Deep learning algorithms for predicting association between antibody sequence, structure, and antibody properties

Thesis

Antibodies are one of the most important classes of pharmaceuticals, with over 100 antibody therapeutics approved against a wide variety of diseases and many more in active development. However, the development of antibody therapeutics remains a time- and cost-intensive process. While computational screening methods have been used to improve the efficiency of this process, they are limited by a lack of accuracy, lack of generalisability or low-throughput nature. In this thesis, I detail the development of deep learning approaches for the prediction of antibody-antigen binding, harnessing state-of-the-art approaches for the analysis of protein structures.

To this end, I start by outlining improvements to the Structural Antibody Database (SAbDab). I then describe several high-throughput deep learning tools for a structure-based antibody virtual screening pipeline using computationally generated antibody structures (DLAB). I demonstrate that convolutional neural network (CNN) models can improve the prediction of antibody-antigen complex structures through the development of the pose rescoring and docking quality assessment tool DLAB-Re. Further, by developing and evaluating the virtual screening tool DLAB-VS, I show that CNNs enable structure-based antibody-antigen binding prediction. Motivated by recent advances in protein structure analysis and equivariant graph neural network (GNN) models, I extend the DLAB framework to GNNs, developing DLAB-EG and showing that these models can offer significant advantages over existing CNN approaches.

This thesis demonstrates the applicability of state-of-the-art machine learning approaches to the antibody-antigen virtual screening task. I provide a proof-of-principle for a structure-based virtual screening pipeline for antibody model libraries as well as a promising avenue for further improvement of antibody screening capabilities using machine learning approaches.

Authors: Schneider, CC

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Structural biology; Machine learning; Immunoinformatics
• Subjects: Deep learning (Machine learning); Immunoinformatics