acceptodds
Under review as a conference paper at ICLR 2027

Shortest-Path Flow Matching with Mixture-Conditioned Bases for Unseen-Condition Generalization

Abstract

Robust generalization under distribution shift remains a key challenge for conditional generative modeling: conditional flow-based methods often fit the training conditions well but fail to extrapolate to unseen ones. This is particularly important for therapeutic prioritization, where the goal is to predict cellular responses to unseen candidate drugs and genetic perturbations in order to select the most effective intervention. We introduce SP-FM, a shortest-path flow-matching framework that improves generalization to unseen conditions by conditioning both the base distribution and the flow field on the condition. Specifically, SP-FM learns a condition-dependent base distribution parameterized as a flexible, learnable mixture, together with a condition-dependent vector field trained via shortest-path flow matching. Conditioning the base allows the model to adapt its starting distribution across conditions, enabling smooth interpolation and more reliable extrapolation beyond the observed training range. We provide theoretical insights into the resulting conditional transport and show how mixture-conditioned bases enhance robustness under shift. Empirically, SP-FM is effective across heterogeneous domains, including predicting responses to unseen perturbations in single-cell transcriptomics and modeling treatment effects in microscopy-based drug screening. Overall, SP-FM provides a simple yet effective plug-in strategy for improving conditional generative modeling and generalization to unseen conditions.

Primary area:
generative models
Keywords:
Flow matching, conditional generative modelling, biology, perturbation prediction, drug screening, out-of-distribution generalization, optimal transport

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dZptOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SP-FM, a conditional flow-matching framework that learns a condition-dependent Gaussian mixture base distribution alongside a conditional velocity field. The method minimizes the transport distance between the learned base and target distributions while training the flow through minibatch optimal transport. The authors provide a theoretical analysis based on mixture Wasserstein distance and evaluate SP-FM on synthetic data, single-cell perturbation prediction, and microscopy-based drug screening. Experimental results demonstrate improvements over standard conditional flow matching and several application-specific baselines, particularly on image-based benchmarks.

Critical strengths
  1. Interesting methodological perspective. Learning condition-dependent base distributions offers an intuitive extension to conventional conditional flow matching. The framework is relatively simple and can be integrated into existing flow-based architectures.
  2. Broad experimental evaluation. The experiments cover synthetic and biological datasets across multiple modalities.
Critical weaknesses
  1. Theoretical correctness. Statement B.5 overlooks the non-uniqueness of OT dual potentials, while Eq. (12) appears to require min rather than max. These issues undermine the identifiability argument.
  2. Theory–implementation gap. The analysis assumes mixture-Wasserstein transport and an auxiliary dual predictor that are not used in the practical algorithm, weakening its theoretical justification.
  3. Limited extrapolation evidence. The experiments primarily consider interpolation or randomly held-out conditions, providing limited evidence for extrapolation beyond the training descriptor distribution.
  4. Insufficient ablation controls. Replacing the learned base only at inference introduces a distribution mismatch. Retrained, capacity-matched adaptive-prior baselines would better isolate the proposed contribution.
Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Authors response

Comment R3-C1

Theoretical correctness. ... an auxiliary dual predictor ....

Thank you for this comment. We modified and corrected our theoretical approach, while obtaining the very same I−JDI-JD degrees of freedom as before, with the auxiliary dual predictor now removed from the analysis. Here at the end of the response (under "New Theory Approach") we provide the summary. For complete proofs, please refer to the updated pdf of the manuscript.

Comment R3-C2

Theory–implementation gap

Please refer to Appendix H of the manuscript, titled "Upper bound on the discrepancy between the distribution distance metrics".

Comment R3-C3

Limited extrapolation evidence. ...

Thank you for this valuable comment. In our single-cell experiments, no test perturbation is seen during training, and in most datasets the majority of perturbations are held out: ~70% for Adamson and Combosciplex, ~45% for Norman, and >90% for Replogle-Jurkat. Furthermore, for example, every single-gene test perturbation in Norman targets a gene never seen during training. In our synthetic letters experiment, we also evaluate a harder setting with blocked rotations (Table 8), where the gap between training rotations is increased and unseen rotations are further from the training conditions, still outperforming the baselines.

Extrapolation to entirely novel (in the sense of out-of-distribution, OOD) regions of descriptor space remains an open problem for the field, which we now note in the Limitations.

Comment R3-C4

Insufficient ablation controls ... .

We agree that replacing a learned base only at inference introduces a distribution mismatch and should not, by itself, be interpreted as a controlled ablation. However, our experimental evidence does not uniquely rely on this comparison. We compare against:

  • CFM with N(0,1) source
  • SP-FM with I=1I=1, and varying II (Fig 3a)
  • CellFlow with control-cell source

To directly address the reviewers' concern, we further compared with an unconditional “global” GMM baseline. For details please refer to our reply to R2 under "Comment R2-C4".

New Theory Approach

All proofs are moved to the updated manuscript pdf, updated Appendix B.

Statement newB.5

Let ρGMM(Γ,q)\rho_{GMM(\boldsymbol{\Gamma}, \boldsymbol{q})} be a GMM, and μGMM(Θ∗,p∗)\mu_{GMM(\boldsymbol{\Theta}^{\ast}, \boldsymbol{p}^{\ast})} be its projection (Def. B.3 of the modified manuscript pdf). Let π∈RI×J\boldsymbol{\pi}\in \mathbb{R}^{I\times J} be the joint plan, and uiju_{ij} be the amount of mass moved from mode jj in ρ\rho to mode ii in μ\mu (note that πij=pivij=qjuij\pi_{ij} = p_i v_{ij} = q_j u_{ij}). Set u∈RI×J\boldsymbol{u}\in \mathbb{R}^{I\times J} and cij:=∥Θi−Γj∥22c_{ij} := \lVert\boldsymbol{\Theta}_i-\boldsymbol{\Gamma}_j\rVert_2^2. At this stage, for simplicity assume that for each jj, the closest element of {Θi}i=1I\{\boldsymbol{\Theta}_i\}_{i=1}^I to Γj\boldsymbol{\Gamma}_j is unique, i.e. there are no equally-close nearest neighbours in the latter set to Γj\boldsymbol{\Gamma}_j. Then the following transport plan {uijhard}i=1,j=1I,J\{u_{ij}^{hard}\}_{i=1,j=1}^{I,J} is an optimal plan between ρ\rho and μ\mu.

1 & \text{if } i = \operatorname{argmin}_{i} c_{ij} \\\\ 0 & \text{otherwise} \end{cases} \tag{newB.2}$$ In Statement newB.5 we assumed that the closest element of $\{\boldsymbol{\Theta}_i\}_{i=1}^I$ to $\boldsymbol{\Gamma}_j$ is unique. In the following statement, we discuss the case where the nearest neighbour is not unique. **Statement newB.6** Assume for some $j$, both $\boldsymbol{\Theta}_{i_1}$ and $\boldsymbol{\Theta}_{i_2}$ have minimal and equal distance to $\boldsymbol{\Gamma}_j$. Then both plans for $j$ are optimal and have equal cost. **Statement newB.7** Assume that $\mu_{GMM(\boldsymbol{\Theta}^{\ast}, \boldsymbol{p}^{\ast})}$ is the projection of $\rho_{GMM(\boldsymbol{\Gamma}, \boldsymbol{q})}$, and that $\left\lvert\{j \mid q_j > 0\}\right\rvert \ge I$. Then: $$\forall i_1,i_2,\ 1\le i_1 < i_2 \le I:\quad \boldsymbol{\Theta}^{\ast}_{i_1} \ne \boldsymbol{\Theta}^{\ast}_{i_2}.$$ **Statement newB.8.** Assume that $\mu_{GMM(\boldsymbol{\Theta}^{\ast}, \boldsymbol{p}^{\ast})}$ is the projection of $\rho_{GMM(\boldsymbol{\Gamma}, \boldsymbol{q})}$, and that $\left\lvert\{j \mid q_j > 0\}\right\rvert \ge I$. Moreover, assume that $\mu$'s mode probabilities are not exactly equal to 0: $\forall i,\ 1\le i \le I,\ p_i > 0$. Then: $$\forall j,\ 1 \le j \le J:\quad \left\lvert\operatorname{argmin}_i \lVert\boldsymbol{\Theta}^{\ast}_{i} - \boldsymbol{\Gamma}_{j}\rVert_2^2\right\rvert = 1.$$ **Statement newB.9.** In the setting of Statement newB.5, the hard pullback assignment $\boldsymbol{u}^{hard}$ is the unique globally-optimal plan. The hard pullback assignme (i.e. 1NN assignment) has $J$ degrees of freedom, which are reduce by the $ID$ barycenter conditions to $J-ID$.

Dear Reviewer,

Thank you again for your thoughtful review of our submission. We have put considerable time and effort into addressing your comments, including additional analyses and clarifications, and have posted a detailed response.

As the discussion period ends soon, we would be grateful if you could look at our response when you have a moment and consider whether it addresses your concerns. If so, we would kindly ask you to consider updating your assessment.

We are happy to answer any further questions or provide additional clarification.

Best regards, The Authors

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

The paper proposes SP-FM, a conditional flow-matching framework that learns a condition-dependent Gaussian mixture base distribution alongside a conditional velocity field. The learned base aims to reduce the source–target transport distance and simplify the subsequent generation process. Experiments on synthetic rotations, single-cell perturbations, and image-based morphological responses show promising results across multiple benchmarks.

Critical strengths
  1. Learning a condition-dependent base distribution is a reasonable and potentially useful extension to standard flow matching. Reducing the initial source–target mismatch offers a plausible way to simplify the transport task.

  2. The evaluation spans multiple data modalities and includes comparisons with relevant baselines, as well as analyses of mixture size, dimensionality, and source–target distance. The results provide encouraging evidence that SP-FM can improve predictive performance in several settings.

Critical weaknesses
  1. The benefit of the proposed base distribution for unseen-condition generalization is not sufficiently established. The method relies on predicting mixture parameters from condition descriptors, so its generalization depends on the quality of these representations. The paper does not clearly explain how entities absent from training are represented across all benchmarks. Moreover, reducing source–target distance does not by itself establish better generalization. A comparison with a conditional velocity field using a jointly learned global GMM base would help isolate the contribution of the condition-dependent base.

  2. The theoretical analysis does not fully support the claimed mechanism. The paper argues that increasing the number of mixture components reduces the remaining transport difficulty, but the reported results do not show consistent improvements as the mixture size increases. Some steps in the theoretical derivation also require correction or stronger assumptions. These issues weaken the explanation of why the proposed design should improve generalization, even though they do not invalidate the method itself.

  3. Some experimental descriptions need clarification. Appendix C.2 states that all letters appear during training and only rotation angles are held out, whereas Section 5.1 describes W and Y as unseen. The distinction between held-out combinations of known entities and conditions involving entities absent from training should also be made explicit.

Questions for authors
  1. How are condition descriptors and embeddings constructed in each benchmark, particularly for entities absent from all training conditions? Could the authors compare SP-FM against a conditional velocity field using a jointly learned global GMM base to isolate the benefit of conditioning the base distribution?

  2. Could the authors report source–target and generated–target distances across different mixture sizes, separately for training and held-out conditions? This would help clarify whether a closer learned base consistently translates into better predictions.

  3. Could the authors clarify the assumptions and conclusions of the theoretical analysis, including the role of mixture size and the soft Gumbel–Softmax sampling procedure? Please also reconcile the descriptions of the held-out letter experiment.

4 comments
AuthorsOpenReview ↗

Authors reply

Comment R2-C1

The benefit of the proposed base distribution ... . ... its generalization depends on the quality of these representations.

We thank the reviewer for this valuable comment. We argue that dependence on condition descriptors is not a limitation specific to SP-FM. Crucially, SP-FM and the descriptor-conditioned baselines are evaluated using the same condition representations, so descriptor quality is controlled across methods and cannot explain differences in performance (see our reply under Comment R1-C1).

We do not claim that SP-FM eliminates dependence on informative condition representations; no descriptor-conditioned method can make such a claim. The relevant question is whether the proposed source improves predictive performance on held-out conditions.

Comment R2-C2

... how entities absent from training are represented ....

How are condition descriptors and embeddings constructed ...?

Thank you for this comment. The details of our perturbation representations are provided in our reply under Comment R1-C1.

Comment R2-C3

Moreover, reducing source–target distance does not by itself establish better generalization.

Thank you for this invaluable comment. We never claim improved generlisation is a result of reduced source-target distance. Instead, we relate the improved generalisation to number of base GMM modes, which itself reduces the source-target distance.

Comment R2-C4

A comparison with ... a jointly learned global GMM base ....

We thank the reviewer for suggesting this ablation. Table 1 in the manuscript already provides 2 competing GMM variants (the exact implementation is reported in Appendix C.4). In Figure 3 and in Ablations, we report a single-mode GMM (I=1I=1) experiment for both transcriptomics and morphological perturbations, and PriorGrad already represents an adapted (although not learned end-to-end like SP-FM with I=1I=1) Gaussian source distribution model. In all these experiments, SP-FM outperformed the baselines.

We further compared SP-FM against a baseline where a global base GMM (shared across populations) is fitted once by EM (full covariances) to the PCA-50 coordinates of training cells, and the decoder flow conditioning on population descriptors and training parameters are consistent with SP-FM.

The results are provided in Table.Response.2.

Table.Response.2: Due to lack of space, only W1 and W2 are reported.

DatasetMethodW2 ↓W1 ↓
AdamsonCellFlow0.155±0.0210.339±0.016
Global GMM0.144±0.0120.327±0.011
SP-FM0.135±0.0140.325±0.012
NormanCellFlow0.082±0.0020.260±0.003
Global GMM0.073±0.0100.245±0.008
SP-FM0.067±0.0040.238±0.001
ComboSciplexCellFlow0.280±0.1090.500±0.068
Global GMM0.204±0.0500.406±0.032
SP-FM0.209±0.0670.411±0.041
JurkatCellFlow3.098±0.2851.531±0.065
Global GMM3.047±0.3311.481±0.069
SP-FM2.985±0.2881.458±0.066
K562CellFlow0.114±0.0130.283±0.012
Global GMM0.108±0.0060.275±0.006
SP-FM0.091±0.0050.265±0.0055

Comment R2-C5

the reported results do not show consistent improvements as the mixture size increases.

Please refer to the manuscript, Table.1 (I=1I=1), and Fig.3a when II is varied.

COmment R2-C6

... theoretical derivation also require correction ....

Thank you for this comment. We modified and corrected our theoretical approach. For details, please refer to our response under "Comment R3-C1".

Comment R2-C7:

Some experimental descriptions need clarification. Appendix C.2 states that ... .

We thank the reviewer for pointing this out. To clarify, the letter experiment evaluates unseen combinations of letters and rotations, rather than entirely unseen letters, as described in Appendix C.2. The benchmark uses six letters (S, Q, W, H, I, Y) and 20 rotations uniformly spaced in [0, 2π]. All six letters are used during training at the 10 even-indexed rotations, while validation (S) and testing (W, Y) use the 10 odd-indexed rotations, which are held out during training. Therefore, every evaluated (letter, rotation) combination is unseen during training, although the letters themselves are seen at other rotations. We will clarify the wording in Section 5.1 and the Table 1 caption. We will also clarify the distinction between unseen combinations of known entities and entirely unseen entities in the main text, and will correct the number of letters in Appendix C.2 from four to six.

Comment R2-C8:

Please also reconcile the descriptions of the held-out letter experiment.

The experiment evaluates unseen letter-rotation combinations rather than entirely unseen letters. We will clarify this in Section 5.1 and the Table 1 caption and correct the number of letters in Appendix C.2 to ensure consistency throughout the paper.

Comment R2-C9:

... the (role of) soft Gumbel–Softmax sampling procedure.

Please refer to our reply under "Comment R1-C6"

Dear Reviewer,

Thank you again for your thoughtful review of our submission. We have put considerable time and effort into addressing your comments, including additional analyses and clarifications, and have posted a detailed response.

As the discussion period ends soon, we would be grateful if you could look at our response when you have a moment and consider whether it addresses your concerns. If so, we would kindly ask you to consider updating your assessment.

We are happy to answer any further questions or provide additional clarification.

Best regards, The Authors

The rebuttal has strengthened the experiments, especially with the added global GMM comparison, and cleared up the confusion on the letter-rotation experiment. Even so, I retain my main reservations regarding the interpretation and limits of unseen condition generalization, the impact of mixture size, and the relationship between the proposed base distribution and generalization

Dear aKcc,
Thank you for acknowledging that the additional experiments strengthen the paper. However, we find your remaining reservations difficult to interpret, as they do not identify any specific unresolved technical issue or engage with the quantitative evidence and clarifications provided in our rebuttal.

We would appreciate a more precise assessment of the three points you raise.

Regarding the “interpretation and limits of unseen-condition generalisation”: we are particularly puzzled by this reservation, since there are several relevant generalisation settings explicitly defined in the manuscript.

Predicting responses to held-out perturbations is an established generalisation task in the literature. See – e.g. – “A systematic comparison of single-cell perturbation response prediction models” by Li et al. 2026 in Science Advances and “Benchmarking algorithms for generalizable single-cell perturbation response prediction” by Wei et al. in Nature Methods. Following this established definition, we have already explained how our experiments evaluate this setting:

  • Approximately 70% of perturbations are held out in Adamson and ComboSciplex, 45% in Norman, and over 90% in Replogle-Jurkat.
  • In Norman, the genes targeted by held-out single-gene perturbations are entirely absent as perturbation targets during training.
  • Similar splits-design has been applied to the morphological perturbations experiments.
  • The blocked-rotation experiment (Table 8) additionally evaluates generalisation with larger gaps between observed and unseen rotation conditions.

Could you clarify which aspect of “our” definition or experimental protocol you are challenging, and on what technical basis?

Regarding the “impact of mixture size”: Section 6 and Figure 3a already investigate this question explicitly. They show that increasing mixture size improves performance until a dataset-dependent optimum, after which improvements saturate or occasionally reverse.

There is no mathematical reason to expect monotonically decreasing test error as model capacity increases. Increasing the number of mixture components expands the representable distribution family, but does not guarantee monotonically improving finite-sample generalisation or optimisation.

Our revised theoretical analysis concerns the structural properties of the transport problem under stated assumptions, not a guarantee of monotonically decreasing test error.

Are you questioning a specific statement or assumption in the revised theoretical analysis, or the empirical trend in Figure 3a? If the former, we would appreciate a concrete counterexample or identification of the problematic step.

With respect to the “relationship between the learned base and generalization”: this is precisely the motivation for the additional global-GMM experiment requested in your original review.

Table.Response.2 compares SP-FM with an unconditional global GMM using the same descriptor-conditioned flow architecture and comparable training settings. SP-FM achieves lower W1 and W2 on four of five datasets.

Moreover, Table 5 reports source–target distances, while Tables 2–4 evaluate generated–target agreement. These experiments distinguish the reduction in initial transport discrepancy from the actual quality of generated distributions under unseen conditions.

Importantly: We do not (and we never did) claim that shorter transport paths universally guarantee better generalisation. Rather, we provide empirical evidence that descriptor-conditioned mixture bases can improve held-out-condition prediction compared with fixed and globally fitted alternatives.

What specific aspect of these results do you believe remains insufficient? Would we appreciate objective and actionable observations.

We welcome substantive criticism, including disagreement with the strength of our evidence. However, the latest response does not identify a numerical contradiction, an invalid assumption, a specific theoretical error, or an experimental result that undermines our conclusions.

As currently formulated, the remaining concerns are essentially restatements of the original reservations, without explaining why our detailed responses and additional experiments fail to address them.

We would therefore appreciate a concrete, paper-specific justification for maintaining these reservations, particularly given the substantial revisions and analyses undertaken in direct response to your requests. This would also allow the Area Chair to assess which issues genuinely remain unresolved.

Reviewer_fmN2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed Shortest-path flow matching to improve the generalization performance of conditional generative model towards unseen perturbation conditions. They introduced condition-dependent Gaussian mixture model as source distribution, where the mean and mixture weights are predicted via the neural network in condition descriptor. In training, there are two alternating objectives. Minibatch optimal transport is used to build flow matching training trajectories to learn conditional velocity field; GMM based distribution is then optimized towards target distribution to shorten the transport distance. Their mixture Wasserstein distance-based analysis suggests increasing the number of mixed components in source distribution can reduce the degrees of freedom in conditional transport identification, thus improves generalization to unseen conditions.

Critical strengths
  1. They proposed a conceptually simple inductive bias that is applicable to improve conditional generative modeling. and it can be relatively easily applied to other methods, e.g., CellFlow, rather than limited to certain application scenarios.

  2. Experiments are carried out thoroughly, on synthetic data, scRNA-seq and microscopy based morphology, where interestingly while as expected, lifted performance more significantly than in other modalities.

Critical weaknesses
  1. The construction of unseen perturbation descriptors are not clear. learned condition embedding in Appendix D doesn't explain for generalization held-out gene or chemical compound. e.g., using a lookup table prevents any effective learning to unseen perturbation.

  2. PCA was conducted using complete data rather than on training split only, which results that train split contains mixture of distributions leaked from test split, since test split cell population is included in representation distribution construction.

  3. There is missing comparison between unconditional GMM, conditional single Gaussian and conditional GMM, whereas replacing the base from the trained model for inference introduced extra distribution mismatch.

  4. Missing comparisons to simple, statistic-based baseline methods. latest benchmark papers revealed that distributional improvements using small dataset doesn't necessarily result in more accurate biological effects estimation to specific perturbation.

  5. There are inappropriate formulations or insufficient proof. e.g., the claim for absolute uniqueness of dual variables doesn't hold by simply fulfilling the subset-sum condition, and assuming k-component GMM while in practice to use weighted sum component means with Gumbel-softmax plus Gaussian noise as alternative.

Questions for authors
  1. Can u provide justifications or fix in equation 12, statement B.10, and see if further adjustment is needed in other sections (likely B.5/6/12/13) accordingly?

  2. Can u use highly variable genes instead of PCs as input, or only conduct PCA on train split then project test split on top of the train split PCA space, and re-examine whether advantages in table4 still holds?

  3. Extend benchmark to null or simple statistical based methods like mean-response or control-only predictors, and evaluate on other metrics like DE genes and delta expression.

  4. Can you provide more evidence on performance lift coming from mixture conditioning, e.g., by comparing to conditional Gaussian and/or unconditional GMM settings?

3 comments
AuthorsOpenReview ↗

Authors reply

Comment R1-C1

The construction of unseen perturbation descriptors are not clear. ....

We thank the reviewer for inquiring about this, allowing us to clarify. SP-FM does not rely on a trainable lookup table of perturbation identities for predicting unseen perturbations. For genetic perturbations, we use externally computed GenePT embeddings, derived from gene descriptions. For chemical perturbations, we use molecular fingerprints computed from SMILES using RDKit.

Thus, population descriptors for held-out genes and compounds remain available at test time. This is consistent with the consensus of sota models in the field, such as cellflow. The trained condition encoder projects the descriptors into the conditioning space, then aggregates them via average pooling, and the resulting embedding is used by the base predictor and the flow module. We agree that Appendix D does not sufficiently describe the procedure and will clarify it in the final version of the manuscript.

Comment R1-C2

PCA was conducted using complete data rather than on training split ... .

Thank you. We would like to clarify a distinction that was insufficiently explicit in the manuscript. The PCA used for model training is fitted exclusively on the training data. The reference in Section 5.3 to PCA fitted on the complete dataset concerns only the evaluation procedure, following the protocol adopted by CellFlow. Specifically, a separate PCA transformation is used to compute distributional metrics in a common 50-dimensional representation. This evaluation transformation is applied after model training. In sum, there is not test-set leakage in our experiments, and we will revise the text to clarify.

Comment R1-C3

There is missing comparison between unconditional GMM, conditional single Gaussian and conditional GMM, ....

Can you provide more evidence ..., e.g., by comparing to conditional Gaussian and/or unconditional GMM settings?

Please refer to our response under "Comment R2-C4".

Comment R1-C4

Missing comparisons to simple methods. ... distributional improvements ... doesn't necessarily result in more accurate biological effects estimation.

Extend benchmark to null or simple ..., and evaluate on ... like DE genes and delta expression.

We thank the reviewer for raising this point. We emphasise that SP-FM is primarily a methodological contribution to conditional flow matching, rather than a task-specific model. The direct assessment of this contribution is the distributional agreement between generated and observed populations. Still, to directly address the reviewer's concern, we have conducted the requested additional evaluation. On top of this, the reviewers may appreciate that more recent work has demonstrated that standard metrics such as MSE and control-referenced Pearson Δ can be insensitive to perturbation-specific information, particularly when effects are sparse or shared expression changes dominate. Miller et al. (Nature Biotechnology, 2026) systematically investigated this issue . Thus, we evaluated two additional baselines: (i) control-only prediction and (ii) mean-response prediction, which assigns the same average training perturbation effect to every held-out condition.

Table.Response.1: Results are averages over the paper’s 4 splits. * means SP-FM is significantly better than both simple baselines.

DatasetMethodPearson Δctrl (DEG) ↑Pearson Δpert (DEG) ↑WMSE ↓WR² Δpert ↑NIR ↑
AdamsonCellFlow + SP-FM (ours)0.5890.3710.1250.0230.582
Mean response0.467-0.0250.128-0.1290.498
Control only-0.0530.0100.136-0.5870.501
NormanCellFlow + SP-FM (ours)0.6170.6310.1620.3450.803
Mean response0.239-0.0090.280-0.1360.500
Control only0.0010.0450.308-0.1960.500
ComboSciplexCellFlow + SP-FM (ours)0.6740.6770.1220.5510.742
Mean response0.5010.0150.259-0.0390.501
Control only0.0200.1620.387-0.8990.500
Replogle K562CellFlow + SP-FM (ours)0.3530.3830.0360.1780.621
Mean response0.2290.0130.043-0.0710.502
Control only0.0110.0910.0460.0150.500
Replogle JurkatCellFlow + SP-FM (ours)0.1990.0760.265-0.1830.521
Mean response0.2150.0440.266-0.2330.500
Control only0.043-0.0820.273-0.1570.501

Comment R1-C5

... inappropriate formulations or insufficient proof. e.g., ...

Thank you for this comment. We modified and corrected our theoretical approach. For details, please refer to our response under "Comment R3-C1".

Comment R1-C6

... assuming k-component GMM while in practice to use weighted ... Gumbel-softmax ....

We use Gumbel-Softmax so that component sampling is differentiable, allowing end-to-end optimisation of the base distribution, and it recovers hard assignment in the limit of the temperature parameter, τ→0\tau \to 0.

Dear Reviewer,

Thank you again for your thoughtful review of our submission. We have put considerable time and effort into addressing your comments, including additional analyses and clarifications, and have posted a detailed response.

As the discussion period ends soon, we would be grateful if you could look at our response when you have a moment and consider whether it addresses your concerns. If so, we would kindly ask you to consider updating your assessment.

We are happy to answer any further questions or provide additional clarification.

Best regards, The Authors

Reviewer_fmN2OpenReview ↗

Increasing score to 3.

Most of my concerns have been addressed -- my thanks to the authors. I increase my score to a 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.