Using the full model with the genome and transcriptome data provided, we demonstrated that the extra layers of structure in the DSPN allowed us to achieve substantially better trait prediction than traditional additive models (Fig. 7C). For instance, a logistic predictor was able to gain a 2.4-fold improvement when including the transcriptome versus using the genome alone (+9.3% for the transcriptome versus +3.8% for the genome, above a 50% random baseline). By contrast, the DSPN was able to gain a larger, 6-fold improvement (+22.9% versus +3.8%), which may reflect its ability to incorporate nonlinear interactions. This result clearly manifests that the transcriptome carries additional information, which the DSPN is able to extract. Moreover, the DSPN allows us to perform joint inference and imputation of intermediate phenotypes (i.e., transcriptome and epigenome) and observed traits from just the genotype alone, achieving a ~3.1-fold improvement over a logistic predictor in this context (Fig. 7C and fig. S46). Overall, these results demonstrate the usefulness of even a limited amount of functional genomic information for unraveling gene-disease relationships and show that the structure learned from
