Methods for structural variant detection with long-read sequencing data

Abstract The advent of third-generation sequencing and particularly long-read sequencing, which produces contiguous DNA fragments thousands of bases long, can greatly improve diploid genome assembly and detection of structural variants (SVs), in principle. With long-read whole genome sequencing data becoming increasingly available, designing efficient and reliable algorithms to identify SVs is critical. As a result, a plethora of SV callers have been recently developed to identify SVs. Alignment-based approaches are most popular because they are less computationally demanding and require substantially less sequencing coverage. Alternative approaches assembling the whole genome based on available reads alone (de novo assembly) and comparing assemblies to the reference genome with assembly-based tools to detect SVs, are much more demanding in data coverage and computational resources. However, the lack of comprehensive benchmarking limits our understanding of these two general approaches and hampers further algorithm development. Here we systematically compared 12 read alignment-based SV callers, and 4 assembly-based SV callers, along with 4 upstream aligners and 7 assemblers based on a rigorous and comprehensive benchmarking design. We provide performance results of these tools for different sizes of SVs, insertions vs. deletions, criteria for breakpoint and novel sequence detection, and alignment-based strategy vs. assembly-based strategies under different evaluation scenarios. Based on these results, we offer guidelines and recommendations for different tools and directions for further method development.