- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes VCLMU, a virtual cell model that represents cellular state using Latent Mechanism Units (LMUs). Each LMU combines a shared identity grounded in multimodal gene evidence with an observation-dependent state. An action-conditioned stochastic transition module predicts changes in these states before decoding the transcriptional response. A separate inverse head identifies perturbations using the same state representation. The model undergoes pseudo-bulk and single-cell pretraining. Evaluation across ten datasets reports improvements in forward prediction and inverse retrieval, alongside component ablations and enrichment analyses of LMU-associated genes.
Critical strengths
-
A concrete and potentially useful modular representation. The separation between shared LMU identities and observation-dependent states provides an explicit structure for organizing perturbation responses. Multimodal gene evidence and module-specific updates offer a reasonable inductive bias for this problem.
-
Promising empirical results across multiple datasets. The reported forward-prediction improvements, particularly in response-gene recovery, motivate further investigation. The component ablations also provide initial evidence that LMU factorization and transition modeling contribute to performance.
Critical weaknesses
-
The distinctive contribution and mechanistic interpretation are not sufficiently established. Biologically informed latent programs, sparse perturbation mechanisms, and object-centric dynamics have substantial precedent, including VEGA, expiMap, and SAMS-VAE. The contribution appears to lie in their particular integration, but the paper does not clearly demonstrate capabilities attributable to this integration. Moreover, enrichment between LMU gene sets and response genes establishes biological association rather than validating the predicted mechanism-state transitions. Comparisons against fixed programs, knowledge-derived clusters, or initialization-only modules would strengthen this interpretation.
-
The experimental controls and uncertainty reporting limit the conclusions. VCLMU combines extensive biological evidence with large-scale perturbation pretraining, whereas several baselines use downstream-only training or frozen encoders. The inverse comparison also contrasts a directly supervised inverse head with forward-consistency retrieval. These comparisons do not adequately isolate the contribution of LMUs. In addition, condition-level mean-expression metrics do not directly validate the claimed modeling of single-cell heterogeneity.
Table 7 contains a separate reporting inconsistency: 42 of 169 nonmissing entries cannot represent mean ± SD over three nonnegative seed-level scores, as stated in its caption, even allowing for rounding. For three such scores, the sample SD cannot exceed times their mean. Scouter’s Fr-Cocu result, , exceeds the maximum possible SD of approximately . VCLMU w/o Transition on Fr-IFN, , also exceeds its maximum of approximately . The aggregation unit requires clarification before these values can be interpreted as variability across seeds.
Questions for authors
-
Which comparisons match VCLMU’s pretraining data, biological evidence, and training budget? Are the LMU ablations capacity-matched, and is there an equally supervised non-LMU inverse baseline?
-
How were downstream test experiments and perturbation identities checked against all pretraining resources?
-
What is the aggregation unit for Table 7’s SDs? Please provide the three seed-level aggregate scores and clarify whether the reported variability instead comes from individual perturbations or pooled observations.
-
Appendix D.6 computes NDCG only over successful top- retrievals. Could the authors report conventional all-query NDCG, including failed retrievals?
5 comments
W2. Experimental controls and uncertainty reporting
Baseline training and LMU contributions. We clarified the baseline training procedures in Section 4.1 and Appendix D.2. Geneformer-FT, scFoundation-FT, Tahoe-x1-FT, and scGPT-FT fine-tune their pretrained encoders on each downstream training split. Tahoe-x1 was pretrained in part on Tahoe-100M, a perturbation compendium. Both stages of VCLMU pretraining exclude all ten downstream evaluation datasets. The cross-model comparisons therefore evaluate complete systems with different pretraining backgrounds under a common downstream protocol.
To test the predictive contribution of the LMUs within the trained model, we added a targeted ablation experiment in Figure 3. For seven perturbations shared by Replogle K562 and RPE1, we compared the drop in corr20DE after ablating the target LMU states with the drop after ablating the same number of randomly selected LMU states. When aggregated across the two cell lines, target ablation caused a larger drop for all seven perturbations (Fig. 3a). For NDUFB6, TIMM13, UQCRB, and UQCRFS1, the drop was larger in each cell line separately (Fig. 3b). This comparison shows that predictions for these perturbations depend more strongly on their associated LMU states than on randomly selected states.
Direct inverse prediction. We added Heimdall and GREmLN, along with linear and MLP scorers trained directly on the same downstream perturbation labels using the same biological-prior embeddings as VCLMU. All methods rank the same candidate perturbations. Table 3 reports macro means ± cross-dataset standard deviations, and Appendix Table 7 gives the per-dataset results for the four added baselines.
| Method | Hit@1 | Hit@5 | Hit@20 | MRR | NDCG@20 |
|---|---|---|---|---|---|
| Linear (multimodal) | 0.0135±0.0229 | 0.0585±0.0492 | 0.1840±0.0965 | 0.0512±0.0344 | 0.3664±0.0533 |
| MLP (multimodal) | 0.0038±0.0039 | 0.0520±0.0396 | 0.2072±0.1078 | 0.0439±0.0156 | 0.3305±0.0237 |
| Heimdall | 0.0110±0.0088 | 0.0492±0.0357 | 0.1412±0.0854 | 0.0435±0.0256 | 0.3936±0.0513 |
| GREmLN | 0.0301±0.0322 | 0.1010±0.0815 | 0.2925±0.1791 | 0.0838±0.0589 | 0.3881±0.0360 |
| VCLMU-FT | 0.1645±0.0827 | 0.2249±0.0874 | 0.3274±0.0907 | 0.2043±0.0819 | 0.6544±0.1129 |
Single-cell heterogeneity. We added a comparison of predicted and observed single-cell expression distributions within perturbation conditions using an unbiased RBF-kernel estimate of MMD². Table 11 reports all ten datasets. The three lowest averages are shown below. VCLMU achieves the lowest average MMD² and the lowest MMD² on Tian CRISPRa, with rankings varying across datasets.
Average MMD² across ten datasets (lower is better)
| Method | VCLMU | GEARS | scFoundation-FT | Geneformer-FT | LPM | Tahoe-x1-FT |
|---|---|---|---|---|---|---|
| MMD² | 0.2209±0.0292 | 0.2229±0.1383 | 0.2240±0.0423 | 0.2343±0.0351 | 0.2450±0.0874 | 0.3877±0.1101 |
| Method | scDFM | scGPT-FT | GenePert | CPA | scGPT-ZS |
|---|---|---|---|---|---|
| MMD² | 0.3877±0.0723 | 0.4234±0.1252 | 0.7938±0.0109 | 0.8025±0.0299 | 0.8255±0.0293 |
Correction to the original Table 7. For the affected Jaccard@20 entries, the reported standard deviations were calculated across individual perturbations, although the caption described variability across three seeds. We recalculated these standard deviations by first averaging the test perturbation scores within each seed and then computing the standard deviation across the three seed-level means. The corrected per-dataset results appear in Table 9 (p. 30), and Table 10 (p. 31) lists the individual seed-level scores for the baselines. The correction affects the reported variability within datasets. The dataset-level means used to compute the main-text Table 1 results are unchanged.
Q1. Matching pretraining, model capacity, and inverse supervision
LMU ablation design. The ablations in Table 2 use the same downstream splits and evaluation protocol. The w/o LMU variant replaces the collection of LMU states with a single global latent state to evaluate the effect of removing LMU factorization.
Equally supervised non-LMU inverse baselines. We added linear and MLP candidate scorers that use the same biological-prior embeddings as VCLMU and train on the same downstream perturbation labels, candidate-gene sets, and splits. Both score candidate genes directly from the observed expression change. This comparison matches the biological-prior inputs and downstream inverse supervision.
Table 3 reports macro means ± cross-dataset standard deviations over the ten evaluation datasets. Appendix Table 7 reports the per-dataset results for both added scorers.
Q2. Exclusion of downstream experiments and test perturbations
We excluded all ten downstream evaluation datasets in their entirety from both Stage 1 pseudo-bulk pretraining and Stage 2 single-cell pretraining. Consequently, neither stage uses experimental observations or perturbation-response measurements from these datasets. Appendix B identifies the data sources used in each pretraining stage and clarifies the exclusion of the downstream benchmarks.
Within each downstream dataset, we split by perturbation identity. For each of three random seeds, we hold out 25% of the perturbation identities for testing and assign all cells associated with each test perturbation to the test set. The remaining identities are divided into disjoint training and validation sets (Appendix D.1). Thus, each test perturbation is unseen during downstream training and validation for that benchmark. The pretraining exclusion applies to the experimental dataset as a whole, so the same target gene may appear in an independent pretraining experiment.
Q3. Aggregation of Jaccard@20 across random seeds
The original Table 7 reported SDs across individual test perturbations for some baseline entries. Some means also used pooled perturbation scores or previously rounded values. We corrected both issues in the revised Table 9 (p. 30). For each method, dataset, and seed, we first average Jaccard@20 over that seed’s test perturbations. We then give the three unrounded seed-level scores equal weight and report their mean and population SD. Each SD in Table 9 therefore measures variation across three random seeds.
We also updated the Jaccard@20 macro means and cross-dataset SDs in main-text Table 1 using the corrected means for the ten datasets. These corrections leave the overall performance comparison and the paper’s main conclusions unchanged.
Table 10 (p. 31), reproduced below, gives the three seed-level scores for all 11 baselines on all ten datasets. Each cell lists seeds 1 / 2 / 3.
| Method | Adamson | Fr-Ctrl | Fr-Cocu | Fr-IFN | Norman |
|---|---|---|---|---|---|
| GenePert | 0.0730/0.0362/0.0746 | 0.0043/0.0074/0.0049 | 0.0031/0.0025/0.0013 | 0.0100/0.0075/0.0031 | 0.0151/0.0380/0.0552 |
| Geneformer-FT | 0.2557/0.2445/0.2572 | 0.0062/0.0061/0.0098 | 0.0024/0.0031/0.0037 | 0.0174/0.0137/0.0140 | 0.1188/0.1160/0.1306 |
| scFoundation-FT | 0.2557/0.2445/0.2572 | 0.0062/0.0061/0.0098 | 0.0024/0.0038/0.0037 | 0.0174/0.0137/0.0140 | 0.1520/0.1117/0.1377 |
| Scouter | 0.2421/0.1950/0.2231 | 0.0043/0.0024/0.0049 | 0.0337/0.0195/0.0396 | 0.0339/0.0233/0.0345 | 0.1391/0.1176/0.1051 |
| scDFM | 0.0325/0.0456/0.0208 | 0.0024/0.0031/0.0031 | 0.0037/0.0031/0.0018 | 0.0049/0.0018/0.0031 | 0.0057/0.0035/0.0029 |
| LPM | 0.0793/0.0735/0.0723 | 0.0081/0.0068/0.0099 | 0.0024/0.0006/0.0024 | 0.0181/0.0117/0.0092 | 0.0390/0.0424/0.0488 |
| Tahoe-x1-FT | 0.0322/0.0186/0.0497 | 0.0061/0.0099/0.0118 | 0.0043/0.0037/0.0024 | 0.0088/0.0086/0.0105 | 0.0595/0.0289/0.0472 |
| CPA | 0.1719/0.1671/0.1559 | 0.0062/0.0056/0.0069 | 0.0102/0.0062/0.0062 | 0.0151/0.0086/0.0105 | 0.0461/0.0399/0.0555 |
| GEARS | 0.1633/0.2267/0.2363 | 0.0174/0.0174/0.0226 | 0.0258/0.0104/0.0252 | 0.0287/0.0372/0.0395 | 0.1233/0.1208/0.1100 |
| scGPT-ZS | 0.1932/0.1819/0.1606 | 0.0069/0.0043/0.0062 | 0.0121/0.0062/0.0068 | 0.0146/0.0074/0.0093 | 0.0422/0.0407/0.0555 |
| scGPT-FT | 0.1107/0.1959/0.1506 | 0.0086/0.0161/0.0079 | 0.0136/0.0141/0.0116 | 0.0130/0.0123/0.0093 | 0.0694/0.0434/0.0898 |
| Method | K562 | RPE1 | CRISPRa | CRISPRi | Xu |
|---|---|---|---|---|---|
| GenePert | 0.0234/0.0179/0.0046 | 0.0053/0.0094/0.0106 | 0.0072/0.0167/0.0135 | 0.0045/0.0045/0.0017 | 0.0036/0.0021/0.0052 |
| Geneformer-FT | 0.0277/0.0275/0.0264 | 0.0279/0.0338/0.0330 | 0.0936/0.0960/0.0749 | 0.0271/0.0329/0.0282 | 0.0469/0.0497/0.0428 |
| scFoundation-FT | 0.0275/0.0265/0.0270 | 0.0279/0.0336/0.0348 | 0.0886/0.0938/0.0748 | 0.0264/0.0322/0.0330 | 0.0469/0.0483/0.0439 |
| Scouter | 0.0610/0.0564/0.0583 | 0.0421/0.0487/0.0460 | 0.0487/0.0471/0.0325 | 0.0154/0.0165/0.0171 | 0.0150/0.0168/0.0203 |
| scDFM | 0.0248/0.0140/0.0162 | 0.0181/0.0083/0.0152 | 0.0043/0.0010/0.0031 | 0.0023/0.0022/0.0034 | 0.0062/0.0057/0.0118 |
| LPM | 0.0370/0.0350/0.0343 | 0.0200/0.0193/0.0180 | 0.0374/0.0394/0.0301 | 0.0097/0.0096/0.0079 | 0.0460/0.0480/0.0468 |
| Tahoe-x1-FT | 0.0358/0.0338/0.0331 | 0.0187/0.0192/0.0174 | 0.0200/0.0062/0.0189 | 0.0125/0.0056/0.0039 | 0.0092/0.0242/0.0195 |
| CPA | 0.0593/0.0556/0.0540 | 0.0265/0.0285/0.0272 | 0.0297/0.0241/0.0190 | 0.0079/0.0091/0.0068 | 0.0443/0.0464/0.0456 |
| GEARS | 0.0247/0.0240/0.0235 | 0.0313/0.0335/0.0376 | 0.0410/0.0713/0.0511 | 0.0101/0.0067/0.0159 | 0.0314/0.0329/0.0314 |
| scGPT-ZS | 0.0551/0.0545/0.0596 | 0.0286/0.0291/0.0270 | 0.0170/0.0185/0.0114 | 0.0084/0.0102/0.0068 | 0.0545/0.0508/0.0480 |
| scGPT-FT | 0.0264/0.0559/0.0544 | 0.0352/0.0390/0.0395 | 0.0094/0.0296/0.0156 | 0.0089/0.0079/0.0097 | 0.0484/0.0270/0.0442 |
W1. Distinctive contribution of VCLMU and interpretation of LMU state transitions
We have expanded Section 2 (Related Work) to clarify how VCLMU builds on prior work. VEGA and expiMap use biological prior knowledge to construct interpretable gene-program representations, and SAMS-VAE models sparse perturbation effects in latent space. VCLMU integrates these ideas with object-centric dynamics for perturbation-response modeling. Each LMU has a reusable identity informed by multimodal gene evidence and a state inferred from the current cellular profile. For forward prediction, the model predicts how a perturbation updates each active LMU and decodes the resulting states into an expression response. For inverse retrieval, it uses the same LMU representation to infer the perturbation from the observed state change.
Table 2 tests the contribution of two parts of this design. Replacing the LMUs with a single latent state reduces mean Jaccard@20 from 0.1648 to 0.1388 and mean ΔPCC from 0.3992 to 0.3536 across ten datasets. Retaining the encoded control state and perturbation representation but removing the explicit LMU transition reduces these metrics to 0.1362 and 0.3422, respectively. These comparisons show that both representing the cell with multiple LMUs and explicitly updating their states contribute to forward prediction.
In Section 4.5 (LMU Interpretability), we clarify that the enrichment analysis is a case study of how learned LMUs relate to biological response programs. We also added a targeted ablation experiment that tests whether the LMUs associated with a perturbation contribute to predicting its response. Across the seven perturbations shared by Replogle K562 and RPE1, ablating the target LMU state caused a larger decrease in corr20DE than size-matched random-LMU ablation for all seven when results were aggregated across cell lines (Fig. 3a). For NDUFB6, TIMM13, UQCRB, and UQCRFS1, this effect was positive in each cell line separately (Fig. 3b).
The ablations establish a predictive role for the learned LMU states and show cross-cell-line consistency for four shared perturbations. The reviewer’s proposed controls would address a complementary question. Comparing learned LMUs with fixed gene programs, knowledge-derived clusters, or modules frozen at initialization would test how much learning the LMU identities adds to the predictive value of the biological priors.