Single nucleotide variant calling protocols routinely discard calls in the regions immediately flanking structural variants (often 50 bp either side) because it is technically challenging to call them accurately. Although there are undoubtedly true variants of interest in these regions, many remain hidden as they are considered too difficult to distinguish from false positive calls. Nevertheless, with advances in both long-read sequencing and deep-learning algorithms, it is increasingly possible to resolve structural variants and their context more accurately. To provide guidance on SNV calling in the regions flanking structural variants (SVs), and to facilitate ongoing method development, we refined data from the Chinese Quartet project to construct a benchmarking set of 1000 SVs (more precisely, 299 deletions and 701 insertions), each on a completely assembled chromosome arm, supported by multiple sequencing technologies, and manually curated. We then used this real data, alongside corroboratory simulated data, to evaluate the performance of 35 short-read and 19 long-read variant calling pipelines at calling SNVs in their vicinity, representing combinations of up to 9 read aligners with up to 10 variant callers. Our datasets extend the scope of human benchmarking resources into these specific regions of the 'dark genome' and our results highlight practical strategies for variant calling within them.
Dang et al. (Mon,) studied this question.