BAMSE: Bayesian model selection for tumor phylogeny inference among multiple tumor samples

Intra-tumor heterogeneity is believed to be a major source of confounding analysis and treatment resistance. In this research we introduce BAMSE, a Bayesian model based tool for intra-tumor heterogeneity analysis of bulk tumor sequencing results across multiple samples. BAMSE takes as input a list of somatic mutations and their corresponding reference and variant read counts, clusters these mutations into sub-clones and outputs a list of high probability evolutionary trees, each representing a scenario for clonal evolution of the tumor. We use a Hierarchical Uniform Prior for clustering of mutations into subclones and a uniform prior over tree topologies describing the evolutionary relations between them. This way, all configurations that have equal number of subclones are assigned equal prior, leading to an unbiased model selection. We show that for this model, to calculate the posterior for a model with K subclones, we need to calculate an integral over a K-1 simplex. These integrals are calculated numerically using a series of convolutions, allowing fast and accurate calculation of the posterior probability. Finally, for the selected high-probable models, we use convex optimization to determine the maximum likelihood cell fraction for each subclone. Both synthetic and experimental data are used to benchmark BAMSE against existing tools for analysis of intra-tumor heterogeneity of bulk samples. Unbiased model selection, accurate calculation of subclonal cell fractions and short runtimes are the main advantages of BAMSE. We will extend BAMSE to account for copy number variations in a future work. BAMSE is available at https://github.com/HoseinT/BAMSE.