A Biologically Interpretable Graph Convolutional Network to Link Genetic Risk Pathways and Imaging Phenotypes of Disease

We propose a novel deep neural network for whole-genome imaging-genetics. Our genetics module uses hierarchical graph convolution and pooling operations that mimic the organization of a well-established gene ontology to embed subjectlevel data into a latent space. The ontology implicitly tracks the convergence of genetic risk across biological pathways, and an attention mechanism automatically identifies the salient edges in our network. We couple the imaging and genetics data using an autoencoder and predictor, which couples the latent embeddings learned for each modality. The predictor uses these embeddings for disease diagnosis, while the decoder regularizes the model. For interpretability, we implement a Bayesian feature selection strategy to extract the discriminative biomarkers of each modality. We evaluate our framework on a population study of schizophrenia that includes two functional MRI (fMRI) paradigms and gene scores derived from Single Nucleotide Polymorphism (SNP) data. Using 10-fold cross-validation, we show that our model achieves better classification performance than the baselines. In an exploratory analysis, we further show that the biomarkers identified by our model are reproducible and closely associated with deficits in schizophrenia. Preprint. Under review. .