Population-level structural variant characterization from pangenome graph

Population-level structural variant (SV) profiling is crucial in the era of pangenomes. However, identifying SVs from genome assemblies and pangenome graphs remains a significant challenge. Here we present Swave, a sequence-to-image, deep-learning based method that accurately resolves both simple and complex SVs, along with their population characteristics, from assembly-derived pangenome graphs. Swave introduces projection waves to summarize the dotplot images that capture mapping patterns between reference and SV-indicating alleles in pangenome. These images are analyzed by a recurrent neural network to distinguishes true SV signals from background noise introduced by genomic repeats. Swave demonstrates superior performance in both SV type classification and genotyping compared to existing methods. When applied to a healthy cohort (n=334) and a rare-disease cohort (n=574), Swave reveals complex and polymorphic SV patterns across human populations and identifies potentially pathogenic SVs. These advancements will facilitate the creation of comprehensive population-level SV catalogs, deepening our understanding of SVs in genetic diversity and disease associations.