acceptodds
Under review as a conference paper at ICLR 2027

Share to Extrapolate: Learning Reusable Response Components for Single-Cell Perturbation Prediction

Abstract

Predicting how genetic perturbations change gene expression is central to understanding gene function and cellular regulation. Yet experimental coverage remains limited, requiring generalization to unseen perturbations. Despite advances in perturbation prediction, matching observed conditional distributions alone leaves responses to unseen perturbations underdetermined. With supervision limited to observed perturbations, models can fit the training distributions equally well yet predict different responses to the same unseen perturbation. Similar patterns of gene-expression change across perturbations suggest that shared response components can provide an inductive bias for extrapolation. We introduce CoShare, a graph-routed factorized Delta Flow Matching framework. CoShare decomposes responses relative to the control mean into mean responses and cell-level residuals, learning from unpaired control and perturbed populations. A component router uses multi-source gene graphs and cell-type context to produce a component coefficient vector. This vector weights shared response vectors in the population mean branch and controls adapters combining shared feature transformations in the cellular residual branch. Experiments on Adamson, Norman, and Replogle demonstrate improved prediction of mean responses and cell distributions, including for unseen gene combinations. CoShare reduces Energy Distance over 2,000 highly variable genes by 44.5% on Adamson and 66.2% on Replogle relative to the best baseline for this metric on each benchmark. Ablations support the complementary contributions of mean–residual factorization and graph-routed component reuse.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Flow matching

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

CoShare predicts single-cell responses to unseen genetic perturbations by splitting the response (relative to the control mean) into: (i) a mean shift, modeled as a weighted sum of M shared basis vectors (initialized by PCA), and (ii) a residual distribution generated by x-prediction flow matching. A router driven by a five-view gene graph encoder (STRING, GO-BP/MF/CC, LLM embeddings) sets both the basis weights and the mixing of low-rank adapters in the generator. CoShare beats nine baselines on Adamson, Norman and Replogle across RMSE, energy distance and Cell-Eval metrics, with ablations on Norman and a coefficient-swap analysis on Adamson.

Critical strengths
  1. Consistent gains across datasets and metric types... The Replogle PDS_cos result (0.77 vs 0.59) is the most convincing; however, PDS is also the least biologically grounded metric, and can be often gamed and hill climbed on.

  2. Regressing the pseudobulk mean directly and generating only residuals is a sensible design; however, prior publications like Rahister from Tahoe and PIE from Arc Institute implement this as well.

Critical weaknesses
  1. Gains are not attributable to component reuse: simple baselines are missing and baselines get less prior information.

The mean branch (b0 = average training response, B = top PCs of training responses) is essentially the linear baseline of Ahlmann-Eltze et al. (Nat. Methods 2025), which matched or beat deep models on these datasets. Other baselines like mean, linear, graph-neighbor, additive (Norman 0/2) and, for Replogle, copy-same-perturbation-in-other-cell-lines baselines are all absent.

The consequence: on Adamson, every method has PDS_cos of 0.52 to 0.65 (chance about 0.5) while CoShare's PDCorr is 0.87... this is a pattern that Systema attributes to predicting systematic variation. Also concerning is that PerturbDiff represents targets by averaging learned embeddings, which carry no information for genes never perturbed in training; PerturbDiff is nonetheless the best baseline on Norman 2/2, suggesting a largely perturbation-agnostic predictor is competitive there, while CoShare alone gets graph and LLM priors for unseen genes. The plain "Base FM" (E-dist 0.68) already matches the best baseline... also some baselines look mis-adapted (GraphVCI RMSE 0.512 on Replogle).

  1. Rigor and mechanism claims are under-supported... single split, single run, no uncertainty; several margins are small. No validation set or selection procedure is described for CoShare. The ablations do not say what replaces each removed component or whether capacity is matched, and there is no shared- vs independent-router comparison.
Questions for authors
  1. Please add more baselines: mean, linear (same prior features), graph-neighbor averaging, additive (Norman 0/2) and, for Replogle, same-perturbation-from-other-lines baselines, and report metrics on perturbation-specific deltas. For distributional metrics, how does CoShare's predicted mean plus centered control cells (or pooled centered training residuals) compare with the generated residuals?

  2. Please report mean + sd over multiple splits/seeds, the size of each Norman group, and held-out singles separately.

  3. How were hyperparameters and checkpoints selected, and how were baselines tuned? What represents a never-seen target in each baseline?

  4. What replaces the graph in "w/o Graph" and the factorization in "w/o Factorization"? Does a capacity-matched model with independent mean/residual routers do worse, and does swapping only the residual adapters degrade distributional metrics?

  5. What fraction of held-out HepG2 targets are perturbed in K562/RPE1/Jurkat, and how do results differ for those vs. globally unseen targets?

  6. How was the LLM graph built (prompt, inputs)? What happens with it removed, with each view alone, and with a degree-preserving shuffled graph?

  7. How many cells were generated and observed per perturbation? The V-statistic in Eq. 17 is biased by sample size, so a real-vs-real floor would help. Do generated residuals have approximately zero mean?

Reviewer_nzMYOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies prediction of previously unobserved genetic perturbation responses in a target cellular context. CoShare decomposes the expression change of a perturbed cell relative to the corresponding control mean into a conditional mean response and a cell-level residual around the perturbation-specific population mean.

The model represents single- or double-gene perturbations using a multi-source graph encoder that integrates STRING, GO, and LLM-derived associations, and combines these perturbation representations with cell-type information to produce component coefficients. The mean branch uses these coefficients to combine a learned dictionary of shared response components. The residual branch uses conditional flow matching to generate residuals from Gaussian noise, while mapping the same coefficients to mixture weights over layer-wise low-rank adapters.

The two branches and the router are trained jointly. At inference time, the predicted mean response, generated residual, and control mean are added to obtain the predicted perturbed-cell population. The method is evaluated on Adamson, Norman, and multi-cell-line Replogle settings.

Critical strengths

S1. The division of labor between the mean and residual branches is clearly implemented, and the shared routing mechanism is supported by component-level ablations

CoShare uses different parameterizations for conditional mean response and cell-to-cell variation. The former is represented through a compact response dictionary, while the latter is modeled with a conditional generative model around the perturbation mean. The same component coefficients affect both the composition of the mean-response dictionary and the adapter weights in the residual network, and receive gradients from both objectives.

S2. The empirical gains cover both distributional fitting and perturbation-response recovery, including challenging compositional generalization settings

CoShare performs strongly across several demanding evaluation settings and on metrics that capture different aspects of perturbation prediction.Figure 5C is particularly useful because it goes beyond the correlation between graph distance and component similarity shown in Figures 5A–B. By replacing the routing coefficients of a test perturbation with those from nearby or distant perturbations, the experiment tests whether the learned routing structure has a functional effect on prediction.

Critical weaknesses

W1. The incremental contribution of cross-branch sharing is not sufficiently separated from general graph conditioning, adapter capacity, and additional prior information

The most distinctive architectural claim is that the same component coefficients coordinate mean prediction and residual generation. The mean–residual decomposition itself is less specific to the proposed method. The "w/o routed adapter" ablation in Figure 4 removes both a conditioning pathway and the associated parameters. It therefore does not answer whether shared routing is better than using two independent routing heads while keeping the same graph encoder, two-branch structure, adapters, parameter budget, and training budget.

The residual generator also directly receives perturbation representations, cell type, and control statistics. Part of the improvement could therefore come from general conditional modeling or additional capacity rather than from cross-branch sharing specifically. Figures 5A–B show that nearby genes tend to use more similar components or operators, but graph relationships are already part of the encoder input. Figure 5C provides stronger functional evidence, but the coefficient replacement affects both the mean and residual branches simultaneously and is evaluated using PDCorr of the mean response. It therefore does not isolate the contribution of shared routing to residual extrapolation.

A related issue is the use of an LLM-derived graph and node features. The current "w/o Graph" setting does not make clear which sources of information are removed. A matched-information comparison is needed to distinguish gains from richer prior information from gains due to the proposed sharing mechanism.

W2. The evidence for improved residual-distribution modeling is still incomplete

Energy distance in the original expression space depends on both differences in population means and differences in distributional shape. The large ED improvement of the full model supports better overall distribution matching, but does not by itself quantify the contribution of the residual generator. Figure 3 provides a positive example of residual standard-deviation recovery for one unseen combination. A test-set-level evaluation of centered distributions, variances, or covariances would provide much stronger evidence for the claimed residual modeling benefit.

From Eqs. 4–5, the training residual is defined as perturbed-cell expression minus the perturbation-specific population mean, so its empirical mean within each condition is zero. This does not guarantee that the residuals generated by a finitely trained model have zero mean.

In Eq. 15, the mean of the final predicted population is calculated. It is therefore important to report the mean offset of generated residuals and the discrepancy between the output of the mean branch and the empirical mean of the final generated population. Appendix A.1.2 distinguishes these quantities conceptually, but the current experiments do not show whether the two branches maintain this intended division of labor in practice.

Questions for authors

Q1. Core architectural validation: does sharing component coefficients outperform independent routing?

Could the authors compare the current shared-coefficient design against a variant in which the mean and residual branches use separate routing heads, while keeping the same multi-source graph encoder, mean–residual decomposition, low-rank adapters, parameter count, and training budget as closely matched as possible?

Please report both mean-response metrics and residual-distribution metrics, rather than only overall ED. This comparison would directly test the central claim of coordinated cross-branch reuse and would not require introducing a new task or rerunning all baselines.

Q2. Residual-function validation: how much of the distributional improvement comes from mean correction versus residual generation?

Using the existing predictions, please report the condition-wise mean offset of the generated residuals and the difference between the mean-branch output and the empirical mean of the final generated population.

After centering the real and predicted populations by their respective means, does CoShare still show consistent improvements in centered ED, variance recovery, or covariance recovery?

Another useful analysis would be to hold the same predicted mean fixed and compare alternative residual-generation mechanisms. Test-set-level summaries would be more informative than a single representative perturbation combination.

Q3. Prior contribution and information scope: what exactly is provided or removed by the LLM graph, node features, and graph ablations?

Please specify the source of the gene text, the prompt or procedure used to obtain the LLM-derived information, the model/version used, and how embeddings are converted into graph edges and edge weights. It would also be important to state whether the source material contains descriptions of perturbation responses from any of the evaluated experiments.

What exactly is removed in "w/o Graph": graph connectivity, semantic node features, or both? A particularly informative matched control would retain STRING and GO while removing only the LLM-derived association graph, with a clear statement of whether LLM-derived node features remain available. If the goal is to measure the contribution of all LLM-derived information, node features should be controlled accordingly.

Q4. Generalization scope and statistical stability: do the gains hold across different types of unseen targets?

Please clarify whether the main-table results are from a single run or aggregated over multiple runs. For at least one representative benchmark and the key ablations, repeated training runs or uncertainty estimates over perturbation conditions would help establish the stability of the conclusions.

Please also report the number of conditions in each of the three Norman combination groups, and the fraction of HepG2 test targets that appear in the training data of other cell lines. Stratifying results into targets seen elsewhere versus globally unseen targets would make the generalization setting easier to interpret. The current Replogle setting is a valid evaluation scenario, but its extrapolation regime should be characterized separately from settings involving a completely unseen cell line or targets that are globally unseen during training.

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

The paper predicts single-cell responses to unseen genetic perturbations. It splits each response into a mean shift and cell-level residuals. A router uses STRING, GO and LLM gene graphs to produce one coefficient vector. This vector weights a small dictionary of response vectors for the mean, and mixes low-rank adapters in a flow-matching model for the residuals. The paper reports the best results on most metrics on Adamson, Norman and Replogle.

Critical strengths

Modeling the mean shift and cell-to-cell variation separately is simple and useful. Figure 3 shows the method can recover the residual spread, which most methods do not test. The paper is clear and easy to follow.

Critical weaknesses
  1. It is unclear what is shared and whether sharing explains the gains. The paper does not show what the shared components are or which ones unseen perturbations use. In the ablation, the graph gives most of the gain, while the routed adapter, which shares the coefficients between the two branches, gives the smallest gain. Figure 5 is hard to read as evidence for sharing, because nearby genes would get similar coefficients anyway when the coefficients are computed from the graph.

  2. The baseline settings may not reflect each baseline’s best performance. All baselines are moved to 2,000 HVGs, while most were designed for about 5,000 genes, and some run without their pretraining or without a way to represent unseen genes. CoShare is the only method that uses STRING, GO and LLM priors together. A plain flow-matching model already reaches the level of the best baseline, so adding simple baselines such as the training mean would help show how large the gains are.

Questions for authors
  1. What does each shared component represent? Could you show its top genes, how far B moves from its PCA start, how many components each perturbation uses, which ones unseen perturbations reuse, and whether they are stable across seeds?

  2. Could you add a train-mean baseline for all datasets, an additive baseline for the Norman 0/2 group, and a cross-cell-line mean baseline for Replogle? How many test HepG2 perturbations are also perturbed in the training cell lines?

  3. Could you run at least scFoundation, scGPT and PerturbDiff with their original settings (original gene space with target genes, and PerturbDiff with pretraining and fixed gene embeddings)? If not, how many test target genes are outside the 2,000 HVGs, and how did each baseline handle them on Adamson and Norman?

  4. Could you add ablations with separate routers for the two branches, with B fixed at its PCA start, and with GO as the only prior? Please also define each ablated model and report the ablations on Replogle and for the 0/2, 1/2 and 2/2 groups.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.