acceptodds
Under review as a conference paper at ICLR 2027

RELAY: Relational In-Context Learning for Cellular Perturbation Prediction

Abstract

Cellular responses to perturbations vary across multiple experimental variables, such as perturbation identity, cell type, donor, and dose, defining a multidimensional condition space in which each measured population occupies one coordinate. Existing methods predict unmeasured conditions using explicit representations of these variable values, generalizing unseen values therefore depends on the availability and quality of such representations. We present RELAY, a relational in-context model that predicts unmeasured coordinates from measured responses at related coordinates. For each target, RELAY uses measured control–perturbed pairs from conditions that share one or more experimental variables with the target, such as the same perturbation in another cell type. The model receives only which variables are shared, not their values, and uses these demonstrations to predict the target response. We evaluate RELAY on chemical, genetic, and cytokine perturbation datasets under two settings: unseen combinations, where test values are individually observed during training but not in their target combinations, and unseen variable values, where test values are introduced only through context at inference. RELAY outperforms existing in-context methods on unseen variable values and supervised perturbation-prediction methods on unseen combinations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single Cell, Virtual Cell, In-Context Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HUU3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RELAY, which predicts cellular perturbation responses from related control–perturbed population pairs rather than explicit perturbation, cell-type, or donor values. The method is evaluated on chemical, genetic, and cytokine datasets under unseen-combination and unseen-variable-value settings.

Critical strengths

RELAY uses measured control–perturbed responses as inference-time context and predicts target-cell expression changes through cross-attention, yielding an architecture that is well aligned with the task. The results are broadly consistent across four datasets, and the ablations of relation embeddings, state embeddings, and context composition support that the model uses the shared relations between context and target conditions.

Critical weaknesses

1.The novelty of relational modeling is not fully clear. Prior context-based methods also transfer information from observed control–perturbed pairs. RELAY adds shared-variable relation embeddings to these pairs, but its benefit beyond standard context conditioning or a simple match-mask embedding is not isolated.

2.The comparison lacks an explicit residual-effect baseline. Except for STATE, the other learned baselines are generative or transport-based. A matched deterministic residual predictor, such as GEARS on the genetic benchmark, would help determine whether RELAY’s gains arise from relational context or from predicting perturbation effects as residuals.

Questions for authors

1.Equations (6)–(8) add shared-variable-type embeddings to context tokens. How does this differ from standard variable-type or segment embeddings? A simpler matched conditional-encoding baseline using the same shared-variable mask would help establish the value of the relational design.

2.The role of the paired-population encoder is unclear. Since control and perturbed cells are unpaired, why does jointly encoding their population summaries represent a perturbation-induced transition rather than a generic joint representation?

3.Could the authors clarify how context is obtained and introduced in Setting 2, and whether all context-based baselines use the same context and information budget?

4.Could the authors add matched Setting 2 ablations across multiple datasets, such as removing relation embeddings, state embeddings, or using a generic context encoder?

5.Figure 4 shows diminishing but still positive gains as m2m_2 increases. Is the maximum context size constrained by the number of available conditions? Could the authors explicitly compare different allocations under a fixed total context budget?

Reviewer_T5M3OpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

they proposed a new method named RELAY for perturbation effect prediction under unseen condition. RELAY introduced 'demonstration' as measured control-perturbed population pairs at inference time。 unlike previous methods relying on perturbation identity, cell type or donor label, it uses shared-variable relations to represent the relationship between each demonstration and target condition. RELAY consists of three major parts, a pair encoder, relation embeddings and context-conditioned prediction. targeting unseen combination and varaible values, their experiments suggest RELAY is overwhelmingly better on preason correlation and MSE.

Critical strengths
  1. RELAY makes predictions utilizing control-perturbed pairs and its shared variable relation with target condition rather than explicit representation learning on each perturbation/celltype/donor variable. it provides a reasonable alternative for unseen or vaguely-described experimental conditions.
  2. it shows performance superiority on all of chemical, genetic and cytokine pertubration criterias, especially getting best pearson correlation and mse on setting1. the model seems indeed utilize target-specific relational information rather than mean perturbation effect from relation embedding ablation and context replacement experiment.
Critical weaknesses
  1. experiment part lacks a simple relational transfer baseline inder information-matched condition. seems more simpler baselines like relation-weighted averaging method or ridge-based transfer method. abaltions are not sufficient to clear my concerns over complex model structures (e.g. paired encoders, multiple self/cross attention modules).
  2. REALY doesn't perform significantly better than other learned baseline on PBMC data (setting2), especially, the increment in their pearsonr lies under noise level. thus it brings doubt on that their major components can bring stable and significant performance gain on unseen variable values.

this is an interesting paper but my final score remain unsettled depending on whether supplementary experiments/analyses can clear my concerns.

Questions for authors
  1. can u add a simple baselin using exactly the same demonstrations and shared-variable relations like ones suggested in weakness?
  2. can u use some ablations to show that how much does RELAY's performance gain comes from relation embeddings and matched context?
  3. can you show that if REALY is able to reconstruct post-perturbation cell distribution rather than average expression change? as the model already uses population-level MMD loss, it is reasonble to report held-out population MMD and/or OT distance, and discuss whether cell type/state heterogeneity is maintained.
Reviewer_CNLVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RELAY, an in-context approach for cellular perturbation prediction.

Rather than encoding perturbation and cellular-context identities directly it uses measured control–perturbed populations and information about which experimental variables they share with the target.

The method is evaluated on four datasets and considers both unseen combinations and entirely unseen variable values.

Critical strengths

I like the formulation of perturbation prediction as inference over related experimental conditions.

Paper provides a useful perspective, especially when explicit perturbation representations are unavailable.

Strong results in the unseen-combination setting.

The ablations also clearly demonstrate the importance of relation embeddings.

Critical weaknesses

The method relies heavily on having relevent experimental measurements available at inference.

The results show that demonstrations sharing two variables are much more useful than those sharing only one.

It is unclear how often such rich context would be available in practice.

Generalization to completely unseen values remains limited in some settings. On PBMC-3D, PDCorr is only 0.124 and the untreated baseline has lower MSE.

Also, the provided PDF refers to Appendices A/B for essential experimental details, but these appendices are not included. This makes the evaluation difficult to verify.

Questions for authors

How well does RELAY perform in Setting 2 when only a few matching demonstrations are available? For example, without double-shared conditions?

Can the authors provide the referenced appendices, complete split definitions, and comparisons against simple relation-aware interpolation baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.