Unsupervised representational learning of hierarchical graphs with graph convolutional networks.

Interpretation of functional genomic data attempts to correlate gene and protein expression with phenotype. While direct analysis of gene relationships associated with phenotype manifestation provides reasonable correlations, their use for direct assessment often fails to capture numerous complex biological phenomena. As a result, tools that cluster sets of genes based solely on set membership often fail to capture knowledge about contextual relationships of individual gene sets. To leverage large scale gene set relationships and gene set metadata to increase accuracy, we developed an approach that uses the network relationship of genes to extract system-level relationships. We employed an approach that segregates an empirically derived gene-gene network graph using deep learning to encode the network structure into low-dimensional embeddings. By applying graph neural network approaches we can generate lower dimensional embeddings that more accurately identify sets of genes related to biological traits of interest. Using well-adopted metadata over-representation techniques, we further demonstrate that our approach produces drastically different results when compared to direct set comparison methods, and more accurate results when subjected to manual analysis.