Mutational signature analysis with a structural topic model

Abstract

Cancer occurs on the accumulation of various mutational processes and each mutational process has its unique mutation patterns on the genome, which is called the mutational signature. Extracting the mutational signatures through analyzing the mutation count data in the cancer genome has received much attention as we can identify the origin of cancer with the extracted signatures. While many researchers use nonnegative matrix factorization and latent Dirichlet allocation as a basic method for mutational signature analysis, the approaches only analyze the co-occurrence pattern of mutations and do not fully utilize sample-level information and contextual information of mutations. To overcome these limitations, we propose the variant of latent Dirichlet allocation for extracting mutational signatures with the help of the information. We show that the model can capture two mutational signatures in Alzheimer’s disease data that are resembled the signatures in initially reported. In addition, through the analysis on Alzheimer’s dataset, we demonstrate that using sample-level information improves qualitative interpretability of sample-wise signature proportions.