Regulatory network hubs guide dynamic human lineage specification

Abstract

Transcription factors (TFs), including DNA binding proteins and epigenetic co-regulators, control the timing of gene programs during lineage commitment1. However, systematically defining the dynamic activities of the human genomes nearly 2000 TFs during development remains challenging2,3. Here, we use Perturb-Seq to map the transcriptional impact of nearly all human TFs and a subset of enhancers during cardiomyocyte differentiation. Our results suggest that developmental regulatory networks are distributed across highly connected hub TFs4-6 that are dynamic across lineage and time, rather than organizing into top-down hierarchies directed by "master regulators". We also identify distinct TF ensembles that regulate sequential decision points during cardiac cell-fate specification by coordinating lineage-specific activation with alternate lineage repression. Separately, analysis of TF-TF cooperativity uncovers a dynamic interaction between MEF2 family TFs7 and members of the Polycomb Repressive complex 18-10 to execute alternative fate repression. Building on these regulatory interactions, we construct a deep learning transformer model to accurately predict perturbed TFs driving altered regulatory networks in patient-derived transcriptomes. Together, our results define the regulatory network architecture of human lineage specification and provide a platform for predicting TF function and interpreting disease mechanisms.