Functional modeling of noncoding risk variants based on convolutional neural networks

Abstract

One of the key tasks for post-GWAS analyses is to identify causal noncoding variants with regulatory function. On the basis of >2,000 functional features, I developed a convolutional neural network framework for combinatorial, nonlinear modeling of complex patterns shared by risk variants scattered across multiple associated loci. I evaluated the performance and validity of this method in various ways in the context of multiple complex diseases, especially psychiatric disorders. A main advantage of the method is its applicability for prioritization of rare variants. Although trained with independent data, the model made positive predictions for candidate rare variants derived from multiplex autism families, including those mapped to CHD5 and FRRS1L. CHD5 knockout mice reportedly show abnormal social behavior. Behavioral experiments conducted in collaboration with Prof. Jin-Hee Han’s laboratory showed that FRRS1L, a component of the AMPA receptor, was specifically involved in recognition of social novelty. In conclusion, I propose a novel approach to discover biological patterns shared by disease-causing regulatory variants based on their regulatory function.