dicast: a machine learning method for accurate structural variant detection from short-read sequencing data

Structural variants are a common cause of human diseases, but their detection from short-read sequencing remains challenging, despite being the technology underlying most clinical workflows. We present dicast, a machine-learning method that scores SV calls from short-read data using alignment and genomic-context features. dicast is trained on a new multi-technology ground truth built from nine samples, with extensive manual curation. It outperforms existing short-read callers and consensus approaches, recovering substantially more true positives at high precision. We also demonstrate dicast's applicability for diagnostics, identifying all pathogenic variants in multiple disease cohorts, and 20% more candidate pathogenic deletions than consensus approaches.