acceptodds
Under review as a conference paper at ICLR 2027

SDFM: Cross-Aware Knowledge-Conditioned Flow Matching for Single-Cell Genetic Perturbation Prediction

10 views

Abstract

Genetic perturbation screens reveal how interventions reshape cell populations, but their limited coverage leaves much of the intervention space unexplored. Predicting unmeasured responses requires generalization across perturbations while preserving cellular diversity. Flow matching provides a natural framework for modeling such population-level transformations, yet its local velocity objective only indirectly supervises the responses induced by complete generative trajectories. We propose SDFM, a cross-aware knowledge-conditioned flow matching framework for perturbation response prediction. Knowledge graph representations from STRING and Gene Ontology condition the velocity field through a cross-aware perturbation representation, allowing related perturbations and combinatorial contexts to influence learned transport dynamics. Mini-batch optimal transport establishes population-level correspondences between unpaired control and perturbed cells, while full-rollout supervision connects local velocity learning with predicted endpoints and population expression shifts. Experiments on Adamson, HepG2, and Norman show that SDFM achieves the highest mean all-gene expression change Pearson correlation among evaluated methods across the three single-gene benchmarks and improves combinatorial response prediction. It also reduces mean sliced Wasserstein and energy distances relative to a uniform mean-shift baseline, supporting its ability to model both average transcriptional effects and heterogeneous population-level responses.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell genomics, Perturbation prediction, Flow matching, Optimal transport, Knowledge-guided generative modeling, Biological machine learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ELFxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SDFM, a knowledge-conditioned flow-matching framework for predicting single-cell genetic perturbation responses. It incorporates perturbation representations derived from STRING and Gene Ontology, multi-head cross-attention conditioning, an interaction-aware composer for multi-gene perturbations, and mini-batch optimal-transport pairing. Beyond local velocity matching, its training objective adds one-step endpoint-based population supervision and differentiable rollout losses. The work validates the method on single-gene and combinatorial perturbation benchmarks with component ablation studies.

Critical strengths
  1. Strong benchmark improvements with substantial gains across most evaluation metrics.

  2. Good reproducibility. Training configurations are well documented, and the released code enables straightforward pipeline reproduction.

Critical weaknesses
  1. Limited methodological novelty. The rationalization for integrating different components is insufficient; the main contribution lies in combining existing techniques.

  2. Aggregate improvements mask the distribution of benefits from the knowledge graph. A subset of perturbations may see degraded performance, limiting the generalizability of the conclusions.

Questions for authors

This work presents a promising framework for single-cell genetic perturbation prediction and demonstrates meaningful benchmark gains, yet I have three main concerns regarding interpretation of the results:

  1. Clarification on Teacher-Δ: The teacher endpoint uses xt+(1−t)vθ(xt,t,… )x_t+(1-t)v_\theta(x_t,t,\dots). This formulation follows the ideal straight conditional path, yet the learned velocity field varies along the generated trajectory. Could the mismatch between this one-step extrapolation and the integrated endpoint introduce systematic bias?

  2. Please report the distribution of gains from knowledge conditioning, including the fraction of test perturbations with improved/worsened predictions and per-perturbation gain statistics. This helps evaluate whether the reported average improvement transfers to genes with limited prior annotations.

  3. How is the default top-20 edge setting selected? Are the main conclusions stable under sparser or denser graph constructions?

5 comments
AuthorsOpenReview ↗

Q1. The Teacher-Δ endpoint is obtained through a one-step extrapolation along the ideal straight conditional path. However, the learned velocity field may vary along the generated trajectory. Could the mismatch between this one-step approximation and the fully integrated ODE endpoint introduce systematic bias?

R1. We thank the reviewer for raising this important question. We agree that the one-step Teacher-Δ endpoint is an approximation and may differ from the endpoint obtained through full ODE integration when the learned velocity field varies along the trajectory.

Specifically, Teacher-Δ estimates the endpoint from an intermediate state as:

x~p=ρ(xt+(1−t)vθ(xt,t,xc,ep))\tilde{x}_p = \rho(x_t + (1-t)v_\theta(x_t,t,x_c,e_p))

Under the ideal conditional velocity field, where the velocity is constant along the straight conditional path, this extrapolation recovers the exact endpoint. However, for a learned state- and time-dependent velocity field, the one-step estimate may introduce approximation errors.

Importantly, SDFM does not treat the Teacher-Δ endpoint as the actual generated endpoint. Instead, it serves as an auxiliary population-level supervision signal. The complete ODE trajectory is independently constrained through two additional objectives:

  • Rollout-Δ loss: Aligns the population-level expression shift obtained from full ODE integration with the observed perturbation response.
  • Endpoint Reconstruction loss: Further constrains the integrated endpoints using paired endpoint supervision.

Therefore, the overall objective combines local velocity matching, approximate teacher endpoint supervision, and supervision based on the complete integrated trajectory.

We acknowledge that the one-step approximation may introduce local discrepancies, but the additional rollout-based objectives explicitly account for the learned velocity dynamics throughout the integration process. We will clarify this distinction in the revised manuscript to avoid implying that the one-step teacher endpoint is identical to the actual ODE-generated endpoint.

AuthorsOpenReview ↗

Q2. Please report the distribution of gains from knowledge conditioning, including the fraction of test perturbations with improved or worsened predictions and per-perturbation gain statistics. This would help determine whether the benefits generalize to perturbations involving genes with limited prior annotations.

R2. We thank the reviewer for this insightful suggestion. We agree that aggregate performance alone may obscure the heterogeneous effects of biological knowledge conditioning across individual perturbations.

To address this concern, we analyzed the per-perturbation performance changes associated with knowledge conditioning across the three datasets. The results are summarized below:

DatasetTest PerturbationsImprovedWorsenedMean GainMedian Gain
Adamson2218 (81.82%)4 (18.18%)0.106530.06145
HepG2594376 (63.30%)218 (36.70%)0.046310.04100
Norman10755 (51.40%)52 (48.60%)0.032470.00580

The results show that knowledge conditioning improves predictions for the majority of test perturbations across all three datasets. The improvement is particularly pronounced on Adamson, where 81.82% of perturbations benefit from biological knowledge. On HepG2, 63.30% of perturbations show improvements, while Norman exhibits a more balanced distribution.

Importantly, we acknowledge that knowledge conditioning does not benefit every perturbation. In particular, approximately half of the Norman test perturbations show reduced performance, and its median gain is relatively small. This indicates that the effectiveness of biological priors depends on the perturbation and dataset, rather than providing uniform improvements.

We further refer to our existing gene-family holdout evaluation (Table 20), which examines generalization beyond gene families represented during training. SDFM achieves Pearson Δ of 0.5987, DE100 of 0.6324, and normalized retrieval of 0.7215, outperforming the evaluated comparison methods in overall response correlation and retrieval.

Together, these analyses provide a more detailed characterization of the benefits and limitations of knowledge conditioning. We will include the per-perturbation statistics and clarify that biological knowledge provides an overall benefit while its effectiveness remains perturbation-dependent.

AuthorsOpenReview ↗

Q3. How was the default top-20 edge setting selected? Are the main conclusions stable under sparser or denser graph constructions?

R3. We thank the reviewer for this helpful question. The top-20 setting was adopted as a practical graph sparsification choice to retain informative local gene neighborhoods while controlling graph density and computational complexity. We agree that its sensitivity should be evaluated rather than assuming it is uniquely optimal.

To examine this, we compared sparser (top-10) and denser (top-40) graph constructions with the default top-20 setting. The results are summarized below:

DatasetGraphPearson ΔNormalized Retrieval
AdamsonTop-100.7696 ± 0.00410.6127 ± 0.0029
AdamsonTop-200.8071 ± 0.00840.6649 ± 0.0081
AdamsonTop-400.8105 ± 0.00830.6697 ± 0.0092
HepG2Top-100.2138 ± 0.00360.5379 ± 0.0057
HepG2Top-200.2634 ± 0.00320.5751 ± 0.0062
HepG2Top-400.2593 ± 0.00560.5802 ± 0.0075
NormanTop-100.2840 ± 0.00580.5201 ± 0.0036
NormanTop-200.3266 ± 0.00430.5649 ± 0.0061
NormanTop-400.3271 ± 0.00620.5564 ± 0.0063

The results show that the top-40 configuration achieves performance comparable to top-20, with slightly different trade-offs across datasets and metrics. Specifically, top-40 achieves slightly higher Pearson Δ on Adamson and Norman, while top-20 performs better on HepG2. Regarding Normalized Retrieval, top-40 performs slightly better on Adamson and HepG2, whereas top-20 performs better on Norman.

In contrast, the sparser top-10 configuration consistently reduces both Pearson Δ and Normalized Retrieval across all three datasets, suggesting that excessive graph sparsification may remove useful biological connectivity.

Overall, these results indicate that SDFM is relatively stable when increasing graph density from 20 to 40 neighbors, while retaining sufficient graph connectivity is important for predictive performance. We will clarify the rationale for the default top-20 setting and report these sensitivity results in the revised manuscript.

Reviewer_zbvvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper provides a novel method, SDFM, for single-cell perturbation prediction that leverages biological knowledge and flow matching to predict the effect of unseen perturbations, as well as unseen combination of perturbations (up to two). The model is validated in three independent datasets under different conditions and different quantitative metrics, such as correlation between predicted and true expression change (under different gene selection conditions), perturbation retrieval and in terms of distributional mapping. The model outperform the assessed models in most conditions. The ablation studies show the importance of every component, although with stronger value to the OT component.

Critical strengths

One of the main strength of the paper is its performance under several quantitative evaluation metrics, that seems to outperform several previous models under different scenarios such as generalization to unseen perturbations, as well as combinations of unseen perturbations. It is also interesting to see that this improvement is also done at the "distribution" level, where de predicted distributions under specific perturbations match better the true ones than those predicted by other models (although I am quite surprise to see the results in the main paper to look less appealing than those in the supplementary.... In the main paper it seems that AttentionPert and GEARS outperform SDFM in the second and third scenario).

The comprehensive ablation study on such a multicomponent model is also highly appreciated as it show the benefits in perturbation prediction across the different components, which can serve as a designing guide for future research.

Critical weaknesses

The critical weakness is the evaluation framework. The authors miss critical baselines, as well as recent models in their evaluation. For example, I strongly suggest that they include linear models as baseline, given that it has been recently shown that they perform very well in perturbation predictions. Also, I am surprised that CFM-GP, while mentioned in the current state-of-the-art as a very relevant model, it is not used in the comparisons. Additionally, there are some unused genes in some of the datasets, which it not explained why they were excluded. Also, the gene splits used for training and evaluation are not clearly identified. While the authors mention the different Seen0, Seen1 and Seen2 splits, it is not clear which genes are in each split. In this regards, I am also surprised on the number of actual gene used in the evaluation with the Norman dataset, since that dataset has significantly larger amount of ground truth on double perturbations.

My second critical concern is the biological validation of the model. While the quantitative metrics (albeit my concerns above) seems to suggest that the model is preforming well while predicting perturbation effects, I am not sure of the value of the work if I cannot learn the underlying mechanisms that are driving those prediction. There are a plethora of models for perturbation prediction using perturb-seq data, so I believe that the biological grounding of the models is of utmost importance to make a model relevant. Section 4.4, which, in my opinion, should be a strong section to show the capacity of the model to recapitulate known biology, it is very difficult to follow and it is not clear the benefits brought by the proposed model.

Questions for authors
  • Why Replogle dataset was not used? I would suggest including this dataset in the assessment.
  • What is the rationale for not including all genes across the different datasets? Otherwise it may look like cherrypicking the results to boost performance.
  • Please clarify the gene splits, as currently it is not clear how these are split across train/val/test, and which genes are truly never seen during training, and which (part of the combinations) are indeed seeing during training. In this regard, I would suggest to try all possible combinations on Norman dataset as Seen2 split, since this directly compares to previous work on prediction of combination of perturbations.
  • Why were linear models not included, given that it has been shown to perform very well in perturbation prediction?
7 comments
AuthorsOpenReview ↗

Q1: Why was the Replogle dataset not included in the original evaluation? Could the authors clarify whether SDFM generalizes to this additional perturbation dataset?

R1. We thank the reviewer for this helpful suggestion. Following the reviewer's recommendation, we have extended our evaluation to the Replogle K562 dataset (Replogle et al., 2022) to examine the generalizability of SDFM on an additional Perturb-seq benchmark.

The dataset contains 162,751 cells, of which 162,246 were retained in our evaluation, covering 5,000 genes and 272 test perturbation conditions. The cell-level split consists of 112,373 training cells, 11,044 validation cells, and 38,829 test cells.

We evaluated SDFM against eight comparison methods: AttentionPert, GEARS, TxPert, scGPT, STATE, scDFM, CPA, and a linear model. The results are summarized below:

MethodPearson ΔSpearman ΔDE20DE100Normalized RetrievalFast Retrieval
SDFM0.38860.28710.48480.54450.61790.6060
AttentionPert0.33310.27700.39210.38410.48930.5033
GEARS0.28330.24920.45340.46120.50070.4741
TxPert0.35080.25560.42730.42140.45510.4842
scGPT0.29090.25890.38660.39230.52580.5567
STATE0.28800.23740.40090.41190.59440.5615
scDFM0.34740.21770.39270.38870.48520.4698
CPA0.28320.23590.49650.50750.45570.4196
Linear Model0.37840.28530.41120.42960.56340.5353

SDFM achieves the best performance on five of the six metrics, including Pearson Δ, Spearman Δ, DE100, Normalized Retrieval, and Fast Retrieval. CPA achieves the highest DE20 score.

These results extend our evaluation to an additional cell line and perturbation dataset, providing further evidence of the predictive performance of SDFM beyond the three datasets considered in the original manuscript.

Ref.

[1] Replogle, J. M., et al. (2022). Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell, 185(14), 2559–2575.e28. https://doi.org/10.1016/j.cell.2022.05.013

AuthorsOpenReview ↗

Q2. What is the rationale for excluding certain genes from the datasets instead of including all available genes? How can the authors ensure that this selection does not introduce bias into the evaluation?

R2. We thank the reviewer for raising this important concern. We agree that the treatment of unused genes should be clearly documented, particularly to distinguish the predefined evaluation scope from any potential performance-based selection.

According to our dataset statistics, the numbers of unused genes are 0, 7, and 20 for Adamson, Norman, and HepG2, respectively. These genes are not included among the perturbation conditions used in our evaluation.

Importantly, the evaluation follows predefined perturbation conditions and fixed train/validation/test splits. Gene selection was not performed according to the prediction performance of SDFM or any comparison method.

We recognize that the original manuscript did not sufficiently explain the distinction between the available genes and those involved in the evaluated perturbation conditions. We will make this distinction explicit in the revised manuscript and report the corresponding gene statistics more clearly.

We appreciate the reviewer's concern, as transparent reporting of gene selection is essential for interpreting benchmark results and ensuring reproducibility.

AuthorsOpenReview ↗

Q3. How are genes and perturbation conditions divided into training, validation, and test sets? Specifically, which genes are completely unseen during training, and which test combinations contain previously observed genes? Why does the Norman Seen2 evaluation include only 19 combinations rather than all possible combinations?

R3. We thank the reviewer for pointing out that the gene-level split definitions and the corresponding evaluation protocol require further clarification.

In our experiments, Norman dual-gene perturbations are categorized into Seen0, Seen1, and Seen2 according to whether their constituent genes have appeared in training perturbations:

  • Seen0: Neither constituent gene has appeared in the training perturbations.
  • Seen1: One constituent gene has appeared in training, while the other has not.
  • Seen2: Both constituent genes have appeared individually in training, but their specific combination is held out.

The Norman dataset contains 138 training, 31 validation, and 107 test perturbation conditions. Among the 71 dual-gene test conditions, the distribution is:

SplitNumber of Test ConditionsDefinition
Seen09Neither gene observed during training
Seen143One gene observed during training
Seen219Both genes observed individually, but the combination is unseen
Total71

There are 33 genes shared between the training and test conditions. This overlap enables us to evaluate generalization to unseen combinations of previously observed genes.

Thus, Seen2 specifically measures combinatorial generalization, rather than generalization to entirely unseen genes.

Regarding the suggestion to evaluate all possible combinations, the current study follows a predefined held-out benchmark containing 19 Seen2 test conditions. We have not performed an exhaustive evaluation of every possible gene combination. We therefore retain the original evaluation protocol and report results on these held-out conditions.

We will clarify the distinction between gene-level overlap and perturbation-level splits, together with the number of conditions in each category, to make the evaluation protocol easier to understand and reproduce.

Reviewer_hFrmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies genetic perturbation prediction from unpaired single-cell data. SDFM conditions flow matching on STRING/GO graph representations and combines optimal transport pairing with endpoint supervision from one-step extrapolation and full ODE integration. The authors evaluate single-gene prediction on three datasets, combinatorial perturbations on Norman, and the distributions of generated cell populations.

Critical strengths
  1. The motivation is clear. Conditioning the velocity field on gene relationships and supervising predictions after a full rollout are both interesting design choices to address the problem.
  2. The experiments are extensive, with Figures 3–5 and Table 22 examining perturbation response structure and cell population distributions.
Critical weaknesses
  1. The methodological innovation in knowledge integration remains unclear (see Q3). The paper is also difficult to follow, with many methodological details deferred to the appendix.
  2. The ablations do not clearly establish what drives the gains, particularly the added value of full rollout over endpoint supervision.
Questions for authors
  1. Which dataset and training configuration does the 3.48 GPU hours in Table 25 correspond to? Please specify the rollout step count and frequency during training, and report training time and predictive performance with and without full rollout, to establish whether the gains justify the additional computation.
  2. In Appendix B.2, PearsonΔ uses the control mean as its reference, while DE20/DE100 remain correlation measures on selected gene subsets. To complement the retrieval results, could the authors report correlation relative to the training perturbation centroid, following Systema[1], and compare DE gene-set overlap between generated and observed cells using the same DE analysis procedure, to test whether recovery of perturbation-specific responses and affected genes also improves?
  3. The title emphasizes “Cross-Aware Knowledge-Conditioned,” but TxPert[2] already encodes perturbations using multiple knowledge graphs, including STRING and GO, while GeneGeoFlow[3] already conditions flow matching on biological network information. What specific limitation of these approaches does the proposed cross-attention conditioning address, and could the authors compare it directly with GeneGeoFlow’s conditioning mechanism under matched supervision to establish its methodological contribution?

References

  1. Viñas Torné et al. (2026). Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation. Nature Biotechnology, 44, 1050–1059.
  2. Wenkel et al. (2026). TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects. Nature Biotechnology.
  3. Li et al. (2026). Control-Anchored Residual Flow Matching Conditioned on Gene Geometry for Virtual Cell Perturbation Modeling. arXiv:2608.06824.
6 comments
AuthorsOpenReview ↗

Q1. Which dataset and training configuration does the 3.48 GPU hours in Table 25 correspond to? Please specify the rollout step count and frequency during training, and report training time and predictive performance with and without full rollout, to establish whether the gains justify the additional computation.

R1. We thank the reviewer for raising this important concern. We agree that clarifying the implementation details and computational cost of Full ODE Rollout is essential for assessing its practical contribution.

The primary motivation for introducing Full ODE Rollout in SDFM is to establish consistency between local velocity learning and the final generated perturbation distribution. Conventional Flow Matching primarily supervises the velocity field at local time points. However, accurate local velocity prediction does not necessarily guarantee that numerical integration of the learned vector field will produce an accurate perturbed-cell population distribution. To address this issue, SDFM incorporates differentiable full ODE rollout into training, allowing the model to optimize not only local transport dynamics but also the population-level response produced by the complete generation process.

(1) Implementation details of Full ODE Rollout

We re-examined the training code, configurations saved in 13 checkpoints, and execution logs to verify the actual rollout settings.

During training, SDFM uses the Heun solver with four integration steps over the time interval [0, 1]. Since Heun evaluates the velocity field twice per integration step, each trajectory requires eight velocity-field evaluations.

Full ODE Rollout is performed at every parameter update, starting from the first training iteration, and is not affected by the learning-rate warmup schedule.

Specifically, at each parameter update, eight perturbation conditions are initially selected, from which two conditions are sampled for rollout. Eight cells are sampled for each selected condition, resulting in 16 complete cell trajectories per parameter update. The two conditions are processed through two separate batched integration calls. Importantly, the rollout-related loss terms share the same integrated endpoints and do not require separate ODE integrations.

During final evaluation and inference, SDFM uses four-step Euler integration, requiring four velocity-field evaluations per trajectory.

(2) Actual rollout execution across datasets

Based on the original training logs, the actual rollout execution statistics are summarized below:

Dataset / VersionTraining EpochsTotal Rollout StagesGenerated Cell Trajectories
Adamson303,00048,000
Norman454,50072,000
HepG2454,50072,000
All three datasets use the same integration configuration: four-step Heun integration during training and four-step Euler integration during inference.

(3) Predictive performance with and without Full ODE Rollout

To evaluate the contribution of complete ODE integration, we compared the full SDFM model with a variant trained without the complete ODE rollout. The results are summarized below:

DatasetMethodPearson ΔNormalized Retrieval
AdamsonFull SDFM0.8071 ± 0.00840.6649 ± 0.0081
Adamsonw/o Full ODE Rollout0.7363 ± 0.00640.6139 ± 0.0075
HepG2Full SDFM0.2634 ± 0.00320.5751 ± 0.0067
HepG2w/o Full ODE Rollout0.2210 ± 0.00320.5411 ± 0.0092
NormanFull SDFM0.3266 ± 0.00430.5649 ± 0.0061
Normanw/o Full ODE Rollout0.2946 ± 0.00870.5257 ± 0.0056

Removing Full ODE Rollout leads to consistent performance degradation across all three datasets.

Specifically, Pearson Δ decreases by 0.0708, 0.0424, and 0.0320 on Adamson, HepG2, and Norman, respectively. Normalized Retrieval decreases by 0.0510, 0.0340, and 0.0392, respectively.

These results indicate that complete ODE integration provides meaningful supervision beyond local velocity learning, improving both perturbation-response prediction and perturbation retrieval.

Regarding computational cost, Table 25 reports 3.48 GPU hours for SDFM. We acknowledge that the current presentation does not sufficiently clarify the exact dataset and configuration associated with this value. We will clarify the measurement scope and corresponding configuration in the revised manuscript.

Overall, although Full ODE Rollout introduces additional computational overhead during training, the ablation results demonstrate consistent predictive improvements across three datasets, supporting the value of incorporating complete generated trajectories into the training objective.

AuthorsOpenReview ↗

Q2. In Appendix B.2, Pearson Δ uses the control mean as its reference, while DE20/DE100 remain correlation measures on selected gene subsets. To complement the retrieval results, could the authors report correlation relative to the training perturbation centroid, following Systema, and compare DE gene-set overlap between generated and observed cells using the same DE analysis procedure, to test whether recovery of perturbation-specific responses and affected genes also improves?

R2. We thank the reviewer for this valuable suggestion. We agree that evaluating expression changes relative to the control mean alone may not fully characterize the model's ability to recover perturbation-specific biological responses. Following the reviewer's suggestion, we provide two complementary analyses to further evaluate SDFM from the perspectives of expression-response correlation and differentially expressed gene (DEG) set consistency.

(1) Training Perturbation Centroid-based Correlation

In our original evaluation, Control Mean is used as the reference for calculating Pearson Δ between predicted and observed perturbation responses.

Following the reviewer's suggestion, we additionally use the Training Perturbation Centroid as the reference and report both Pearson and Spearman correlations. This provides a complementary evaluation of the consistency between predicted and observed perturbation responses under an alternative reference baseline.

The results are summarized below:

DatasetN Test ConditionsPearson (Training Centroid)Spearman (Training Centroid)Pearson (Control Mean)
Adamson220.4280 ± 0.00820.2758 ± 0.00330.8071 ± 0.0084
HepG25940.2562 ± 0.01390.1826 ± 0.01340.2634 ± 0.0032
Norman (Single)360.1049 ± 0.01110.1654 ± 0.00920.3266 ± 0.0043
Norman (Dual)710.5242 ± 0.01050.2563 ± 0.01240.6820 ± 0.0076
Overall7230.2751 ± 0.01330.1913 ± 0.00750.3242 ± 0.1407

The results show that the correlation values vary across reference baselines and datasets. In particular, the Training Perturbation Centroid-based correlations are lower than the original Control Mean-based Pearson Δ on Adamson and Norman.

This observation highlights that different reference baselines capture different aspects of perturbation-response consistency. Rather than relying exclusively on control-referenced expression changes, the additional centroid-based evaluation provides a complementary perspective on the model's ability to recover perturbation-specific response patterns.

(2) DE Gene Set Overlap

We further compare the differentially expressed gene sets identified from Generated Cells and Observed Perturbed Cells.

Specifically, we apply the same DE analysis procedure to both generated and observed perturbed populations and quantify the overlap between their resulting DEG sets.

We report the following metrics:

  • Overlap: Number of shared genes between the observed and generated DEG sets.
  • Jaccard Index: Intersection over union of the two DEG sets.
  • Precision: Fraction of generated DEGs that are also identified as observed DEGs.
  • Recall: Fraction of observed DEGs recovered by the generated cells.
  • F1: Harmonic mean of precision and recall.

The results are summarized below:

DatasetN PerturbationsDE Genes (Observed)DE Genes (Generated)OverlapJaccard Index ↑Precision ↑Recall ↑F1 ↑
Adamson221131.35 ± 13.023958.73 ± 10.88849.19 ± 13.120.1999 ± 0.00280.2149 ± 0.00300.7882 ± 0.00370.3316 ± 0.0040
HepG2594153.94 ± 0.72616.93 ± 4.9350.20 ± 0.690.0280 ± 0.00030.0395 ± 0.00020.1241 ± 0.00370.0483 ± 0.0005
Norman (Single)36194.63 ± 1.281747.63 ± 4.5075.67 ± 1.200.0418 ± 0.00100.0468 ± 0.00120.3731 ± 0.00970.0790 ± 0.0019
Norman (Dual)71273.49 ± 0.832646.32 ± 10.56183.26 ± 0.920.0744 ± 0.00090.0793 ± 0.00100.6435 ± 0.00300.1360 ± 0.0014
Overall723197.45 ± 0.66974.21 ± 4.9388.85 ± 0.530.0392 ± 0.00040.0498 ± 0.00030.2132 ± 0.00290.0683 ± 0.0006

The results reveal differences in DEG recovery across datasets and perturbation types.

On Adamson, SDFM achieves a Jaccard Index of 0.1999 and a Recall of 0.7882, indicating that a substantial proportion of the observed differentially expressed genes can be recovered from the generated cells.

Similarly, on Norman dual-gene perturbations, the Recall reaches 0.6435, suggesting that SDFM recovers a considerable fraction of the genes affected by the observed combinatorial perturbations.

At the same time, the number of DEGs identified in Generated Cells is generally higher than that in Observed Perturbed Cells. Consequently, Precision and Jaccard Index remain relatively low on some datasets. This indicates that, although the generated responses recover a fraction of observed DEGs, they may also introduce additional differential-expression signals that are not detected in the observed populations.

Overall, the Training Perturbation Centroid-based Correlation and DE Gene Set Overlap analyses provide two complementary perspectives for evaluating SDFM: consistency in perturbation-response prediction and recovery of affected genes. These results further characterize both the strengths and the dataset-dependent limitations of SDFM.

AuthorsOpenReview ↗

Q3. Given that TxPert already incorporates knowledge graphs (e.g., STRING and GO) and GeneGeoFlow integrates biological networks into Flow Matching, what distinguishes SDFM's cross-attention conditioning, and can it be directly compared with GeneGeoFlow under matched supervision to demonstrate its methodological contribution?

R3. We thank the reviewer for pointing out this important issue. We acknowledge that SDFM is not the first method to incorporate biological knowledge graphs into genetic perturbation prediction, nor is it the first to use biological network information in Flow Matching. We will clarify the distinction between SDFM and existing approaches and better articulate the specific methodological contributions of our work.

(1) Difference from TxPert

While TxPert uses knowledge graphs such as STRING and GO to construct perturbation representations, SDFM further integrates these representations with generation-time information through multi-head cross-attention within the Velocity Network.

Specifically, the current cellular state serves as the Query, while perturbation-time conditions serve as the Key and Value. This enables the model to dynamically aggregate biological knowledge according to the current cellular state, rather than relying solely on static perturbation representations.

(2) Difference from GeneGeoFlow

Although GeneGeoFlow also incorporates biological network information into Flow Matching, SDFM explicitly models the interaction between perturbation knowledge, cellular states, and generation time.

Specifically, SDFM constructs Perturbation-Time Condition Tokens by combining perturbation embeddings epe_p and time representations τ(t)\tau(t):

C=reshape(Wc[ep,τ(t)]+bc)C = \mathrm{reshape}\left(W_c[e_p,\tau(t)] + b_c\right)

These tokens are injected into the Velocity Network through cross-attention:

hl+1=hl+CrossAttn(hl,C)h^{l+1} = h^l + \mathrm{CrossAttn}(h^l,C)

This mechanism enables the velocity field to dynamically adapt to cellular states throughout the generation trajectory. Furthermore, it is integrated with Optimal Transport Coupling and Full-Rollout Supervision to connect local dynamics with population-level responses.

(3) Conditioning Ablation Results

The effectiveness of the proposed conditioning mechanism is further supported by the conditioning ablation results reported in Table 3 of the manuscript.

Specifically, replacing cross-attention with simple concatenation (Concat), removing cellular-state information (w/o Cell State), or removing time embeddings (w/o Time Embedding) consistently reduces performance across the three datasets.

These results support the importance of dynamically integrating cellular-state information and perturbation-time conditions through cross-attention.

(4) Direct Comparison with GeneGeoFlow

Regarding the suggested direct comparison, we investigated GeneGeoFlow but could not identify publicly available code suitable for reproduction. Therefore, a controlled comparison under matched data splits and supervision settings was not feasible.

We will clarify this limitation and strengthen the methodological discussion to distinguish SDFM from existing approaches.

In summary, SDFM's contribution lies in integrating Perturbation-Time Cross-Attention with OT-coupled Flow Matching and trajectory-consistent supervision, enabling dynamic knowledge conditioning throughout the continuous generation process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.