The full interaction between genotype and phenotype involves many levels, beyond those encapsulated by the regulatory network. We addressed this by embedding our regulatory network into a larger multilevel model. In particular, we developed an interpretable deep-learning framework, the Deep Structured Phenotype Network (DSPN) (21). This model combines a Deep Boltzmann Machine architecture with conditional and lateral connections derived from the regulatory network (50). Traditional classification methods such as logistic regression predict phenotype directly from genotype, without using intermediates such as the transcriptome (Fig. 7A). In contrast, the DSPN is constructed via a series of intermediate models that add layers of structure. We included layers for intermediate molecular phenotypes associated with specific genes (i.e., their gene expression and chromatin state) and predefined gene groupings (cell-type marker genes and coexpression modules), multiple higher layers for inferred groupings (hidden nodes), and a top layer for observed traits (psychiatric disorders and other brain phenotypes). Finally, we used sparse inter- and intralevel connectivity to integrate our knowledge of QTLs, regulatory networks, and coexpression modules from the sections above (Fig. 7B). By using a generative
