CnnSV-Typer: Calling of Structural Variation Genotype Based on CUDA Acceleration

摘要

The current methods used for genotype calling are still limited to the length of structural variations and suffer from poor precision and sensitivity. Thus, this paper proposes CnnSV-typer, a deep learning approach which takes two-dimensional images as inputs and calls the variation genotypes. CnnSV-typer is tested and compared with current tools, including CNVnator, Pindel, LUMPY+SVTyper, Delly, GINDEL. The experimental results indicate that CnnSV-typer surpasses other methods on next generation sequencing data of various coverage, and could detect a wider length range of deletion genotypes with the highest precision and sensitivity reaching 99.3% and 99.6% respectively. The compression strategy based on CUDA parallel acceleration can achieve a maximum acceleration ratio of up to 381.3 times, which greatly improves the calling efficiency.