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

Gene regulatory networks (GRNs) dynamically regulate gene activation and repression, driving cellular differentiation. Despite advancements in GRN inference, challenges remain in capturing differentiation dynamics and causal inference. To address these limitations, we developed TRIGON, a Transformer-based model that infers dynamic GRN through learning temporal causality among genes. TRIGON achieved state-of-the-art performance, improving accuracy by 204% over the latest methods across four developmental datasets. When applied to well-established paradigms, including mouse embryonic stem cell and hematopoietic stem cell differentiation, TRIGON identified key transcription factors (TFs) that were not detectable by differential expression analysis and revealed potential TFs associated with primitive endoderm differentiation. Furthermore, TRIGON constructed dynamic GRN across varying time resolutions, enabling the analysis of GRN dynamics from diverse time scales. Through in silico perturbation, TRIGON accurately recapitulated cell fate changes following Gata1 and Spi1 knockout and accurately predicted gene expression changes.