- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces SPARC, a gene-level single-cell perturbation prediction model based on conditional flow matching. Instead of attention, the model propagates information across genes through co-expression adjacency graphs split into positive and negative components, complemented by a global pooling channel and weight-shared recurrent blocks. To model genetic perturbations, an action vector is injected into each target gene node, and a sampling strategy ensures that target genes appear in every training batch. The training objective combines a flow-matching loss with a Sinkhorn endpoint loss weighted by differentially expressed genes (DEGs).
Critical strengths
-
The approach is simple and computationally efficient. It uses the graph directly as the mixing operator between genes, without attention, and normalizes the positive and negative channels separately, which prevents negative neighbors from being diluted by positive ones.
-
The study reports two notable findings: (i) performance drops substantially when the source distribution is too close to the control distribution; and (ii) the Sinkhorn endpoint generates fewer off-support cells than MMD.
Critical weaknesses
-
The ablation studies suggest that the main performance gains likely stem from the DEG-weighted endpoint loss. This loss concentrates the objective on genes with strong responses during training, which are precisely the genes on which Spearman LFC and direction match are computed.
-
The ablation analysis is incomplete. The unsigned variant simultaneously changes the edge signs, edge weights, and sampling order, so the contribution of each factor cannot be isolated.
-
The model applies to its input both during training and at every solver step. By the authors’ own derivation, for genes with , is negative with probability 1/2 for all . After rectification, the network cannot see the magnitude of these values, so the velocity along these coordinates is not identifiable. This design choice is neither motivated nor ablated. Notably, the public scDFM code clips only the final generated samples, not the network inputs.
-
The paper does not report the number of positive versus negative edges or the proportion of empty rows in the negative channel. It also provides no analysis of the gating mechanism or of the relative contributions of the positive and negative channels. Moreover, the TRRUST evaluation ranks candidate pairs by the absolute correlation computed on generated cells and ignores the annotated activation/repression direction. The claim that signs convey the direction of the response therefore remains empirically unverified.
Questions for authors
See my Critical Weakness.
2 comments
Dear Reviewer 4GQg,
We posted our detailed rebuttal and revised manuscript, including a capacity-matched signed-versus-unsigned comparison, independently retrained input-rectification controls, and additional channel-masking and gating diagnostics.
These experiments were specifically designed to address your concerns about the independent contribution of signed propagation, potential confounding factors, and the interpretation of the model components.
With the discussion period nearing its end, we would appreciate your feedback on whether these controlled experiments resolve your principal concerns. If any technical objections remain, please let us know so that we can address them explicitly.
We ask that these new results and the corresponding manuscript revisions be considered in your final assessment.
Best regards,
The Authors
Response to Reviewer 4GQg
Thank you for the precise technical review. We ran every experiment you proposed (Appendix E, blue). The results support the design.
W1. DEG-weighted endpoint loss. The weights are computed from training conditions only and enter only the loss; held-out conditions have no weights. Two experiments separate the loss from the architecture.
First, we removed DEG weighting entirely (Table 9, ComboSciPlex):
Without DEG weighting, SPARC with unweighted Sinkhorn achieves 0.896 Spearman LFC, 0.925 direction match, 0.922 Pearson delta, and 2.24×10−3 MSE, outperforming the published scDFM benchmark scores on all four metrics (Table 9). These are separate runs from the single-seed curves in Fig. 3A. The original DEG-weighted benchmark configuration achieves 0.908 Spearman LFC; this benchmark reference is from a separate training cohort and is not a within-seed matched loss-only comparison. With an unweighted MMD endpoint, SPARC retains a Spearman LFC of 0.894, compared with scDFM's published 0.829. The MMD implementations retain their original bandwidth conventions, as detailed in Appendix E.3.
Second, we fixed the DEG-weighted loss and changed only the routing (next point).
W2. Confounded unsigned ablation. Agreed, so we supplemented the original compound ablation with a capacity-matched control. Both arms keep two learned branches and gates, the same graph support, absolute edge magnitudes, neighbor ordering, sampled genes, parameter count and objective. The unsigned arm feeds the same magnitude-weighted adjacency to both branches (Table 8, seven conditions, 300k steps):
Sign is the only factor that differs, and signed routing wins on both metrics in every seed.
W3. Input rectification. We formalized your argument (Eq. 3) and ran the single-change experiment: the lower clamp is removed in training and at every solver evaluation; the upper cap and final-output clipping are unchanged (Table 10, ten Norman additive conditions, 300k steps):
Preserving negative intermediate states improves Spearman LFC in all three matched training pairs, by 0.023 on average. Table 10 uses independently retrained matched pairs, whose rectified arm averages 0.848, whereas the original benchmark checkpoints average 0.884 in Table 1. As clarified in Appendix E.4, these absolute scores are not interchangeable, and the controlled conclusion is the consistent within-pair improvement from removing the lower clamp.
W4a. Graph statistics, gating and channel contributions. Table 12 reports edge counts and empty rows for every dataset. Norman additive has 16,229 positive and 1,613 negative directed entries with 611 empty negative rows; ComboSciPlex has 17,832 and 2,186 with 496. Both channels are available for 38.7% and 50.4% of evaluated genes. In the trained Norman model at t=0.75, mean positive/negative gates are 0.852/0.767 with gated message norms 14.98/5.03. Masking one channel at inference, with weights, noise, controls and solver fixed, lowers Spearman LFC from 0.825 to 0.813 (negative masked) and 0.735 (positive masked). Both gates are open and both channels are used.
W4b. TRRUST and the meaning of sign. The signs encode co-variation type and route messages through separate learned transforms; we do not equate negative correlation with repression. TRRUST precision ranks by absolute correlation and is now stated as TF-target association recovery. The claim that sign informs response direction is tested directly: direction match rises from 0.923 to 0.946 under the capacity-matched control (Table 8), down-DEG direction accuracy is 0.870 against 0.783 and 0.640 for the simple predictors (Table 11), and top-20 direction-correct candidate precision is 16.5% for signed against 12.0% for unsigned routing (E.5).
Please let us know if any further experiments or clarification are required during the discussion period or before the camera ready version. We will be happy to discuss and perform further experiments.