Joint Learning of Node Semantics and Graph Topology using a Transformer in the sparse network regime

Abstract

The human interactome is a valuable tool for unraveling disease mechanisms, advancing precision medicine, facilitating drug discovery, and identifying biomarkers. Yet, current interactomes are incomplete, in part due to limited experimental coverage. Therefore, augmenting the human interactome by predicting missing links in the Protein-Protein interaction network (PPI), is a core challenge for precision medicine. This study proposes an end-to-end trainable transformer-based neural network for enhanced aggregation of Gene Ontology (GO) terms features. We augment the models predictive capabilities by incorporating semantic anc2vec features, complementing the structural node2vec embeddings specifically designed for sparse PPIs. As a result, by integrating semantic and graph features, we demonstrate superior performance in link prediction.