SVision-pro: comparative sequence-to-image representation and instance segmentation for de novo and somatic structural variant discovery

Abstract Long-read based de novo and somatic structural variant (SV) discovery remains challenging, necessitating genomic comparison between samples. Here, SVision-pro visually represents genome-to-genome-level sequencing differences and comparatively discoveries SV between genomes by a neural-network-based instance segmentation framework without prerequisite of SV inference models. SVision-pro outperforms state-of-the-art approaches, particularly in resolving complex SV (CSV), with low Mendelian error rates and high sensitivity of low-frequency SVs. Moreover, SVision-pro successfully discoveries 26 high-quality de novo SVs in six family datasets and retrieves eight somatic CSVs in normal-tumor-paired samples.