Neural networks applied to antibody structure determination

点击翻译按钮获取中文摘要

Antibodies are important proteins of the immune system. They recognise potentially harmful molecules, binding to them and initiating their removal from the body. Their ability to bind with high affinity and specificity to almost any antigen means they can be used as therapeutics - in fact they are the most successful class of biologics, with 93 approved for clinical use to date.Since it is the three-dimensional structure of the antibody that determines its binding properties, knowledge of this structure is very useful. However, experimental structure determination is low-throughput and therefore cannot be used routinely during therapeutic development. Computational modelling tools have hence become increasingly important, allowing researchers to predict large numbers of antibody structures which can then be used to infer and improve binding properties.This project will build on previous research into loop modelling carried out by the Oxford Protein Informatics Group (OPIG) and will explore the use of machine learning methods in the context of protein structure prediction. It will explore 3 potential areas of research involving computational antibody structure prediction. The first area involves using deep residual neural networks to assist in predicting the structure of challenging antibody segments. In particular the structure and the dynamics of CDRH3 loops.Other potential areas of research would involve the incorporation of multiple sequence information into methods for CDRH3 prediction, examination of the interplay of the CDR loops and improving the prediciton of sidechain conformations.This project falls within the EPSRC Synthetic Biology research area, but also contributes to the Artificial Intelligence Technologies and Synthetic Organic Chemistry. . It is a collaboration with F.Hoffmann-La Roche AG.