Graph Embedding and Node Features for Drug-Target Interaction Prediction

Abstract

One of the most important tasks in the drug discovery process is to identify possible drug-target interactions (DTIs). The research and development of novel drugs frequently consumes billions of dollars and more than a decade of effort with high failure frequency. As a result, it is critical for pharmaceutical companies to identify novel drug–target interactions (DTIs) by leveraging known DTIs. Existing drugs have well-known qualities and are confirmed to be safe. However, biochemical tests for identifying novel DTIs have limitations in terms of coverage and throughput. As a result, computer approaches for the prediction of DTIs have garnered considerable interest. Computational prediction of DTI has been a prominent topic in the bioinformatics sector for the past decade, and it has substantially sped up drug development. Existing methods can be divided into two categories: network-based and classification-based methods. This thesis focuses on the network-based category that employs methods from the graph representation learning field. We propose two different models fulfilling the tasks of the drug-target binding affinity (DTA) prediction (regression) and the drug target interaction prediction (classification), utilizing graph embeddings and graph convolutions. Under these specifications, a data collection process is described for retrieving drug and protein features from public biological databases. Specifically, we extract SMILES of drugs and amino acid sequences of proteins in FASTA form. Then using several programming tools, we transform the drugs to graphs (where nodes correspond to atoms and edges to atom bonds), and the proteins to vectors using Word2Vec. The DTA approach leverages graph convolutions and 1D convolutions to transform drug and target features respectively, and concatenates the two outputs to consequently predict the binding affinity. In the DTI approach, drug features are extracted using graph convolutions, and then both drugs and targets are used to form a heterogeneous graph. This graph is then transformed by a graph auto-encoder, which generates the predicted interactions. Finally, we thoroughly present the models and their results, and compare them with recent state-of-the-art methods, which demonstrates the effectiveness of our approaches.