Genome-driven cancer site characterization: An overview of the hidden genome model

A large and growing body of research has documented strong links between somatic mutations and different cancer types. This has put forward an emerging field aiming to characterize cancer types through tumor mutation fingerprints. However, the vast majority of somatic mutations in a typical cancer are either extremely rare or have never been recorded in existing databases. These make up a hidden genome buried underneath the surface of a small number of commonly occurring mutations that have been extensively studied. This chapter reviews the hidden genome model, a multilevel statistical model that uses a projection-based approach to contextualize and extract information from this ultra-sparse and ultra-high-dimensional space of hidden-genome mutations. By leveraging meta-features quantifying DNA sequence and epigenetic context as a hierarchical layer, the model can effectuate impressive signal condensation in these rare variants. A formal discussion of the model, its assumption, implementation strategy, and various modeling consequences are provided. Applications of the model to three large-scale genome sequencing databases are noted and various interesting insights are discussed.