Fig. 7. DSPN deep-learning model links genetic variation to psychiatric disorders and other traits.

(A) The schematic outlines the structure of the following models: logistic regression (LR), conditional Restricted Boltzmann Machine (cRBM), conditional Deep Boltzmann Machine (cDBM), and DSPN. Nodes are partitioned into four layers (L0 to L3) and colored according to their status as visible, visible or imputed (depending on whether nodes were observed or not at test time), or hidden. (B) DSPN structure is shown in further detail, with the biological interpretation of layers L0, L1, and L3 highlighted. The GRN structure learned previously (Fig. 5A) is embedded in layers L0 and L1, with different types of regulatory linkages and functional elements shown. Co-expr. mods., coexpression modules. (C) The performance of different models is summarized, with comparisons of performance across models of different complexity and of transcriptome versus genome predictors, corresponding to being with or without imputation for the DSPN (colors highlight relevant models for each comparison). Performance accuracy is shown first, with variance explained on the liability scale in brackets. All models were tested on identical data splits, which were balanced for predicted trait and covariates (including gender, ethnicity, age, and assay). RNA-seq, cell fraction, and H3K27ac data were binarized by thresholding at median values (per gene, cell type, and enhancer, respectively), as was age (median, 51 years) when predicted. LR-gene and LR-trans are logistic models using genetic and transcriptomic predictors, respectively; DSPN-impute and DSPN-full are models with imputed intermediate phenotypes (genotype predictors only) and fully observed intermediate phenotypes (transcriptome predictors), respectively. Differential performance is shown in terms of improvement above chance, with liability variance score increases in brackets. GEN, gender; ETH, ethnicity; AOD, age of individual at death.