Fig 1.
Optimisation of read depth, DNA quantity, and unique alternate observation threshold.
A: The detection rate of reference standard variants ranging from 0.0008 to 0.1 VAF at 500-5,000 × depth. Variants ≥ 0.02 VAF could be detected at 97% as low as 500 × , whereas 5,000 × was required to detect two-thirds of variants at 0.0008 VAF. A target depth of 2,000-3,000 × detected 95% of variants ≥~0.01 VAF. B: The relationship between the detection rate of very small (~0.001 VAF) variants and sequencing depth. A target depth of 3,000-4,000x detected 60-80% of variants at ≥~0.001 VAF. C: The association between expected VAF and observed VAF according to read depth. Lower read depth may contribute to an overestimated VAF of very small variants, although differences did not reach significance. D-F: Blood samples from 383 community-dwelling subjects were sequenced at a target depth of 3,000-4,000 × . Variants with a prevalence of ≥ 10% within this cohort were excluded as likely sequencing artefacts, which excluded 10.5% of variants (D). The mean number of variants detected per sample above the standard CHIP threshold of ≥ 0.02 VAF increased linearly with subject age (E). Greater than 99.9% of variants detected were below ≥ 0.02 VAF, potentially including those with prognostic or predictive clinical relevance (F). G: Libraries were prepared with 5–400ng of input DNA. Using somatic variants detected in the 400ng sample as “ground truth”, the mean detection rate was determined for VAF ranges with 0.01–0.02, 0.02–0.05, and ≥ 0.05. Almost all variants at ≥ 0.02 VAF were detected using an input DNA of ≥ 50ng, however detection of variants at < 0.02 VAF benefitted from increasing input DNA amounts. H: The detection rate of variants detected in samples from 4 AML patients from which duplicate library preparations were performed. Across both independent replicates 96% of variants were detected at VAF ≥ 0.02, 85% for 0.01–0.02 VAF, 60% of variants at 0.005-0.01 VAF, and 18% of variants at 0.001–0.005 VAF. I: The CV of variant size quantitation across independent replicates, which displayed a strong negative correlation to variant size (dotted lines indicate 95% confidence interval). J-L: Increasing unique alternative observation (UAO) thresholds for variant calls increased detection rate of rare variants (J), but markedly decreased the number of variants passing the threshold (K) and therefore reduced detection of true positive variants in reference standards (L). Error bars represent standard deviation.
Table 1.
Expected and observed VAF, read depth and UAO for reference standard variants. Reference standard HD829 was serially diluted with leukocyte-derived DNA from a healthy 24-year-old subject. Undiluted (100%) and diluted samples (25%, 6.25%, 3.13% and 1.56%) were used to compare expected VAF (manufacturer’s supplied, ddPCR) to observed VAF for 15 accessible variants, using the optimised parameters illustrated in Fig 1. ND, not detected.
Fig 2.
Analytical range of DNA reference standards.
A-B: Overall, Horizon reference standards with an average read depth of 3121x showed linearity (R2 = 0.9784) across dilutions with a strong Pearson correlation (p < 0.0001) to manufacturer’s expected VAF (ddPCR to an average read depth of 582x) (A). C. Assessment of individual variants identified an over-estimation of ASXL1:c.1934dup at VAF ≤ 0.025. D. The lower limit of detection (LLOD) for ASXL1:c.1934dup was set at 0.031 VAF, which is below the VAF of reference standard and reference sample true positives and above observed VAF in reference samples not called by the reference laboratory. E. Variant FLT3:c.2503G > T showed consistent negative bias of 0.3-fold compared to expected VAF. F-G: FLT3-ITD showed a linear relationship across dilutions (F), and a 100% detection rate was observed as low as 0.0031% VAF at a mean depth of 3128 × , though this decreased rapidly at lower VAFs (G). H-I: Variants other than ASXL1:c.1934dup and FLT3:c.2503G > T showed a strong linear relationship (R2 = 0.9979) and Pearson correlation (p < 0.0001, H), with a positive bias at VAF ≤ 0.004 (I). J-K: Limiting VAF to ≥ 0.004 revealed a bias of 0.89-fold. Error bars represent standard deviation. Vertical dotted line indicates 0.004 VAF. Horizonal solid and dotted lines in Bland-Altman plots (C, E, I, K) indicate bias, and the upper limit (UL) and lower limit (LL) of 95% confidence intervals.
Fig 3.
Precision and lower limit of detection.
A: Repeatability of optimised assay showing intra-technologist and -sequencer machine variability as VAF vs CV. Average CV increased with decreased VAF. B-C: Reproducibility inter-sequencer machine (B) and inter-technologist (C) demonstrated high linearity (R2 = 0.9674, 0.9444) and Pearson correlation (both p < 0.0001). D-E: CV varied according to technologist, and consistently showed higher CV with decreasing VAF. F: Detection rate of the optimised assay was 100% for the reference standards at VAF ≥ 0.004 but fell rapidly < 0.004 VAF. LLOD was therefore set at VAF ≥ 0.004. G: The detection of five rare clones (<0.005 VAF) identified in blood samples from community-dwelling subjects sequenced at a mean depth of 3485 × with a UAO filter of ≥ 3 were independently validated by ultra-sensitive ddPCR. Of these, 3 (60%) were successfully detected, consistent with the moderate potential false-positivity rate at < 0.005 VAF. Error bars represent standard deviation. Vertical dotted line mark the LLOD of 0.004 VAF. Dotted lines on CV indicate 95% confidence interval.
Table 2.
Reference standard and clinical sample measures of trueness. Trueness validation of optimised assay with a lower limit of detection (LLOD) of ≥ 0.004 VAF (and LLOD 0.031 for ASXL1:c.1934) using reference standards HD829 and HD752 true positives (TP, 23 assessed variants) and true negatives (TN, wild type sequences of 29 assessed variants), and clinical samples from a reference laboratory included 16 participants to determine concordance of 36 variants reported above the reference laboratory LLOD and 3 participants without detectable variants (Presumed Negative clinical samples) to determine concordance for absence of variants. Results showed a 100% concordance between 59 true positive variants and 29 true negatives (concordance for absence of variants). Sensitivity, specificity, positive predictive value, negative predictive value and accuracy were determined to be 100%, which exceed the acceptable criteria of ≥ 98%.
Table 3.
Detect ion of variants in presumed negative clinical samples. A total of 193 variants were detected in Presumed Negative reference laboratory samples. However, 108 of these were not covered in the reference laboratory panels, and the remaining 85 variants were detected below the reference laboratory LLOD at the time of testing. Hence 100% concordance for absence of variants was achieved.
Table 4.
Utility of optimised assay with a lower limit of detection of ≥ 0.004. For a target VAF range the optimal read depth and DNA quantity is indicated to achieve 100% accuracy. Indicated relative total cost and time include raw data file size for 144 samples and the cost for data storage, cost and time for library preparation and quantification, sequencing, and bioinformatic analysis (refer to S7–S8 Tables). Also included are the number (mean and range) of somatic variants observed in community dwelling individuals, and the coefficient of variance (CV) in reference standard and community dwellers at each VAF range.