Graph representation learning: algorithm development and its application to early drug discovery

Yue, Yang (2025). Graph representation learning: algorithm development and its application to early drug discovery. University of Birmingham. Ph.D.

Preview

Yue2025PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (28MB) | Preview

Abstract

Drug discovery is a time- and cost-consuming process because of the complex biological mechanisms of the diseases and interactions between the drug and biological system. In the early (candidate selection) stage of drug discovery, machine learning, especially the graph representation learning (GRL)-based approaches have drawn wide-range attention. This is due to their powerful capability to flexibly handle various molecular and biological interaction information at different data scales, enabling effective reduction of the search space of following wet experiment validations based on efficient candidate screenings. Focusing on one of the most popular GRL approaches graph neural networks (GNNs) and relevant techniques, this thesis provides a comprehensive examination of a series of specialized designs on multiple crucial early drug discovery tasks, with the following contributions:
• Based on the large-scale heterogeneous drug-disease-target (DDT) relationship graphs, the examination of feasibility of non-Euclidean hyperbolic representation learning for identifying candidate targets for given drug and disease types.
• Based on the large-scale drug-target interaction (DTI) graphs, the examination of feasibility of meta-path-based representation learning for drug-drug therapeutic synergy effect predictions with the help of adverse effect (AE) information.
• Based on the small-scale individual protein molecular graphs, the examination of feasibility of a pre-trained geometric equivariant framework for predicting the protein binding affinity change caused by amino acid (AA) mutations with the mixture of atom- and residue-scale molecular representation information.
• Based on the small-scale protein molecular graphs, the examination of feasibility of a more general and efficient geometric framework based on coarse grained (CG) molecular representation learning for broader candidate protein screening tasks.
In summary, this thesis contributes towards a greater understanding of GRL applications on early drug discovery, ranging from learning macro-biological interaction graphs to micro-molecular structure graphs, to enable further potential biological mechanism exploration and experimental validations.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):