Introducing deep learning -based methods into the variant calling analysis pipeline

Abstract

Biological interpretation of the genetic variation enhances our understanding of normal and pathological phenotypes, and may lead to the development of new therapeutics.

However, it is heavily dependent on the genomic data analysis, which might be inaccurate due to the various sequencing errors and inconsistencies caused by these errors. Modern analysis pipelines already utilize heuristic and statistical techniques, but the rate of falsely identified mutations remains high and variable, particular sequencing technology, settings and variant type.

Recently, several tools based on deep neural networks have been published. The neural networks are supposed to find motifs in the data that were not previously seen.

The performance of these novel tools is assessed in terms of precision and recall, as well as computational efficiency. Following the established best practices in both variant detection and benchmarking, the discussed tools demonstrate accuracy metrics and computational efficiency that spur further discussion.