Statistical Methods for Improving Low Frequency Variant Calling in Cancer Genomics

Abstract

Cancer is not a single disease, but a family of genomic diseases characterized by a set of initiating genomic variants accumulated in a single cell that allows that cell to begin dividing uncontrollably. Tumors grow by cell division, and each cell division generates a new set of variants that are passed along to its offspring. As a result, at the time of diagnosis a typical tumor of approximately 100,000,000 cells contains hundreds of millions of genomic variants, whose frequency in the population is a function of the time that they arose. Mutation accumulation through both inheritance and de novo variant production results in a final tumor in which the vast majority of variants are present at low frequency. Current methods used to identify variants have difficulty identifying low frequency variants. Here I will describe two algorithms aimed at improving low frequency variant calling in two settings. Patient-Derived Xenografts (PDXs) serve as avatars for individual patient disease as well as invaluable models for studying basic cancer biology. Molecular character- ization of PDXs is common, but the extensive homology between human and mouse genes present special challenges in sequencing tumors grown in mice. In Chapter 2 I describe an algorithm and R implementation called MAPEX that allows labs study- ing PDXs to use commercial sequencing technologies and locally filter false positive variants caused by sequence homology. Detecting somatic mutations within tumors is key to understanding treatment re- sistance, patient prognosis, and tumor evolution. In Chapter 3 I present BATCAVE (Bayesian Analysis Tools for Context-Aware Variant Evaluation), which extends cur- rent state-of-the-art statistical models for tumor variant calling. I also present an R implementation of the algorithm, and show using simulations that the BATCAVE algorithm improves variant detection.