MEDICC2: whole-genome doubling aware copy-number phylogenies for cancer evolution

Abstract

Chromosomal instability (CIN) and somatic copy number alterations (SCNA) play a key role in the evolutionary process that shapes cancer genomes. SCNAs comprise many classes of clinically relevant events, such as localised amplifications, gains, losses, loss-of-heterozygosity (LOH) events, and recently discovered parallel evolutionary events revealed by multi-sample phasing. These events frequently appear jointly with whole genome doubling (WGD), a transformative event in tumour evolution, which generates tetraploid or near-tetraploid cells. WGD events are often clonal, occuring before the emergence of the most recent common ancestor, and have been associated with increased CIN, poor patient outcome and are currently being investigated as potential therapeutic targets. While SCNAs can provide a rich source of phylogenetic information, so far no method exists for phylogenetic inference from SCNAs that includes WGD events. Here we present MEDICC2, a new phylogenetic algorithm for allele-specific SCNA data based on a minimum-evolution criterion that explicitly models clonal and subclonal WGD events and that takes parallel evolutionary events into account. MEDICC2 can identify WGD events and quantify SCNA burden in single-sample studies and infer phylogenetic trees and ancestral genomes in multi-sample scenarios. In this scenario, it accurately locates clonal and subclonal WGD events as well as parallel evolutionary events in the evolutionary history of the tumour, timing SCNAs relative to each other. We use MEDICC2 to detect WGD events in 2778 tumours with 98.8% accuracy and show its ability to correctly place subclonal WGD events in simulated and real-world multi-sample tumours, while accurately inferring its phylogeny and parallel SCNA events. MEDICC2 is implemented in Python 3 and freely available under GPLv3 at https://bitbucket.org/schwarzlab/medicc2.

CNETML: Maximum likelihood inference of phylogeny from copy number profiles of spatio-temporal samples

“…Given different input data and aims, trees reconstructed from CNAs called from a single patient are of four major types: 1) mutation tree when the order and evolutionary history of mutational events are of interest [15], as in SCICoNE [5] and CONET [7], where each tip represents copy number events and cells are attached to each node; 2) clone tree when clonal deconvolution is feasible, as in CNT-MD [4] and DEVOLUTION [8], where each tip represents a clone; 3) cell tree when CNAs can be called for each cell, as in FISHtree [16, 17], sitka [6], and NestedBD [10], where each tip represents a cell; 4) sample tree when each sample is assumed to be homogeneous, as in MEDICC [18], MEDICC2 [9] and PISCA [3], where each tip represents a sample.…”

“…However, only a few reliable methods exist to detect CNAs from sWGS data, especially absolute copy numbers [27]. Most of the previous sample phylogeny inference methods are designed for absolute allele-specific integer copy numbers which are often called from SNP arrays and high-coverage NGS data, such as MEDICC [18], MEDICC2 [9], and PISCA [3]. To better understand cancer progression from these sWGS data, it is important to have methods that can build sample trees based solely on (relative) total copy numbers, which will be addressed in this paper.…”

“…Other methods represent the genome as a vector of copy number values, often called copy number profile (CNP) [4]. Based on CNPs, some methods build trees without a model, such as the maximum parsimony (MP) method with the Fitch algorithm [23, 32] and distance matrix methods based on Euclidean [33] or Manhattan [34] distances, and hence they may underestimate the true evolutionary distance as no correction of hidden changes is applied [9, 31]. Other methods use copy number transformation (CNT) models that allow the computation of minimum evolutionary distance between CNPs, which is the shortest sequence of events that transform one CNP to the other.…”

“…This model deals with horizontal dependencies caused by overlapping CNAs and hence is less likely affected by convergent evolution. It has been extended to allow weights on CNAs of different position, size, and type (duplication/deletion) [35] and WGD [9, 36]. The weighted versions of both models allow the estimation of CNA rates in term of event probabilities [17, 35], but mutation rates by calendar time cannot be estimated.…”

“…In addition, most phylogeny inference methods based on CNPs cannot handle multiple scales of chromosomal changes due to the inherent complexity. Notable exceptions are FISHtree, which was designed for FISH data and is not scalable for longer CNPs [16, 17], and MEDICC2 which only considered segment duplication/deletion and WGD but excluded chromosome or arm level gain/loss [9].…”