acceptodds
Withdrawn from ICLR 2027

Mechanism-Grounded Flow Matching for Interpretable Single-Cell Perturbation Prediction with Adaptive Transport Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-cell perturbation prediction is essential for systems biology and therapeutic discovery, where interpretability is crucial for both grounding transcriptional responses in biological mechanisms and understanding how mechanisms shape cellular state transitions. Recently, flow matching has emerged as an effective approach for modeling continuous transitions between cellular states. However, existing methods primarily use it to predict perturbed expression profiles, without explicitly grounding the learned transitions in underlying biological mechanisms, thus limiting interpretability. To solve this problem, in this paper, we propose MechFlow, a mechanism-grounded flow matching framework that adaptively models the dynamics of transport from untreated to perturbed states. Specifically, we propose three modules: target-anchored mechanism grounding for perturbation drift propagates perturbation effects through directed biological graphs to derive a mechanism drive toward transcriptional changes, termed the perturbation drift; graph metric adaptation for state-responsive cellular geometry calculates a graph metric to define an adaptive geometry that assigns different costs to directions of transcriptional change according to gene associations, and transport dynamics for interpretable distribution flow integrates the drift and graph geometry to construct interpretable transition flows for supervising velocity field learning in flow matching. Together, these modules enable MechFlow to ground predicted transcriptional responses in biological mechanisms and characterize how they unfold along cellular state transitions. Experiments on Tahoe-100M show that MechFlow achieves not only strong predictive performance under various settings but also interpretability at both the response and transition levels.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9uqyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce a method called MechFlow. It combines three modules, i.e. a directed perturbation graph, an adaptive gene-graph metric, and flow matching to predict untreated-to-treated single-cell distributions. In the experiments, they show that on Tahoe-100M, it improves metrics for unseen perturbations and unseen cellular contexts, and also provide some interpretability.

Critical strengths

The empirical performance of the method seems good on the two Tahoe exps, compared to the simple baselines and an FM baseline. The ablations in Tab 6 and 7 show that corrupting target priors or graph topology reduces effect prediction, so it seems that structured inputs drive the sources of gain.

Critical weaknesses

In Tab 6, the ablations are based on the 512-condition proxy, but in Tab 1 the full model is evaluated on 2998 conditions. Regarding the point on interpretability, this is rather descriptive. What I mean by that is that the analysis is mostly based on graph routes and flow patterns, but there is no independent biological evidence.

Questions for authors
  1. See the first point in Weaknesses.
  2. Can the selected target/pathway routes be checked against independent perturbation evidence?
Reviewer_XDuhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MechFlow, a flow-matching model that incorporates biological priors. It aims to explain changes in cell states under drug perturbations by learning the velocity field of these states, while also providing biologically interpretable insights. The authors conducted experiments on the Tahoe100M dataset and compared the model against other methods, demonstrating its advantages in predicting cell fate and elucidating underlying mechanisms.

Critical strengths
  1. The authors incorporate biological priors and establish a connection with the learned velocity fields. Experiments demonstrate a competitive result of this approach in perturbation predicting and interpreting underlying mechanisms.

  2. On the Tahoe100M subset used in the study, the method yields competitive results. In the P-holdout setting, it outperforms other methods across all four perturbation-effect metrics as well as MMD²; in the C-holdout setting, it achieves superior performance across all six metrics.

Critical weaknesses
  1. The experiments cover only a subset of Tahoe-100M, without other validation on independent drug-perturbation datasets. It therefore remains unclear whether the reported performance advantages generalize across datasets. Moreover, the ablation results show relatively small performance differences, which do not sufficiently establish the necessity of individual components in this complex framework.

  2. Interpretability is a central contribution of this work, but the current analyses primarily rely on visualizations of model-generated trajectories and velocity fields, without independent evidence supporting the biological relevance of the results. Furthermore, the paper does not show whether mechanistic conclusions remain consistent across independent training runs with different random seeds.

Questions for authors

The following questions arise from the weaknesses above, and I would appreciate clarification from the authors.

  1. Do the performance advantages persist on independent drug-perturbation datasets such as sci-Plex? I suggest additional experiments on other datasets to substantiate the findings.

  2. What independent biological evidence supports the “early” and “late” pathway changes shown in Figure 3? Presenting these patterns alone is insufficient to establish the reliability of the inferred biological mechanisms.

  3. Given that mechanistic interpretability is a central contribution of this method, which biological interpretations remain stable when the model is retrained with different random seeds? Can their consistency or uncertainty be quantified?

  4. The current ablation study does not sufficiently establish the necessity and role of each module. Could the authors jointly ablate two or more modules to identify which components contribute most to performance, or design alternative experiments that better establish the role of each module?

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

The paper proposes MechFlow for predicting single-cell responses to perturbations. It combines a graph-based mechanism drift with a context- and state-dependent metric in a two-stage flow-matching framework. Results on held-out drugs and cell lines are promising.

Critical strengths
  1. The evaluation includes unseen drugs and cell line context (both hard and useful tasks), with promising results against both simple and recent baselines.
  2. The model exposes gene-level drift and transport quantities that are useful for inspecting its predictions.
Critical weaknesses
  1. The paper is difficult to follow in some places. Several concepts and variables are introduced before their motivation is clear, such as the state-dependent susceptibility in Section 3.3 and the residual tangent in Section 3.4. It is also unclear why these particular designs are needed.
  2. The interpretability experiments mainly visualize model outputs. They provide little independent evidence that the selected entry nodes, drift decomposition, or pathway dynamics are biologically meaningful.
Questions for authors
  1. Why is the graph-geometric cost applied to the residual velocity ut−mp,cu_t - m_{p,c}, rather than the total velocity utu_t? Could you clarify the motivation and compare against applying the cost to utu_t? It would also be useful to remove Stage A entirely and train the same prediction architecture on straight interpolation paths.

  2. In Stage A, the path and mechanism drift are jointly learned, so both sides of the residual ut−mp,cu_t - m_{p,c} change during training. Could this lead to unstable optimization? Have you compared joint training with freezing one component or alternating their updates?

  3. Section 3.3 refers to Appendix C.2 for the parameterization of ℓi(x,c)\ell_i(x,c), but it seems to be missing. Could you provide the complete formula?

  4. Why are chemCPA and PerturbDiff reported only for P-holdout, while STATE, CellFlow, and scDFM appear only for C-holdout? Please specify the limitation preventing each omitted evaluation.

  5. As noted in Weakness 2, the interpretability analyses mainly describe model outputs. Could you validate at least one key interpretation using independent biological evidence?

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

Related papers

Discussion (0)

Sign in to comment.