Machine-Learning Identifies a Strategy for Differentiation Therapy in Solid Tumors

Despite advances in artificial intelligence (AI) within cancer research, its application toward realizing differentiation therapy in solid tumors remains limited. Using colorectal cancer (CRC) as a model, we developed a machine learning (ML) framework, CANDiT (Cancer Associated Nodes for Differentiation Targeting), to selectively induce differentiation and death of cancer stem cells (CSCs)--a key obstacle to durable response. Centering on one node, CDX2, a master differentiation factor lost in high-risk, poorly differentiated CRCs, we built a transcriptomic network to identify therapeutic strategies for CDX2 restoration. Network-based prioritization identified PRKAB1, a stress polarity sensor, as a top target. A clinical-grade PRKAB1 agonist reprogrammed transcriptional networks, induced crypt differentiation, and selectively eliminated CDX2-low CSCs in CRC cell lines, xenografts and patient-derived organoids (PDOs). Multivariate analyses in PDOs revealed a strong therapeutic index, linking efficacy (IC) to the biomarker-defined CDX2-low state. A 50-gene response signature--derived from an integrated analyses of all three models and trained across multiple datasets--revealed that CDX2 restoration therapy may translate into a [~]50% reduction in recurrence and mortality risk. Mechanistically, treatment activated a differentiation-associated stress polarity signaling axis while dismantling Wnt and YAP-driven stemness programs essential to CSC survival. Thus, CANDiT offers a scalable path to CSC-directed therapy in solid tumors by translating transcriptomic vulnerabilities into precision treatments. Graphic Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=174 SRC="FIGDIR/small/557628v4_ufig1.gif" ALT="Figure 1"> View larger version (52K): org.highwire.dtl.DTLVardef@16eeb1eorg.highwire.dtl.DTLVardef@10e308borg.highwire.dtl.DTLVardef@9512f0org.highwire.dtl.DTLVardef@10e74eb_HPS_FORMAT_FIGEXP M_FIG C_FIG One sentence summaryIn this work, Sinha et al. introduce a machine learning-guided framework to identify and target transcriptomic vulnerabilities in colorectal cancer, demonstrating that differentiation therapy selectively eliminates cancer stem cells and reduces recurrence risk. HighlightsO_LIAn ML framework (CANDiT) identifies target for differentiation therapy for CRCs C_LIO_LITherapy induces crypt differentiation and CSC-specific cytotoxicity C_LIO_LICDX2-low state predicts therapeutic response; restoration improves prognosis C_LIO_LITherapy dismantles stemness via reactivation of stress polarity signaling C_LI