Speaker
Description
RNA velocity has emerged as a popular approach for modeling cellular change along the phenotypic landscape but routinely omits regulatory interactions between genes. Conversely, methods that infer gene regulatory networks (GRNs) do not consider the dynamically changing nature of biological systems. To integrate these two currently disconnected fields, we present RegVelo, an end-to-end dynamic, interpretable, and actionable deep learning model. RegVelo learns a joint model of splicing kinetics and gene regulatory relationships and allows us to perform in silico perturbation predictions. When applied to datasets of the cell cycle, human hematopoiesis, and murine pancreatic endocrinogenesis, RegVelo provides reliable predictive power for terminal states, gene interaction and perturbation simulations. To leverage RegVelo’s full potential, we studied the dynamics of zebrafish neural crest development and underlying regulatory mechanisms using deep full-transcript-length Smart-seq3 dataset and shared gene expression and chromatin accessibility measurements. Using RegVelo's in silico perturbation predictions, supported by CRISPR/Cas9-mediated knockout and single-cell Perturb-seq, we establish transcription factor tfec as an early driver and elf1 as a novel regulator of pigment cell fate. Together, RegVelo provides a powerful framework for quantitatively bridging gene regulation and cell fate decisions.
