Scalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumors

Abstract

Whole-exome sequencing of cell-free DNA (cfDNA) could enable comprehensive profiling of tumors from blood but the genome-wide concordance between cfDNA and tumor biopsies is uncertain. Here we report ichorCNA, software that quantifies tumor content in cfDNA from 0.1× coverage whole-genome sequencing data without prior knowledge of tumor mutations. We apply ichorCNA to 1439 blood samples from 520 patients with metastatic prostate or breast cancers. In the earliest tested sample for each patient, 34% of patients have ≥10% tumor-derived cfDNA, sufficient for standard coverage whole-exome sequencing. Using whole-exome sequencing, we validate the concordance of clonal somatic mutations (88%), copy number alterations (80%), mutational signatures, and neoantigens between cfDNA and matched tumor biopsies from 41 patients with ≥10% cfDNA tumor content. In summary, we provide methods to identify patients eligible for comprehensive cfDNA profiling, revealing its applicability to many patients, and demonstrate high concordance of cfDNA and metastatic tumor whole-exome sequencing.

Conflict of interest statement

T.R.G., L.A.G., N.W. are consultants and equity holders in Foundation Medicine, Inc. M.M. was previously a consultant and equity holder in Foundation Medicine, Inc. C.-Z.Z. is a consultant and equity holder in Pillar Biosciences. The authors have filed a patent application on methods described in this manuscript. The remaining authors declare no competing financial interests.

Figures

Fig. 1

Copy number and tumor fractions from ULP-WGS. a cfDNA workflow. b Genome-wide copy number from 0.1× ULP-WGS of cfDNA from a healthy donor. c Genome-wide copy number from 25× WGS and 0.1× WGS of cell-free DNA from a metastatic breast cancer patient (MBC_315), and 1× WGS and WES of matched tumors from this patient. SCNA for tumor WES and cfDNA 25× coverage WGS were predicted using TITAN (“Methods”). d Comparison of copy ratios between ULP-WGS of cfDNA with deep (>10×) WGS of the same cfDNA sample, WGS (1×) of matched tumors from 22 metastatic breast cancer (MBC) patients, and WES (average mean target coverage 173×) of matched tumors from 41 MBC and prostate cancer (CRPC) patients. Log2 copy ratios were computed as normalized read coverage for each 1 Mb (WGS/ULP-WGS) and the mean of overlapping 50 kb bins (WES) after adjustment for tumor fraction/purity. The correlation of copy ratios between tumor and cfDNA was computed using Spearman rank correlation (coefficient ρ). F-measure (F1) is the harmonic mean of the CNA positive predictive value (precision) and sensitivity (recall) performance. Recall is defined as the proportion of SCNA gain/loss in tumor biopsy also observed in ULP-WGS of cfDNA (“Methods”). e Comparison of tumor fractions estimated from ULP-WGS and WES of cfDNA. Samples (n = 35) with similar tumor ploidy (difference < 0.75 and ploidy ≥1.5) estimated in both ULP-WGS and tumor WES are shown. The correlation between the two data types was calculated using Pearson correlation (coefficient r). Red line denotes y = x. WES tumor fractions were estimated using ABSOLUTE (shown) and TITAN (Supplementary Fig. 15, Supplementary Data 6)

Fig. 2

Comparison of whole-exome sequencing of cfDNA to whole-exome sequencing of matched tumor biopsies. a Fraction of clonal (≥0.9 cancer cell fraction, CCF) and subclonal (<0.9 CCF) SSNVs detected by MuTect in WES of tumor biopsies and confirmed (i.e., supported by ≥3 variant reads) in WES of cfDNA. Sites with <3 reads that had power <0.9 for mutation calling were not included when computing the fraction of SNVs confirmed (“Methods”). b Fraction of clonal and subclonal SSNVs detected in WES of cfDNA and confirmed in WES of tumor biopsies. For 18 patients with WES of cfDNA at a second time point t2, SSNVs not detected in the matched tumor biopsy but confirmed at t2 are indicated with black. c Analysis of clonal dynamics in an ER+ breast cancer patient diagnosed with metastatic disease 1.5 years (yrs) prior to biopsy and cfDNA collection (t1, Day 0). Clustering analysis of CCF for SSNVs between matched tumor biopsy and cfDNA (t1) is shown in the left panel. The right panel shows the CCF of four mutation clusters, one containing ESR1 L536P (Subclonal Cluster 1, orange) and the other containing ESR1 D538G (Subclonal Cluster 2, light blue), at t1 and t2 (51 days apart) from a patient with ER+ metastatic breast cancer being treated with a SERD. The lymph node biopsy was taken at the same time as cfDNA t1. Mutations were clustered by the CCFs for each pair of samples using Phylogic (“Methods”). Error bars represent the 95% credible interval of the joint posterior density of the clusters. Mutations, excluding indels, having ≥90% estimated power based on coverage in both samples are shown; clusters with fewer than three mutations are excluded. The number of mutations in each cluster is indicated in the legend in parentheses

Fig. 3

Genomic alterations of known significance and applicability to large cohorts. a, b The alteration status of significantly mutated genes predicted by MutSig2CV, (a), focal SCNAs (b), and known cancer-associated genes are shown for cfDNA and tumor biopsies from 27 metastatic breast cancer (MBC) patients. Mutated genes with MutSig2CV q-value < 0.1 are statistically significant. Mutations that were exclusively detected in one sample may be present at low CCF in the other matched sample but were excluded from the frequency calculation. SCNA frequencies were computed for oncogenes (MYC to ERBB2) and tumor suppressors (BRCA1 to ATM) using only amplification and deletion status, respectively. Mutations were predicted using MuTect and SCNAs were predicted using ReCapSeg and ABSOLUTE. Red dot indicates distinct mutations in tumor and cfDNA. c Mutational signatures in whole-exome sequencing of cfDNA and tumor biopsies were predicted using a Bayesian non-negative matrix factorization (NMF) approach (“Methods”). Samples with predicted biallelic inactivation of BRCA1/2 are indicated in red and blue. Black line denotes y = x; blue line denotes model fit using linear least squares regression. d Neoantigen burden, defined as the number of predicted neoantigen SSNVs, was calculated using NetMHCpan (“Methods”). Black line denotes y = x; blue line denotes model fit using linear least squares regression. e Applicability to many patients with metastatic cancer. Tumor fractions estimated from ULP-WGS of cfDNA from 903 blood samples from 391 patients with metastatic breast cancer and 536 blood samples from 129 patients with metastatic prostate cancers. The earliest blood drawn for each patient is shown. Samples with coverage <0.05× were excluded