Prevalence and detection of low-allele-fraction variants in clinical cancer samples

Abstract

Accurate detection of genomic alterations using high-throughput sequencing is an essential component of precision cancer medicine. We characterize the variant allele fractions (VAFs) of somatic single nucleotide variants and indels across 5095 clinical samples profiled using a custom panel, CancerSCAN. Our results demonstrate that a significant fraction of clinically actionable variants have low VAFs, often due to low tumor purity and treatment-induced mutations. The percentages of mutations under 5% VAF across hotspots in EGFR, KRAS, PIK3CA, and BRAF are 16%, 11%, 12%, and 10%, respectively, with 24% for EGFR T790M and 17% for PIK3CA E545. For clinical relevance, we describe two patients for whom targeted therapy achieved remission despite low VAF mutations. We also characterize the read depths necessary to achieve sensitivity and specificity comparable to current laboratory assays. These results show that capturing low VAF mutations at hotspots by sufficient sequencing coverage and carefully tuned algorithms is imperative for a clinical assay.

Fig. 1

Prevalence of point mutations with low variant allele fraction (VAF) in cancer specimens. VAF distributions of the four most frequently mutated actionable genes in our data: EGFR, KRAS, PIK3CA, and BRAF. Each dot corresponds to a sample, with the violin plot showing the estimated density; the red vertical dotted lines are at 5, 10, and 20% VAF. 24% of the EGFR T790M, 17% of PIK3CA E545, and 12% of KRAS G12 mutations are below 5%. Two important non-frameshift indels (indicated by asterisk) are also included for EGFR

Fig. 2

Features of actionable mutations with low VAF. a Concordance between VAFs estimated from panels and dPCRs for low VAF variants. The Pearson correlation coefficient based on 59 actionable variants is 0.86. Variants with higher coverage (colors correspond to sequencing depths) tend to show higher correlation; also see Supplementary Data 4. b Differences in VAF distributions among EGFR mutations in eight refractory lung cancer samples harboring EGFR C797S. All samples had an activating mutation (EGFR exon19 non-frameshift (NFS) deletion or L858R) and two resistance mutations (EGFR T790M and C797S), with the latter occurring at lower VAF. Dotted lines indicate mutations belonging to the same sample. The P-values were calculated using the paired t-test. c, d Browser view of the case in which EGFR C797S occurred in cis with EGFR T790M (only a subset of the reads are shown). e An acquired EGFR C797S mutation is found after AZD9291 therapy. f Comparison of VAF of EGFR variants of lung cancer samples (n = 141) pre- and post-EGFR-TKI therapy (afatinib, erlotinib, or gefitinib). The P values were calculated using the Wilcoxon rank sum test

Fig. 3

Impact of sequencing depth on single nucleotide variant (SNV) detection. a Limit of detection (LOD) is defined as the VAF for which 95% sensitivity is achieved for a given depth (Methods section). On the asis of subsampling analysis of computational mixture data, the LOD of 2% is achieved at 1085×, 5% at 294×, 10% at 94×, 20% at 40×, and 40% at 18×. b 72 samples harboring EGFR T790M were down-sampled from the original depth, and detection rates were measured at each depth (10 iterations), assuming that all variants were identified with the full data (1500×). Detection rate for this variant is <75% even for 100× data. Error bar: s.e.m

Fig. 4

Distribution of tumor purity in clinical samples. a The distributions of tumor purity estimates in our cohort and the TCGA samples are shown after kernel density smoothing. It was possible to reliably estimate purity (Supplementary Data 3) for about half the cases in our cohort (Methods section); purity estimates for the TCGA samples are taken from Aran et al.. Below 30%, the estimates are less reliable, as indicated by the dotted line. b Histological tumor purity estimates of 3697 lung cancer biopsy specimens obtained in daily clinical practice described in a related cohort. The cumulative distribution (inset) shows that half of the samples had purity <40%

Fig. 5

Variant classification and illustration of clinically-relevant low VAF mutation. a Frequency of genes with actionable alterations, colored by alteration type. b Distribution of SNVs/indels, CNVs, and fusions. c Proportions of patients with actionable alterations, non-actionable but known alterations (as catalogued by COSMIC), and without known alterations, among the samples sequenced by CancerSCAN V2 (n = 3598). d Examples of a metastatic lung cancer patient harboring low VAF (3.5%) EGFR T790M who had partial remission (PR) to the targeted therapy. e Comparison of the progression-free survival curves for lung cancer patients (n = 65) receiving third generation EGFR-TKI therapy with EGFR T790M. There is no statistically significant difference in survival between those with high and low VAF variants. SD, stable disease; PD, progressive disease