GWANN: Implementing deep learning in genome wide association studies

Abstract

MotivationGenome wide association studies (GWAS) are extensively used across species to identify genes that underlie important traits. Most GWAS methods apply modifications and extensions to a linear regression model in order to detect significant associations between genetic variation and a trait. Despite their popularity, these statistical models tend to suffer from high false positive rates, especially when utilized on large variant datasets or complex demographic scenarios. To overcome this, aggressive statistical corrections are applied which frequently diminish true associations. ResultsHere we consider a deep learning approach, and present an implementation of a convolutional neural network (CNN) to identify genetic variation that is associated with a trait of interest. To exploit the strength of CNNs in visual recognition, the genotype information is represented as an image, which enables the model to correctly classify genetic variants with respect to the trait, even when a population structure is present. Our proposed approach was implemented in a package called GWANN which exhibited solid performance. Overall, GWANN outperformed popular GWAS tools on both simulated and real datasets, and enabled the identification of association signals with increased sensitivity and speed. Availability and implementationThe package is available at: https://github.com/hubner-lab/GWANN