Abstract
Structural variants (SVs) in microbial genomes play a critical role in phenotypic changes, environmental adaptation, and species evolution, with deletion variations particularly closely linked to phenotypic traits. Therefore, accurate and comprehensive identification of deletion variations is essential. Although long-read sequencing technology can detect more SVs, its high error rate introduces substantial noise, leading to high false-positive and low recall rates in existing SV detection algorithms. This paper presents an SV detection method based on graph convolutional networks (GCNs). The model first represents node features through a heterogeneous graph, leveraging the GCN to precisely identify variant regions. Additionally, a knowledge-augmented activation layer (KANLayer) with a learnable activation function is introduced to reduce noise around variant regions, thereby improving model precision and reducing false positives. A clustering algorithm then aggregates multiple overlapping regions near the variant center into a single accurate SV interval, further enhancing recall. Validation on both simulated and real datasets demonstrates that our method achieves superior F1 scores compared to benchmark methods (cuteSV, Sniffles, Svim, and Pbsv), highlighting its advantage and robustness in SV detection and offering an innovative solution for microbial genome structural variation research.