Third, we developed a comprehensive deep-learning model, the DSPN, and used it to illustrate how functional genomics data could improve the link between genotype and phenotype. In particular, by integrating regulatory-network connectivity and latent factors, the DSPN improves trait prediction over traditional additive models. Moreover, it takes into account dependencies between gene expression levels not modeled by univariate eQTL methods. In this study, we kept our eQTL methods very standard, closely following the GTEx paradigm. This separation we make between univariate eQTL detection and multivariate integrative modeling allows us to compare our eQTLs directly with those from previous analyses, such as the CMC study. However, multivariate-based methods for QTLs have been used elsewhere and, in the future, may be combined with our approach (54, 55).
