Automatic multi-target clinical classification and biomarker discovery in cancer

Abstract

Precision medicine relies on accurate and interpretable biomarker and subtype discovery. Many multi-omics subtyping algorithms have been developed to manage subtype identification across platforms but have yet to be evaluated with respect to identification of clinically prognostic subtypes. Further, many comprehensive characterization studies of cancer, which have identified multi-omics subtypes or molecular subtype signatures, have done so through the use of manually-derived expert-designed trees. Despite interpretability, current decision tree approaches are unable to explainably reproduce subtyping findings, owing to the complex nature of molecular and clinical factors driving the disease. Current machine learning (ML) approaches do not achieve interpretability (explainability) across disease endpoints, and models constructed manually by trained experts can be subjective. We develop a multi-objective decision tree (MuTATE) framework which performs automated, explainable, and multi-outcome segmentation to construct interpretable trees, simultaneously identifying biomarkers and subtypes of clinical relevance across disease endpoints. Molecular, clinical, and survey data may be input to identify prognostic biomarkers with either preventive or therapeutic implications. We provide a proof-of-concept for multi-objective, quantitative, explainable trees, enabling interpretable, automated molecular insights for precision medicine. This comprehensive approach can improve therapeutic decisions and has applications across complex diseases, and the availability of our method as an R package enables improved access to comprehensive and quantitaive disease modeling to identify those who may benefit from different treatment plans.