Machine Learning to improve the outputs of an antibody synthetic library

点击翻译按钮获取中文摘要

Antibodies are essential proteins of the adaptive immune system that bind to their target proteins, called antigens, with great specificity and affinity. Antibodies are one of the most important classes of pharmaceuticals with over 100 antibody therapeutics approved. However, the majority of therapeutic antibody candidates fail before regulatory approval and the antibodies that do eventually progress cost more than $2bn to develop and can typically take around 10 years to bring to market. This time-consuming and cost-intensive process of developing these therapeutics would benefit from computational and machine learning-driven methods for predicting antibody properties. This offers an immense promise for the successful development of next-generation biologics. This project will focus on machine learning techniques to select antibodies with the best biophysical properties. The project will be performed alongside experts in the field of applying machine learning to antibody development at the Oxford Protein Informatics Group of Professor Charlotte Deane and in collaboration with Fusion Antibodies. Fusion Antibodies have been experts in the antibody space for more than 20 years. Besides from harnessing this expertise, Fusion Antibodies will provide a wealth of valuable data linking antibody sequences to expression and various biophysical properties. This data will be used to design and create machine learning protocols and will be beneficial for this project as these types of datasets are rarely available to the public. Furthermore, experimental validation of algorithms could be performed by the company. This project will initially focus on the biophysical property immunogenicity and expression. Animal models are often used to derive therapeutic antibodies potentially leading to an immune response in humans. Humanising is the process of making to antibody more human-like while maintaining antibody expression levels and its ability to bind their target epitope with high affinity. To humanise the antibody the complementarity-determining region, the antibody variable domain involved in binding the target, of the animal derived antibody is grafted into a human framework. Fusion Antibodies has performed this task, and applied structure- and expertise-driven back mutations, for a set of targets. In order to generate a predictive algorithm of this antibody property of interest based on this dataset the value of the dataset should be evaluated. This should provide insights into volume, diversity, and the type of data needed to generalise predictions. This project will contribute to maximizing efficacy in a shorter timeframe and at a reduced cost by guiding antibody therapeutic development towards antibodies with better biophysical properties. Machine learning and artificial intelligence technologies would be used to create predictive algorithms for biophysical properties. The project is interdisciplinary and involves immunoinformatics, machine learning, (structure-based) antibody design, as well as experimental validations. Therefore, this project falls within the EPSRC research areas: Chemical biology and biological chemistry, Synthetic biology, Biological informatics, and Artificial intelligence technologies.