CauFinder: Steering Cell-State and Phenotype Transitions by Causal Disentanglement Learning

Understanding and controlling cell-state and phenotype transitions is central to biological discovery and therapeutic development, yet identifying true causal regulators from observational transcriptomic data remains challenging because of confounding and correlated signals. CauFinder is a framework that integrates causal disentanglement modeling with network control to prioritize causal regulators of state and phenotype transitions from observed data. By leveraging causal reasoning based on do-calculus and optimizing information-flow metrics, CauFinder separates putative causal factors from spurious associations, quantifies transition-relevant states, and nominates master regulators. Across simulations and multiple real-world datasets, CauFinder identifies regulators associated with diverse transitions, including differentiation, adenocarcinoma-to-squamous transdifferentiation, and shifts between drug-sensitive and drug-resistant states. In epidermal growth factor receptor (EGFR) tyrosine kinase inhibitor (TKI) resistance, CauFinder prioritizes DAAM1 as a previously unrecognized driver. Small interfering RNA (siRNA)-mediated knockdown of DAAM1 enhances sensitivity to osimertinib, providing functional support for this causal prediction. Overall, CauFinder enables actionable target nomination and testable hypotheses for intervening in disease-relevant state transitions using observational transcriptomic data.