Inductive inference of novel protein-molecule interactions using Heterogeneous Graph Transformer (HGT) AutoEncoder

Abstract

1AbstractProtein-molecule interactions are promoted by the physicochemical characteristics of the actors involved, but structural information alone does not capture expression patterns, localization and pharmacokinetics. In this work we propose an integrative strategy for protein-molecule interaction discovery that combines different layers of information through the use of convolutional operators on graph, and frame the problem as missing link prediction task on an heterogeneous graph constituted by three node types: 1) molecules 2) proteins 3) diseases. Physicochemical information of the actors are encoded using shallow embedding techniques (SeqVec, Mol2Vec, Doc2Vec respectively) and are supplied as feature vectors to a Graph AutoEncoer (GAE) that uses a Heterogeneous Graph Transformer (HGT) in the encoder module. We show in this work that HGT Autoencoder can be used to accurately recapitulate the proteinmolecule interactions set and propose novel relationships in inductive settings that are grounded in biological and functional information extracted from the graph.