Reproducibility of SNV-calling in multiple sequencing runs from single tumors

Abstract

We examined 55 technical sequencing replicates of Glioblastoma multiforme (GBM) tumors from The Cancer Genome Atlas (TCGA) to ascertain the degree of repeatability in calling single-nucleotide variants (SNVs). We used the same mutation-calling pipeline on all pairs of samples, and we measured the extent of the overlap between two replicates; that is, how many specific point mutations were found in both replicates. We further tested whether additional filtering increased or decreased the size of the overlap. We found that about half of the putative mutations identified in one sequencing run of a given sample were also identified in the second, and that this percentage remained steady throughout orders of magnitude of variation in the total number of mutations identified (from 23 to 10,966). We further found that using filtering after SNV-calling removed the overlap completely. We concluded that there is variation in the frequency of mutations in GBMs, and that while some filtering approaches preferentially removed putative mutations found in only one replicate, others removed a large fraction of putative mutations found in both.

SAVANA: reliable analysis of somatic structural variants and copy number aberrations in clinical samples using long-read sequencing

“…The key point here is that true somatic SVs should be detected in all replicates, whereas false positive calls resulting from library preparation or sequencing errors should only be detected in one replicate. The use of replicates has been widely established as an efficient strategy to benchmark algorithms for the detection of point mutations and indels [36][37][38] . More broadly, both technical and biological replicates represent a cornerstone across various research domains to assess the reproducibility and robustness of experimental and computational results, including sequencing data analysis 37,39 .…”

“…A potential limitation of using replicates to benchmark mutation detection algorithms is that tumours often show high levels of intra-tumour genetic heterogeneity. As a result, true somatic SVs present at low allele frequency (AF) in a tumour might only be detected in one replicate if the number of sequencing reads in other replicates is not enough to meet the threshold for SV calling 38,39 . Thus, we next sought to investigate whether the variable number of SVs detected by different algorithms is the result of variable sensitivity for low-AF SVs.…”

Accuracy and reproducibility of somatic point mutation calling in clinical-type targeted sequencing data

“…Also problematically, they defined gold standard mutations by intersecting variant calls from GATK and SAMtools, two of the four programs they benchmarked. Derryberry et al [14] focused on the technical reproducibility of variant calls, studying 55 replicate pairs of data from gliobastoma tumours. However, this was whole genome sequencing data, where coverage was substantially lower than for targeted panels.…”