- 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
- 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.
- Broad experimental evaluation. The experiments cover synthetic and biological datasets across multiple modalities.
Critical weaknesses
- 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.
- 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.
- Limited extrapolation evidence. The experiments primarily consider interpolation or randomly held-out conditions, providing limited evidence for extrapolation beyond the training descriptor distribution.
- 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
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
Authors response
Comment R3-C1
Thank you for this comment. We modified and corrected our theoretical approach, while obtaining the very same I−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
Please refer to Appendix H of the manuscript, titled "Upper bound on the discrepancy between the distribution distance metrics".
Comment R3-C3
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
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:
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) be a GMM, and μGMM(Θ∗,p∗) be its projection (Def. B.3 of the modified manuscript pdf). Let π∈RI×J be the joint plan, and uij be the amount of mass moved from mode j in ρ to mode i in μ (note that πij=pivij=qjuij). Set u∈RI×J and cij:=∥Θi−Γj∥22. At this stage, for simplicity assume that for each j, the closest element of {Θi}i=1I to Γj is unique, i.e. there are no equally-close nearest neighbours in the latter set to Γj. Then the following transport plan {uijhard}i=1,j=1I,J is an optimal plan between ρ and μ.
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 $J-ID$.