Topical Hidden Genome: Discovering Latent Cancer Mutational Topics using a Bayesian Multilevel Context-learning Approach

Statistical inference on the cancer-site specificities of collective ultra-rare whole genome somatic mutations is an open problem. Traditional statistical methods cannot handle whole-genome mutation data due to their ultra-high-dimensionality and extreme data sparsity -- e.g., >30 million unique variants are observed in the ~1700 whole-genome tumor dataset considered herein, of which >99% variants are encountered only once. To harness information in these rare variants we have recently proposed the "hidden genome model", a formal multilevel multi-logistic model that mines information in ultra-rare somatic variants to characterize tumor types. The model condenses signals in rare variants through a hierarchical layer leveraging contexts of individual mutations. The model is currently implemented using consistent, scalable point estimation techniques that can handle 10s of millions of variants detected across thousands of tumors. Our recent publications have evidenced its impressive accuracy and attributability at scale. However, principled statistical inference from the model is infeasible due to the volume, correlation, and non-interpretability of the mutation contexts. In this paper we propose a novel framework that leverages topic models from the field of computational linguistics to induce an *interpretable dimension reduction* of the mutation contexts used in the model. The proposed model is implemented using an efficient MCMC algorithm that permits rigorous full Bayesian inference at a scale that is orders of magnitude beyond the capability of out-of-the-box high-dimensional multi-class regression methods and software. We employ our model on the Pan Cancer Analysis of Whole Genomes (PCAWG) dataset, and our results reveal interesting novel insights.