Dissecting dynamic gene regulatory network using transformer-based temporal causality analysis

The AI Virtual Cell (AIVC) framework promises to revolutionize biological research through high-fidelity simulations of cellular behaviors and responses to perturbations. Central to realizing this vision is the ability to model cell differentiation dynamics, which requires accurate inference of gene regulatory network (GRN) that govern cell fate decisions. However, existing computational approaches rely on static GRN models that fail to capture the dynamic changes of regulatory relationships during differentiation, limiting their utility for simulating developmental processes and predicting perturbation outcomes. Here, we present CellProphet, an interpretable AI model that infers dynamic GRN by integrating temporal causality with transformer self-attention mechanism. CellProphet captures time-lagged dependencies between transcription factor (TF) expression and target gene activation while providing interpretable regulatory weights, enabling both accurate prediction and mechanistic insight. When benchmarked against nine state-of-the-art methods across seven differentiation datasets, CellProphet achieves superior performance in all evaluation metrics. Applied to mouse embryonic stem cell differentiation, CellProphet identifies both well-known and potentially novel TFs with substantially high sensitivity and successfully reconstructs dynamic regulatory relations validated through multi-modal epigenomic data. In mouse hematopoietic differentiation, CellProphet accurately predicts cell fate transitions and gene expression changes following in silico perturbation of key TFs Gata1 and Spi1, demonstrating its capability for virtual experimentation. These results establish CellProphet as a foundational tool for the AIVC framework, enabling researchers to decode the dynamic regulatory logic of differentiation, accelerate discovery of key regulatory factors, and design targeted cellular interventions for widespread applications.