Analysis and Development of Graph-Embedding Techniques for Biomedical Knowledge Graphs

The drug discovery process is highly complex requiring knowledge from a variety of domains including human diseases, genetics, structural biological and medicinal chemistry. These factors have prompted the search for reliable virtual (i.e. computer-based) screening methods to select likely drug candidates. While machine learning methods are now well established for the prediction of biological activities of small molecules, they have generally been applied to predict binding to a single biological target and do not consider more complex relationships such as the potential for side effects. Knowledge graphs are a recent construct that provide a way of organising data from multiple heterogeneous sources and have recently begun to be used in drug discovery. This project will develop graph theory techniques for extracting information-rich feature vectors from knowledge graphs for use with machine learning methods. The aim will be to provide more accurate predictions of the physicochemical and biochemical properties of small molecules.