A Combined Multi-Omics and Supervised Learning Strategy Uncovers an Epithelial Signature of Radiotherapy Response in Colorectal Cancer

Colorectal cancer (CRC) is the third most diagnosed cancer globally, accounting for 9.6% of all cancer cases, and the second leading cause of cancer deaths. Radiotherapy is a common treatment, but can demonstrate dangerous side effects and varying patient outcomes that reflects on CRC cancer heterogeneity at genetic, epigenetic, transcriptomics and proteomic levels. Identifying biomarkers capable of effectively predicting CRC patient responses to radiotherapy remains paramount, and an unmet need. We channeled both unsupervised and supervised approaches to assess radiotherapy response for 233 patients of the S:CORT Consortium. Splitting the cohort, we first integrated matched RNA, CNA, mutation, and methylation profiles of 117 patients using multi-omics factor analysis (MOFA). We identified a new radiotherapy signature of 101 biomarkers associated with patients who demonstrate a complete response to radiotherapy, and validated the signature using a random forest classifier on the internal validation dataset, and an independent testing cohort. Our signature effectively predicted treatment outcomes, achieving 89% accuracy with strong discriminatory performance (ROC_AUC = 0.85; PR_AUC = 0.71) to differentiate patients with complete response to radiotherapy compared to incomplete responders. Assessing human and murine scRNAseq datasets underscores that the signature is predominantly expressed in CRC epithelial cells, which underpin CRC heterogeneity and cellular diversity. Our identified signature enables pre-treatment identification of CRC patients that are unlikely to achieve a complete response to radiotherapy, thereby sparing these patients from unnecessary radiation exposure and off-target damage effects.