The stochastic nature of errors in next-generation sequencing of circulating cell-free DNA

Abstract

Challenges with distinguishing circulating tumor DNA (ctDNA) from next-generation sequencing (NGS) artifacts limits variant searches to established solid tumor mutations. Here we show early and random PCR errors are a principal source of NGS noise that persist despite duplex molecular barcoding, removal of artifacts due to clonal hematopoiesis of indeterminate potential, and suppression of patterned errors. We also demonstrate sample duplicates are necessary to eliminate the stochastic noise associated with NGS. Integration of sample duplicates into NGS analytics may broaden ctDNA applications by removing NGS-related errors that confound identification of true very low frequency variants during searches for ctDNA without a priori knowledge of specific mutations to target.

Fig 1. Comparison of sequencing metrics between adapter types.

The total number of reads where both read 1 and read 2 were present was similar between the singleton and duplex adapter groups (a). However, after consensus sequence determination, there were significantly fewer consensus sequences (b) and larger overall family sizes (c) in the duplex adapter group. Read depth (d) was greater in the duplex adapters at larger family sizes. Bar and whiskers represent mean±SD. Data points shown in (d) represent the mean value from the seven control samples.

Fig 2. Use of singleton and duplex adapters to reduce noise.

Compared to singleton adapters, consensus sequences derived from duplex adapters provided greater error correction (a) and lower error (b) at family size (FS) ≥2. The gain in error correction using duplex adapters was similar regardless of family size and increments in family size reduced error regardless of adapter type (c). Although errors most commonly occurred at an allele frequency <0.1%, a substantial portion of errors had an allele frequency >0.1% (d). Bar and whiskers represent mean±SD. Data points shown in (c) represent the mean value from the seven control samples.

Fig 3. Sources of error and effects of error correction in ccfDNA.

Removal of potential CHIP-related artifacts had a relatively small impact on error (a), particularly when compared to removal of highly patterned error (i.e., positions with errors in all seven controls; b). The greatest reduction in error occurred with application of sample duplicates (c). The effects on error from using duplex adapters (X), accounting for CHIP-related artifacts (C), removing positions with highly patterned error (P), and applying data from samples duplicates (D) are shown individually in (d). Accounting for all of the different sources of noise yielded the lowest error (d, pink), which continually decreased with increments in family size. Data points shown in (d) represent the mean value from the seven control samples. FS = family size; CHIP = clonal hematopoiesis of indeterminate potential.

Fig 4. Distribution of NRA types associated with ccfDNA.

All data are shown for family size (FS) ≥2. The number of counts for each type of the twelve possible base pair changes is shown in (a) for all observed NRAs. In (b), the percent of the total NRAs for each base pair change is shown. The distribution for base pair changes associated with CHIP-related artifacts and highly patterned error (NRAs common to all seven samples) is shown in (c) and (d), respectively. For CHIP-related artifacts, only results from duplex adapters are shown because buffy coat DNA was sequenced only with duplex adapters. In (d), error bars are absent because the NRAs were present in all seven samples for each adapter type. In (e), the number of counts associated with each type of base pair change present in both of the full library duplex adapter duplicates is depicted. After using full library duplex adapter duplicates to reduce error and removing CHIP-related artifacts and highly patterned error, the distribution of each type of base pair change is shown in (f). Error bars represent SD. CHIP = clonal hematopoiesis of indeterminate potential.