Towards standards for analysing structural variants from short-read whole genome sequencing data

No widely accepted standard exists for analysing structural variants from short-read whole genome sequencing (WGS) data. We use comprehensive SV truth sets (HG002 T2T-Q100, Platinum Pedigree) and long-read sequencing databases (1KGP-ONT, CoLoRS) to benchmark SV breakpoints and sizes at base pair resolution. Using 30X Illumina WGS data, only five of sixteen individual tools tested reliably output deletions, insertions and duplications that perfectly match the truth sets. Ensemble pipelines (e.g. GATK-SV, MOPline, Parliament2), frequently used in population-scale datasets, reduced perfect matching accuracy by up to 98% compared to their highest-performing constituent tools. This inaccuracy propagates into a short-read database (gnomAD-SV) constructed from an ensemble pipeline (GATK-SV) and likely affects other databases (TOPMed, All of Us, Biobank Japan). As a proof-of-concept, we demonstrate an intra-sample merging strategy that generates the highest F1-scores for deletions and insertions on the HG002 sample, recovering more perfect matches than total matches from the best performing ensemble pipeline tested. Common HG002 truth SVs (allele frequency > 1%) showed extremely high concordance for exact matches between our ensemble callset and two orthogonal long-read sequencing databases (CoLoRS, ONT-1KGP). This study provides a vital step towards robust standards for SV analysis, guiding short-read WGS ensemble pipeline development and improving cross-study comparisons.