Towards building a World Model to simulate perturbation-induced cellular dynamics by AlphaCell

Abstract Predicting cellular responses to perturbations is crucial for therapeutic discovery, yet experimental screening is severely constrained by the combinatorial vastness of biological space. While computational simulations offer a scalable alternative, current models are limited by the incomplete latent representation——mainly relying on highly variable genes in feature representation; the poor genome-wide reconstruction fidelity; and the ungeneralizable dynamic laws across diverse contexts. Consequently, they fail to mechanistically transfer learned dynamics to unseen cellular contexts. To address these systemic flaws, we introduce AlphaCell, a generative Virtual Cell World Model that unifies genome-wise representation with continuous state transition modeling. AlphaCell achieves three synergistic innovations: (1) Latent Manifold Rectification, processing the full protein-coding transcriptome to construct a differentiable Virtual Cell Space, effectively filtering noise while preserving intrinsic cellular topology; (2) Biological Reality Reconstruction, utilizing a massive, knowledge-rich decoder to translate abstract latent states back into high-fidelity, genome-wide expression profiles; and (3) Universal State Transition, applying Optimal Transport Conditional Flow Matching to model perturbations as continuous, deterministic vector fields. By abstracting perturbation mechanisms into generalized dynamic laws, AlphaCell makes robust prediction of perturbation responses in a compositional generalization scenario and enables zero-shot prediction of cellular dynamics in entirely unseen cellular contexts, providing a foundational engine for cellular-context-generalizable perturbation prediction and perturbation-induced cellular dynamics simulation.