acceptodds
Withdrawn from ICLR 2027

RELAY: Transferring Dense Response Contrasts to Distinguish Unseen Perturbations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predicting the effects of unseen perturbations requires capturing gene-level response tendencies while distinguishing the effects of different perturbations on the same readout gene. On PerturbQA, a gene prior (each readout gene's majority training label) attains Macro-F1 of 0.8814 for differential expression (DE), and every language-model and expression-based predictor we evaluate falls below it. Input ablations identify retrieved outcomes as the source of a retrieval LLM's within-gene signal; descriptions of biological mechanisms contribute little. Yet labels cover only about 1% of measured perturbation–gene pairs. We introduce RELAY, which transfers dense measured responses without a language model. Kernel ridge regression over a biological knowledge kernel transfers responses from related training perturbations. Task-specific readouts then combine each gene's tendency with the transferred contrast to predict DE and direction of change (DIR), retaining the gene prior's global accuracy while distinguishing perturbations that act on the same readout gene. Across four PerturbQA cell lines, RELAY Pareto-dominates every evaluated baseline on both global and within-gene metrics. It reaches DE and DIR Macro-F1 of 0.8941 and 0.8722 and improves within-gene balanced accuracy over the strongest baseline by 7.48 and 6.53 points. RELAY also extends to a held-out cell line and to unseen compounds, retaining its advantage over the competing methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cellular perturbation prediction, Perturb-seq, large language models, biological reasoning, model evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_aLtbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RELAY does not rely on an LLM. Instead, it performs kernel ridge regression over a knowledge-graph-derived kernel, transferring dense measured responses from related training perturbations to unseen ones to make DE and DIR predictions.

Critical strengths
  • This paper is easy to read and follow

  • The finding is compelling: a simple gene-majority-label prior outperforms all evaluated LLMs and expression-based models on DE prediction.

Critical weaknesses

1.The comparison with existing perturbation-prediction methods is not sufficiently comprehensive. Recent baselines such as CellFlow, PerturbDiff, and scDFM should be included.

2.No code is provided, which limits reproducibility.

3.The gain in global accuracy stems primarily from DIR, whereas on DE, RELAY only marginally outperforms the simple prior. In the compound setting—which is closer to real-world applications—its global metrics actually fall below the prior, with modest within-gene improvements of only ~4.6 points (DE) and ~5.7 points (DIR).

4.Information asymmetry in model inputs: Both the dense responses rr and effect statistics (F/SEF/\text{SE}) used by RELAY are derived from the exact same screen as the DE tests used to generate the ground-truth labels. Specifically, the standardized effect E=log⁡(1+∣F∣/SE)E = \log(1 + \vert{}F\vert{}/\text{SE}) serves as a continuous proxy for the discrete DE label. By contrast, the LLM baselines only have access to roughly 1% of the discrete labels. Consequently, the supervision signal available to RELAY and the baselines is not comparable, making the comparison unfair.

Questions for authors

see weakness

Reviewer_idcAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies perturbation prediction response. The authors propose RELAY which uses a biological knowledge kernel to identify related perturbations and to transfer the measured response contrasts from training perturbations to unseen ones. Unlike prior methods where the main objective was continuous gene expressions, in this work the authors focus on DE/DIR classification. The primary task used in this work is genetic perturbation on perturbQA dataset. RELAY improves both global and within-gene metrics on PerturbQA and is additionally evaluated on a held-out cell line and unseen compounds.

Critical strengths
  • The observation that a simple gene prior achieves very high Macro-F1 while having no perturbation specific predictive ability is interesting.
  • The proposed method is relatively simple and well motivated.
Critical weaknesses
  • The methodological contribution is incremental.
  • The paper is confusing in a sense that it sometimes gives the impression that biological prior enables prediction of perturbation effects, while the experiments suggest a more limited conclusion. -The main tasks are DE and direction classification while than quantitative prediction of perturbation-induced expression changes is also important and has been studied in prior works.
Questions for authors
  • How sensitive are the unseen-compound results to the availability and quality of known drug targets?
  • The paper shows relatively weak correlation between knowledge similarity and response similarity. Could the authors explain where the knowledge kernel actually works?
Reviewer_7PAtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As I have explained bellow, the paper is misstargeted: it clearly does not belong in an ML conference but would be better suited for a audience specialized in bioinformatics or genetics. The paper makes little to no efforts to relate its findings to topics of interest to the ML community, does not introduce or explain its core concepts or elements, and is in my opinion heavily generated using AI. I strongly advise to reject this paper.

Critical strengths

I cannot assess the strenghts of this paper. The experiments appear to show good performance of the model over chosen metrics, but due to the poor writing of the paper and its lack of introduction or explanation, I am unable to assess the soundness of the paper's claims.

Critical weaknesses

I believe the introduction misses the point of introducting the reader to the questions and terms the paper is adressing. The terms gene prior, PertubQA, majority training label, readout gene are all left undefined. Furthermore, I do not believe that the introduction is the right place to introduce metrics.

The introduction further features a long bloc of text "Kernel ridge regression over a knowledge profile of curated annotations [...] how strongly the gene usually responds" which is, in my opinion, absolutely unintellegible. The statement of contributions is also quite confusing. Sections 2 and 3 are equally confusing.

I do not believe that this paper is suited for publication at ICLR, but should be retargeted at a bioinformatics or genetics journal or conference.

Questions for authors

None.

Reviewer_Ubj7OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RELAY, which predicts differential expression (DE) and direction of change (DIR) for unseen perturbations. RELAY uses kernel ridge regression over biological knowledge profiles to transfer measured response contrasts, followed by task-specific readouts that combine these predictions with gene-level response tendencies. The paper also investigates why existing predictors fall short in either global accuracy or within-gene discrimination. Evaluation covers four PerturbQA cell lines, a separate transfer setting with RPE1 held out, and unseen compounds. RELAY improves both global and within-gene performance on DE and DIR when averaged across the PerturbQA cell lines, with smaller within-gene gains on compounds.

Critical strengths
  1. The evaluation shows that strong aggregate performance can arise from gene-level tendencies without distinguishing the effects of different perturbations on the same gene. This distinction makes the paper valuable beyond the proposed method by clarifying what predictive performance should demonstrate.

  2. RELAY achieves strong average performance against the evaluated baselines on both primary metrics. The contrast-removal and readout ablations provide additional evidence that the model captures perturbation-specific information rather than merely reproducing gene-level priors.

Critical weaknesses
  1. A comparison with methods such as PertOmni [1], which contrastively aligns biological text embeddings with transcriptomic perturbation profiles and supports retrieval-based perturbation prediction, would strengthen the evaluation. Although RELAY includes a GenePT-kernel substitution and neighbor-averaging controls, these do not evaluate a retrieval representation trained through text–response alignment. A comparison under matched perturbation splits, comparable training-response access, and comparable DE/DIR readouts and calibration would help establish RELAY’s advantages relative to learned retrieval representations. This gap limits the breadth of the comparative evidence, although it does not negate the improvements over the existing baselines.

[1] Long, Wenxin, et al. “Learning Perturbation Effects Through Contrastive Alignment of Multimodal Biological Embeddings.” bioRxiv (2026), 2026.06.23.734145. doi:10.64898/2026.06.23.734145.

Questions for authors

Clarification requests: Appendix D.1, page 23, lines 1214–1222, describes resampling mixed-label genes for within-gene metrics while retaining their metric values and class-support weights, but resampling query perturbations and all associated rows for accuracy and Macro-F1. What motivates these different resampling units, and what sources of sampling uncertainty are the resulting intervals intended to capture? In particular, how does the gene-level procedure account for dependence among readout genes affected by the same perturbation? Clarifying this would help interpret the reported significance of the performance differences.

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

Related papers

Discussion (0)

Sign in to comment.