Abstract
Structural variants (SVs) are major sources of human genetic diversity. Despite advances in SV discovery, most single-nucleotide polymorphism (SNP)-centric frameworks overlook sequence-level information and encode overlapping variants as independent features, potentially reducing power in association studies, biasing heritability estimation, and limiting prediction accuracy. Here, we present LAVrefiner, a scalable, sequence-aware method that decomposes overlapping sequence-resolved SVs and indels into nonoverlapping refined sequence components and generates genotype matrices compatible with standard SNP-based analyses. In tandem repeat (TR) regions, LAVrefiner further supports motif-aware allele comparison and repeat copy-number estimation. When applied to variants from the 1000 Genomes Project, LAVrefiner identifies 224,708 refined variants. Simulations show that LAVrefiner is fast and memory-efficient while maintaining robust genotype representations under TR-associated sequencing noise. In gene expression analyses, models combining LAVrefiner-derived copy-number estimates with refined variants yield the highest heritability estimates and prediction accuracy in short tandem repeat (STR) and variable number tandem repeat (VNTR) regions, outperforming models based on the original SV/indel representation and existing TR copy-number tools. Cis-expression quantitative trait locus (cis-eQTL) mapping further shows that refined variants can localize regulatory signals to specific sequence components, including an independent VNTR-derived signal at the TFRC locus. Together, LAVrefiner provides an efficient, context-aware strategy for integrating sequence-resolved SVs and indels into existing genetic analysis frameworks.