The goals of this proposal are to develop novel statistical methods, more accurate inference procedures, and interactive software tools to perform mutational signature deconvolution in cancer samples. Mutational signatures are patterns of co-occurring mutations that can reveal insights into a cancer's etiology and evolution. Currently, non-negative matrix factorization (NMF) is the “gold-standard” for mutational signature deconvolution. However, NMF has several deficiencies in that it cannot do the following things: 1) predict signatures in new samples, 2) perform joint learning of known and novel signatures at the same time, 3) alleviate problems from signature “bleeding”, 4) cluster tumors into subgroups based on mutational signature profiles, and 5) characterize uncertainty in model fit. In this proposal, we will develop a novel Bayesian hierarchical models that overcome the limitations of NMF. Furthermore, there is a lack of interactive software for mutational signature inference and visualization for non-computational users. We will also develop an R/Shiny interface on top of our R package to facilitate data preprocessing, inference, and visualization of large-scale datasets. This interface will have a cloud backend to facilitate computationally intensive operations. Overall, this software will streamline mutational signature analysis for noncomputational researchers and will have the capability to interface with other projects from the Informatics Technology for Cancer Research (ITCR) program. Finally, we will analyze a novel targeted sequencing dataset from Chinese patients and perform a meta-analysis of all publicly available variants to generate a novel reference set of mutational signatures for investigators to use in their own studies. Overall, our tools will be of great interest to the cancer community as it will provide greater insights into mutational signature patterns and will be useful in clinical settings to reveal insights into cancer etiology.

Characterizing mutational signatures can reveal insights into tumor etiology and evolution. Current computational methods that perform mutational signature inference such as non-negative matrix factorization (NMF) have several limitations including a lack of ability to predict signatures in individual samples, perform joint learning of known and novel signatures, cluster tumors into subgroups, and perform model selection. We will develop novel statistical methods, more accurate variational inference procedures, and an interactive R/Shiny interface with a cloud backend to allow both computational and non-computational users to better characterize mutational patterns in their tumor cohorts and predict signatures in single clinical samples.