Towards a better understanding of the low recall of insertion variants with short-read based variant callers

Since 2009, numerous tools have been developed to detect structural variants (SVs) using short read technologies. Insertions > 50 bp are one of the hardest type to discover and are drastically underrepresented in gold standard variant callsets. The advent of long read technologies has completely changed the situation. In 2019, two independent cross technologies studies have published the most complete variant callsets with sequence resolved insertions in human individuals. Among the reported insertions, only 17 to 37% could be discovered with short-read based tools. In this work, we performed an in-depth analysis of these unprecedented insertion callsets in order to investigate the causes of such failures. We have first established a precise classification of insertion variants according to four layers of characterization: the nature and size of the inserted sequence, the genomic context of the insertion site and the breakpoint junction complexity. Because these levels are intertwined, we then used simulations to characterize the impact of each complexity factor on the recall of several SV callers. Simulations showed that the most impacting factor was the insertion type rather than the genomic context, with various difficulties being handled differently among the tested SV callers, and they highlighted the lack of sequence resolution for most insertion calls. Our results explain the low recall by pointing out several difficulty factors among the observed insertion features and provide avenues for improving SV caller algorithms and their combinations.