acceptodds
Under review as a conference paper at ICLR 2027

VCLMU: Mechanism-Centric Virtual Cell World Modeling for Perturbation Response

Abstract

Predicting cellular responses to genetic perturbations is a central capability of virtual cell models and an important foundation for computational modeling of biological perturbations. Many existing perturbation models achieve accurate prediction of observed responses without explicitly representing the internal state transitions or biological mechanisms underlying them. We introduce VCLMU, a mechanism-centric world model for virtual cells. VCLMU brings the object-centric principle of world modeling to cellular systems by representing internal cellular state as a collection of Latent Mechanism Units (LMUs),which represent latent biological mechanisms corresponding to objects in an object-centric world model. Each LMU has a reusable identity and a cell-specific state. Its identity is biologically grounded in multimodal gene evidence, while its state is obtained from the current cellular observation to capture cell-specific mechanism states. Given the current cellular state and a perturbation action, the world model predicts the perturbation-induced transition of each active LMU, then aggregates the resulting mechanism states and decodes them into the transcriptional response. Built on the same LMU state representation, VCLMU supports both forward perturbation response prediction and inverse perturbation inference. The forward model predicts the perturbed cellular response from the current state and perturbation action, whereas the inverse model infers the perturbation action from the observed state transition. VCLMU is trained through two-stage pretraining. The first stage uses approximately 165K pseudo-bulk perturbation profiles to learn stable population-level responses for individual perturbation conditions and establish robust perturbation dynamics. The second stage further pretrains on large-scale, gene-aligned single-cell perturbation data to capture cell-specific heterogeneity and stochastic variation beyond the population mean. We evaluate VCLMU through forward response prediction, inverse perturbation inference, mechanistic interpretation, and ablation studies, demonstrating its ability to generalize to unseen perturbations while learning biologically structured latent dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual cell modeling, World models, Latent biological mechanisms, Perturbation response prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_e7JoOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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 3\sqrt{3} times their mean. Scouter’s Fr-Cocu result, 0.0309±0.13970.0309 \pm 0.1397, exceeds the maximum possible SD of approximately 0.05350.0535. VCLMU w/o Transition on Fr-IFN, 0.0020±0.00500.0020 \pm 0.0050, also exceeds its maximum of approximately 0.00350.0035. The aggregation unit requires clarification before these values can be interpreted as variability across seeds.

Questions for authors
  1. 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?

  2. How were downstream test experiments and perturbation identities checked against all pretraining resources?

  3. 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.

  4. Appendix D.6 computes NDCG only over successful top-kk retrievals. Could the authors report conventional all-query NDCG, including failed retrievals?

5 comments

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.

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.

MethodHit@1Hit@5Hit@20MRRNDCG@20
Linear (multimodal)0.0135±0.02290.0585±0.04920.1840±0.09650.0512±0.03440.3664±0.0533
MLP (multimodal)0.0038±0.00390.0520±0.03960.2072±0.10780.0439±0.01560.3305±0.0237
Heimdall0.0110±0.00880.0492±0.03570.1412±0.08540.0435±0.02560.3936±0.0513
GREmLN0.0301±0.03220.1010±0.08150.2925±0.17910.0838±0.05890.3881±0.0360
VCLMU-FT0.1645±0.08270.2249±0.08740.3274±0.09070.2043±0.08190.6544±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)

MethodVCLMUGEARSscFoundation-FTGeneformer-FTLPMTahoe-x1-FT
MMD²0.2209±0.02920.2229±0.13830.2240±0.04230.2343±0.03510.2450±0.08740.3877±0.1101
MethodscDFMscGPT-FTGenePertCPAscGPT-ZS
MMD²0.3877±0.07230.4234±0.12520.7938±0.01090.8025±0.02990.8255±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.

MethodAdamsonFr-CtrlFr-CocuFr-IFNNorman
GenePert0.0730/0.0362/0.07460.0043/0.0074/0.00490.0031/0.0025/0.00130.0100/0.0075/0.00310.0151/0.0380/0.0552
Geneformer-FT0.2557/0.2445/0.25720.0062/0.0061/0.00980.0024/0.0031/0.00370.0174/0.0137/0.01400.1188/0.1160/0.1306
scFoundation-FT0.2557/0.2445/0.25720.0062/0.0061/0.00980.0024/0.0038/0.00370.0174/0.0137/0.01400.1520/0.1117/0.1377
Scouter0.2421/0.1950/0.22310.0043/0.0024/0.00490.0337/0.0195/0.03960.0339/0.0233/0.03450.1391/0.1176/0.1051
scDFM0.0325/0.0456/0.02080.0024/0.0031/0.00310.0037/0.0031/0.00180.0049/0.0018/0.00310.0057/0.0035/0.0029
LPM0.0793/0.0735/0.07230.0081/0.0068/0.00990.0024/0.0006/0.00240.0181/0.0117/0.00920.0390/0.0424/0.0488
Tahoe-x1-FT0.0322/0.0186/0.04970.0061/0.0099/0.01180.0043/0.0037/0.00240.0088/0.0086/0.01050.0595/0.0289/0.0472
CPA0.1719/0.1671/0.15590.0062/0.0056/0.00690.0102/0.0062/0.00620.0151/0.0086/0.01050.0461/0.0399/0.0555
GEARS0.1633/0.2267/0.23630.0174/0.0174/0.02260.0258/0.0104/0.02520.0287/0.0372/0.03950.1233/0.1208/0.1100
scGPT-ZS0.1932/0.1819/0.16060.0069/0.0043/0.00620.0121/0.0062/0.00680.0146/0.0074/0.00930.0422/0.0407/0.0555
scGPT-FT0.1107/0.1959/0.15060.0086/0.0161/0.00790.0136/0.0141/0.01160.0130/0.0123/0.00930.0694/0.0434/0.0898
MethodK562RPE1CRISPRaCRISPRiXu
GenePert0.0234/0.0179/0.00460.0053/0.0094/0.01060.0072/0.0167/0.01350.0045/0.0045/0.00170.0036/0.0021/0.0052
Geneformer-FT0.0277/0.0275/0.02640.0279/0.0338/0.03300.0936/0.0960/0.07490.0271/0.0329/0.02820.0469/0.0497/0.0428
scFoundation-FT0.0275/0.0265/0.02700.0279/0.0336/0.03480.0886/0.0938/0.07480.0264/0.0322/0.03300.0469/0.0483/0.0439
Scouter0.0610/0.0564/0.05830.0421/0.0487/0.04600.0487/0.0471/0.03250.0154/0.0165/0.01710.0150/0.0168/0.0203
scDFM0.0248/0.0140/0.01620.0181/0.0083/0.01520.0043/0.0010/0.00310.0023/0.0022/0.00340.0062/0.0057/0.0118
LPM0.0370/0.0350/0.03430.0200/0.0193/0.01800.0374/0.0394/0.03010.0097/0.0096/0.00790.0460/0.0480/0.0468
Tahoe-x1-FT0.0358/0.0338/0.03310.0187/0.0192/0.01740.0200/0.0062/0.01890.0125/0.0056/0.00390.0092/0.0242/0.0195
CPA0.0593/0.0556/0.05400.0265/0.0285/0.02720.0297/0.0241/0.01900.0079/0.0091/0.00680.0443/0.0464/0.0456
GEARS0.0247/0.0240/0.02350.0313/0.0335/0.03760.0410/0.0713/0.05110.0101/0.0067/0.01590.0314/0.0329/0.0314
scGPT-ZS0.0551/0.0545/0.05960.0286/0.0291/0.02700.0170/0.0185/0.01140.0084/0.0102/0.00680.0545/0.0508/0.0480
scGPT-FT0.0264/0.0559/0.05440.0352/0.0390/0.03950.0094/0.0296/0.01560.0089/0.0079/0.00970.0484/0.0270/0.0442
Reviewer_vhqxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces VCLMU, a mechanism-centric virtual cell model that approaches genetic perturbation prediction through the lens of object-centric world modeling. It proposes Latent Mechanism Units (LMUs) to represent biologically structured cellular states and model perturbation-induced transitions, supporting both forward response prediction and inverse perturbation inference. The paper is generally well-written, presents an interesting and sufficiently novel architectural perspective, and reports promising results across multiple datasets with extensive benchmarking and ablation studies. However, I would like to raise several concerns remain regarding the fairness of baseline comparisons, the separation between pretraining and evaluation data, and the extent to which the learned LMUs genuinely capture biological mechanisms rather than structured predictive representations.

Critical strengths

The paper is well-written and easy to follow, with a clearly motivated and illustrated architecture. The authors also provide ablation studies evaluating the contributions of LMU factorization, stochastic dynamics, transition modeling, and GRN information.

The authors compare against multiple state-of-the-art perturbation prediction methods across ten datasets, covering different cellular contexts and perturbation types (CRISPRa/CRISPRi). VCLMU achieves the best average performance in both forward and inverse perturbation prediction, with particularly strong improvements in Jaccard@20 and inverse Hit@1.

The authors evaluate whether learned LMUs capture biologically meaningful response programs, identifying associations with known cellular processes. They further validate these associations against matched random gene modules, providing additional evidence supporting the proposed architecture beyond predictive performance alone.

Critical weaknesses

Forward prediction is evaluated using ΔPCC and Jaccard@20, which do not directly assess expression-change magnitude errors or the quality of predicted single-cell distributions. Including RMSE on condition-level responses and distributional metrics such as MMD would provide a more comprehensive evaluation, particularly given the model's stochastic formulation.

Although the paper reports strong inverse perturbation prediction results, most baselines rely on forward-consistency retrieval rather than directly trained inverse models. Several relevant approaches explicitly address reverse perturbation prediction, including CIPHER, Heimdall, GREmLN, and PDGrapher. Comparisons against some of these methods under matched experimental settings would strengthen the evaluation and help establish whether VCLMU's gains arise from its proposed architecture rather than the choice of inverse learning objective.

Although the authors provide LMU enrichment analyses against random gene modules, stronger biologically informed controls and direct validation of LMU state transitions are missing. Additionally, the evaluation focuses on average responses without assessing how well the stochastic model captures single-cell heterogeneity.

Questions for authors
  • Although the authors describe perturbation-disjoint splitting procedures, it remains unclear how held-out perturbations are excluded across both pretraining stages and downstream evaluation splits. Could the authors clarify this?

  • The authors discuss Heimdall and GREmLN in the related work as approaches addressing reverse perturbation prediction, yet neither is included in the experimental comparisons. Could the authors clarify the rationale for excluding these methods?

3 comments

W1. Distributional evaluation of forward perturbation predictions

We added a distributional evaluation to assess how closely the predicted single-cell expression profiles match the observed profiles. We compute MMD² using an unbiased estimator with an RBF kernel. Lower values indicate closer agreement between the predicted and observed distributions. This complements PCC and Jaccard@20 by evaluating the distribution of cellular responses within a perturbation condition.

Across ten datasets, VCLMU achieves the lowest mean MMD² (0.2209), followed by GEARS (0.2229) and scFoundation-FT (0.2240).

The full results are shown below and in the appendix table 11 titled “Per-dataset MMD² (RBF, unbiased) for forward perturbation prediction.” Each dataset column reports the mean ± standard deviation over three seeds. The final column summarizes performance across the ten datasets. Bold indicates the lowest mean and italics the second lowest.

MethodAdamsonFr-CtrlFr-CocuFr-IFNNormanK562RPE1CRISPRaCRISPRiXuAverage
GenePert0.8045±0.00320.7983±0.00680.8012±0.00320.8044±0.00390.7956±0.01150.7957±0.00630.8000±0.00350.7716±0.03720.7843±0.02150.7820±0.01710.7938±0.0109
Geneformer-FT0.1852±0.01150.2544±0.00920.2831±0.01490.2696±0.00620.2396±0.02400.2158±0.00720.2137±0.00780.2368±0.04070.1803±0.01200.2649±0.01980.2343±0.0351
scFoundation-FT0.1886±0.01510.2539±0.00930.2832±0.01400.2686±0.00730.1544±0.05820.2155±0.00650.2137±0.00960.2159±0.03540.1821±0.00920.2643±0.01930.2240±0.0423
scDFM0.3608±0.01690.3270±0.01320.3339±0.03340.3353±0.01050.4019±0.02230.4936±0.05560.5392±0.02420.3453±0.02140.3808±0.03380.3591±0.03990.3877±0.0723
LPM0.4704±0.02890.2221±0.00650.1800±0.00420.2229±0.00510.2337±0.02380.1686±0.00250.2604±0.00270.2784±0.00990.2414±0.00740.1720±0.02080.2450±0.0874
Tahoe-x1-FT0.5298±0.03480.3616±0.01890.3757±0.02940.3756±0.02440.3813±0.03490.1715±0.00340.2685±0.00100.5347±0.00510.4329±0.03160.4456±0.07830.3877±0.1101
CPA0.8345±0.00580.7830±0.00350.7933±0.00300.7982±0.00330.7689±0.00240.8626±0.02170.8305±0.00750.7778±0.00510.7864±0.00480.7896±0.04400.8025±0.0299
GEARS0.2925±0.06960.1621±0.05870.1114±0.00180.1573±0.01880.1333±0.02260.1374±0.00450.1562±0.05730.4769±0.06820.4595±0.09110.1425±0.03240.2229±0.1383
scGPT-ZS0.8523±0.01340.7980±0.00540.8120±0.00320.8177±0.00470.8019±0.00200.8912±0.03010.8451±0.01010.8021±0.00210.8121±0.00130.8227±0.01390.8255±0.0293
scGPT-FT0.3228±0.01840.4633±0.02370.4504±0.01720.4854±0.02530.2798±0.18480.5792±0.28150.1842±0.01180.4182±0.02280.4933±0.01480.5571±0.19230.4234±0.1252
VCLMU0.2147±0.13130.1954±0.00190.2232±0.04770.2059±0.00890.2447±0.00930.2598±0.06080.2269±0.05520.1838±0.03960.2669±0.07440.1873±0.05990.2209±0.0292

W2. Comparison with direct inverse perturbation models

VCLMU is designed to support forward response prediction and inverse perturbation retrieval within a unified framework. Many models developed for forward perturbation-response prediction can also perform inverse retrieval by predicting the response for each candidate perturbation and ranking candidates by agreement with the observed response. This motivated the inclusion of forward-consistency retrieval in our original comparison.

We added comparisons with Heimdall and GREmLN under the same downstream evaluation protocol. We also added linear and MLP scorers trained directly on the same downstream perturbation labels. Both scorers use the same biological-prior embeddings as VCLMU to represent candidate genes.

The full comparison is in Table 3, and Appendix Table 7 reports the per-dataset results.

MethodHit@1Hit@5Hit@20MRRNDCG@20
Linear (multimodal)0.0135 ± 0.02290.0585 ± 0.04920.1840 ± 0.09650.0512 ± 0.03440.3664 ± 0.0533
MLP (multimodal)0.0038 ± 0.00390.0520 ± 0.03960.2072 ± 0.10780.0439 ± 0.01560.3305 ± 0.0237
Heimdall0.0110 ± 0.00880.0492 ± 0.03570.1412 ± 0.08540.0435 ± 0.02560.3936 ± 0.0513
GREmLN0.0301 ± 0.03220.1010 ± 0.08150.2925 ± 0.17910.0838 ± 0.05890.3881 ± 0.0360
VCLMU-FT0.1645 ± 0.08270.2249 ± 0.08740.3274 ± 0.09070.2043 ± 0.08190.6544 ± 0.1129

GREmLN achieved the highest Hit@1 and MRR among the four added baselines (0.0301 and 0.0838), while VCLMU-FT achieved 0.1645 and 0.2043, respectively. The advantage therefore extends to comparisons with models trained directly for inverse retrieval. The downstream supervision is matched for the linear and MLP scorers, while differences in pretraining remain relevant when interpreting the contribution of the architecture.

W3. Functional validation of LMU states and evaluation of single-cell heterogeneity

We added a targeted ablation experiment to test whether perturbation-associated LMU states contribute to predicting the corresponding expression response. For the seven perturbations shared by Replogle K562 and RPE1, we compared the decrease in corr20DE\mathrm{corr}_{20\mathrm{DE}} after ablating the target LMU state with the decrease after ablating a size-matched set of random LMU states. When the results were aggregated across the two cell lines, target ablation produced a larger decrease for all seven perturbations (Fig. 3a). This complements the gene-set enrichment analysis by testing how the model’s prediction changes when the associated LMU state is removed.

We then examined whether this contribution was consistent across cell lines. For each perturbation, we calculated the LMU-specific effect separately in K562 and RPE1 as the target-ablation drop minus the random-ablation drop. NDUFB6, TIMM13, UQCRB, and UQCRFS1 showed positive effects in both cell lines, with similar effect magnitudes for UQCRB and UQCRFS1 (Fig. 3b). We report this case study in Section 4.5, “LMU Interpretability.” It directly tests the predictive contribution of the learned states. Independent biological validation of the underlying transitions remains an open question.

To assess single-cell heterogeneity, we also compared the distributions of predicted and observed single-cell expression profiles within perturbation conditions using an unbiased RBF-kernel estimator of MMD². Lower MMD² indicates closer agreement between the two distributions. Across ten datasets, VCLMU achieved the lowest mean MMD² (0.2209), followed by GEARS (0.2229) and scFoundation-FT (0.2240). The appendix table titled “Per-dataset MMD² (RBF, unbiased) for forward perturbation prediction” reports results for each dataset as the mean ± standard deviation over three seeds.

Q1. Exclusion of held-out perturbations during pretraining and downstream evaluation

We clarify the exclusion protocol at two levels. First, each of the ten downstream benchmark datasets is excluded in its entirety from both Stage 1 pseudo-bulk pretraining and Stage 2 single-cell pretraining. Thus, neither pretraining stage uses any response observations from a downstream benchmark. This exclusion applies at the dataset level. A gene held out for downstream testing can still appear as a perturbation target in a different pretraining dataset.

Second, within each benchmark 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 held-out perturbation to the test set. We partition the remaining perturbation identities into disjoint training and validation sets. Downstream fine-tuning uses only the training split, so every test perturbation is held out from both parameter updates and validation in that benchmark.

The revised manuscript lists the data sources used in each pretraining stage in Appendix B and specifies the downstream splitting procedure in Appendix D.1.

Q2. Rationale for the original baseline selection

Our original benchmark focused on models that could be evaluated on both forward response prediction and inverse perturbation retrieval, reflecting VCLMU’s design as a unified framework for the two tasks. For inverse retrieval, a forward model predicts the response to each candidate perturbation and ranks candidates by how closely their predictions match the observed response. This focus on evaluating both tasks guided our original baseline selection.

Following the reviewer’s suggestion, we added Heimdall and GREmLN to the inverse retrieval benchmark using the same downstream datasets, perturbation-disjoint splits, and candidate gene sets. We discuss the comparison in our response to W2. Table 3 reports the results averaged across all ten datasets, and Appendix Table 7 reports the results for each dataset.

Your constructive suggestions helped us identify where additional evidence and clearer explanations would strengthen the paper. We appreciate the opportunity to make these improvements and hope the revised manuscript addresses your concerns. We would be grateful if you could take the new results into account and consider a higher score. Thank you for helping us improve our work.

Reviewer_U7icOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a virtual-cell world model that represents cells as Latent Mechanism Units. Genetic perturbations act on these units whose updated states are used to predict expression changes.

The model supports both forward perturbation prediction and inverse perturbation identification. It uses two-stage pretraining on pseudo-bulk and single-cell perturbation data.

Critical strengths

I like the idea of explicitly factorizing cellular states into reusable units. Supporting forward and inverse predictions in one framework is interesting.

Strong empirical performance across ten datasets.

The inverse prediction results are also promising.

Critical weaknesses

I am not convinced that LMUs actually correspond to biological mechanisms. Gene-set enrichment is useful, but does not demonstrate that the learned units capture causal mechanisms or their dynamics. The "world model" feels overstated given that the model predicts endpoint responses.

The baseline comparisions are not entirely convincing. VCLMU has large perturbation-specific pretraining, while baselines use frozen encoders or train from scratch.

For inverse prediction VCLMU has a dedicated classifier while most baselines use forward-consistency retrieval. This makes it difficult to isolate the benefits of the LMU representation.

Questions for authors

How would a standard inverse classifier perform when trained with the same supervision?

Why not compare against direct inverse models such as Heimdall or GREmLN?

Can the authors show that learned LMUs generalize as stable biological mechanisms across cell types, beyond gene-set enrichment?

4 comments

W1. Biological interpretation of LMUs and scope of the world model

We thank the reviewer for raising this important question. Our motivation is that perturbation-induced expression changes reflect the combined effects of multiple biological processes. VCLMU incorporates this idea by representing cellular state with LMUs informed by multimodal gene evidence and learning how their latent states change under perturbations. The enrichment analysis tests whether the LMUs highlighted by the trained model are associated with known biological response programs. We have clarified this interpretation in the revision. To test whether the LMUs associated with a perturbation also contribute to predicting its response, we added an LMU ablation experiment in Figure 3. For seven perturbations shared by Replogle K562 and RPE1, we compared the drop in corr20DE after ablating the target LMUs with the drop after ablating the same number of randomly selected LMUs. When the results were aggregated across the two cell lines, target ablation caused a larger drop for all seven perturbations (Fig. 3a). For UQCRB, UQCRFS1, NDUFB6, and TIMM13, we also observed a larger drop in each cell line separately (Fig. 3b). These results show that the model relies more strongly on the associated LMU states than on randomly selected LMU states to predict these perturbation responses.

We have also clarified what the world-model formulation means in VCLMU. The encoded control LMU states represent the initial state, the perturbed gene and its regulatory context define the action, and the model predicts an action-conditioned update to the LMU states. The decoder then maps the updated states to the predicted endpoint expression response. During Stage 1 pretraining, we supervise the predicted LMU updates by aligning them with differences between the LMU states encoded from the control and perturbed profiles. The term “world model” thus refers to an explicit, action-conditioned latent transition from the control state to the perturbed endpoint, followed by decoding into expression space. The supervision comes from the two endpoints, so the learned transition represents the overall change across the experiment. Resolving intermediate states or the timing of individual mechanisms would require time-resolved observations.

W2. Pretraining and downstream fine-tuning of baseline models

We have clarified which baselines update their pretrained encoders during downstream training. As specified in Section 4.1 and Appendix D.2, the pretrained encoders of Geneformer-FT, scFoundation-FT, Tahoe-x1-FT, and scGPT-FT are fine-tuned on each downstream training split. scGPT-ZS is evaluated without dataset-specific optimization. Our baseline suite also includes methods trained from scratch or without an external pretrained single-cell encoder, including GenePert, scDFM, LPM, CPA, and GEARS. Tahoe-x1 was pretrained in part on Tahoe-100M, a perturbation compendium.

All methods use the same downstream datasets, perturbation-disjoint splits, candidate gene sets, and common evaluation gene space. All downstream evaluation datasets are excluded from both stages of VCLMU pretraining. Because the methods use different pretraining corpora, the observed performance differences reflect the combined effects of pretraining and model design.

Q1. Inverse classifiers trained with the same downstream supervision

We added linear and MLP multimodal candidate scorers trained using the same perturbation labels, candidate gene sets, and downstream splits. Both scorers take the observed expression change and a candidate gene as inputs and use the same biological-prior embeddings as VCLMU to represent each candidate gene.

The linear scorer uses a rank-64 bilinear interaction. It projects the expression change and the candidate-gene embedding into a shared 64-dimensional space through separate bias-free linear layers and scores the candidate by their dot product multiplied by a learned scalar. The MLP scorer uses separate two-layer networks to encode the expression change and the candidate-gene embedding into 128-dimensional vectors. An interaction head takes the concatenation of these vectors and their elementwise product and outputs one score per candidate.

Table 3 reports all five inverse retrieval metrics across ten evaluation datasets.

MethodHit@1Hit@5Hit@20MRRNDCG@20
Linear (multimodal)0.0135 ± 0.02290.0585 ± 0.04920.1840 ± 0.09650.0512 ± 0.03440.3664 ± 0.0533
MLP (multimodal)0.0038 ± 0.00390.0520 ± 0.03960.2072 ± 0.10780.0439 ± 0.01560.3305 ± 0.0237
VCLMU-FT0.1645 ± 0.08270.2249 ± 0.08740.3274 ± 0.09070.2043 ± 0.08190.6544 ± 0.1129

The MLP retrieved the true perturbation in the top 20 more often than the linear scorer, while the linear scorer achieved higher Hit@1 and MRR. Under the same downstream supervision, VCLMU-FT placed the true perturbation first in 16.45% of queries on average across datasets, compared with 1.35% for the linear scorer and 0.38% for the MLP. VCLMU-FT also achieved higher values on the other four metrics.

Q2. Comparison with Heimdall and GREmLN on inverse perturbation retrieval

VCLMU is designed to support forward response prediction and inverse perturbation retrieval within a unified framework. Many perturbation-response models developed for forward prediction can also perform inverse retrieval by predicting the response for each candidate perturbation and ranking candidates according to their agreement with the observed response. Our original comparison therefore evaluated these models under a shared bidirectional protocol.

Following the reviewer’s suggestion, we added comparisons with Heimdall and GREmLN under the same downstream evaluation protocol. Table 3 reports all five inverse retrieval metrics across the ten evaluation datasets.

MethodHit@1Hit@5Hit@20MRRNDCG@20
Heimdall0.0110 ± 0.00880.0492 ± 0.03570.1412 ± 0.08540.0435 ± 0.02560.3936 ± 0.0513
GREmLN0.0301 ± 0.03220.1010 ± 0.08150.2925 ± 0.17910.0838 ± 0.05890.3881 ± 0.0360
VCLMU-FT0.1645 ± 0.08270.2249 ± 0.08740.3274 ± 0.09070.2043 ± 0.08190.6544 ± 0.1129

GREmLN was closest to VCLMU-FT on Hit@20 (29.25% versus 32.74%). The gap was larger at rank one, where VCLMU-FT retrieved the true perturbation in 16.45% of queries on average across datasets, compared with 3.01% for GREmLN and 1.10% for Heimdall.

Q3. Cross-cell-line consistency of LMU contributions to perturbation-response prediction

We added a cross-cell-line ablation study using the seven perturbations shared by the Replogle K562 and RPE1 datasets (FXN, NDUFA8, NDUFB6, TIMM13, UQCRB, UQCRC1, and UQCRFS1). For each perturbation, we compared the decrease in corr20DE\mathrm{corr}_{20\mathrm{DE}} after ablating the corresponding target LMU state with the decrease after ablating a size-matched set of random LMU states. When the ablation results were aggregated across the two cell lines, target-LMU ablation produced a larger decrease for all seven perturbations (Fig. 3a).

To assess consistency across cell lines, we calculated the LMU-specific effect separately in K562 and RPE1 as the target-LMU ablation drop minus the random-LMU ablation drop (Fig. 3b). Four of the seven perturbations (NDUFB6, TIMM13, UQCRB, and UQCRFS1) showed positive effects in both cell lines. For UQCRB and UQCRFS1, the effects were also similar in magnitude across the two cell lines. These results, reported in Section 4.5 and Figure 3, show that the target LMU states contribute to predicting the same perturbation response in both cellular contexts for these four perturbations.

Per-dataset results for the additional inverse retrieval baselines

The table below reports the detailed results for the four additional inverse retrieval baselines on each of the ten evaluation datasets. These results are also available in Appendix Table 7. Each entry is the mean ± standard deviation over three seeds. Hit-conditioned NDCG@20 is averaged over queries for which the true perturbation ranks in the top 20. Higher is better for all five metrics.

DatasetMethodHit@1Hit@5Hit@20MRRNDCG@20
AdamsonLinear (multimodal)0.0794±0.02240.1905±0.03890.4127±0.08090.1434±0.00240.4682±0.0398
AdamsonMLP (multimodal)0.0000±0.00000.1587±0.05940.4762±0.07780.0829±0.01700.3237±0.0290
AdamsonHeimdall0.0159±0.02240.0794±0.02240.2698±0.08980.0703±0.01620.3758±0.0371
AdamsonGREmLN0.0648±0.00820.2217±0.01420.5700±0.02090.1635±0.01030.4213±0.0078
Fr-CtrlLinear (multimodal)0.0000±0.00000.0079±0.01120.1190±0.05140.0259±0.00790.2766±0.0301
Fr-CtrlMLP (multimodal)0.0079±0.01120.0238±0.01940.1508±0.07360.0395±0.01950.3234±0.0623
Fr-CtrlHeimdall0.0000±0.00000.0159±0.02240.1111±0.04490.0251±0.00660.3034±0.0383
Fr-CtrlGREmLN0.0125±0.00120.0566±0.00130.2233±0.00700.0562±0.00100.3559±0.0031
Fr-CocuLinear (multimodal)0.0000±0.00000.0238±0.00000.1587±0.02240.0326±0.00120.3098±0.0061
Fr-CocuMLP (multimodal)0.0000±0.00000.0397±0.01120.1349±0.04890.0340±0.00460.3357±0.0208
Fr-CocuHeimdall0.0238±0.00000.0873±0.02970.1508±0.04890.0581±0.00420.4833±0.0761
Fr-CocuGREmLN0.0175±0.00120.0792±0.00610.2774±0.01260.0716±0.00320.3712±0.0057
Fr-IFNLinear (multimodal)0.0000±0.00000.0317±0.02970.1032±0.02240.0331±0.00880.3580±0.0551
Fr-IFNMLP (multimodal)0.0000±0.00000.0397±0.01120.1429±0.01940.0372±0.00750.3466±0.0358
Fr-IFNHeimdall0.0159±0.01120.0476±0.01940.1508±0.02970.0450±0.01460.3902±0.0731
Fr-IFNGREmLN0.0165±0.00180.0685±0.00490.2523±0.00680.0652±0.00280.3648±0.0051
NormanLinear (multimodal)0.0095±0.01350.0744±0.03560.2694±0.04050.0678±0.01650.3636±0.0382
NormanMLP (multimodal)0.0093±0.01310.0744±0.01440.3147±0.00770.0610±0.01210.3227±0.0343
NormanHeimdall0.0095±0.01350.0834±0.03940.2033±0.06740.0653±0.01970.4009±0.0447
NormanGREmLN0.1114±0.00170.2724±0.00220.5806±0.00570.2081±0.00150.4812±0.0043
K562Linear (multimodal)0.0059±0.00270.0382±0.00630.1028±0.01080.0271±0.00140.3848±0.0146
K562MLP (multimodal)0.0073±0.00450.0374±0.00310.1263±0.01200.0326±0.00410.3714±0.0180
K562Heimdall0.0007±0.00100.0073±0.00210.0191±0.00100.0067±0.00110.3704±0.0409
K562GREmLN0.0016±0.00030.0075±0.00080.0258±0.00210.0090±0.00060.3683±0.0037
RPE1Linear (multimodal)0.0066±0.00130.0425±0.00690.1322±0.01140.0336±0.00340.3741±0.0085
RPE1MLP (multimodal)0.0066±0.00350.0585±0.00580.1841±0.01890.0428±0.00240.3660±0.0108
RPE1Heimdall0.0038±0.00130.0076±0.00480.0246±0.00350.0097±0.00240.4178±0.0482
RPE1GREmLN0.0046±0.00070.0179±0.00300.0539±0.00480.0178±0.00190.3900±0.0111
CRISPRaLinear (multimodal)0.0133±0.01890.0800±0.03270.2667±0.04990.0678±0.01850.3636±0.0450
CRISPRaMLP (multimodal)0.0000±0.00000.0533±0.03770.2533±0.11470.0457±0.00860.2988±0.0200
CRISPRaHeimdall0.0267±0.01890.1067±0.04990.2800±0.11780.0884±0.01630.4501±0.0553
CRISPRaGREmLN0.0366±0.00460.1330±0.00860.4134±0.00890.1101±0.00550.3908±0.0071
CRISPRiLinear (multimodal)0.0000±0.00000.0290±0.01020.1087±0.01770.0278±0.00220.3291±0.0147
CRISPRiMLP (multimodal)0.0072±0.01020.0145±0.01020.1087±0.03550.0291±0.00850.3194±0.0633
CRISPRiHeimdall0.0072±0.01020.0435±0.01770.1159±0.05420.0392±0.01360.4193±0.0424
CRISPRiGREmLN0.0111±0.00180.0530±0.00590.1942±0.01240.0516±0.00350.3608±0.0048
XuLinear (multimodal)0.0200±0.01630.0667±0.00940.1667±0.05250.0533±0.00940.4365±0.0848
XuMLP (multimodal)0.0000±0.00000.0200±0.01630.1800±0.01630.0340±0.00490.2971±0.0347
XuHeimdall0.0067±0.00940.0133±0.01890.0867±0.05250.0276±0.01410.3246±0.0544
XuGREmLN0.0242±0.00250.1000±0.00710.3337±0.01280.0846±0.00420.3764±0.0047

We greatly appreciate the time and effort you have devoted to reviewing our submission. We have carefully considered each of your concerns and addressed them in our responses and the revised manuscript. In light of the new experiments and clarifications, we would be grateful if you would consider raising your score. Thank you again for your careful reading and thoughtful suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.