Pharmaceutical drug discovery is expensive, time consuming and scientifically challenging. In order to increase efficiency of the pre-clinical drug discovery pathway, computational drug discovery methods and most recently, machine learning-based methods are increasingly used as powerful tools to aid early stage drug discovery.

In this thesis, I present three complementary computer-aided drug discovery methods, with a focus on aiding hit discovery and hit-to-lead optimization. In addition, this thesis particularly focuses on exploring different molecular representations used to featurise machine learning models, in order explore how best to capture valuable information about protein, ligands and 3D protein-ligand complexes to build more robust, more interpretable and more accurate machine learning models.

First, I developed ligand-based models using a Gaussian Process (GP) as an easy-to-implement tool to guide exploration of chemical space for the optimization of protein-ligand binding affinity. I explored different topological fingerprint and autoencoder representations for Bayesian optimisation (BO) and showed that BO is a powerful tool to help medicinal chemists to prioritise which new compounds to make for single-target as well as multi-target optimisation. The algorithm achieved high enrichment of top compounds for both single target and multiobjective optimisation when tested on a well known benchmark dataset of the drug target matrix metalloproteinase-12 and a real, ongoing drug optimisation dataset targeting four bacterial metallo-β-lactamases.

Next, I present the development of a knowledge-based approach to drug design, combining new protein-ligand interaction fingerprints with a fragment-based drug discovery approach to understand SARS-CoV-2 Mpro-substrate specificity and to design novel small molecule inhibitors in silico. In combination with a fragment-based drug discovery approach, I show how this knowledge-based interaction fingerprint-driven approach can reveal fruitful fragment-growth design strategies.

Lastly, I expand on the knowledge-based contact fingerprints to create a ligand-shaped molecular graph representation (Protein Ligand Interaction Graphs, PLIGs) to develop novel graph-based deep learning protein-ligand binding affinity scoring functions. PLIGs encode all intermolecular interactions in a protein-ligand complex within the node features of the graph and are therefore simple and fully interpretable. I explore a variety of Graph Neural Network architectures in combination with PLIGs and found Graph Attention Networks to perform slightly better than other GNN architectures, performing amongst the best known protein-ligand binding affinity scoring functions.