Determinants of patient response to cancer immunotherapy are multi-factorial, depending on factors such as whether the tumor has immune-evading mutations, whether the immune system can effectively recognize the tumor, and the composition of the tumor microenvironment. Current practices in precision oncology devise personalized treatment strategies based on each tumor’s unique genomic mutation or protein biomarker, often one at a time i.e., EGFR mutation or HER2 expression. However, this single-marker paradigm neglects all other sources of information that may inform treatment response. As a result, considerable fractions of patients with the targeted biomarker still fail to respond and improved methods to select patients are still required.

Recent works by us and others have shown that response prediction can be enhanced by developing machine-learning (ML) models to combine biomarkers found within an omic layer such as gene mutations found by DNA-Sequencing (DNA-Seq). Nevertheless, clinical trials now routinely generate -omic data at many molecular layers, i.e., DNA-Seq, RNA-Seq, single-cell data, etc. Multi-view learning is a category of ML techniques to solve prediction tasks when each sample has multiple views of different features.

Herein, we develop a multi-view learning technique particularly suited to multi-omic data and demonstrate its application to predicting immunotherapy response in cancer. We (1) develop multi-view learning techniques particularly suited for multi-omic data, (2) benchmark its performance against conventional approaches, and (3) utilize explainable AI to derive parsimonious models combining selected markers across omic layers to predict treatment response.