MACHINE LEARNING FOR MULTI-OMICS BIOMARKER DISCOVERY IN LUNG, KIDNEY, AND HEAD AND NECK CANCERS

Abstract

Lung, kidney and head and neck cancer early detection still pose significant clinical challenges due to delayed diagnosis of many patients when treatment is limited and prognosis is poor. In this paper, the authors introduce BiomarkMachine, a machine learning framework for the discovery of multi-omics biomarkers with potential for use in early cancer detection across multiple tumor types. The framework combines genomic, transcriptomic, epigenomic, proteomic and clinical data to uncover biologically relevant biomarker signatures related to early stage malignancy. BiomarkMachine is designed to learn to classify cases with cancer from healthy controls or benign cases, as well as to identify which molecular features are most important to the classification. The proposed strategy is particularly useful in the case of complex cancers like lung cancer, kidney cancer, and head and neck cancers, where a single-biomarker approach doesn't adequately reflect the heterogeneity of the disease. Multi-omics integration could provide a more comprehensive understanding of mechanisms of tumor initiation, molecular dysregulation, and alterations of pathways involved in cancer. Furthermore, the framework can be used to assess biomarker predictive performance, biological relevance and interpretability, which facilitate biomarker prioritization. BiomarkMachine offers an approach that not only is data-driven but is also explainable for finding candidate biomarkers that could enhance early diagnosis and help with personalized screening, as well as future cancer translational research.