acceptodds
Withdrawn from ICLR 2027

CELLKAFC: KNOWLEDGE-AWARE SINGLE-CELL PERTURBATION PREDICTION WITH FLOW CORRECTION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-cell perturbation-response modeling is essential for decoding cellular mechanisms and accelerating drug discovery. However, predicting perturbation-induced gene expression distributions remains a formidable machine learning challenge: the data is inherently high-dimensional, sparse, and heterogeneous, and target conditions frequently exhibit substantial distribution shifts from cell distributions. While existing methods improve conditional prediction, they often underutilize gene-level biological priors and lack explicit mechanisms to correct condition-dependent distribution shifts. In this work, we introduce CellKAFC, a Knowledge-Aware Flow Correction framework for single-cell perturbation modeling. CellKAFC integrates biological priors into a conditional transport formulation and couples this process with an explicit correction for distributional mismatch across varying conditions. Through extensive evaluation on public benchmarks—including Norman, RPE1, and ComboSciPlex—CellKAFC consistently outperforms prior state-of-the-art methods. Notably, it reduces mean squared error by 12.5–40.5% relative to corresponding uncorrected baselines. Furthermore, an orthogonal decomposition analysis reveals that 21.7–42.4% of this net improvement stems from components orthogonal to the decoded dynamic space, validating the efficacy of our correction mechanism. These results demonstrate that knowledge-aware transport coupled with explicit shift correction provides a principled and highly transferable representation for single-cell perturbation modeling.

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

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_S7FrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study introduces CellKAFC, a framework for predicting single-cell gene-expression responses to unseen genetic and drug perturbations. It combines biological prior information with conditional flow matching in a low-dimensional effect space, then applies kernel-based residual correction to adjust the predicted population mean. Its main contribution is to separate cell-dependent response modeling from condition-level mean correction, supported by an analysis of prediction errors within and outside the flow model’s update space. The authors evaluate the framework on Norman, RPE1, and ComboSciPlex, covering single and combinatorial perturbations. They report improved mean-expression prediction accuracy over established baselines and provide component ablations, error-decomposition analyses, and selected biological case studies to examine the sources of these improvements.

Critical strengths

The method addresses a concrete limitation of low-dimensional prediction. A fixed decoder prevents the flow model from recovering expression changes outside its update space. CellKAFC introduces a separate correction in gene-expression space to address this limitation, providing a clear rationale for the design.

The correction mechanism is supported by an interpretable quantitative analysis. The orthogonal error decomposition and projected-correction comparison distinguish improvements within and outside the decoder space. This analysis helps explain where the gains come from and shows that the additional correction directions contribute to prediction accuracy.

The reported performance is competitive across multiple perturbation settings. The evaluation covers three public datasets, genetic and chemical interventions, and single and combinatorial perturbations. CellKAFC achieves the best results on all six reported metrics for Norman Additive and five of six for RPE1, supporting its practical value for mean-expression prediction.

The matched ablations provide evidence for the contributions of individual components. On RPE1, the authors compare variants under the same split, training budget, and three training seeds, refitting the residual correction for each applicable variant. These comparisons support incremental benefits from the biological priors and flow model while identifying residual correction as the largest source of improvement.

Critical weaknesses
  1. The reported benchmarks were used to select model configurations, limiting the evidence for generalization. The authors explicitly acknowledge selecting model configurations using the reported benchmarks (p. 9, lines 476–479). They also chose RPE1 for the main ablation study after inspecting the benchmark results, where the adopted configuration ranked best on five of six metrics (p. 18, lines 958–965). Because these results informed the selection of the reported configurations, they do not provide an independent assessment of generalization. The disclosure is welcome, but the claimed advantage needs to be evaluated on untouched data after the configurations and selection rules have been fixed.

  2. The main evaluation does not adequately assess differences between individual cells. All six main metrics use condition-level mean expression (p. 6, lines 284–291). These metrics only assess the average response. They cannot tell whether the model captures how responses differ from cell to cell. This matters because residual correction, which produces the largest reported gains, shifts every predicted cell by the same amount and leaves the differences between cells unchanged (p. 5, lines 243–248). Appendix E provides some additional distributional tests, but these do not cover all final model configurations and mainly compare the authors’ own Flow and Direct variants (p. 20, lines 1030–1040). The reported results therefore support more accurate average responses, while evidence that the final models capture cellular variation better than strong external baselines remains limited.

Questions for authors

Clarification on model selection: Which model choices were made after inspecting the reported benchmark results? Were any test results kept separate from all configuration-selection decisions? If so, please identify them, as they would provide independent evidence for the claimed generalization advantage.

Evidence for capturing cellular variation: Do the final model configurations capture differences between cells better than the strongest external baselines? Please provide any available comparisons that assess cellular variation separately from average expression. Appendix E mainly compares Flow and Direct and does not cover all final configurations, so it does not fully answer this question.

Reviewer_uy2gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies single-cell expression prediction under unseen genetic and chemical perturbations. CellKAFC uses perturbation priors, updates effect coordinates while keeping background coordinates fixed in a PCA–ICA representation, and corrects residual mean-expression errors through kernel regression. The method is evaluated on Norman, RPE1, and ComboSciPlex through benchmark comparisons, component ablations, and error decomposition.

Critical strengths
  1. Learning cellular responses in a low-dimensional space and correcting the remaining mean bias in gene space is a well-motivated design.

  2. The PCA–ICA background–effect decomposition is an interesting design. It offers a way to separate background and effect coordinates, allowing the flow model to condition on the fixed background while updating the effect coordinates.

Critical weaknesses
  1. The writing is uneasy to follow, and Figure 1 is AI-generated.
  2. The ablations do not clearly show whether the gains come from the PCA–ICA representation and background–effect split or simply from limiting model flexibility. This is my main concern about the methodological contribution.
Questions for authors
  1. How does performance change when PCA–ICA is removed entirely and prediction operates directly in gene expression space/latent space? (retaining the same prior conditioning, flow matching, and residual correction)
  2. Could the authors provide a biological interpretation of the PCA/ICA components? For example, are these components correlated with specific gene programs?
  3. CINEMA-OT[1] already combines PCA, FastICA, and Chatterjee-based component selection to separate confounding and treatment-associated variation. What specific changes does §3.2 introduce to these procedures?
  4. Following Systema[3], Miller et al.[2], and the VCC[4], could the authors report Pearson correlation relative to the training perturbation centroid, WMSE, and DE gene-set overlap to assess whether the reported gains translate into more accurate perturbation-specific responses and DE gene identification?

References

  1. Dong, M., Wang, B., Wei, J., de O. Fonseca, A. H., Perry, C. J., Frey, A., Ouerghi, F., Foxman, E. F., Ishizuka, J. J., Dhodapkar, R. M., & van Dijk, D. (2023). Causal identification of single-cell experimental perturbation effects with CINEMA-OT. Nature Methods, 20(11), 1769–1779. DOI: 10.1038/s41592-023-02040-5
  2. Miller, H. E., Mejia, G. M., Leblanc, F. J. A., Swain, B., Wang, B., & de Lima Camillo, L. P. (2026). Deep learning perturbation models can outperform baselines on calibrated metrics. Nature Biotechnology. Advance online publication. DOI: 10.1038/s41587-026-03307-w
  3. Viñas Torné, R., Wiatrak, M., Piran, Z., Fan, S., Jiang, L., Teichmann, S. A., Nitzan, M., & Brbić, M. (2026). Systema: A framework for evaluating genetic perturbation response prediction beyond systematic variation. Nature Biotechnology, 44(6), 1050–1059. DOI: 10.1038/s41587-025-02777-8
  4. The virtual cell challenge 2026 metric suite by Arc Institute.
Reviewer_gpqgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces CellFafc a framework for single cell perturbation prediction. CellFafc combines biological prior, conditional flow matching, and post hoc mean correction. The prior used in this model includes scGPT, co-expression features, ESM2, and target based features. It then learns a conditional flow in a PCA derived space and then improve the predictions using a residual correction mechanism. They evaluated their model on different datasets of genetic and chemical perturbations and show strong results against baselines in this field.

Critical strengths
  • The residual correction is simple but effective.
  • The paper shows strong empirical performance compare to the reported baselines.
Critical weaknesses
  • My main concern is that the paper is framing biological prior as an important part of the framework while its contribution appears substantially weaker than what it is framing. I appreciate the ablation study that shows the contribution of prior knowledge however, a major motivation in the paper is that existing perturbation models do not use biological prior information and the prior fusion is presented as a major contribution while the ablation study do not support such claims. For example, on RPE1, removing the expression graph or ESM2 only modestly worsens performance, while removing the residual correction causes a much larger effect. Similarly, the comparison involving scGPT, graph, and ESM2 shows very small differences between prior configurations.
  • Another concern is that the residual correction may be doing most of the work and in that case while this is a useful information it is only a post hoc mean correction and makes the contribution of the paper less significant.
  • Residual correction relies on similarity between a query perturbation and training perturbations. This makes sense when nearby perturbations exist in the learned representation, but it is less clear how the approach behaves for genuinely distant OOD perturbations and the paper is not showing results for this.
  • While the framework shows results for both molecular and genetic perturbations, the chemical perturbation evaluation is quite small. The ComboSciPlex experiment contains only seven held-out conditions (two single drugs and five drug combinations).
  • Although CellKAFC is presented as a single-cell distribution prediction model, most metrics are calculated from mean expression. I would therefore prefer the paper to distinguish more clearly between improvements in mean response prediction and improvements in modeling heterogeneous single-cell responses.
Questions for authors
  • Can the authors provide a simpler baseline using the same flow and correction framework but without biological priors?

  • Why does prior fusion constitute a central contribution when the prior ablations appear much smaller than the effect of residual correction?

  • How does the method perform on a larger unseen-drug benchmark (such as Tahoe) with substantially more held-out molecules?

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

Related papers

Discussion (0)

Sign in to comment.