VCLMU: Virtual Cell World Modeling with Latent Mechanism Units

Predicting cellular responses to unseen perturbations is a central capability of virtual cells, yet most existing models directly map an unperturbed molecular profile and perturbation to the resulting observation without explicitly modeling the latent cellular transition that generates it. We introduce a mechanism-centric virtual cell world model that represents cellular state as a set of Latent Mechanism Units (LMUs) and treats genetic perturbations as actions on these latent states. Each LMU combines a reusable identity grounded in multimodal biological evidence with an observation-specific state, enabling perturbations to induce mechanism-specific stochastic transitions before decoding the resulting transcriptional response. We train the model through two-stage pretraining, first on approximately 200K pseudo-bulk perturbation profiles and then on gene-aligned single-cell perturbation data. We evaluate VCLMU on perturbation-disjoint held-out single-gene prediction benchmarks, perform stage-wise analysis of pseudo-bulk and single-cell pretraining, and assess both global and response-gene-focused prediction quality. Across six perturbation-disjoint benchmark datasets, VCLMU consistently outperforms state-of-the-art perturbation prediction methods. We further analyze the learned LMUs through perturbation-LMU enrichment and gene-program characterization to examine whether the latent states capture structured biological response programs. Together, these experiments evaluate whether mechanism-level latent state transition provides a useful and interpretable formulation for virtual cell modeling.