Machine learning driven antibody-antigen complex modelling and its applications to antibody therapeutics design

抗体是最重要的药物之一，已有100多种抗体疗法被批准用于治疗各种疾病，还有更多的抗体疗法正在积极开发中。抗体疗法的发展是一个时间和成本密集型的过程，使其成为使用计算和ML驱动的筛选和设计技术加速的有希望的目标。预测抗体与抗原靶标的复合体(也称为抗体-抗原对接)的结构仍然是一个具有挑战性的问题，近年来几乎没有重大进展。与蛋白质结构预测的许多其他领域一样，分离抗体结构的建模自2020年以来一直受到现代机器学习模型的应用，在许多情况下受到AlphaFold2成功的启发。AlphaFold-MULTIMER在一般蛋白质-蛋白质对接方面显示了良好的结果，但由于抗体结合区的高度多样性，在抗体-抗原对接方面表现不佳。该项目旨在利用扩散生成模型和等变图神经网络结构的最新进展，开发一个用于预测抗体-抗原相互作用结构的框架，这些模型和结构已被证明能够预测蛋白质的阶跃变化，特别是抗体结构预测。近年来，扩散模型极大地提高了人工智能生成的图像质量，并已成功地应用于包括蛋白质工程在内的生物任务。目前最先进的小分子蛋白质对接是DiffDock，一种扩散模型。DiffDock-PP是执行刚体蛋白质-蛋白质对接的改编版本。在我的轮换项目中，我研究了DiffDock-PP在增加抗体结构数据池中进行训练时执行抗体-抗原对接的能力。在我的DPhil期间，我将开发一个新的基于扩散的蛋白质-蛋白质对接模型，具体目标是对接抗体。这将包括改变空间表示法、更有效地使用内存、注意力模型，以及结合类似物理学的术语来指导反向扩散过程。该模型将在实验得出的对接3D结构上进行测试，如果成功，还将在计算预测的结构上进行测试。为通过计算预测的3D结构创建可靠的抗体-抗原对接模型将使研究人员能够纯粹根据测序数据了解抗体如何相互作用。这将大大加快治疗性抗体工程的步伐，并提高我们对未来大流行期间免疫逃逸的理解。该项目属于EPSRC的“人工智能技术”、“生物信息学研究”和“计算和理论化学”等研究领域。这项研究将在牛津大学统计系的夏洛特·迪恩教授和埃斯卡里亚大学的康斯坦丁·施奈德博士的监督下进行。

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. The development of antibody therapeutics is a time- and cost-intensive process, making it a promising target for acceleration using computational and ML-driven screening and design techniques. The prediction of the structure of antibodies in complex with antigen targets (also referred to as antibody-antigen docking) remains a challenging problem, with little significant progress published in recent years. The modelling of antibody structures in isolation, like many other areas of protein structure prediction, has been overhauled since 2020 by the application of modern machine learning models, inspired in many cases by the success of AlphaFold2. AlphaFold-multimer has shown promising results for general protein-protein docking but performs poorly on antibody-antigen docking due to the high diversity of the antibody binding regions. The proposed project would aim to develop a framework for the prediction of antibody-antigen interaction structures drawing on the recent advances in diffusion generative models and equivariant graph neural network architectures which have proven capable of step changes in protein and specifically antibody structure prediction. In recent years, diffusion models have drastically improved AI-generated image quality and have been successfully applied to biological tasks including protein engineering. The current state-of-the art for small molecule protein docking is DiffDock, a diffusion model. DiffDock-PP is an adapted version which performs rigid body protein-protein docking. During my rotation project I investigated the ability of DiffDock-PP to perform antibody-antigen docking when trained on an increased pool of antibody structural data. During my DPhil I will develop a new diffusion-based model for protein-protein docking, with the specific aim of docking antibodies. This will include changes to the spatial representations, more efficient memory usage, an attention model, and incorporation of physics-like terms to guide the reverse diffusion process. The model will be tested on experimentally derived docked 3D structures and, if successful, on computationally predicted structures. The creation of a reliable antibody-antigen docking model for computationally predicted 3D structures would allow researchers to understand how antibodies interact purely from sequencing data. This would greatly improve the pace of therapeutic antibody engineering as well as improve our understanding of immune escape during future pandemics. This project falls within the EPSRC research areas of "artificial intelligence technologies", "biological informatics research", and "computational and theoretical chemistry". It will be performed under the supervision of Professor Charlotte Deane from the Statistics department at the University of Oxford as well as Dr Constantin Schneider from Exscientia.