Single-cell tumor phylogeny inference with copy-number constrained mutation losses

Abstract

Motivation: Single-cell DNA sequencing enables the measurement of somatic mutations in individual tumor cells, and provides data to reconstruct the evolutionary history of the tumor. Nearly all existing methods to construct phylogenetic trees from single-cell sequencing data use single-nucleotide variants (SNVs) as markers. However, most solid tumors contain copy-number aberrations (CNAs) which can overlap loci containing SNVs. Particularly problematic are CNAs that delete an SNV, thus returning the SNV locus to the unmutated state. Such mutation losses are allowed in some models of SNV evolution, but these models are generally too permissive, allowing mutation losses without evidence of a CNA overlapping the locus. Results:We introduce a novel loss-supported evolutionary model, a generalization of the infinite sites and Dollo models, that constrains mutation losses to loci with evidence of a decrease in copy number.We design a new algorithm, Single-Cell Algorithm for Reconstructing the Loss-supported Evolution of Tumors (SCARLET), that infers phylogenies from single-cell tumor sequencing data using the losssupported model and a probabilistic model of sequencing errors and allele dropout. On simulated data, we show that SCARLET outperforms current single-cell phylogeny methods, recovering more accurate trees and correcting errors in SNV data. On single-cell sequencing data from a metastatic colorectal cancer patient, SCARLET constructs a phylogeny that is both more consistent with the observed copynumber data and also reveals a simpler monooclonal seeding of the metastasis, contrasting with published reports of polyclonal seeding in this patient. SCARLET substantially improves single-cell phylogeny inference in tumors with CNAs, yielding new insights into the analysis of tumor evolution.Availability: Software is available at

CellPhy: accurate and fast probabilistic inference of single-cell phylogenies from scDNA-seq data

“…For example, in our re-analysis of the dataset from Leung et al . 31 CellPhy was able to recover a monoclonal metastasis -in line with a recent analysis by Satas et al 33 that accounts for copy number variants-, while the competing methods implied a polyclonal seeding. Importantly, the CellPhy bootstrap analyses illustrate the importance of explicitly considering phylogenetic uncertainty, as in the absence of it one cannot really assess if different parts of a tree are equally supported by the data.…”

“…In the original study, after performing custom targeted sequencing of 186 single-cells sampled from primary and metastatic lesions, the authors derived a cell tree using SCITE and inferred a polyclonal seeding of liver metastases (i.e., distinct populations of tumor cells migrated from the primary tumor towards the liver). However, their findings have been recently re-evaluated in two different studies 32,33 . In particular, the former applied the newly developed SiCloneFit, which relaxes the infinite-sites assumption, to the original SNV dataset and also proposed a polyclonal seeding of the metastases, while the latter performed a joint analysis between SNVs and copy-number variants (CNVs) with SCARLET to suggest that the liver metastasis was instead seeded by a single clone.…”

Methods for copy number aberration detection from single-cell DNA-sequencing data

“…Departing from this assumption, methods like Sifit [30], PhiSCS [32], and SiCloneFit [33] use finite-sites models, where gain and loss of mutations are allowed to occur at any site any number of times. While all these methods use SNV data alone, the recently devised method SCARLET [79] also takes into account copy number loss as a cause of a variant loss.…”

Identifying tumor clones in sparse single-cell mutation data

“…We note that this problem is very different from existing methods for analyzing mutations in single-cell sequencing data. Specifically, existing methods ( Borgsmueller et al , 2020 ; Ciccolella et al , 2018 ; El-Kebir, 2018 ; Jahn et al , 2016 ; Malikic et al , 2019 ; McPherson et al , 2016 ; Ross and Markowetz, 2016 ; Roth et al , 2016 ; Satas et al , 2019 ; Singer et al , 2018 ; Zafar et al , 2019 ) attempt to directly model the error rates of observing individual mutations, or rely on distances between cells according to mutations or distances between mutations according to cells. In our case, because we have ultra-low-coverage data with very few mutations recorded as present in individual cells and no confidence in the absence of a mutation in an individual cell, such approaches are unlikely to work well.…”

“…To address the limitations in identifying mutations in single-cell sequencing data, a number of computational methods have been developed to improve mutation calling by grouping cells with similar mutational profiles ( Borgsmueller et al , 2020 ; Roth et al , 2016 ) or shared cellular lineage ( Ciccolella et al , 2018 ; El-Kebir, 2018 ; Jahn et al , 2016 ; Malikic et al , 2019 ; McPherson et al , 2016 ; Ross and Markowetz, 2016 ; Satas et al , 2019 ; Singer et al , 2018 ; Zafar et al , 2019 ). These methods have primarily been applied to analyze single-cell DNA sequencing data obtained from limited genomic regions: e.g.…”