Improving Neural Networks for Genotype-Phenotype Prediction Using Published Summary Statistics

Phenotype prediction is a necessity in numerous applications in genetics. However, when the size of the individual-level data of the cohort of interest is small, statistical learning algorithms, from linear regression to neural networks, usually fail due to insufficient data. Fortunately, summary statistics from genome-wide association studies (GWAS) on other large cohorts are often publicly available. In this work, we propose a new regularization method, namely, main effect prior (MEP), for making use of GWAS summary statistics from external datasets. The main effect prior is generally applicable for machine learning algorithms, such as neural networks and linear regression. With simulation and real-world experiments, we show empirically that MEP improves the prediction performance on both homogeneous and heterogeneous datasets. Moreover, deep neural networks with MEP outperform standard baselines even when the training set is small.