Optimal sequencing strategies for identifying disease-associated singletons

Abstract

With the increasing focus of genetic association on the identification of trait-associated rare variants through sequencing, it is important to identify the most cost-effective sequencing strategies for these studies. Deep sequencing will accurately detect and genotype the most rare variants per individual, but may limit sample size. Low pass sequencing will miss some variants in each individual but has been shown to provide a cost-effective alternative for studies of common variants. Here, we investigate the impact of sequencing depth on studies of rare variants, focusing on singletons—the variants that are sampled in a single individual and are hardest to detect at low sequencing depths. We first estimate the sensitivity to detect singleton variants in both simulated data and in down-sampled deep genome and exome sequence data. We then explore the power of association studies comparing burden of singleton variants in cases and controls under a variety of conditions. We show that the power to detect singletons increases with coverage, typically plateauing for coverage > ~25x. Next, we show that, when total sequencing capacity is fixed, the power of association studies focused on singletons is typically maximized for coverage of 15-20x, independent of relative risk, disease prevalence, singleton burden, and case-control ratio. Our results suggest sequencing depth of 15-20x as an appropriate compromise of singleton detection power and sample size for studies of rare variants in complex disease.

Large-scale whole-genome sequencing of three diverse Asian populations in Singapore

“…In contrast, the mutation burden did not have an obvious trend with sequencing depth for variants with MAF>0.001 (Pearson's r<0.07, Figure 1B). Both simulation studies 24 and our empirical data (Supplementary Figure 6) suggested the power to detect extremely rare variants plateaus at ~25×. By assuming a linear model and a 0.99 non-reference sensitivity at 25× sequencing depth, we estimated 0.9329, 0.9225, and 0.8887 non-reference sensitivities at detecting variants with MAF<0.001 in our data for Chinese, Malays, and Indians, respectively ( Table 2).…”

“…The linear model enabled us to estimate the non-reference sensitivity in our call set by assuming 25× WGS achieved 0.99 non-reference sensitivity for detecting variants with MAF<0.001. 24 We reported the Pearson's correlation r between the number of non-reference variants and the sequencing depth per sample, and tested the null hypothesis of r=0 using the Student's t test.…”

“…We chose a medium coverage design of ~15× to maximize the statistical power for rare variant association studies given a fixed sequencing budget. 24,25 In this article, we describe the Phase 1 data of SG10K and initial findings based on 4,810 whole genomes, including 2,706 healthy samples without major diseases and 2,104 patient samples with heart failure, Parkinson's disease, and obesity. We characterize genetic mutations segregating in the population and in personal genomes and discuss the implications of our findings for precision medicine.…”

Extremely rare variants reveal patterns of germline mutation rate heterogeneity in humans

“…We identified and removed 156 samples which appeared to be technical outliers, resulting in a final call set of 35,574,417 autosomal ERVs from 3560 individuals (Methods). Due to the relatively low coverage of our sample, we likely failed to detect millions more ERVs—a recent study 26 estimated the discovery rate for singletons in a sample of 4000 whole genomes at 10× coverage to be ~65–85%. Quality control measures indicate that the ERVs we detected are high quality, with a transition/transversion (Ts/Tv) ratio of 2.00, within the commonly observed range for single nucleotide variants (SNVs) from WGS data 27 (Supplementary Table 1 ).…”

Extremely rare variants reveal patterns of germline mutation rate heterogeneity in humans

“…These studies have produced 24 accurate estimates of the genome-wide average mutation rate (~1 − 1.5 × 10 −8 mutations per base 25 pair per generation), and uncovered the aforementioned mutagenic effects of genomic features. 26 However, given the inherently low germline mutation rate, family-based WGS studies detect only 40-80 27 de novo mutations for each trio sequenced 9,10,12 . Due to the sparsity of these observed mutations, it is 28 difficult to accumulate a large dataset to precisely estimate mutation rates and spectrum at a fine scale 29 and identify factors that explain genome-wide variability in mutation rates.…”