Long-read-based human genomic structural variation detection with cuteSV

Abstract

Long-read sequencing is promising for the comprehensive discovery of structural variations (SVs). However, it is still non-trivial to achieve high yields and performance simultaneously due to the complex SV signatures implied by noisy long reads. We propose cuteSV, a sensitive, fast, and scalable long-read-based SV detection approach. cuteSV uses tailored methods to collect the signatures of various types of SVs and employs a clustering-and-refinement method to implement sensitive SV detection. Benchmarks on simulated and real long-read sequencing datasets demonstrate that cuteSV has higher yields and scaling performance than state-of-the-art tools. cuteSV is available at https://github.com/tjiangHIT/cuteSV .

Keywords: Long-read sequencing; Scaling performance; Structural variants detection.

Fig. 1

Schematic illustration of the cuteSV approach. cuteSV uses sorted BAM file as input to detect SVs in 3 major steps. In step 1 (“discovering SV signatures”), cuteSV collects various types of SV signatures comprehensively from inter- and intra-alignments. In step 2 (“clustering of SV signatures”), a heuristic clustering-and-refinement method is employed to sensitively discover accurate SV alleles. In step3 (“SV calling and genotyping”), cuteSV generates the SV callsets and assigns genotypes

Fig. 2

Benchmark results of the SV callers on various simulated datasets. F1 scores of a deletion, b insertion, c duplication, d inversion, e translocation at breakpoint level, and f translocation at breakend level, for the simulated datasets in various coverages and w/o genotyping. In the figure, “N×” and “N×-GT” indicate the statistics without and with genotyping, respectively

Fig. 3

Benchmark results of the SV callers on various HG002 PacBio sequencing datasets. a Precisions, recalls, and F1 scores on the whole and down-sampled HG002 PacBio CLR datasets. b Precisions, recalls, and F1 scores on the whole and down-sampled HG002 PacBio CCS datasets. c Recall rate of homozygous parental variants. d Mendelian-Discordance-Rates (MDRs) for the variants unique to the offspring

Fig. 4

Benchmark results of the SV callers on various of HG002 ONT sequencing datasets. a Precisions, recalls, and F-scores on the whole and down-sampled HG002 ONT datasets. b The Venn diagram of SV calls produced by cuteSV from HG002 PacBio CLR, CCS, and ONT PromethION datasets (indicated by “CLR”, “CCS,” and “ONT”, respectively). c The Venn diagram of SV calls produced by different tools on HG002 PacBio CLR data. d The Venn diagram of SV calls produced by different tools on HG002 PacBio CCS data. e The Venn diagram of SV calls produced by different tools on HG002 ONT PromethION data

Fig. 5

Performance of the benchmarked SV callers. The a runtimes and b memory footprints of cuteSV, cuteSV, Sniffles, and PBSV with 1, 2, 4, 8, and 16 CPU threads. “Skip GT” indicates the statistics without genotyping. SVIM was benchmarked with single CPU thread only since it does not support multiple thread computing