acceptodds
Withdrawn from ICLR 2027

P2P: Cross-View Population Denoising for Unpaired Single-Cell Perturbation Response Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AIVC (AI Virtual Cell) is a learned simulator of cellular behavior across conditions. Predicting how a cell population responds transcriptionally to a genetic perturbation is a core task. Perturb-seq records that response by destructive sequencing, so a control cell and a perturbed cell are never observed as a pair, and cells under one condition remain heterogeneous and noisy. Regression on individual cells absorbs sampling variation into the estimated effect, whereas interpretation requires the reproducible population effect. P2P (Perturbation-to-Perturbation) takes a stochastic cell-set view as its supervision unit. Two views drawn from the same condition share a reproducible population effect and differ by view-specific variation. A permutation-invariant set encoder summarizes the control population, a structured encoder represents perturbation tokens, cellular context, dose, and combination interactions, and a gate blends empirical condition-effect memory with a neural residual. A heteroscedastic head predicts the population mean and gene-wise response variance. Under one protocol and five seeds, P2P attains the lowest expression RMSE and the highest Effect Pearson, DEG F1, and DEG average precision on each of Adamson, Norman, Replogle K562, and Replogle RPE1 relative to GenePert, LinearPert, SLIM, Scouter, and scPILOT. On Replogle K562, Effect Pearson rises from 0.643 to 0.702 and DEG F1 rises from 0.067 to 0.178 relative to Scouter, the strongest baseline on both metrics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual cells, perturbation prediction, low-rank perturbation modeling, effect denoising, single-cell genomics

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_D7xrOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes P2P, a population-level framework for predicting single-cell perturbation responses from unpaired control and perturbed cell populations. Instead of constructing control-treated cell pairings, P2P samples two stochastic views from each condition and uses cross-view supervision to recover reproducible population-level perturbation effects. The model combines a permutation-invariant set encoder for control populations, a structured perturbation/context encoder, an empirical condition-effect memory blended with a neural residual through a learned gate, and a heteroscedastic output head for gene-wise response variation.

Critical strengths

Thearchitecture is coherent and well aligned with the population-level formulation of the problem. The use of a permutation-invariant set encoder is a natural choice for unordered cell populations, and the explicit separation of control-population representation, perturbation/context encoding, empirical condition memory, and neural residual makes the model relatively interpretable. The memory-gated design is also a sensible way to combine a stable empirical condition-level prior with a learned correction, while the heteroscedastic head extends the model beyond point estimation to capture response variability.

Critical weaknesses
  1. The novelty framing around “unpaired” perturbation modeling is incomplete, and a highly relevant prior work appears to be missing. A central motivation of the paper is that destructive single-cell measurements prevent observing the same cell before and after perturbation, resulting in inherently unpaired control and perturbed populations. However, this formulation is not new. To my knowledge, the unpaired nature of single-cell perturbation data was already explicitly characterized in prior work [1], although that study did not focus on perturbation prediction. More importantly, prior work [2] has directly considered the unpaired nature of control and perturbed cells in the context of single-cell perturbation prediction and developed a dedicated framework for perturbation effect estimation under this setting. Therefore, the manuscript should more clearly distinguish the methodological contribution of P2P from the existing formulation of the unpaired perturbation problem.

  2. There is a significant mismatch between the paper's “perturbation prediction/generalization” framing and the actual evaluation protocol. The benchmark splits cells within each condition approximately into 50% train, 20% validation, 10% support, and 20% test roles, rather than holding out perturbation conditions altogether. More importantly, the direct readout uses target-condition training-role cells to construct the empirical perturbation memory, and the appendix explicitly states that the benchmark is transductive with respect to token identity. Thus, for a test condition, the model has already observed perturbed cells from the same condition during training and explicitly incorporates their empirical effect through the memory branch. This setting is valid if the task is within-condition population estimation/denoising, but it does not establish performance on genuinely unseen perturbation conditions. The paper should either substantially narrow its claims or provide a perturbation-held-out evaluation in which no target-condition perturbed cells are available.

[1] Learning single-cell perturbation responses using neural optimal transport

[2] Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation

Questions for authors

In addition to the concerns raised above, I have the following questions:

  1. What are the learned gate values on each dataset? Since the gate appears to be a single global scalar, how much does the final prediction actually rely on the empirical memory versus the neural residual?

  2. What additional information does the learned control-set representation provide beyond the empirical control mean that is already explicitly used in the final prediction? Could the authors provide an ablation that removes the SetEncoder while retaining the control mean?

  3. How sensitive is the method to view size and view overlap? Does the benefit of cross-view supervision decrease as overlap increases?

  4. The heteroscedastic head predicts only gene-wise variances, and the Gaussian sampling formulation appears to imply independent sampling across genes (i.e., a diagonal covariance structure). If so, how does the generated population preserve gene-gene correlations or coordinated transcriptional programs? Could the authors clarify whether any dependence structure across genes is modeled during sampling?

Reviewer_9RnhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes P2P (Perturbation-to-Perturbation), a framework for predicting population-level transcriptional responses to genetic perturbations from unpaired single-cell measurements. The main motivation is that single-cell perturbation experiments do not provide one-to-one correspondences between control and perturbed cells, while individual cells exhibit substantial biological and technical variability. The authors argue that population-level perturbation effects are more reproducible targets than individual-cell responses. P2P introduces cross-view population supervision, in which stochastic subsets are sampled from the same control and perturbed populations. A permutation-invariant set encoder summarizes the control population, while a structured perturbation encoder incorporates perturbation identity, cellular context, dose, and potential combination interactions. The model additionally predicts gene-wise variance and is trained using a combination of likelihood, effect-regression, correlation, moment-matching, and cross-view consistency objectives.

Critical strengths

S1. The paper addresses a meaningful statistical challenge in single-cell perturbation modeling. The absence of paired control and perturbed cells is a fundamental characteristic of Perturb-seq experiments. Learning reproducible population-level effects rather than treating arbitrary cell correspondences as ground truth is a reasonable and biologically motivated formulation.

S2. The proposed model provides a coherent integration of population statistics and learned representations.

S3. The reported experimental results are promising under the stated protocol. The evaluation spans four commonly studied perturbation datasets and includes multiple complementary metrics.

Critical weaknesses

W1. The evaluation relies on observed responses from the target perturbation condition, substantially weakening the main predictive claim. This is my most significant concern. According to Section 3.2.3 and Appendix C, the model incorporates an empirical condition-effect memory computed from observed perturbed and control cells. More importantly, Appendix C, Eq. (19), specifies that the direct readout uses the mean of the target perturbation's training-role cells together with the mean of the control test cells. Therefore, although the held-out test cells themselves are not used to construct the memory, the model has already observed other cells under the exact perturbation condition being evaluated.

This is not necessarily conventional train–test leakage, because the cell-level roles are separated. However, it makes the task substantially easier than prospective perturbation prediction, where the response of the target perturbation is unavailable. In fact, combining Eqs. (11) and (19) shows that when the memory gate approaches one, the predicted population mean approaches the empirical mean of the observed target-condition training cells. In this regime, the task is largely to estimate another sample from an already observed condition, rather than to infer the response to an unmeasured intervention.

W2. The comparisons do not sufficiently isolate the value of the learned model beyond empirical condition-level estimation. Given the availability of target-condition training cells, a natural baseline is simply to predict the mean expression of those cells or use their estimated perturbation effect without a learned neural model. Other appropriate comparisons include a shrinkage estimator combining the target-condition empirical mean with a global response prior, as well as a population-level regression model supplied with exactly the same condition-specific observations. Although the paper includes a no-memory ablation, this does not answer whether the complete model substantially improves upon the empirical memory alone. Furthermore, the manuscript states that baseline evaluation outputs were supplied under the fixed benchmark protocol, but it is not sufficiently clear that the baselines received the same target-condition response information available to P2P.

W3. There is an apparent inconsistency in the pair-interaction ablation. Section 3.2.2 states that the pair-interaction correction is disabled for single-token perturbations. However, Table 2 reports that removing the pair-interaction module on the Replogle RPE1 dataset increases RMSE from 0.0910 to 0.1123 and decreases Effect Pearson from 0.6609 to 0.6096. This is among the largest degradations reported in the ablation table. Since Replogle RPE1 is described as an essential-gene CRISPRi screen, the authors should clarify whether the evaluated conditions contain genuine multi-perturbation inputs. If all evaluated conditions contain a single perturbation token, removing an inactive module should not substantially affect predictions under otherwise matched settings. This result raises concerns about either the implementation, the ablation construction, or the description of the architecture.

W4. Although the paper motivates population denoising and models gene-wise variance, most quantitative results concern population means and differential-expression statistics. The optional Gaussian sampling mechanism does not explicitly model gene–gene covariance, and the current evaluation offers limited quantitative evidence for capturing complex changes in cellular heterogeneity.

Finally, the DEG direction results are weaker than those of some baselines. For K562, the reported direction agreement is 0.8414 for P2P versus 0.9944 for GenePert, suggesting a trade-off that deserves more discussion.

Questions for authors

Q1. Can the authors evaluate P2P under a strict perturbation-holdout setting in which all cells corresponding to a test perturbation are excluded from training and from empirical memory construction?

Q2. How much of the performance improvement comes directly from the observed target-condition memory? Please report results for empirical-memory-only prediction, the neural branch alone, and the complete gated model under identical information access. Reporting the learned gate value would also be informative.

Q3. Do all compared baselines have access to the same target-condition training cells used by P2P? If not, can the authors provide comparisons under matched observational budgets?

Reviewer_w5d2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a model, which aims to reduce the variation in pairing unrelated cells across conditions in perturbation prediction models. This is achieved by generating randomly sampled subsets of cells from the same conditions. These samples share a reproducible effect while differing only in sampling noise.

Critical strengths

The cell-set view approach provides a new supervision scheme for perturbation models from single time point RNA-seq datasets.

Critical weaknesses
  1. The paper uses perturb-seq data and generates cell-set views within the same condition. Not all cells from the population will respond similarly to perturbations. Random sampling could select biased cells within a particular view
  2. The data used does not account for other contributors to variation within the population ( such as the cell-cycle state etc ). Perturbations also affect these factors that could create ambiguity in the predictions.
  3. The models tested do not cover other existing foundation models.
Questions for authors
  1. What are the effects of variations due to cell-cycle phase?
  2. Were off target effects due to integrating CRISPR based screens taken into consideration during model training?
  3. Why were other existing baselines such as GEARS, Geneformer, and STATE, fine-tuned to the specific task not included in the benchmarking?

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

Related papers

Discussion (0)

Sign in to comment.