Running SV detection tools and benchmarking.

Severus (version 0.1.2), SAVANA (version 1.0.3), Manta (version 1.6.0), SvABA (version 1.2.0), and GRIDSS (version2.13.2) followed by GRIPSS (version 2.2) were run with tumor and normal bam pairs using default settings. For Severus, inputs also included a phased vcf described above and tandem repeat annotations file was downloaded from https://github.com/PacificBiosciences/pbsv. “nanomonsv parse” (version 0.6.0) was run with default parameters, and “nanomonsv get” was run with --use_racon and --single_bnd options. For Sniffles (version 2.0.7), default settings were used for the normal alignment, while the --non-germline option was used for the tumor alignment. The normal and tumor “snf” output files were merged into a single vcf and parsed for calls only found in the tumor sample. Both tumor and normal Sniffles runs also used the tandem repeats annotation file as an input.

As the GIAB HG002 benchmark was originally designed for the grch37 human reference, we lifted over the benchmarking Tier1 confidence intervals to grch38 and created a ground truth SV set from a curated HG002 assembly (^{Liao et al. 2023}) using hapdiff (^{Kolmogorov et al. 2023}).

For the complex SV analysis Severus (v1.0) was run with ‘--resolve-overlap’ and ‘--single-bp’ parameters for improved clustering.