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

Area of Science:

• Oncology
• Computational Biology
• Molecular Biology

Background:

• Cancer stem cells (CSCs) are a therapeutic challenge due to their resistance to conventional treatments.
• Reactivating cell differentiation is a promising strategy to eliminate CSCs.
• Colorectal cancer (CRC) often involves the loss of key lineage factors like CDX2 in aggressive subtypes.

Purpose of the Study:

• To develop a machine learning framework (CANDiT) for identifying therapeutic targets to induce CSC differentiation in CRC.
• To identify specific molecular vulnerabilities associated with CSCs in CRC.
• To evaluate the therapeutic potential of targeting identified vulnerabilities for CSC elimination.

Main Methods:

• Development of CANDiT, a machine learning framework analyzing transcriptomic data to find differentiation vulnerabilities.
• Identification of PRKAB1, a stress polarity sensor, as a key target, particularly in CDX2-low CRC cells.
• Testing a PRKAB1 agonist in various CRC models (cell lines, xenografts, patient-derived organoids).

Main Results:

• The PRKAB1 agonist successfully reactivated intestinal lineage programs and dismantled stemness pathways (Wnt/YAP).
• Selective elimination of CDX2-low CSCs was observed across all tested CRC models.
• A strong therapeutic index was linked to the CDX2-low CSC state.
• A 50-gene predictive signature indicated a potential ~50% reduction in CRC recurrence and mortality risk.

Conclusions:

• CANDiT provides a precision framework for targeting CSCs by inducing lineage restoration.
• PRKAB1 activation represents a viable therapeutic strategy for eliminating CSCs in CRC.
• The identified gene signature may predict treatment response and patient outcomes in CRC.