contrastive encoder with nearest-prototype retrieval and a straightforward MLP retriever trained with the same candidate-level objective. I believe this could help isolate the benefits of the proposed effect memory and identity-grounding mechanisms.

Q2. Baseline adaptation and experimental fairness. Could the authors clarify how the forward-model baselines in C1–C3 were adapted to inverse retrieval? Were these models directly optimised for retrieval, or were their predicted responses used for candidate matching?

Q3. Component ablations under distribution shifts. Could the authors extend the component ablations to C3 and C5–C6, including independently trained identity-only models? Given that the main motivation for effect pretraining and grounding is improved generalisation, these experiments could help establish where the additional components provide the greatest benefit. A comparison with joint end-to-end training would also be interesting.

Reviewer_kng1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents RePert, a framework for inverse perturbation retrieval that aims to identify genetic or chemical perturbations from observed transcriptional responses. The method first uses contrastive learning to obtain effect representations from responses sharing the same perturbation or annotated target. It then builds a prototype-based effect memory and aligns learned perturbation identity embeddings with these prototypes. The final retrieval score combines identity-based and effect-based similarities. The authors evaluate the method on six tasks covering gene and compound retrieval, held-out cell lines, and cross-platform transfer from LINCS L1000 to single-cell datasets. RePert achieves consistent improvements over the reported baselines, particularly for gene-level retrieval and cross-cell-line generalization.

Critical strengths

S1. The paper addresses an important problem in perturbation modeling. Compared with perturbation prediction, inverse retrieval has received relatively less attention, particularly when transferring across cellular contexts. The authors correctly identify cell line dependent transcriptional responses as a major challenge. The evaluation also includes heldout cell lines and independent single-cell platforms rather than relying entirely on random train/test splits.

S2. The reported improvements are substantial in several settings. For example, recall@10 increases from 0.0719 to 0.1325 in the heldout cell line experiment, and the model also improves compound retrieval on external single-cell datasets without target platform finetuning. The authors provide component ablations and comparisons between the identity and effect branches, which help explain where some of the improvements come from. The method is also relatively straightforward to implement, making it potentially useful for large-scale perturbation retrieval.

Critical weaknesses

W1. My main concern is whether the proposed components are really responsible for the reported improvements. The method combines contrastive representation learning, class prototypes, an embedding alignment loss, and two retrieval scores. These are all fairly established techniques, and the novelty seems to lie mainly in their combination. More importantly, the ablation results suggest that the additional components contribute relatively modest improvements over the identity only model. In Table 5, recall@10 on C1 increases from 0.0603 to 0.0790, while on C4 it actually decreases slightly from 0.4895 to 0.4838. Yet both variants substantially outperform the external baselines. I am not sure whether most of the improvement comes from the underlying retrieval model rather than the proposed effect-grounded learning. I would also like to see a simple baseline that averages the training expression signatures for each perturbation and retrieves candidates using cosine similarity or Pearson correlation. Since RePert explicitly constructs an effect memory from training responses, this would be a natural comparison to establish the added value of the learned representations and functional grounding.

W2. I am not fully convinced that the experiments establish the claimed generalizability of perturbation effects. The heldout cell line experiment is useful, but the model still retrieves perturbations or targets that have already been observed during training. Similarly, the cross-platform experiments use the same compound vocabulary as LINCS rather than testing unseen compounds. This distinction matters because recognizing a previously observed perturbation under a different cellular context is not necessarily equivalent to learning a generalizable representation of its biological effects. The paper also does not sufficiently examine how the number and quality of training responses per perturbation affect retrieval performance. Since the effect memory is constructed directly from these responses, frequently observed perturbations may have more reliable prototypes. In particular, I would like to know whether the improvements persist for perturbations with weak or poorly reproducible transcriptional responses. An additional experiment about perturbation strength, replicate consistency, and the number of available training signatures would help determine whether the method captures transferable biological information or mainly benefits from repeated observations of the same perturbation.

Questions for authors

Q1. In Table 5, the identity only model already substantially outperforms the external baselines. Can the authors explain what accounts for this gap? How would RePert compare with a simple baseline based on the mean training expression signature of each perturbation?

Q2. How sensitive is RePert to the number and reproducibility of training signatures for each perturbation? Do the improvements remain consistent for perturbations with weak transcriptional responses?

Q3. In Table 10, the effect-only branch performs substantially worse than the identity-only branch on the external single-cell datasets. Does this suggest that effect grounding mainly serves as a regularizer for identity learning rather than providing directly transferable effect representations?

Reviewer_fDh9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RePert, a framework for identifying genetic or chemical perturbations from their observed transcriptional responses. The central idea is to learn a representation of perturbation effects that is less sensitive to cellular context, and then use this representation to ground perturbation identities during retrieval. The method combines a learned identity-based score with similarity to prototypes constructed from observed responses.

The authors evaluate the approach on genetic and chemical perturbation datasets, including held-out cell lines and transfer from L1000 to single-cell transcriptomic measurements. The reported improvements over existing methods are encouraging, particularly under changes in cellular context. However, I am not yet convinced that the experiments establish the central methodological contribution, or that the formulation captures the settings in which inverse perturbation retrieval would be most scientifically useful.

Critical strengths

The problem is interesting and timely. As perturbational atlases grow, being able to connect an observed cellular response to plausible interventions or molecular targets could become an important capability. I appreciate the authors' emphasis on generalization across cellular contexts rather than treating performance on randomly held-out responses as sufficient evidence of progress.

The proposed architecture is also appealing in its simplicity. Building a memory of observed perturbation effects is intuitive, computationally inexpensive, and potentially useful beyond the particular retrieval setting studied here. The distinction between learning perturbation effects and learning their identities is a reasonable organizing principle.

The empirical study is reasonably broad, and I particularly appreciate the experiments on unseen cell lines and different transcriptomic platforms. The results on primary cells are interesting, and the consistent improvements across individual cell lines suggest that the method has learned something useful. The representation analysis also provides some evidence that effect pretraining reduces cell-line-dependent similarity. These are worthwhile observations, even if their implications for the full retrieval framework are less clear.

Critical weaknesses

My main concern is that the experiments do not clearly establish what is gained from the proposed effect-grounding framework.

The improvements over existing methods initially appear substantial. However, the ablation results tell a more complicated story. A relatively straightforward identity-based retriever, without the proposed effect pretraining or grounding, already accounts for much of the improvement. On compound retrieval, this simpler model actually slightly outperforms the full method at recall@10, although the full method performs better at larger retrieval cutoffs. On held-out cell lines, the additional gain from incorporating effect information is relatively modest.

This does not mean that the proposed components are ineffective. There is evidence that they improve some metrics, and the representation analysis is interesting. But I would have expected a clearer demonstration that effect grounding is responsible for the claimed generalization, particularly since that is the principal motivation of the paper. The results leave open whether the more consequential improvement comes from the discriminative retrieval formulation, the input normalization, or the proposed representation-learning components.

I would encourage the authors to distinguish these contributions more carefully. In particular, comparing the full method against a well-tuned identity-only model under exactly the same training and evaluation conditions would help establish how much of the improvement is attributable to the central idea.

I am also surprised by the absence of a stronger comparison to classical transcriptional signature matching.

There is a long history of using perturbation-induced gene-expression signatures to identify related interventions, most notably through the Connectivity Map. The proposed effect memory is, in some sense, a learned consensus representation of such signatures. This makes direct comparison with consensus signatures in the original expression space particularly important.

A simple baseline that averages training responses for each perturbation and retrieves candidates using cosine similarity or a connectivity-style score would provide a useful point of reference. One could also ask whether learning a representation before averaging improves retrieval relative to a simpler supervised encoder.

This matters especially because several of the gene-level baselines are primarily designed for forward perturbation prediction rather than inverse retrieval. I appreciate that the authors include PDGrapher and retrain DrugReflector for the compound experiments, but these comparisons do not fully answer whether the proposed approach improves on the most natural form of response-based matching.

Without this comparison, I find it difficult to assess how much methodological progress is being made beyond learning a discriminative embedding for a familiar signature-matching problem.

A broader question concerns the intended scientific application of inverse perturbation retrieval.

The experiments largely assume that the correct perturbation is already represented in the training vocabulary. This is a well-defined retrieval problem, and I can see its value in matching responses against a reference library. However, the paper motivates the method in terms of biological discovery and generalization, where the more interesting cases often involve interventions or mechanisms not previously characterized.

For example, suppose one observes the transcriptional response to a newly discovered compound and wishes to infer its molecular targets. This seems like a compelling application of inverse retrieval. Interestingly, the target-level experiments in this paper could potentially support precisely this use case: the model could retrieve known target genes from the response of a previously unseen compound. Yet the reported evaluation does not isolate this setting by holding out compounds during training.

I would be particularly interested in this experiment. It would make the biological relevance of the task much more concrete and help distinguish generalization to new cellular environments from generalization to new perturbations.

The distinction is important for interpreting the cross-platform experiments as well. These demonstrate transfer to new measurement platforms while retaining the known compound vocabulary, which is valuable, but considerably narrower than generalizing to previously unseen interventions.

Relatedly, the candidate representations do not explicitly incorporate chemical structure, molecular targets, or other biological relationships. This is a reasonable design choice for closed-set retrieval, but it limits the ability to reason about new perturbations or related compounds. It also raises an interesting question about what the learned space captures. Are compounds with similar mechanisms placed near one another, or does the discriminative objective primarily encourage separation of individual identities?

The current evaluation regards retrieval of a different compound as an error, even when that compound may act through the same pathway. I would find it informative to examine whether the highest-ranked incorrect predictions are biologically related to the true perturbation. This could reveal more about the representation than exact-identity recall alone.

Another concern is whether the model is learning meaningful perturbation effects in cases where the measured response is weak or absent.

Not every perturbation produces a substantial transcriptional response. Some compounds are inactive under particular conditions, and some genetic perturbations have little measurable effect on the assayed transcriptome. In such cases, there may be insufficient information in the expression profile to identify the intervention reliably.

Nevertheless, the present formulation always ranks the available candidates and does not explicitly distinguish a meaningful perturbation response from a null or uninformative one. I would like to understand how the method behaves on matched controls, non-targeting guides, or compounds producing very weak signatures.

This is more than a question of adding an abstention mechanism. It also bears on what is learned during training. If nearly indistinguishable responses are assigned different perturbation identities, the model may exploit technical or experimental correlations rather than reproducible biology.

It would be useful to stratify retrieval performance by perturbation strength and to evaluate the false-assignment rate on null responses. I would also like clarification on the independence of the training and evaluation splits, particularly whether signatures from the same experimental plate or closely related replicates can occur in both. These analyses would help establish when the learned representations are biologically informative rather than merely discriminative.

Finally, I think the evidence for cross-platform generalization deserves a more careful interpretation.

The external single-cell experiments are a useful addition, and the improvements on primary cells are encouraging. At the same time, these evaluations involve relatively few independent perturbation conditions compared with the large LINCS training dataset, and some of the reported advantages are small relative to their variability.

An important missing detail is how transcriptional responses from the different assay platforms are made comparable. The construction of pseudobulk profiles, normalization, selection of landmark genes, and matching of treatment and control conditions could all materially affect retrieval performance. These choices deserve a clearer description, especially when cross-platform transfer is one of the central claims.

I would also avoid describing the combination of identity and effect scores as reliability-calibrated without a more direct analysis of reliability. As presented, the procedure selects a global weighting coefficient using validation performance. This is a sensible practical choice, but does not itself establish that the model recognizes when a prediction or an effect prototype is unreliable.

Questions for authors

Questions:

  1. I would be interested in a more careful analysis of what drives the improvements. How much is gained by effect pretraining and grounding beyond a well-tuned discriminative retriever, particularly on held-out cell lines? Could the authors also compare against a simple consensus-signature retrieval baseline?

  2. What application do the authors ultimately envision for this framework? In particular, could they evaluate target retrieval on compounds entirely excluded from training? This seems like a natural and scientifically meaningful test of the proposed approach.

  3. What does the model learn about biological relationships between perturbations? Are highly ranked incorrect compounds enriched for shared targets or mechanisms, and does the contrastive training objective preserve these relationships?

  4. How does retrieval performance depend on the strength of the transcriptional response? Can the method distinguish informative responses from controls or near-null signatures, and how sensitive are the results to the experimental splitting and cross-platform preprocessing procedures?

Overall assessment

I appreciate the authors' effort to address a genuinely interesting problem, and there is merit in the simplicity of the proposed approach and the breadth of the evaluation. The experiments suggest that discriminative retrieval of perturbation identities can work across different cellular contexts, and the effect-representation analysis offers some useful observations.

My reservation is that the paper does not yet convincingly establish its central claim: that learning transferable perturbation effects and grounding identities in those effects provides a substantial advance over simpler retrieval approaches. The strongest results are difficult to separate from the performance of the identity-only baseline, and the absence of direct signature-matching comparisons makes the contribution harder to judge. More fundamentally, the evaluation remains centered on identifying known interventions, without demonstrating the more compelling biological uses of inverse retrieval.

I think there is a potentially useful idea here, but the current evidence does not sufficiently establish either its methodological necessity or its broader scientific value. For these reasons, I cannot recommend acceptance in its present form.

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

Related papers

Discussion (0)

Sign in to comment.