SVcnn: A Novel Convolutional Neural Network Approach for Accurate Structural Variant Detection

SVcnn is a general structural variant (SV) caller method based on a convolutional neural network (CNN). It can accurately detect deletions (DELs), insertions (INSs), and inversions (INVs). While SVcnn does not directly detect duplications (DUPs), these can be identified by comparing INS sequences to the reference genome. SVcnn utilizes a simple CNN and is divided into three main parts: candidate SV region identification from BAM files, conversion of candidate regions into images for the LetNet model, and filtering false SVs using the LetNet model. Evaluations across three read datasets (CHM13, HG002, HG00733) show that SVcnn outperforms existing methods, with F1-scores improved by 2-8% when read depth exceeds 5x. Notably, SVcnn excels at detecting multi-allelic SVs and reduces false positives. Overall, SVcnn is a promising SV caller method with advantages in accuracy, particularly for detecting multi-allelic SVs, and reduced false positives, especially at higher read depths. Its reliance on a simple CNN makes it a valuable tool for SV detection, although its inability to directly detect DUPs may be a limitation in some cases.