Enter Note Done

Session


Keywords: Bioinformatics; Cancer; Cell-free DNA; Copy number/structural variation; Diagnostics

Authors:
C. Melton; P. Freese; S. Bagaria; E. Hubbell; D. Filippova; R. Calef; M.C. Maher; V. Nicula; S. Gross; O. Venn; A. Valouev; A. Aravanis; A. Jamshidi

Affiliation: GRAIL, Menlo Park, CA


Cancer can be detected through whole genome sequencing (WGS) of cell-free DNA (cfDNA). The most commonly derived set of features from WGS data, fragment counts in bins tiled across the genome, enable detection of somatic copy number alterations (CNAs); however, other features can be leveraged to classify cancer status. In the first substudy of Circulating Cell-free Genome Atlas (CCGA; NCT02889978) we evaluated three prototype assays for discriminating cancer vs non-cancer; we determined that fragment methylation-based features from whole genome bisulfite sequencing had the best sensitivity (at high specificity), followed by small variants detected from targeted sequencing and somatic CNAs from WGS read depth. An exploratory post hoc analysis of the first substudy evaluated the impact of alternative features derived from WGS data on classification performance. We assessed the utility of counting fragments originating at cancer-enriched fragment endpoint positions and explored alternative means of detecting CNAs by observing (a) allelic imbalance at heterozygous SNPs and (b) changes in fragment length distributions in tiled bins. Optimization of features and learning algorithms were performed in 10-fold cross-validation on a training set (561 controls, 863 cancer participants [pts], 20 solid tumor types) and final assessment performed on an independent test set (362 controls, 464 cancer pts, 20 solid tumor types). At 98% specificity, fragment endpoints had 18.1% (95% CI, 14.7-21.9) sensitivity vs 34.1% (29.7-38.6) sensitivity for methylation and 29.3% (25.2-33.7) for binned counts. 29.1% sensitivity (25.0-33.5) for fragment lengths and 21.8% (18.1-25.8) for allelic imbalance was also observed. All classifiers showed strong stage and tumor fraction dependence. An ensemble WGS classifier combining endpoints, allelic ratios, fragment lengths, and binned counts outperformed binned counts alone (+2.2% sensitivity, p=0.044) but had lower sensitivity than the methylation-based classifier (-2.6% sensitivity, p=0.045). WGS and methylation features combined failed to improve upon the methylation classifier alone (-0.6% sensitivity, p=0.546). Overall, combining WGS features improved sensitivity at high specificity in the absence of methylation features but was inferior to methylation-based classification; these results, in part, motivated development of an improved methylation assay assessed in the second substudy of CCGA.1

1. J Clin Oncol 37, 2019 (suppl; abstr 3049).