BioNet+: A Comprehensive Interaction Model of Finding Therapeutics for Diseases with Graph Deep Learning

The prediction of interactions between biomedical entities is crucial for the discovery of therapeutics for diseases. The fundamental step is to build an interaction model with a comprehensive coverage on literature and curated databases. The massive amount of heterogeneous data in interaction networks of biomedical entities poses significant challenges to relevant prediction tasks. Previous research has mainly focused on constructing networks with limited types of entity relations. In this paper, we construct a large-scale heterogeneous graph model, BioNet+, using a data-driven approach to extract latent information and predict various relations among biomedical entities. BioNet+ provides a crucial update to our previous BioNet model, with the addition of disease entities. The model employs a graph convolution encoder to learn embedding information from the intricate interactions among chemicals (C), genes (G), diseases (D), and biological pathways (P). Furthermore, a decoder that applies two tensor decomposition methods is deployed to predict multiple types of relations, including CG, CD, and GD interactions. The model constructs a graph network consisting of 81,987 entity nodes and 206,195,703 interaction edges. To handle this massive model, BioNet+ reduces the computational complexity via data categorization and proposes a parallel model training algorithm for cross-node GPUs. Evaluation experiments demonstrate that BioNet+ achieves superior predictive performance, as demonstrated by its highest area under the receiver operating characteristic curve (AUROC) of 0.964.