- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper asks when to combine several external "priors" (compound or gene embeddings) when predicting single-cell responses to unseen perturbations. Instead of representation-level fusion, it proposes "response-level fusion": train one predictor per prior, then combine the M predicted values of each gene with a rule shared across genes and perturbations (Eq. 3-5). is a mean, median, trimmed mean, learned weighted mean, or a small "Shared MLP". Evaluation uses SciPlex3 and Replogle with cross-validation over perturbations and top-50 DEG metrics.
Critical strengths
- The comparison is set up the way I'd want it to be. All arms share backbone, folds and prior coverage (Sec 5.2, 5.3) so for the param-free rules only the stage of fusion changes.
- Tab. 2 has the rows needed for calibration, also where they're unflattering like on random control and perturbed cells, one-hot, a per-perturbation oracle, a Systema-style centered Pearson, and fold stds. Credit to the authors!
Critical weaknesses
- The biggest numbers (37.4% MMD gain, 11 of 12 cells, L348-350) come from the Shared MLP, and it reads to me like repair of an error shared by all base models, not fusion of priors? only sees M numbers per gene (Eq. 5). On SciPlex3 K562 its six inputs have MMD 0.329-0.347, worse than random perturbed cells (0.284), and sign agreement 0.535-0.549, below random control. The MLP turns these into 0.206 and 0.971 (Tab. 2). Picking the best prior per test perturbation (the oracle) only gets 0.299, and learned global weights equal the plain mean (0.299 vs 0.300). could still use how the six values agree but nothing shows that that is what it's doing. The authors do call the combiner error correction, but then the arms aren't treated equally. No row applies the same MLP to one prior or to the concatenation model. In their favour on RPE1 the inputs beat both trivial baselines and the MLP still winsbut the control is missing there too
- Without the Shared MLP, the evidence for response-level over representation-level fusion is mixed. The 12-of-12 and 10-of-12 counts compare the best method per class in each cell, picked on the test folds (L338-341). The median alone gets 7 of 12 (L343-344). Their largest gain (sign agreement) I can't really interpret on MCF7 the mean gives 0.581 and the median 0.928, with no input above 0.584, while MMD barely moves.
Clean question, careful table. I just can't attribute the main result to combining priors.
Questions for authors
- (experiment) What does the Shared MLP, with the same loss and fitting data, give with M=1, on the concatenation model, and on M seeds of one prior, in all three settings?
- (clarification) Why does the median lift sign agreement from 0.58 to 0.93 on MCF7? If it removes single-prior outliers, that's real fusion. What fraction of the top-50 DEGs is down-regulated?
- (experiment) Top-50 DEGs are standard, but does the ranking in Tab. 2 hold on all 5,000 genes? That would also move me up.
5 comments
Response to Reviewer kJZ6 (Part 2)
Question 3: Evaluation across all 5,000 genes
Reviewer question
“(experiment) Top-50 DEGs are standard, but does the ranking in Tab. 2 hold on all 5,000 genes? That would also move me up.”
Response
We have additionally evaluated all methods across all 5,000 genes in both pseudo-bulk and single-cell settings. Response-level fusion remains competitive in pseudo-bulk prediction and in single-cell genetic perturbation prediction, although its performance is more mixed for single-cell chemical perturbations.
However, recent studies have concluded that all-gene evaluation is fundamentally limited because of the noisy nature of the large number of non-differentially expressed genes. Because these metrics and models weight genes uniformly, the majority of genes that are not differentially expressed can mask biologically meaningfully differences in gene expression and lead to misleading model rankings and metrics (Miller et al., 2026; Vollenweider and Bühlmann, 2026).
[1] Henry E. Miller, Gabriel M. Mejia, Francis J. A. Leblanc, Brendan Swain, Bo Wang, and Lucas Paulo de Lima Camillo. Deep learning perturbation models can outperform baselines on calibrated metrics. Nature Biotechnology, October 2026. ISSN 1546-1696. doi: 10.1038/s41587-026-03307-w. URL https://doi.org/10.1038/s41587-026-03307-w.
[2] Michael Vollenweider and Peter B¨uhlmann. Signal, Bounds, and Baselines: Principles for Evaluating Virtual Cell Perturbation Models. bioRxiv, 2026. doi: 10.64898/2026.04.20.719650. URL https://www.biorxiv.org/content/early/2026/05/27/2026.04.20.719650.
Response to Reviewer kJZ6 (Part 3)
Table 6. All-5,000-gene single-cell results
Single-cell prediction on all 5,000 genes, with the lowest MMD biological prior shown for each dataset. Entries are mean ± SD across cross-validation folds. Bold marks the best displayed mean for each dataset and metric, including ties.
| Dataset | Fusion type | Method | MMD ↓ | Sign ↑ | Pearson ↑ |
|---|---|---|---|---|---|
| SP3 K562 | Learning-free | Random control cell | 0.528 ± 0.003 | 0.785 ± 0.002 | 0.082 ± 0.022 |
| SP3 K562 | Learning-free | Random perturbed cell | 0.526 ± 0.002 | 0.785 ± 0.002 | 0.030 ± 0.007 |
| SP3 K562 | Best single prior | ChemBERTa | 0.529 ± 0.002 | 0.309 ± 0.002 | 0.120 ± 0.025 |
| SP3 K562 | Representation | Concatenation | 0.529 ± 0.002 | 0.312 ± 0.004 | 0.124 ± 0.021 |
| SP3 K562 | Representation | Soft-argmax | 0.530 ± 0.002 | 0.307 ± 0.002 | 0.091 ± 0.013 |
| SP3 K562 | Representation | Perceiver-style | 0.530 ± 0.002 | 0.312 ± 0.004 | 0.121 ± 0.016 |
| SP3 K562 | Representation | PRESAGE-style | 0.529 ± 0.002 | 0.309 ± 0.003 | 0.109 ± 0.021 |
| SP3 K562 | Response | Mean | 0.536 ± 0.007 | 0.309 ± 0.002 | 0.116 ± 0.022 |
| SP3 K562 | Response | Median | 0.536 ± 0.008 | 0.356 ± 0.022 | 0.117 ± 0.021 |
| SP3 K562 | Response | Trimmed mean | 0.536 ± 0.008 | 0.344 ± 0.017 | 0.117 ± 0.021 |
| SP3 K562 | Response | Learned weighted mean | 0.536 ± 0.007 | 0.309 ± 0.002 | 0.116 ± 0.021 |
| SP3 K562 | Response | Shared MLP | 0.586 ± 0.005 | 0.363 ± 0.024 | 0.108 ± 0.018 |
| SP3 MCF7 | Learning-free | Random control cell | 0.494 ± 0.011 | 0.850 ± 0.001 | 0.231 ± 0.046 |
| SP3 MCF7 | Learning-free | Random perturbed cell | 0.492 ± 0.010 | 0.847 ± 0.001 | 0.045 ± 0.013 |
| SP3 MCF7 | Best single prior | ECFP | 0.488 ± 0.010 | 0.263 ± 0.004 | 0.275 ± 0.035 |
| SP3 MCF7 | Representation | Concatenation | 0.487 ± 0.009 | 0.252 ± 0.002 | 0.298 ± 0.038 |
| SP3 MCF7 | Representation | Soft-argmax | 0.496 ± 0.007 | 0.246 ± 0.005 | 0.266 ± 0.032 |
| SP3 MCF7 | Representation | Perceiver-style | 0.488 ± 0.009 | 0.255 ± 0.008 | 0.297 ± 0.048 |
| SP3 MCF7 | Representation | PRESAGE-style | 0.489 ± 0.009 | 0.260 ± 0.006 | 0.291 ± 0.043 |
| SP3 MCF7 | Response | Mean | 0.519 ± 0.011 | 0.261 ± 0.001 | 0.287 ± 0.029 |
| SP3 MCF7 | Response | Median | 0.516 ± 0.012 | 0.371 ± 0.003 | 0.295 ± 0.030 |
| SP3 MCF7 | Response | Trimmed mean | 0.518 ± 0.012 | 0.365 ± 0.005 | 0.295 ± 0.030 |
| SP3 MCF7 | Response | Learned weighted mean | 0.519 ± 0.011 | 0.261 ± 0.001 | 0.287 ± 0.029 |
| SP3 MCF7 | Response | Shared MLP | 0.505 ± 0.011 | 0.285 ± 0.044 | 0.256 ± 0.050 |
| Rep. K562 | Learning-free | Random control cell | 0.564 ± 0.002 | 0.499 ± 0.001 | 0.045 ± 0.009 |
| Rep. K562 | Learning-free | Random perturbed cell | 0.562 ± 0.001 | 0.528 ± 0.001 | -0.009 ± 0.002 |
| Rep. K562 | Best single prior | DepMap | 0.560 ± 0.001 | 0.552 ± 0.004 | 0.216 ± 0.006 |
| Rep. K562 | Representation | Concatenation | 0.559 ± 0.001 | 0.556 ± 0.003 | 0.231 ± 0.006 |
| Rep. K562 | Representation | Soft-argmax | 0.562 ± 0.001 | 0.550 ± 0.004 | 0.197 ± 0.010 |
| Rep. K562 | Representation | Perceiver-style | 0.561 ± 0.001 | 0.550 ± 0.001 | 0.206 ± 0.007 |
| Rep. K562 | Representation | PRESAGE-style | 0.560 ± 0.001 | 0.554 ± 0.004 | 0.222 ± 0.006 |
| Rep. K562 | Response | Mean | 0.601 ± 0.003 | 0.559 ± 0.004 | 0.233 ± 0.005 |
| Rep. K562 | Response | Median | 0.586 ± 0.002 | 0.546 ± 0.004 | 0.228 ± 0.006 |
| Rep. K562 | Response | Trimmed mean | 0.593 ± 0.002 | 0.550 ± 0.004 | 0.231 ± 0.005 |
| Rep. K562 | Response | Learned weighted mean | 0.600 ± 0.003 | 0.559 ± 0.004 | 0.234 ± 0.004 |
| Rep. K562 | Response | Shared MLP | 0.590 ± 0.013 | 0.525 ± 0.009 | 0.197 ± 0.021 |
| Rep. RPE1 | Learning-free | Random control cell | 0.645 ± 0.003 | 0.501 ± 0.001 | 0.115 ± 0.015 |
| Rep. RPE1 | Learning-free | Random perturbed cell | 0.615 ± 0.002 | 0.570 ± 0.002 | 0.091 ± 0.007 |
| Rep. RPE1 | Best single prior | DepMap | 0.601 ± 0.002 | 0.596 ± 0.003 | 0.349 ± 0.026 |
| Rep. RPE1 | Representation | Concatenation | 0.597 ± 0.003 | 0.604 ± 0.004 | 0.386 ± 0.015 |
| Rep. RPE1 | Representation | Soft-argmax | 0.609 ± 0.004 | 0.591 ± 0.003 | 0.264 ± 0.019 |
| Rep. RPE1 | Representation | Perceiver-style | 0.603 ± 0.003 | 0.596 ± 0.004 | 0.334 ± 0.013 |
| Rep. RPE1 | Representation | PRESAGE-style | 0.600 ± 0.003 | 0.598 ± 0.004 | 0.355 ± 0.012 |
| Rep. RPE1 | Response | Mean | 0.630 ± 0.005 | 0.606 ± 0.002 | 0.338 ± 0.010 |
| Rep. RPE1 | Response | Median | 0.618 ± 0.004 | 0.604 ± 0.002 | 0.332 ± 0.010 |
| Rep. RPE1 | Response | Trimmed mean | 0.624 ± 0.004 | 0.605 ± 0.002 | 0.335 ± 0.010 |
| Rep. RPE1 | Response | Learned weighted mean | 0.630 ± 0.004 | 0.607 ± 0.002 | 0.347 ± 0.010 |
| Rep. RPE1 | Response | Shared MLP | 0.650 ± 0.021 | 0.585 ± 0.006 | 0.319 ± 0.026 |
Response to Reviewer kJZ6
We thank the reviewer for their careful examination of our results and constructive feedback.
Critical weakness 1 and Question 1: Shared MLP and error correction
Reviewer comment — Critical weakness 1
Reviewer question — Question 1
Response
Please see our general response, where we provide the ablations to compare fusing the same prior with multiple seeds and fusing multiple priors. This disentangles the contributions of learned error correction and complementary biological priors, providing empirical evidence explaining when and why each contributes to the gains of response-level fusion.
To test the concatenation model, we ensemble multiple concatenation models using our response-level fusion techniques. Here, in order to keep comparison fair, we use PCA to reduce the dimensionality of the concatenation to normalize for model parameter count. Here we note that fusing a concatenation model is worse than both fusing multiple seeds of the best prior and fusing multiple priors.
Critical weakness 2: Best-of-class comparisons and aggregation
Reviewer comment
Response
Simple response-level aggregation methods, such as mean and median, introduce no additional learnable parameters, whereas sophisticated representation-level fusion methods could involve greater model complexity. Their ability to outperform any representation-level approaches on several metrics is therefore a meaningful result. Moreover, the Shared MLP achieves the strongest overall performance and is also favored by validation-based model selection, so excluding it would not reflect the performance of response-level fusion under practical model selection.
Additional reviewer comment: Attribution to priors
Reviewer comment
Response
Our new ablations (Tables 1–3) explicitly try to separate the contributions of learned error correction from complementary information provided across biological priors. Although we found that the original genetic prior set provided limited gains over same-prior ensembling, this does not necessarily mean that combining heterogeneous priors offers no additional benefit. To show this, we expand the set to seven biological priors and found that multi-prior fusion outperformed a seven-seed best single-prior ensemble on four of six metrics across the two genetic datasets (Table 3). These results suggest that when priors provide complementary biological information, the multi-prior ensembling still plays a major role.
Question 2: Median aggregation on MCF7
Reviewer question
Response
On MCF7, 89.1% of the top 50 DEGs are down-regulated, while individual predictors tend to underestimate perturbation effects. Median aggregation is less sensitive to unusually high predictions than mean aggregation, potentially shifting small predicted effects toward downregulation, which explains the improvement we observed in sign agreement.