acceptodds
Under review as a conference paper at ICLR 2027

COHERE: Coherent Response Transfer for Cross-Context Perturbation Prediction

Abstract

Predicting perturbation responses across cellular contexts requires capturing more than average transcriptional changes. Populations with similar means can differ in expressing fractions and positive-expression distributions, yielding distinct differential-expression (DE) evidence. We introduce COHERE, a framework that jointly transfers response means and marginal shape while enforcing mean–shape compatibility. To predict unobserved perturbation–context responses, COHERE uses target controls and responses to other training perturbations in the target context to adapt source mean changes, expressing fractions, and positive-expression quantiles. A shared low-rank transformation updates these statistics while preserving source components outside the response subspace. Constrained generation constructs finite, nonnegative cell populations that preserve feasible predicted gene-wise means. Across six genetic, chemical, and cytokine perturbation benchmarks, COHERE demonstrates broad performance advantages over evaluated baselines in mean-response prediction, perturbation discrimination, and DE recovery. Under the benchmark DE scoring protocol, learned shape improves the area under the precision–recall curve (AUPRC) at fixed predicted means on all six datasets, including a relative 32.5% gain on Nadig. Matched retraining shows that preserving source components outside the response subspace improves DE recovery and marginal fidelity on PBMC and Nadig. These findings support coherent transfer of marginal response shape as a complement to accurate mean prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Cross-context response transfer, Perturbation response modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2VjQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

COHERE is a few-shot perturbation predictor. It predicts the response to perturbation (p) in a target context (c) when same-perturbation responses is seen in other contexts and the new context is also seen with other perturbations. The process is as followed: (i) Each source response is summarized per gene by a mean change, a residual occurrence logit, and five positive-expression quantiles. These are Huber-aggregated across source contexts and projected onto a 128-dimensional basis from an SVD of training mean responses. (ii) A per-context low-rank affine map, shared across the three descriptor families and fitted on the target training panel, adapts the basis coordinates. Only the coordinate change is back-projected, so source components outside the basis are retained. (iii) A residual from ESM2 or Morgan-fingerprint features corrects the mean. A deterministic realization step then converts mean, occurrence, and quantiles into N non-negative cells that conserve the predicted gene-wise mean, with values assigned to control anchors by a pretrained ordering network.

The evaluation covers six benchmarks (PBMC cytokines, Replogle-Nadig CRISPRi, SciPlex3, Frangieh21, OP3, McFaline) with 13 Cell-Eval metrics against grouped means, CPA, CellFlow, scGen, scPILOT, TxPert, STATE, and PerturbDiff. COHERE has the best point estimate on 11 of 13 metrics on PBMC and 9 of 13 on Nadig, and the best AUPRC on three of four external datasets. Fixed-mean interventions show that replacing source with learned shape descriptors raises AUPRC on all six datasets, driven by the occurrence channel. Matched retraining shows that source-preserving reconstruction beats subspace-only reconstruction in DE recovery and marginal Wasserstein error on PBMC and Nadig.

Critical strengths
  • The fixed-mean factorial design isolates what marginal shape contributes to DE recovery. Because realization conserves the predicted mean exactly, swapping occurrence and quantile descriptors at fixed means separates shape effects from mean effects, which most distributional methods cannot do. On Nadig, learned occurrence adds ~0.08 AUPRC. The effect survives raw-q and unadjusted-p rankings, and matched retraining with mean-only adaptation reproduces it. The equal-positive control shows that DE ranking and marginal fidelity can move in opposite directions. This matters for my assessment because it is the paper's most defensible finding, and it is useful to anyone building or interpreting Wilcoxon-based perturbation benchmarks.

  • The reporting is transparent. The appendix documents results that weaken the main story: negative transfer on PBMC, score saturation in baseline DE metrics, STATE's lower Nadig W1, and the sensitivity of the PBMC shape effect to score clipping. This makes the claims auditable, and most of my concerns below come from the paper's own tables.

Critical weaknesses
  • The titular claim does not survive the paper's own simplest strong baseline, and the DE comparison is confounded by cell realization. Direct source transfer (source descriptors plus COHERE's realization, no adaptation) appears only as an internal control. On PBMC it beats COHERE and every listed baseline on DEOver, DEPrec, LFCSpear, AUPRC, and AUROC. Including it would cut COHERE's PBMC wins from 11 to at most 8 of 13 and remove its DE advantage there. Direct transfer also has lower MSE than COHERE on OP3 and Frangieh21 and higher AUPRC on McFaline.

  • Separately, Table 19 shows eight PBMC baselines at the label prevalence because most of their genes hit the 1e-10 FDR floor, so their DE scores reflect how each method materializes cells, not what it predicts. Much of COHERE's DE advantage may therefore come from realization and source evidence rather than learned adaptation.

  • The mean-response claim is also mixed: COHERE's MSE is worse than a grouped-mean baseline on SciPlex3 and OP3, and worse than four baselines on Frangieh21. This is consequential because the abstract's "broad performance advantages" in mean prediction, discrimination, and DE recovery is the paper's main empirical claim. Also, SciPlex3 is tested but not Tahoe-100M.

  • Component claims are based on single-seed fits and small effects, and several components contribute nothing measurable.

Questions for authors
  1. Realization control: If the pseudobulk means from STATE, PerturbDiff, and mean/perturbation are passed through COHERE's realization with source shape, how do their DE metrics compare with COHERE's? This would separate gains from adaptation and gains from realization.

  2. Seed variance: Please report at least three training seeds for COHERE (and ideally STATE) on PBMC and Nadig, with the seed standard deviation for the Table 1 headline metrics and the Table 2 contrasts.

  3. Baseline fidelity: STATE was retrained from scratch with an adapted architecture, and PerturbDiff uses released checkpoints. Do the PBMC and Nadig numbers match those reported for these models on the same partitions in the PerturbDiff comparison?

  4. Minimal method: Given that the identity residual and allocator have no measurable effect on any reported metric, which components do you consider essential? What does a minimal version (mean plus occurrence adaptation, residual skip, realization) score in Table 1?

  5. PBMC DE target: Adaptation lowers pooled PBMC AUPRC but raises within-context AP, and H.2 attributes the pooled loss to new false positives among non-DE genes. Which DE target do you consider correct for PBMC, and why should the pooled protocol be primary?

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

This paper introduces cohere which is a model for predicting single cell perturbation response. The difference between cohere and previous models in this field is that it represents perturbation responses through different components: source mean changes, expressing fractions, and positive-expression quantiles. The authors argue that predicting mean response is not sufficient for recovering DE since populations with similar means might have different expressing fractions and positive expression distributions. They evaluate their method on multiple genetic and chemical datasets and they show cohere has competitive performance against several baselines.

Critical strengths
  • The discussion about why mean response prediction is insufficient for recovering DE is interesting and novel.
  • The authors conduct a couple experiments to isolate the contribution of distributional shape to DE recovery which is useful and interesting.
  • The authors considered both genetic and molecular setting and showed that the framework works on both settings.
Critical weaknesses
  • The task that this work is trying to solve is less interesting and easier than prior works. Cohere does not work in the zero-shot setting for novel cell lines since it requires both target control and target responses to other perturbations to fit their context adaptations.

  • The contribution of several architectural components are not clear and the paper needs a better ablation study. Cohere has multiple components and it is not clear if all parts are essential or if we could use a simplified model that could achieve similar results.

  • Cohere is conceptually very similar to Rhaister [1] that it should explicitly discuss and compare against this method.

[1] https://www.biorxiv.org/content/10.64898/2026.06.09.731197v1

Questions for authors
  • Have you tested Cohere on a setting where the target cellular context has no observed perturbation response?
  • Since the gains from the identity-residual component is very small why we need molecular fingerprints and protein embeddings in the current architecture?
  • Could an occurrence-aware mean-response predictor, without positive-quantile adaptation and identity correction, provide most of the reported benefits at substantially lower computational and implementation complexity?
Reviewer_AZtCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

COHERE predicts a perturbation’s response in a new context by transferring its observed responses from other contexts. It adapts mean, occurrence, and quantile statistics with a low-rank map fitted on the target’s training perturbations, then builds cells that match the predicted means.

Critical strengths

The motivation is clear: equal means can give different DE results. The fixed-mean interventions are a useful analysis tool, and the appendix reports negative results openly.

Critical weaknesses
  1. The learned adaptation does not help consistently. On PBMC, direct transfer without learning beats COHERE on DE and beats all baselines, yet it appears only in the appendix.

  2. The DE gains seem to come from cell realization and the evaluation protocol rather than from transfer. The occurrence effect is expected for a rank-based test on sparse data, and the realization relies on untested heuristic choices.

  3. The comparison with STATE and PerturbDiff is not fair. Perturbation inputs are not aligned, and only gene-wise marginals are evaluated, which skips the joint structure these models target.

  4. All results use a single seed, key margins are small, and the method needs source observations and a target panel.

Questions for authors
  1. Direct transfer beats COHERE on DE on PBMC but not on Nadig. When does recipient adaptation help over direct transfer? Can the gain be related to the distance between source and target contexts or to the size of the target panel?

  2. In UMI data, the zero fraction is largely determined by the mean. Does learned occurrence improve DE over a simple baseline that derives the zero fraction from the predicted mean, for example using a mean–zero-fraction curve fitted on target control cells? This would show whether the occurrence channel adds information beyond the mean.

  3. The current metrics depend only on gene-wise marginals, so they cannot separate COHERE from joint-distribution models. How does COHERE compare with STATE and PerturbDiff on cell-level joint metrics, such as energy distance or gene–gene correlation? Including COHERE cells permuted within each gene would show whether the allocator adds any joint structure.

  4. All results come from single-seed fits, and the intervals reflect only the sampling of test perturbations. How stable are the main comparisons, especially against STATE on Nadig, across training seeds?

1 comment

Thank the authors for their detailed response and the large number of additional experiments. I have read the revised paper, the authors' response, and the comments from other reviewers. Here is my feedback:

Overall, I like this simple and effective method. It separates the transfer of the mean and the distribution shape, and then unifies them using a mean-fixed realization. The idea is clear. The fixed-mean experiment cleanly attributes the gains on DE to the shape, which current distributional models cannot do. However, I still have a few concerns:

  1. Easier Experimental Setting: Inspired by Reviewer JEvs, the motivation is good, but the evaluation setting is a bit easy. It requires a panel of observations sharing the same perturbation source and target context (few-shot). Without a panel, it mostly degrades to Direct. We need further comparison of zero-shot capabilities.
  2. The gains are not consistent across metrics. The direction of change is predicted well (PDCorr), but the magnitude (MSE) is not better than several baselines, and the DE gains rely largely on realization in the few-shot setting.
  3. Oracle-Mean Test: If the mean is fixed to the true value, does the conclusion about the DE shape still hold?

I believe the greatest value of this method is that it can act as a plug-in. Like the authors' hybrid experiments ( A.2 in Reviewer AZtC), the realization and fixed-mean analysis do not depend on COHERE itself and can be attached to any mean prediction model. A very natural combination would be: a model with good mean prediction (like STATE) plus a mean-zero proportion curve fitted by the target control, plus the realization. Such combination does not need a panel, and it holds strong promise for achieving both a good mean and a good DE shape. This could be a useful reference for future cross-context zero-shot methods.

Therefore, I believe this paper has value, and I am maintaining my score.

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their rebuttals. I would be grateful if you could take a moment to review them carefully and respond as appropriate. Thank you for your ongoing engagement and support.

Your AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.