- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3(+1)
Read review
Paper summary
The paper introduces a novel framework, named scCasFlow, that aims to improve upon flow matching-based single-cell perturbation prediction approaches by explicitly modeling intermediate regulatory steps in its generative process. Specifically, scCasFlow aims to first predict population-level perturbation responses and then generates single-cell gene expression profiles conditioned on these predictions, treating the two as separately parameterized flows. The authors demonstrate that scCasFlow outperforms several other perturbation prediction models in generalization to unseen genes and perturbation combinations across a large battery of evaluation metrics. They also perform several ablation analyses to confirm the merits of leveraging intermediate supervision as well as various design choices aimed at formulating the regulatory cascade and corresponding generation procedure.
Critical strengths
- The paper is clearly written and does a good job of contextualizing its contributions with respect to the broader literature on single-cell perturbation prediction methodologies. The Methodology section is written in a manner that is easy to understand, and the related modeling decisions appear to be generally accompanied by convincing rationales.
- The authors leverage a comprehensive suite of evaluation metrics to assess the performance of scCasFlow against a number of state-of-the-art models, demonstrating consistently better performance than all benchmarked models. The diversity of the evaluation metrics (covering reconstruction, differential expression, and discrimination tasks) provides strong evidence for scCasFlow’s outperformance in the evaluated settings.
- The authors introduce a number of creative ablations that help to assess the advantages of the various modeling choices used in scCasFlow. The results of these ablation studies provide support for the utility of the core components of this framework (e.g., use of intermediate flow, intermediate supervision, generation order).
Critical weaknesses
- It seems as if a cell’s intermediate response representation is defined as the difference between the Stack representations of a control state and a corresponding steady-state, post-perturbation state. It’s not clear that this framing is appropriate, as the post-perturbation Stack representation more closely mirrors the downstream gene expression target than an intermediate state.
- While the paper includes a number of ablation studies that appear to show the benefits of various model design choices, the evaluation metrics used in these analyses only cover a fraction of those that are used in earlier parts of the paper. In addition, the ablation studies focus only on the Norman-Additive dataset.
- A large portion of the Appendix is not cited in the main text of the paper. For example, Table S2 includes what appears to be additional benchmarking results that supplement Table 1’s results, but scCasFlow’s performance on the single perturbation setting is not as consistently strong relative to other baseline methods, especially on the MSE and MAE metrics.
Questions for authors
- Is the intermediate response representation appropriately defined, given that it captures the difference between the final post-perturbation state and a control state?
- What is the rationale behind selecting the Stack encoder to be used for representing intermediate response, while the Stack tokenizer is used for representing expression-flow states?
- How do the various ablations affect the broader set of evaluation metrics used in Table 1? Which modeling decisions are most beneficial for each category of metrics? How do the ablation results look for datasets besides the Norman-Additive dataset?
- Table S2 appears to indicate that scCasFlow’s outperformance of other methods is larger and more robust in the double perturbation setting compared to the single perturbation setting. Can the authors discuss this result?
4 comments
Q3. How do the various ablations affect the broader set of evaluation metrics used in Table 1? Which modeling decisions are most beneficial for each category of metrics? How do the ablation results look for datasets besides the Norman-Additive dataset?
We thank the reviewer for pointing this out. As shown in Table 1, we expanded the ablation evaluation on Norman-Additive to include additional metrics, covering expression reconstruction (MSE and MAE), perturbation discrimination (DS), perturbation-effect correlations (P20 and P50), and differential-expression directional agreement (Dir).
We initially focused on PΔ because it directly measures the agreement between predicted and observed perturbation-induced gene-expression changes, which is central to our task. Unlike expression-level errors such as MSE and MAE, it focuses on perturbation effects rather than overall expression reconstruction. Other metrics provide complementary assessments, but their differences are generally less pronounced or more metric-dependent for some ablations.
Table 1. Additional ablation metrics on Norman-Additive.
| Method | MSE ↓ | MAE ↓ | DS ↑ | P20 ↑ | P50 ↑ | Dir ↑ |
|---|---|---|---|---|---|---|
| Monolithic Flow | 0.000919 | 0.013578 | 0.9355 | 0.8954 | 0.8977 | 0.9400 |
| Deterministic Rewire | 0.000731 | 0.012060 | 0.9477 | 0.9137 | 0.9193 | 0.9481 |
| Decoupled Decoder | 0.000817 | 0.012806 | 0.9467 | 0.9107 | 0.9077 | 0.9165 |
| w/o Tokenizer | 0.005318 | 0.028159 | 0.7399 | 0.7860 | 0.7477 | 0.8750 |
| scCasFlow | 0.000683 | 0.011581 | 0.9646 | 0.9296 | 0.9374 | 0.9587 |
We further evaluated the same ablation variants on Norman-Holdout to examine whether these findings transfer to another evaluation setting. The results (Table 2) on Norman-Holdout show that scCasFlow achieves the best results on most metrics. Overall, these results support the contributions of the proposed components beyond Norman-Additive.
Table 2. Ablation on Norman-Holdout.
| Method | MSE ↓ | MAE ↓ | DS ↑ | PΔ ↑ | P20 ↑ | P50 ↑ | DE LFC ↑ | Dir ↑ | O50 ↑ |
|---|---|---|---|---|---|---|---|---|---|
| Monolithic Flow | 0.001443 | 0.014732 | 0.6911 | 0.5773 | 0.6630 | 0.6843 | 0.6484 | 0.8534 | 0.0487 |
| Deterministic Rewire | 0.001336 | 0.014436 | 0.7184 | 0.5997 | 0.6774 | 0.6925 | 0.6621 | 0.8573 | 0.0507 |
| Decoupled Decoder | 0.001588 | 0.016710 | 0.6803 | 0.5540 | 0.6743 | 0.6904 | 0.5895 | 0.8230 | 0.0483 |
| w/o Tokenizer | 0.001344 | 0.013808 | 0.6211 | 0.5764 | 0.6661 | 0.6744 | 0.6205 | 0.8349 | 0.0703 |
| scCasFlow | 0.001326 | 0.014201 | 0.7436 | 0.6054 | 0.7175 | 0.7078 | 0.6671 | 0.8462 | 0.0800 |
Q4. Table S2 appears to indicate that scCasFlow’s outperformance of other methods is larger and more robust in the double-perturbation setting compared to the single-perturbation setting. Can the authors discuss this result?
We thank the reviewer for pointing out this observation. As shown in Table S2, scCasFlow achieves more consistent improvements in the double-perturbation setting, outperforming all evaluated methods across all nine metrics. In contrast, its improvements in the single-perturbation setting are more modest, with CellFlow achieving lower MSE and MAE. This shows that single-gene perturbation responses are relatively easier to capture using existing predictive models, whereas double perturbations involve more complex combinatorial effects. The population-level intermediate supervision and ordered generation in scCasFlow may provide a more useful inductive bias in these challenging settings by explicitly modeling perturbation-associated changes before generating expression responses.
Thank you to the authors for their thorough responses to my questions, the additional ablation experiments, and the clearer interpretation of the results. These responses clearly describe the modeling design choices and have strengthened my confidence in the work. I have no further comments and have updated my score to a 4.
Thank you again for your thoughtful feedback and for taking the time to reassess our work.
We thank the reviewer for the constructive feedback and helpful suggestions.
Q1. Is the intermediate response representation appropriately defined, given that it captures the difference between the final post-perturbation state and a control state?
We thank the reviewer for this valuable comment. The intermediate is not the difference between paired control and perturbed states of individual cells. Instead, it represents a perturbation-induced population-level shift in pretrained Stack embeddings, estimated from randomly sampled perturbed-cell cohorts relative to the control population (Eq. 5). The cells used to construct this target are sampled independently of those used for individual-cell expression generation.
Importantly, this intermediate is a predictive response representation, not a directly observed biological intermediate state. Although derived from post-perturbation transcriptomic measurements during training, it does not directly identify upstream regulatory mechanisms. Its purpose is to separate the prediction of population-level perturbation effects from the generation of individual-cell expression profiles. At inference, it is predicted solely from control-cell information and perturbation identity, without access to post-perturbation measurements.
Our experiments support this formulation. Shuffling and randomizing intermediate targets reduce PΔ from 0.8488 to 0.7819 and 0.7395, respectively (original manuscript Table 4a). Removing the intermediate flow or using simultaneous generation also reduces performance (original manuscript Table 4b). Furthermore, the associations between learned Rewire representations and inferred TF activities (Fig. 5) provide additional support. These results support the intermediate as a useful perturbation-associated response surrogate rather than a direct measurement of causal regulatory rewiring.
Q2. What is the rationale behind selecting the Stack encoder to be used for representing intermediate response, while the Stack tokenizer is used for representing expression-flow states?
We appreciate the reviewer’s thoughtful feedback. The two components serve distinct purposes. The Stack encoder produces contextualized cell representations suitable for estimating population-level perturbation shifts. We therefore use its embeddings to construct the intermediate supervision. In contrast, the Stack tokenizer projects gene expression into compact gene-module tokens, preserving expression information while reducing the dimensionality of the generative process. These tokens are used by the Express flow, with a jointly trained decoder mapping the generated latents back to gene expression.
This separation follows our modeling formulation: the intermediate representation captures population-level perturbation effects, while the expression flow generates individual-cell responses. The shuffled and random target controls (original manuscript Table 4a) support the encoder-derived intermediate supervision, while the tokenizer and decoder ablations (original manuscript Fig. 4) support the latent-space expression generation design.