ConDoR: tumor phylogeny inference with a copy-number constrained mutation loss model

Abstract

A tumor contains a diverse collection of somatic mutations that reflect its past evolutionary history and that range in scale from single nucleotide variants (SNVs) to large-scale copy-number aberrations (CNAs). However, no current single-cell DNA sequencing (scDNA-seq) technology produces accurate measurements of both SNVs and CNAs, complicating the inference of tumor phylogenies. We introduce a new evolutionary model, the constrained k-Dollo model, that uses SNVs as phylogenetic markers but constrains losses of SNVs according to clusters of cells. We derive an algorithm, ConDoR, that infers phylogenies from targeted scDNA-seq data using this model. We demonstrate the advantages of ConDoR on simulated and real scDNA-seq data.

Keywords: Cancer; Dollo model; Intra-tumor heterogeneity; Single-cell DNA sequencing; Tumor phylogeny.

Fig. 1

Overview of the ConDoR algorithm. ConDoR takes as input: a a clustering of cells based on copy-number profiles and b the observed variant and total read counts from scDNA-seq data. ConDoR employs the constrained k-Dollo model to construct the c constrained k-Dollo phylogeny with mutation losses (dashed box) allowed only between cells from distinct copy-number clusters and the d inferred mutation matrix

Fig. 2

ConDoR outperforms existing methods in recovering the mutation matrix and the tumor phylogeny on simulated data. a Normalized mutation matrix error and b pairwise ancestral relation accuracy for each method compared to the simulated ground truth. Box plots show the median and the interquartile range (IQR), and the whiskers denote the lowest and highest values within 1.5 times the IQR from the first and third quartiles, respectively

Fig. 3

ConDoR provides insights into the evolution and spatial clonal architecture of a pancreatic ductal adenocarcinoma tumor using scDNA-seq data from two different regions of the tumor. a t-SNE plot showing results of clustering (details in the “Methods” section) of cells into 3 clusters (C0, C1, and C2) according to copy number profiles. b Constrained 1-Dollo phylogeny computed by ConDoR with edges labeled by the gain or loss of mutations and vertices labeled by the copy-number cluster and the fraction of cells from samples S1 and S2 that are attached at that vertex. c Reduction in normalized total read count for amplicon AMPL257637 (which contains mutations MGMT_1 and MGMT_2) in cells from cluster C2 compared to cells in cluster C1 ($p<5.8\times {10}^{-33}$, a one-sided KS test), supporting the loss of these mutations in the cells belonging to copy-number cluster C2. d Observed mutation matrix obtained by discretizing read counts of the 7 mutations, with cells grouped by copy number cluster as indicated in the first column. Box plots show the median and the interquartile range (IQR), and the whiskers denote the lowest and highest values within 1.5 times the IQR from the first and third quartiles, respectively

Fig. 4

ConDoR infers a phylogeny that is consistent with the copy-number clones in a metastatic colorectal cancer dataset. a The ConDoR phylogeny shows loss of bridge mutations FHIT and ATP7B and suggests monoclonal origin of the liver metastasis. b Losses inferred by ConDoR are supported by copy-number profiles from whole genome sequencing data of 42 cells from the same patient in the original study [37] (heatmap showing copy-number profiles adapted from [37]). Mutations LRP1B, LINGO2_1, and NR4A3 lie in regions (black boxes) that decrease in copy-number between the clusters that label the vertices on the edge in the phylogeny where the corresponding mutation (bold text in (a)) is lost : P$\to$M1 for LRP1B and LINGO2_1, and M1$\to$M2 for NR4A3