SVision: A deep learning approach to resolve complex structural variants

Abstract Complex structural variants (CSVs) encompass multiple breakpoints and are often missed or misinterpreted by state-of-the-art variant detection algorithms. We developed SVision, a deep-learning based multi-object recognition framework, to automatically detect and characterize CSVs from long-read data. SVision outperforms current variant callers at identifying internal structure of complex events and revealed 80 high-quality CSVs with 25 distinct structures from an individual genome. SVision directly detects CSVs without pattern matching against a database of known structures, allowing sensitive detection of both common and previously uncharacterized complex rearrangements.