Quantification of tumour evolution and heterogeneity via Bayesian epiallele detection

Abstract

Background: Epigenetic heterogeneity within a tumour can play an important role in tumour evolution and the emergence of resistance to treatment. It is increasingly recognised that the study of DNA methylation (DNAm) patterns along the genome - so-called 'epialleles' - offers greater insight into epigenetic dynamics than conventional analyses which examine DNAm marks individually.

Results: We have developed a Bayesian model to infer which epialleles are present in multiple regions of the same tumour. We apply our method to reduced representation bisulfite sequencing (RRBS) data from multiple regions of one lung cancer tumour and a matched normal sample. The model borrows information from all tumour regions to leverage greater statistical power. The total number of epialleles, the epiallele DNAm patterns, and a noise hyperparameter are all automatically inferred from the data. Uncertainty as to which epiallele an observed sequencing read originated from is explicitly incorporated by marginalising over the appropriate posterior densities. The degree to which tumour samples are contaminated with normal tissue can be estimated and corrected for. By tracing the distribution of epialleles throughout the tumour we can infer the phylogenetic history of the tumour, identify epialleles that differ between normal and cancer tissue, and define a measure of global epigenetic disorder.

Conclusions: Detection and comparison of epialleles within multiple tumour regions enables phylogenetic analyses, identification of differentially expressed epialleles, and provides a measure of epigenetic heterogeneity. R code is available at github.com/james-e-barrett.

Keywords: Epigenetics; Heterogeneity; Phylogenetics.

Fig. 1

a An example of a genomic locus (chr1:1,145,478-1,145,614) in which each row corresponds to a sequencing read. Black and white circles represent methylated and unmethylated CpGs respectively. Note that some CpG measurements are missing. b The four epialleles that are inferred from the observed sequencing reads. c The Akaike Information Criterion score versus the total number of epialleles. The inferred number of epialleles corresponds to the minimum AIC score. d The proportion of observed reads attributed to each epiallele after marginalisation over the parameter w (see main text for details)

Fig. 2

Estimation of tumour sample purity for region 2 of the tumour. The parameter ξ was calculated at all eligible loci across the genome and the empirical distribution is plotted here. The sample purity is equal to the maximum value of ξ which is interpreted to occur at the rightmost maximum at ξ=0.53. The distribution of ξ is ‘smoothed’ due to the fact that at each locus ξ is estimated from a finite sample of sequencing reads

Fig. 3

A genomic locus (chr1:2,603,277-2,603,489) composed of seven CpGs. The distribution of five epialleles – inferred using the Bayesian model – are plotted for seven tumour regions (R1 to R7) and one normal sample (N). In a the tumour samples have not been corrected for normal tissue contamination whereas in b they have been. The tumour samples are shifting towards an unmethylated profile in comparison to the normal tissue. The locus lies in a large intronic region in the gene TTC34

Fig. 4

a Heatmap of the top 200 most variable epialleles across the seven tumour samples (labelled R1 to R7) and matched normal sample (labelled N). A proportion of 1.0 (dark blue) means that that epiallele accounted for all observed methylation patterns at the corresponding locus. These data have not been decontaminated of normal tissue. b The phylogenetic tree inferred before correction for contaminating normal tissue. In c and d are the same figures for the decontaminated epiallele profiles. In the top annotation track green denotes a CpG island, yellow a shore, and blue otherwise. In the bottom track dark purple denotes a gene promoter, otherwise light pink. A promoter was defined as between 2kb upstream and 50bp downstream from a transcription start site

Fig. 5

Box plots of the Shannon entropy of the epiallele distribution across normal tissue (N) and the seven tumour regions (R1–R7)