Advancing computational methods for the functional characterisation of antibodies using structural and flexibility data

Thesis

This doctoral thesis presents advancements in the computational characterisation of antibodies through the integration of structural and flexibility data. I first introduce SPACE2 (Chapter 2), a novel antibody epitope profiling method. In contrast to conventional approaches that establish functional equivalence based on sequence identity, I demonstrated that structural comparison enables the identification of a broader range of antibodies engaging the same epitope.

Findings from this initial project identified conformational diversity as a primary limitation in protein structure prediction. Consequently, the subsequent chapters focus on developing methods to accurately predict antibody structural flexibility at scale.

ITsFlexible, presented in Chapter 3, represents an initial approach to this problem. Rather than modelling structures of specific states, I employed a deep learning framework to predict whether or not an antibody CDR is likely to exchange between multiple conformations (exhibits structural flexibility).

Following on from this work, I focussed on methods to structurally characterise the antibody conformational ensemble. In Chapter 4, I describe Calvados 3-Fv, a high-throughput molecular dynamics (MD) workflow. Calvados 3-Fv accurately reproduced CDR ensemble properties at a significantly reduced computational cost compared to all-atom simulations while some limitations regarding the exact conformational landscapes covered remain. Leveraging this compute efficiency, I generated and released FlAbDab and FTCRDab, two databases containing more than 150,000 simulated antibody and TCR structures. These resources provide a substantial foundation for studying CDR flexibility and training the next generation of conformation-aware structure prediction models.

Finally, I present ABB4-STEROIDS (Chapter 5), a generative protein structure prediction model designed to sample the equilibrium distribution of antibody conformations. A key advantage of this framework is that it allows the training on a combination of data sources. I showed that pre-training on large-scale coarse-grained MD data, followed by fine-tuning on a small set of all-atom trajectories, significantly mitigates the biases inherent in coarse-grained data.

The thesis concludes (Chapter 6) by synthesising these contributions and outlining future avenues for research in antibody and protein conformation modelling.

Authors: Spoendlin, FC

• DOI: 10.5287/ora-yrpkexd5j
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: molecular dynamics; conformational dynamics; antibody; structure prediction; machine learning
• Subjects: machine learning; structural biology; computational biology