Abstract

Tumours contain genetically related subpopulations that can differ in their metastatic potential and response to treatment. This heterogeneity also provides a basis to reconstruct the evolutionary history of individual tumours thereby providing insights into how cancers develop. I will discuss our work over the past few years on designing algorithms to reconstruct tumour phylogenies by deconvolving data from cancer genome sequencing efforts.

Often the first step in characterizing tumour subpopulations is clustering somatic mutations (both point and structural) according to their inferred population frequencies. This is only a partial characterization: these clusters correspond to cell lineage and these lineages can include multiple subclonal populations. Determining whether or not different groups of mutations are present in the same subpopulations requires reconstructing the tumour phylogeny. We have developed a series of algorithms (PhyloSub, PhyloWGS, PhyloSpan) that use Bayesian inference in non-parametric models to do this. These methods explicitly represent uncertainty thereby distinguishing ambiguous and unambiguous portions of the reconstruction. We are currently applying these methods to more than 2,000 tumors as part of the ICGC pan-cancer whole genome analysis effort.