Predictive modeling and inference using deep transfer learning in genetic data analysis

Transfer learning has been widely applied in text and image classification, demonstrating its effectiveness in numerous applications. In this paper, we propose a transfer learning procedure for both prediction and association testing between genotypes and phenotypes in a smaller primary data set with an available larger source data set. Specifically, we training a deep neural network model in the source data, transfer a part of the trained weight parameters to the model in the primary data, and complete the training process in the primary data with the remaining free parameters. Furthermore, we develop a permutation-based association test using the trained transfer learning model to identify significant genes in the primary data set. We apply the proposed procedure to two case studies for the investigation of nicotine dependence. These two case studies show that transfer learning can not only improve prediction accuracy but also the power of detecting candidate genes compared to those results without transfer learning.