Inferring tumor evolution from longitudinal samples

Abstract

Background: Determining the clonal composition and somatic evolution of a tumor greatly aids in accurate prognosis and effective treatment for cancer. In order to understand how a tumor evolves over time and/or in response to treatment, multiple recent studies have performed longitudinal DNA sequencing of tumor samples from the same patient at several different time points. However, none of the existing algorithms that infer clonal composition and phylogeny using several bulk tumor samples from the same patient integrate the information that these samples were obtained from longitudinal observations. Results: We introduce a model for a longitudinally-observed phylogeny and derive constraints that longitudinal samples impose on the reconstruction of a phylogeny from bulk samples. These constraints form the basis for a new algorithm, Cancer Analysis of Longitudinal Data through Evolutionary Reconstruction (CALDER), which infers phylogenetic trees from longitudinal bulk DNA sequencing data. We show on simulated data that constraints from longitudinal sampling can substantially reduce ambiguity when deriving a phylogeny from multiple bulk tumor samples, each a mixture of tumor clones. On real data, where there is often considerable uncertainty in the clonal composition of a sample, longitudinal constraints yield more parsimonious phylogenies with fewer tumor clones per sample. We demonstrate that CALDER reconstructs more plausible phylogenies than existing methods on two longitudinal DNA sequencing datasets from chronic lymphocytic leukemia patients. These findings show the advantages of directly incorporating temporal information from longitudinal sampling into tumor evolution studies. Availability: CALDER is available at https://github.com/raphael-group.