CANDiT: A machine learning framework for differentiation therapy in colorectal cancer

Reactivating lineage commitment to differentiate, and hence eliminate, cancer stem cells (CSCs) remains a therapeutic challenge. Here, we present CANDiT (cancer-associated nodes for differentiation targeting), a machine learning framework that identifies transcriptomic vulnerabilities for differentiation therapy in colorectal cancer (CRC). Centering on CDX2-a master intestinal lineage factor lost in high-risk, poorly differentiated CRCs-we identify PRKAB1, a stress polarity sensor, as a top therapeutic target. A clinical-grade PRKAB1 agonist reactivates lineage programs, dismantles Wnt/YAP-driven stemness, and selectively eliminates CDX2-low CSCs across CRC cell lines, xenografts, and patient-derived organoids (PDOs). Multivariate analysis reveals a strong therapeutic index tied to the CDX2-low state. A 50-gene response signature, derived from integrated modeling across all platforms, predicts ∼50% reduction in recurrence and mortality risk. Like immunotherapy, CANDiT resurrects a physiologic program-differentiation-to selectively eliminate CSCs, offering a scalable, precision framework for lineage restoration in solid tumors.