acceptodds
Withdrawn from ICLR 2027

ActFlow: Activation-Aware Flow Matching for Single-Cell Perturbation Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Single-cell perturbation prediction is central to building virtual cell models for gene function analysis and drug discovery. However, single-cell expression profiles are highly sparse, with most genes inactive in each cell, causing conventional regression objectives to be dominated by zero-expression entries. This may obscure sparse yet biologically informative activation signals. To address this challenge, we propose ActFlow, a diffusion transformer-based framework that models perturbation-induced cellular transitions via continuous-time flow matching. During training, ActFlow constructs a cell-specific mask for each perturbed cell and jointly learns two velocity fields within a unified model: a target velocity field that predicts velocities in gene expression space under perturbation conditions, and an auxiliary mask velocity field that models gene activation states. The mask velocity field contributes activation-aware supervision exclusively during training, encouraging the model to focus on biologically relevant activated genes; at inference, only the target expression trajectory is retained as the predicted output. Experiments on genetic and drug perturbation benchmarks show that ActFlow consistently outperforms state-of-the-art methods across both global reconstruction and distribution-level metrics, demonstrating that explicitly modeling gene activation states provides an effective inductive bias for single-cell perturbation prediction.

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

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 4 reviews, above 27% of the others.
Reviewer_QbFcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ActFlow predicts single-cell perturbation responses by jointly learning expression and activation-mask velocity fields within a shared DiT backbone. The masks preserve cell-specific nonzero expression patterns, with positive values replaced by perturbation–gene medians computed from training data. Training combines expression and mask velocity losses with MMD regularization. At inference, both trajectories evolve jointly, while only the expression output is retained. Experiments on Norman and ComboSciPlex demonstrate competitive performance, particularly on distributional metrics.

Critical strengths
  • The paper addresses an important problem and explores a useful approach to modeling sparse single-cell expression data.
  • Evaluation covers both genetic and drug perturbations, using complementary metrics for mean reconstruction, perturbation effects, distributional agreement, and differentially expressed genes.
  • Improvements on distributional metrics are substantial. Four random splits and condition-level significance tests on Norman provide additional support for the reported results.
Critical weaknesses

W1. Insufficient comparison with directly related methods.

Conditional flow matching and MMD have already been used in scDFM, while mask-based modeling of sparse single-cell expression has been explored in Doloris. The contribution requiring validation is therefore the shared-backbone joint generation of expression and masks, together with the proposed mask construction.

Although the paper discusses Doloris and Departures, neither is included in the experimental comparison. These methods directly address the sparsity problem motivating ActFlow. Their omission makes it difficult to establish the advantages of the proposed joint-generation design over existing masking approaches. The current results demonstrate improvements over the selected baselines but do not fully establish the additional value of the central design.

W2. The connection between the stated problem and the proposed mechanism lacks direct evidence.

The paper largely motivates its approach through the high proportion of zero-expression entries, arguing that these entries dominate training and obscure informative perturbation signals. However, under its flow-matching formulation, the target velocity for an entry with �0��0� is �1��1�, rather than zero. The frequency of zero-expression entries alone therefore does not establish that they dominate the velocity-regression loss or parameter gradients. Direct analysis of their optimization contributions is needed.

Table 3 shows meaningful improvements in distributional metrics from the mask branch, but the ablation covers only one additive split, leaving the robustness of this effect insufficiently established. Analyzing whether the mask trajectory recovers reliable structure earlier during generation would also help explain the value of joint evolution.

Furthermore, �2��2� indicates detected expression, not necessarily perturbation-induced activation. Constitutively expressed genes can be nonzero, and perturbation responses also include downregulation and silencing. The paper should distinguish expression-detection modeling from perturbation-induced activation modeling and moderate its biological interpretation accordingly.

W3. Missing task-specific simple baselines.

In the Norman additive setting, the constituent single-gene perturbations of each test combination are observed during training. An explicit additive baseline is therefore applicable: μa+b=μa+μb−μctrl\mu_{a+b}=\mu_a+\mu_b-\mu_{ctrl} This comparison would directly test whether ActFlow improves upon adding known single-gene effects. A training-set perturbed-mean baseline would also help assess whether improvements in mean reconstruction and perturbation-effect metrics depend on condition-specific predictions. The importance of such simple baselines is supported by Ahlmann-Eltze et al. (2025).

Although the paper includes a bilinear Linear baseline, it does not sufficiently explain whether its representations and prediction structure cover the explicit additive formulation. The authors should clarify this equivalence or report the additional baseline. Comparisons on distributional metrics should also distinguish deterministic mean predictions from generated cell populations.

Questions for authors

Q1 Related methods. How does ActFlow compare with Doloris and Departures under the same data splits and evaluation protocol? What advantages does the shared-backbone joint-generation design offer over their approaches to modeling sparse expression?

Q2 Motivation and mechanism. Can the authors quantify the contributions of zero- and nonzero-expression entries to the velocity-regression loss and parameter gradients? Does this analysis support the claim that zero-expression entries dominate optimization?

Q3 Robustness and trajectory behavior. Do the improvements from the mask branch persist across additional random splits? Is there evidence that the mask trajectory recovers reliable activation structure earlier than the expression trajectory, thereby providing useful guidance during generation?

Q4 Biological interpretation. Since �0��0� measures detected expression rather than perturbation-induced activation, what evidence supports interpreting the learned masks as perturbation-specific activation signals rather than general expression patterns?

Q5 Simple baselines. Does the bilinear Linear baseline cover the explicit prediction �1��1�? If not, how does ActFlow compare with this additive baseline on Norman’s additive setting, and with a training-set perturbed-mean baseline on mean reconstruction and perturbation-effect metrics?

Reviewer_S1dnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ActFlow, a conditional flow-matching model based on a Diffusion Transformer (DiT) designed for single-cell perturbation prediction. Its core design involves constructing a "cell-specific activation mask" for each perturbed cell, which indicates whether the expression level of a given gene exceeds zero. The model takes three sets of gene-level tokens—control cells, noisy masks, and noisy target transcriptomes—as input into a single DiT, with perturbation information injected via cross-attention. The model simultaneously learns both the target velocity and the mask velocity. The objective function primarily consists of a target velocity loss, a mask velocity loss (with a cosine annealing schedule), and a Maximum Mean Discrepancy (MMD) term.

Critical strengths

The dominance of sparsity in training signals is a well-documented challenge in single-cell generative modeling. Rather than introducing a separate gating model for the inference phase, ActFlow incorporates activation information as an auxiliary trajectory for joint generation. This avoids the risk inherent in two-stage approaches, where errors in mask prediction could inadvertently erase genuine biological signals.

Critical weaknesses
  • It is unclear whether the performance gains are truly attributable to the mask design. The paper asserts that a binary {0, 1} mask is unsuitable, yet it fails to provide an ablation study to support this claim. Furthermore, given that the model generates an activation mask, why is this mask not used to gate the expression output during inference (as is done in methods like SDD)?
  • Metrics such as E-distance and EMD are highly sensitive to the mass concentrated at zero. Since the paper does not apply post-hoc sparsification (e.g., thresholding) to the baselines, it is difficult to determine how much of the improvement in distributional metrics stems from better modeling of the perturbation response itself.
  • It is plausible that zero-valued entries dominate the training loss, leading to insufficient learning of active genes. However, the proposed method does not directly address this issue; the "zero pattern" of the mask is identical to that of the representation itself, and its auxiliary loss is likewise dominated by zero-valued entries, failing to reweight the active genes. Furthermore, the mask is discarded during inference, and there is no mechanism in place to ensure exact zeros in the output. Since the mask is a deterministic function of the target representation and provides no new information, the source of its effectiveness remains unanalyzed.
Questions for authors

See my Critical Weaknesses.

Reviewer_Gcx3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces ActFlow, a DiT-based framework that models single-cell perturbation with continuous-time flow matching. Compared to most existing methods, it features an additional velocity field that predicts gene activation states. The authors claim this new velocity field helps to mitigate the negative impact of the high sparsity of the single-cell expression data, and justify their claim through notable improvement across two datasets (Norman and ComboSciPlex).

Critical strengths
  • S1. The paper is clearly written and very easy to follow. Readers can quickly understand the main research gap that the authors wish to close and see the results that supports the author's claims. The research gap discussed in this paper, high sparsity of single-cell data, is indeed an important existing gap that many wish to close. The detailed formula and confirgurations are also presented nicely.
Critical weaknesses
  • W1. The baseline numbers, especially for scDFM, are much worse than the numbers published for the same setting, and this changes the main conclusion. Appendix A.4 says ActFlow uses "the same processed expression matrices, preprocessing procedures, and data splits as scDFM", and the Control rows of the two papers are almost identical (Norman additive L2 3.9837 vs 3.9937, PDS 0.5135 vs 0.5135), so the evaluation pipeline appears to be the same.

    However, on the Norman additive setting (Table 1), this paper reports scDFM at L2 2.5815, MSE 0.0069 and Pearson Δ\Delta 0.7977, while the scDFM paper ([1], Table 1) reports L2 1.7043, MSE 0.00315 and Pearson Δ\Delta 0.8853. ActFlow's own numbers (L2 1.7620, MSE 0.0034, Pearson Δ\Delta 0.8850) are about equal to, or slightly worse than, the published scDFM numbers. Moreover, the same pattern holds in the holdout setting (Table 1). For single-gene perturbations, the reproduced scDFM L2 (2.6424) is even worse than Control (2.6308), while the published ([1], Table 2) scDFM L2 is 1.6186 vs ActFlow's 1.6170. For double-gene perturbations, ActFlow's L2 (2.1608) is worse than the published scDFM (2.0309). CellFlow shows the same gap: L2 3.7186 here vs 1.7064 in the scDFM paper ([1], Table 1). Finally, the Additive baseline reported in the scDFM paper (Pearson Δ\Delta 0.9024) is not included ([1], Table 1), although it beats ActFlow on Pearson Δ\Delta.

    Without explaining these gaps, it is hard to say that ActFlow outperforms state-of-the-art methods on reconstruction metrics.

    [1] scDFM https://arxiv.org/pdf/2602.07103

Questions for authors
  • Why are the reproduced scDFM and CellFlow results much worse than those reported in the scDFM paper under what is described as the same preprocessing, splits and evaluation genes? Can you report ActFlow next to the published scDFM numbers, or rerun scDFM with its released configuration (instead of batch size 48 / 100k steps), and include the Additive baseline?
Reviewer_3icSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ActFlow for single-cell perturbation prediction. It uses a shared transformer to jointly learn flow-matching trajectories for gene expression and an auxiliary activation mask. The mask preserves nonzero expression patterns and replaces positive values with perturbation–gene medians. During inference, both trajectories are generated, but only the expression output is used. Experiments on Norman and ComboSciPlex show improvements over existing baselines, especially on distribution-level metrics.

Critical strengths

Strong performance on distribution-level prediction: ActFlow achieves good results on both genetic and drug perturbation datasets, especially on E-distance and EMD. The main results include multiple data splits and statistical tests, which provide useful evidence for the reported improvements over the selected baselines.

Critical weaknesses
  1. The experiments do not clearly establish the benefit of the proposed joint expression–mask framework.

The main motivation is that zero-expression entries dominate training and that activation-aware supervision can address this problem. However, the paper does not analyze whether zeros actually dominate the training loss. Also, the proposed mask in Eq. (11) represents observed nonzero expression, which is not necessarily perturbation-induced activation.

The ablation study does not fully explain where the improvements come from. Removing the mask branch also removes the additional input states and tokens. Therefore, it is unclear whether the gains come from activation modeling or simply from additional auxiliary learning. A matched auxiliary-target baseline would help answer this question.

The benefit of the specific mask construction is also limited. In Table 3, the simplified mask already achieves about 76% of the full E-distance improvement and 91% of the EMD improvement over the no-mask model. These results are reported on only one additive split, without uncertainty estimates. It is also unclear whether the simplified-mask median includes zero entries, which could make the target identically zero.

Furthermore, Appendix A.9 does not clearly explain how generated masks are matched with observed cells to compute the activation prediction metrics. The comparison with only a global random baseline is also insufficient, since a model could achieve good results by learning gene-specific expression frequencies.

Overall, the results show that the auxiliary branch is useful, but they do not clearly establish that the proposed activation modeling is responsible for the improvements.

  1. The paper does not establish a clear advantage over closely related activation-aware methods.

The paper motivates ActFlow by discussing the limitations of existing masking approaches. However, Doloris [1] and Departures [2], which are directly related to this problem, are not included in the experiments. The current comparisons with CellFlow, scDFM, and PerturbDiff do not show whether ActFlow improves over methods that already model gene activation or expression sparsity.

The novelty also needs to be explained more carefully. Departures already jointly trains expression and activation models. Therefore, jointly learning these two components is not entirely new. ActFlow differs by using a shared backbone, a continuous mask trajectory, and no final multiplicative masking. These are useful design differences, but the experiments do not show whether they provide better performance than existing activation-aware approaches.

Without comparisons under the same data splits and evaluation settings, it is difficult to assess the actual contribution of ActFlow over these closely related methods. This is a main concern for my assessment.

[1] Chi et al. Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation. ICLR 2026.

[2] Chi et al. Departures: Distributional Transport for Single-Cell Perturbation Prediction with Neural Schrödinger Bridges. AAAI 2026.

Questions for authors
  1. How is the activation evaluation in Table 6 performed? Please clarify whether observed target expression is used, how generated masks are matched with observed cells, and how the scores are aggregated. Can the authors also compare against a simple baseline that predicts each gene's detection frequency estimated from the training data, without using perturbation information?

  2. What evidence shows that zero-expression entries dominate the velocity loss or gradients? Would a zero/nonzero-balanced expression loss without the mask branch achieve similar improvements?

  3. Can the authors compare the proposed mask with alternative auxiliary targets using the same backbone, token count, and training budget? In Table 3, the simplified mask already retains about 76% of the DE E-distance improvement and 91% of the DE EMD improvement. Are these additional gains consistent across splits? Also, does the simplified-mask median include zero entries?

  4. Can the authors compare ActFlow with Dolori [1] and Departures [2] under matched evaluation settings? Since Departures already jointly trains expression and activation models, what is the specific advantage of ActFlow?

  5. Can held-out Norman perturbation genes appear in other training combinations? Were the 1,000 evaluation genes selected using test expression data, and were test-condition evaluations used for checkpoint selection or hyperparameter tuning?

  6. Can the authors explain why Table 1 reports scDFM holdout L2 errors of 2.6424 for single-gene and 2.8532 for double-gene perturbations, while Table 4 reports a pooled error of 2.2794? Please provide per-split results and clarify any differences in data splits, gene sets, or aggregation.

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

Related papers

Discussion (0)

Sign in to comment.