Bayesian Non-Negative Matrix Factorization with Correlated Mutation Type Probabilities for Mutational Signatures

Somatic mutations, or alterations in DNA of a somatic cell, are key markers of cancer. In recent years, mutational signature analysis has become a prominent field of study within cancer research, commonly with Nonnegative Matrix Factorization (NMF) and Bayesian NMF. However, current methods assume independence across mutation types in the signatures matrix. This paper expands upon current Bayesian NMF methodologies by proposing novel methods that account for the dependencies between the mutation types. First, we implement the Bayesian NMF specification with a Multivariate Truncated Normal prior on the signatures matrix in order to model the covariance structure using external information, in our case estimated from the COSMIC signatures database. This model converges in fewer iterations, using MCMC, when compared to a model with independent Truncated Normal priors on elements of the signatures matrix and results in improvements in accuracy, especially on small sample sizes. In addition, we develop a hierarchical model that allows the covariance structure of the signatures matrix to be discovered rather than specified upfront, giving the algorithm more flexibility. This flexibility for the algorithm to learn the dependence structure of the signatures allows a better understanding of biological interactions and how these change across different types of cancer. The code for this project is contributed to an open-source R software package. Our work lays the groundwork for future research to incorporate dependency structure across mutation types in the signatures matrix and is also applicable to any use of NMF beyond just single-base substitution (SBS) mutational signatures.