Comparison and benchmark of long-read based structural variant detection strategies

Abstract

BackgroundRecent advances in long-read callers and assembly methods have greatly facilitated structural variants (SV) detection via read-based and assembly-based detection strategies. However, the lack of comparison studies, especially for SVs at complex genomic regions, complicates the selection of proper detection strategy for ever-increasing demand of SV analysis. ResultsIn this study, we compared the two most widely-used strategies with six long-read datasets of HG002 genome and benchmarked them with well curated SVs at genomic regions of different complexity. First of all, our results suggest that SVs detected by assembly-based strategy are slightly affected by assemblers on HiFi datasets, especially for its breakpoint identity. Comparably, though read-based strategy is more versatile to different sequencing settings, aligners greatly affect SV breakpoints and type. Furthermore, our comparison reveals that 70% of the assembly-based calls are also detectable by read-based strategy and it even reaches 90% for SVs at high confident regions. While 60% of the assembly-based calls that are totally missed by read-based callers is largely due to the challenges of clustering ambiguous SV signature reads. Lastly, benchmarking with SVs at complex genomic regions, our results show that assembly-based approach outperforms read-based calling with at least 20X coverage, while read-based strategy could achieve 90% recall even with 5X coverage. ConclusionsTaken together, with sufficient sequencing coverage, assembly-based strategy is able to detect SVs more consistently than read-based strategy under different settings. However, read-based strategy could detect SVs at complex regions with high sensitivity and specificity but low coverage, thereby suggesting its great potential in clinical application.