acceptodds
Withdrawn from ICLR 2027

PerturbRIFT: Response-Informed Flow Matching for Predicting Single-Cell Responses to Unseen Genetic Perturbations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Predicting single-cell responses to unseen genetic perturbations requires modeling both population-level expression changes and cell-to-cell variability. Here we introduce PerturbRIFT, a generative framework that incorporates measured response statistics into a perturbation-specific source distribution for conditional flow matching. Gene embeddings and knowledge graphs guide the learned aggregation of gene-wise means, zero probabilities, and nonzero-expression variances from related training perturbations. These statistics parameterize a factorized zero-inflated normal (ZINormal) prior, providing a query-specific starting point for generation. Residual latent flow matching then transforms samples from this prior into predicted single-cell response populations, with a latent maximum mean discrepancy objective aligning transported and observed populations during training. Across four within-dataset unseen-perturbation benchmarks in K562, RPE1, HepG2, and Jurkat, PerturbRIFT improves mean-response accuracy and multiple distributional metrics relative to the published prediction baselines evaluated here. K562 ablations show that biologically guided prior estimation contributes the largest gains in mean-response accuracy, while residual flow matching with latent distribution alignment improves multiple measures of single-cell distributional fidelity. Together, our findings support using measured response statistics to define generative source distributions for predicting single-cell responses to unseen perturbations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation prediction, conditional flow matching, generative modeling, response-informed priors, biological representation learning, single-cell transcriptomics

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_S7uMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PerturbRIFT predicts single-cell expression distributions for genetic perturbations whose response cells are unavailable during training. It first retrieves 16 related training perturbations using biological resources, including STRING, CollecTRI, and pathway memberships. A learned weighting network aggregates their measured gene-wise means, zero probabilities, and conditional nonzero variances. These statistics define a factorized zero-inflated normal prior.

Samples from this prior are encoded using a pretrained, frozen autoencoder. A conditional flow model then transports the encoded populations toward the latent distributions of observed responses, using flow matching and an endpoint MMD objective. The decoder generates expression profiles while calibrating their expected population mean to the retrieved prior mean. Finally, Cube sampling couples output-mask decisions across cells to reduce finite-population imbalance. These components are trained in separate stages.

The evaluation covers unseen perturbations within four datasets—K562, RPE1, HepG2, and Jurkat—with five shared perturbation-level splits per dataset. The reported results show consistent improvements in Top100 DEG mean-response metrics, perturbation discrimination, and several distributional metrics. K562 ablations attribute most mean-prediction gains to biological anchor selection and learned weighting, while flow improves PCA-space distributional agreement and recovery of correlation changes.

Critical strengths

On the perturbations that do have the necessary information, PerturbRIFT performs well. The Top100 results improve consistently over strong linear baselines across all four datasets. I think the experiments and ablations are well-done and relatively comprehensive overall as well. There are four datasets and five shared perturbation splits, and the ablations help explain where the gains come from. In particular, KG-based anchor selection substantially improves the prior mean prediction, while the flow improves correlation-change recovery and PCA-space distributional agreement. These are useful findings.

Critical weaknesses

It seems like a complicated method, with three phases of training, but I am not yet convinced that the experiments establish the need for the proposed source distribution and generative refinement. The retrieved response statistics enter the source, the conditioning, and the decoder’s mean calibration. Table 3 keeps the source fixed, so it does not show whether this source is better than control cells or noise when the same statistics are available through conditioning. There is also no comparison against simply sampling cells from the same retrieved perturbations and rescaling them to the same predicted mean. That would preserve measured gene dependencies and help show what the flow contributes beyond retrieval and a good mean predictor.

Evaluation is restricted to perturbations covered by the required resources, and it is unclear how much coverage is lost or what happens when useful related perturbations are unavailable. Also, some distributional wins are very small, e.g. Jurkat MMD is 4.228 versus 4.256 for SLIM in Table 2, and the main tables provide no uncertainty estimates. There are five data splits, but repeated training seeds on fixed splits are not documented.

Questions for authors
  1. Could the authors compare against latent flow matching with MMD and the same OT pairing, but starting from control cells or a generic noise source? This would directly test whether putting the response information into the source distribution is useful.

  2. Could the authors report uncertainty across the existing splits and run additional training seeds for the key comparisons?

  3. How many perturbations are excluded from each dataset by the resource-coverage filtering? How would the method handle a query with missing annotations or few related training perturbations?

  4. Also, compute permitting, it would be great to see performance on slightly larger datasets like PBMC

  5. Could the authors provide the missing citations for the baseline methods?

  6. Why is endpoint MMD computed against the OT-resampled target batch, which can contain duplicate cells, instead of the original observed batch?

  7. What values are used for the prior-loss coefficients, the probability and variance floors, and the minimum nonzero-observation count?

  8. What are the training and inference costs of the three-stage pipeline compared with the strongest simple baselines?

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

PerturbRIFT retrieves related measured perturbations using gene embeddings and knowledge graphs, and aggregates their response statistics into a query-specific ZINormal prior. A latent residual flow then models the target perturbation distribution. The approach aims to improve prediction for unseen perturbations by combining empirical response information with generative modeling.

Critical strengths
  1. The paper addresses the important problem of predicting single-cell responses to unseen perturbations. Using retrieved experimental responses to construct a query-specific generative prior is a reasonable and potentially useful design.

  2. The evaluation considers multiple aspects of response prediction, including mean expression, marginal distributions, and gene-correlation changes. The ablations on K562 also examine the effects of anchor selection, residual flow, and Cube sampling.

Critical weaknesses
  1. The contribution of the full response-informed ZINormal prior is not sufficiently established. Anchor weights are learned using mean-response supervision but are also used to transfer zero-expression probabilities and nonzero variances. The paper does not clearly demonstrate that these transferred statistics are reliable or that they improve over simpler alternatives, such as a mean-only prior.

  2. The incremental benefit of residual flow remains unclear. Although the ablation shows improvements in some PCA-space and correlation metrics, the analysis is limited to a subset of K562 and a single seed. The prior alone remains competitive on several distributional metrics, including standard-deviation error. Broader and more controlled comparisons are needed to establish when residual flow provides consistent gains.

  3. The evaluation does not fully establish the robustness of the reported improvements. The baseline comparisons and zero-expression metrics require further clarification, particularly given the identical ZeroMAE values reported for several methods. The rationale for baseline selection and the omission of relevant competing approaches should also be better explained.

Questions for authors
  1. Could the authors compare the full prior with matched variants that retain the retrieved mean while replacing the transferred zero probabilities and/or nonzero variances with global or control-derived estimates? Reporting conditional nonzero variance error would help isolate the contribution of each component.

  2. Since Eq. 11 constrains the decoded population mean to match (m_0) in expectation before clipping, could the authors evaluate an uncalibrated or softly calibrated variant and report its mean-response metrics? This would help clarify how much of the performance comes from the prior and the calibration constraint.

  3. Could the query-specific control-shift baseline in Table 3 be evaluated across all four datasets, with comparisons against SLIM-style rescaling using the same retrieved mean? This would help determine whether the residual flow offers gains beyond simpler response-based estimates.

  4. Could the authors clarify baseline selection and explain the identical ZeroMAE values for GEARS, Scouter, CellFlow, and scGPT? It would also be helpful to justify the omission of methods such as LangPert, scDFM, and GeneGeoFlow, and to report the GenePT-KNN baselines in Table 1 across all four datasets where applicable.

Reviewer_yd4qOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed PerturbRIFT, which predicts single-cell responses to unseen genetic perturbations by combining biologically guided retrieval of response statistics with conditional flow matching to generate cell populations.

Critical strengths

The use of responses from related perturbations to construct an informative prior is a reasonable approach.

Critical weaknesses
  1. The evaluation is limited to the Replogle-Nadig perturbation dataset.
  2. The evaluation does not include metrics measuring the recovery of differentially expressed genes, such as DE overlap (as used in STATE). Including these metrics would help understand whether the model correctly identifies genes affected by a perturbation.
Questions for authors

The authors discuss reference-neighborhood coverage in the Appendix. How does PerturbRIFT's performance vary with the biological similarity between test perturbations and their retrieved training perturbations? In particular, does performance degrade when no closely related perturbations are available? How does the model perform on other publicly available datasets?

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

Related papers

Discussion (0)

Sign in to comment.