Computational approaches to antibody library design and property prediction

Thesis

Over the past few decades, antibodies have established themselves as powerful therapeutics, used to treat viral infections, autoimmune diseases, and many cancers. However, designing and optimising such antibody drugs is still costly and time-consuming. In this thesis, we aim to demonstrate how computational methods- both old and new- can improve this process.

First, we introduce Paragraph, our graph-based paratope prediction tool that offers accurate residue-level predictions in just a tenth of a second. Paragraph can help guide computational docking experiments or focus researchers’ attention on where optimising mutations should be made. Next, we explore how large language models and other tools can design antibody libraries enriched in high-affinity variants from just a single known binding sequence. We also demonstrate that once a few hundred of these variants have been tested experimentally, we can train simple machine-learning methods on this data to help screen future variants and further enrich our antibody library. Finally, we introduce Humatch, our humanness classifier and humanisation tool. Humatch uses a combination of three highly accurate CNNs and germline data to offer rapid experimental-like humanisation and ensure final heavy-light designs remain well-matched.

All tools and methods presented in this thesis are easily accessible to other researchers. We hope that others may use and build upon these developments to continue to improve antibody therapeutic development.

Authors: Chinery, L

• DOI: 10.5287/ora-keoeekeqd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: affinity; AI; EGNN; paratope; antibody; optimisation; baselining the buzz; paragraph; machine learning; CNN; humanisation; drug discovery; trastuzumab; humatch
• Subjects: Structural bioinformatics; Artificial intelligence; Trastuzumab; Machine learning; Immunoinformatics; Bioinformatics