Embedding Knowledge Graphs to Predict Polypharmacy Side Effects

Abstract

Many biomedical datasets are inherently interconnected. Knowledge graphs, and embeddings thereof, have recently emerged as a powerful tool for modelling and analysing such datasets. This thesis explores how these techniques may best be applied to the problem of polypharmacy side effect (PSE) prediction. After an introductory examination of the associated history and current state of the field, the first piece of analysis takes a look at the hyperparameters of knowledge graph embedding models and the relationship that they have with the quality of the outputted vectors. We find that the relative importance of the hyperparameters differs greatly by dataset, which motivates our employment of large-scale grid searches when optimising embeddings for graphs in later chapters. In chapter 3, the problem of PSE prediction is described and tackled using tensor factorisation (TF) methods. In contrast to a lot of literature on the subject, we find that TF models can reach levels of performance equivalent to that achieved by much more expensive graph neural networks. Chapters 4 and 5 then seek to build on this finding by addressing the problem of predicting side effects for emerging drugs - a notable deficiency ofTF models. Results from chapter 4 indicate that a simple approach based on aggregation of statically embedded neighbours does work, but not with the same level of success as seen in the transductive case. In chapter 5 we address this by re-embedding the graph, employing a technique from recent research that enables such aggregation operations to be considered as part of the learning process, and measure the change in performance on the same task. The thesis is then brought together in a final discussion chapter which places these analyses within the wider context of the field. By looking towards at the particular strengths and limitations of the work, future directions for study in the field of PSE modelling are suggested.

Date of Award	4 Feb 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Tom R Gaunt (Supervisor), Yi Liu (Supervisor) & Patrick Rubin-Delanchy (Supervisor)