Developing machine learning methods using antibody structural and sequence data to accelerate vaccine design

该项目旨在使用机器学习（ML）技术来提高疫苗设计的速度并降低成本。该项目是牛津蛋白质信息组（OPIG）和GSK疫苗之间的合作。这项工作的目标属于EPSRC分析科学和数学生物学研究领域。目前，疫苗和抗体疗法通常需要5-10年时间和大约10亿英镑才能推向市场。临床前试验占开发时间的很大一部分，因为研究人员的目标是在人类志愿者中评估疫苗之前将风险最小化并将收益最大化。这个临床前试验阶段可以通过使用计算方法来更好地选择候选疫苗进入实验室进行测试来显着加快。此外，在这一早期阶段改进疫苗选择可以导致最终生产出更高功效的产品。高通量测序技术和精确的蛋白质结构建模工具的发展提供了大量的数据来寻找有前途的抗体线索。然而，搜索这个空间仍然是一个挑战，因为由于所涉及的巨大计算复杂性，还不可能精确地建模抗体-抗原相互作用。机器学习技术也很难准确地搜索这个空间，因为目前只有有限数量的标记数据可用于训练。这种标记的数据主要由使用X射线晶体学成像的抗体-抗原复合物组成-这是一种昂贵且耗时的技术。这些成本意味着与现在可用的数十亿个抗体序列相比，只有几千个复合物的结构数据存在。该项目旨在最大限度地利用可用于训练深度神经网络的结构和序列数据，以改善我们对抗体和抗原如何相互作用的预测。然后，这些预测方法将被开发成强大的开源软件工具，这些工具将成为SAbPred-OPIG抗体预测工具箱的一部分。这项工作将区分和改进现有的技术，通过使用物理上重要的特征来标记数据，结合从最先进的Transformer模型中获得的描述性特征嵌入。

This DPhil project aims to use machine learning (ML) techniques to increase the speed and reduce the cost of vaccine design. The project is a collaboration between the Oxford Protein Informatics Group (OPIG) and GSK Vaccines. The goals of this work fall within the EPSRC Analytical Science and Mathematical Biology research areas. Currently, vaccines and antibody therapeutics typically take 5-10 years and approximately £1bn to bring to market. Pre-clinical trials comprise a large proportion of this development time as researchers aim to minimise risks and maximise benefits before assessing vaccines in human volunteers. This pre-clinical trial stage can be sped up significantly by using computational methods to better select vaccine candidates to take into the lab for testing. Furthermore, improved vaccine selection at this early stage can result in higher efficacy products being produced at the end. The development of high-throughput sequencing techniques and accurate protein structure modelling tools have given access to large amounts of data in which to search for promising antibody leads. However, searching this space is still a challenge as it is not yet possible to exactly model antibody-antigen interactions due to the great computational complexity involved. Machine learning techniques also struggle to accurately search this space as only a limited amount of labelled data for training is currently available. This labelled data is largely comprised of antibody-antigen complexes imaged using X-ray crystallography - an expensive and time-consuming technique. These costs mean structural data exists for only a few thousand of complexes compared to the billions of antibody sequences that are now available. This project aims to maximise the utility of the structural and sequence data that is available to train deep neural networks to improve our predictions of how antibodies and antigens interact. These predictive methods will then be developed into robust, open-source software tools that will form part of SAbPred - OPIG's antibody prediction toolbox. This work will differentiate and improve upon existing techniques by using physically important characteristics to label data combined with descriptive feature embeddings obtained from state-of-the-art transformer models.