Gene regulatory networks govern how cells respond to signals and environmental conditions, and reconstructing them can reveal the control principles cells use to maintain homeostasis and execute cell-state transitions. We developed D-SPIN (Jiang et al., Cell 2026), a maximum entropy framework that infers mechanistically interpretable and generative models of gene regulatory networks from single-cell mRNA-seq datasets collected across thousands of perturbation conditions. Treating each cell as a set of interacting genes or gene programs, the models explain how perturbations modulate cell-state proportions by reconfiguring underlying regulatory interactions.
Using large Perturb-seq and drug response datasets, D-SPIN models reveal key regulators of cell fate decisions and the coordination of distant cellular pathways in response to gene knockdowns and drug treatments. The models further elucidate how combinations of immunomodulatory drugs induce combinatorial cell states through additive recruitment of gene expression programs, and simulate shifts in immune cell population structures across unobserved drug dosage combinations. In neural progenitor cell differentiation, we discovered that EGF/FGF2, BMP4, and Wnt control neuron and glia proportions following a log-additive rule. Guided by D-SPIN network models, we propose an annealed noise-driven switch as a candidate mechanism for this tunable population diversity. High early noise lets signals tune fate choices, and declining noise later allows commitment (Jiang et al., bioRxiv 2025). Together, these results show how maximum entropy models of gene regulatory networks can reveal principles of cellular information processing and physiological control.
