A pipeline of machine learning-driven multi-modal data fusion methods for prognostic risk analysis in bevacizumab-treated metastatic colorectal cancer.

We introduce a comprehensive multi-step machine-learning driven pipeline which fuses multi-modal omics datasets and clinical outcomes with survival and treatment response to predict patient outcome following anti-angiogenic therapy in the metastatic colorectal cancer (mCRC) setting. The approach encompasses the following steps: a) employing a sparse Bayesian factor analysis method (PhenMap) to select significant mapping-variables (MVs) and associated biomarkers (features) through joint multivariable modelling of copy number aberrations (CNA) and clinical covariates, including mutations and clinical demographics/outcomes; b) utilizing Cox-proportional hazard analysis on the MVs to select those associated with progression-free survival (PFS); and c) employing elastic net Cox regression analysis on the prognostic features selected by PhenMap to delineate risk groups associated with bevacizumab (BVZ) response. Through this approach, we have identified three putative features (CNA—15q21.1 and 1p36.31 deletions and BRAF mutation) from two prognostically-significant MVs to stratify N = 117 mCRC patients, who received BVZ combination therapy, into 3 (low, medium, high) prognostic risk groups that were significantly associated with PFS and treatment response. Mortality risk was significantly greater in the high-risk group with 100% (n = 12) of patients showing no response to BVZ, compared to the low-risk group where 10 out of 12 patients (88%) showed a response to BVZ. The risk groups were independently negative predictors of survival in BVZ-treated mCRC patients. Overall, Wwe have established a machine learning pipeline integrating multi-modal omics data with response, and have implicated a putative combined CNA/mutation candidate biomarker with associated risk scores that could, in the future, help stratify mCRC patients unlikely to benefit from BVZ combination therapy. Our novel precision medicine approach applies disruptive advancements in artificial intelligence and bioinformatics methodologies to tumour biology datasets.