Machine learning for in silico optimisation and design of therapeutic antibodies

Thesis

Antibodies are essential proteins in our immune systems, defending against foreign pathogens. Their unique ability to bind strongly and specifically to theoretically any target has also made them one of the most important classes of therapeutics. While target binding affinity lies at the heart of therapeutic antibody efficacy, a wide range of properties affecting safety and developability must be considered. Machine learning (ML) offers great promise to overcome the bottlenecks of laborious and trial-and-error experimental optimisation of these properties. In this thesis, I describe the development of ML models for in silico antibody optimisation.

I begin by detailing efforts to predict the effects of mutations on antibody-antigen binding affinity. Using experimental and synthetic data, and an equivariant graph neural network architecture, I demonstrate that there are currently orders of magnitude too little experimental data available for accurate, generalisable prediction. I also investigate the role of dataset diversity and suggest guidelines for robust ML model development and evaluation in this area.

In the next chapter, I explore the interpretability of graph neural network affinity predictions by examining the weighting of interface components. Overall, current methods were unable to provide meaningful insights into the factors most important for model predictions.

As antibody development requires solving a complex, multi-objective optimisation problem beyond affinity, I have also used ML to investigate additional properties. I outline our Random Forest-based approach, trained on millions of sequences, which can distinguish human from non-human antibodies with near-perfect accuracy. These models form the basis of Hu-mAb, our antibody humanization tool.

I then describe our fine-tuning strategy to produce an antibody inverse folding model. AntiFold can guide antibody optimisation by identifying mutations that are predicted to maintain the structure and, therefore, structure-related properties of an antibody.

In my DPhil, I have evaluated the applications and limitations of ML to accelerate multiple steps in the antibody design pipeline. These contributions set the foundation for simultaneous multi-objective optimisation, as well as biasing antibody design towards favourable properties.

Authors: Hummer, AM

• DOI: 10.5287/ora-qymdmgjgo
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; binding affinity; protein design; deep learning; antibody design
• Subjects: Computational biology; Protein engineering; Machine learning