Abstract

Mammalian genomes harbor many more enhancers than genes, which greatly complicates the elucidation of cell-state-specific regulatory networks. Here, we developed a computational framework for learning enhancer-based gene networks from joint data on enhancer activity and transcript abundance. Dissecting the developmental plasticity of T helper (Th) cells with this approach, we uncovered a highly connected enhancer-gene network that supports graded Th-cell differentiation states, rather than mutual exclusivity of type-1 and type-2 immunity. Machine learning identifies a small number of regulatory enhancer types as network hubs. Hub enhancers in Th1 cells integrate as inputs the expression level of the master-regulator transcription factor, T-bet, and STAT signals governed by the cytokine environment. The quantitative balance between cell-intrinsic T-bet, driving phenotypic stability, and environmental cues enabling plasticity explains the heterogeneous reprogramming capacities of individual Th1 cells differentiating during natural infections in vivo. Moreover, we provide a framework for elucidating genome-scale regulatory networks based on enhancer activity.