A key aspect of the DSPN is its interpretability. In particular, we examined the specific connections learned by the DSPN between intermediate and high-level phenotypes. Here, we included coexpression modules in the model, referring to this modification as “DSPN-mod” (fig. S48). Using it, we determined which modules were prioritized, as well as the sets of genes associated with latent nodes that were found at each hidden layer (Fig. 8A and table S10) (15, 21). Broadly, we take an unbiased view of all 5024 modules and higher-order groupings constructed from these and then prioritize a subset of ~180 modules and groupings for each psychiatric disorder, showing these to be enriched in specific functional categories and to intersect substantially with the modules from more disease-focused analyses (Fig. 8, B and C, and fig. S49) (22). [For completeness, we provide a full table showing the prioritization and functional categories for all possible modules associated with various traits (fig. S50).] In particular, we found that cross-disorder prioritized modules are associated with functional categories such as “immune processes,” “synaptic activity,” and “splicing,” consistent with the
