Dual-SVF: a robust knowledge-guided multimodal method for structural variant filtering in long-read sequencing

Structural variants (SVs) are key drivers of genomic diversity and disease, yet their accurate detection from long-read sequencing remains challenged by high false-positive rates caused by sequencing errors and alignment artifacts. Current filtering approaches predominantly rely on alignment structure, often overlooking sequence content and genomic context knowledge, which undermines the robustness of SV detection. To address these issues, we present Dual-SVF, a robust knowledge-guided multimodal method for SV filtering that jointly models genomic semantics (from raw sequence content) and syntactics (from sequence alignment topology). Dual-SVF integrates genomic prior knowledge, including sequence entropy, GC content, and mapping quality, through a confidence-gated cross-attention mechanism that dynamically weights modality reliability and enables mutual error correction. Validation across diverse sequencing platforms and multiple species demonstrates that Dual-SVF consistently achieves superior performance compared with state-of-the-art methods. Dual-SVF is an open-source, VCF-compatible tool, which seamlessly complements existing pipelines to ensure reliable SV filtering across noisy genomic data.