Presentation Information

[SS21-04]Towards AI Virtual Cell Through Dynamical Generative Modeling of Single-cell Omics Data

*Peijie Zhou1 (1. Peking University (China))

Keywords:

AI Virtual Cells,Single-cell transcriptomics,Schrodinger Bridge

Reconstructing continuous cellular dynamics from sparse, high-dimensional single-cell omics data remains a fundamental challenge in systems biology. While classical dynamical models offer interpretability and predictive power for perturbation studies, their utility is constrained by the curse of dimensionality and data sparsity. Recently, a paradigm shift has been witnessed by leveraging artificial intelligence—specifically, dynamical generative modeling—to pioneer the development of an AI virtual cell, a predictive digital twin capable of simulating cellular behavior across time and space. Here we introduce our recent attempts to integrate flow-based generative models with partial differential equations (PDEs) to infer latent dynamics from scRNA-seq data. This approach enables the resolution of continuous cell-state transitions, proliferation, and apoptosis through a dimensionless neural solver, effectively bridging discrete snapshots into a unified trajectory. For spatial transcriptomics data, we extend this method with stVCR, a generative model that aligns transcriptomic snapshots across biological replicates and temporal stages. By reconstructing universal spatial coordinates, stVCR generates a coherent "developmental video" from static spatial transcriptomic images, simulating tissue morphogenesis, cell migration, and regeneration dynamics in axolotl brain and Drosophila embryogenesis. Together, these works suggest how generative AI could have the potential to unify dynamical modeling, spatial reconstruction, and stochastic inference—transforming fragmented omics data into a predictive virtual cell.