Benchmarking long-read aligner-caller pipelines for structural variant detection across ten angiosperm species with PacBio high-fidelity sequencing

Structural variants (SVs) play a central role in genome evolution and phenotypic diversity, yet the large genome size, high repeat content, and structural complexity hinder SV detection in plants. Despite extensive benchmarking on human data, the performance and cross-species generalizability of HiFi long-read SV detection pipelines for plants have been overlooked. Here, we use multiple assembly-based variant calling strategies to construct SV truth sets at base-level resolution for ten representative angiosperm species, followed by systematic benchmarking of 91 HiFi-based SV calling pipelines. Cross-species evaluation reveals pronounced heterogeneity in performance across pipelines and species with deletion detection generally outperforming insertion detection. Using a stability score to quantify the across-species generalizability, we identify NGMLR/Sniffles2 and NGMLR/DeBreak as optimal cross-species pipelines for deletion detection, while Minimap2/Sniffles1, Minimap2/cuteSV2, and pbmm2/cuteSV1 exhibit robust insertion detection. We also find that detection accuracy declines sharply for large SVs and variants located within repetitive regions, particularly simple repeats. Moreover, SV calling precision shares a nonlinear relationship with sequencing depth for each individual haplotype, with F 1 scores peaking at ~ 15 × coverage. This study establishes a benchmarking framework for plant SV detection and provides high-confidence SV truth set resources that can, together, guide the selection of robust, scalable SV calling pipelines for plant genomics.