Development of machine learning tools to improve the developability of therapeutic antibodies.

抗体是一种流行的治疗形式，因为它们能够特异性地靶向任何分子。治疗性抗体已经成功地被开发出来，以针对包括癌症和传染病在内的各种疾病。然而，治疗性抗体的开发是一项复杂而昂贵的任务。即使一个候选者是一个强大和特定的结合体，它也可能会受到发展性问题的影响--例如高免疫原性、不稳定、自结合、高粘度、多特异性或表达能力差--这可能使其无法作为治疗性药物。因此，尽早识别和标记药物发现过程中的任何问题是很重要的。许多这样的问题是由于抗体的基于结构的生物物理性质引起的。目前存在能够计算这些属性并标记这些可开发性问题的工具，例如治疗性抗体分析器(TAP)。然而，目前还没有工具能够对这些标记的候选基因进行多目标结构优化，在保留抗原结合特性的同时消除可发展性问题。事实证明，开发这样一种工具在防止被标记为不可开发的高亲和力抗体从药物发现管道中丢弃方面将是非常宝贵的。这将节省资金，有助于使铅疗法多样化，并有可能打开新的靶点。该项目旨在创建这样一种工具，能够建议突变以克服抗体的可发育性问题，同时还保持其目标表位的结合亲和力。对于像这样的多优化问题，机器学习(ML)模型将是非常合适的。开发的ML模型将利用最先进的深度学习进展，如等变图神经网络和转换器。在抗体计算结构模型的准确性方面的进展(例如ABlooper)允许采取结构感知的方法。基于突变空间限制的规则也可以集成到模型中，以防止抗原结合亲和力的降低。该项目属于EPSRC化学生物学和生物化学主题。该项目位于牛津蛋白质信息学集团统计系，与阿斯利康英国公司合作，由夏洛特·迪恩教授进行学术监督，并由丽贝卡·克罗斯代尔-伍德博士进行工业监督。阿斯利康正在提供抗体合成和检测能力，允许对模型预测进行实验验证。

Antibodies are a popular therapeutic format due to their ability to specifically target any molecule. Therapeutic antibodies have been successfully developed to target a variety of diseases, including cancer and infectious diseases. However, therapeutic antibody development is a complex and expensive task. Even if a candidate is a strong and specific binder, it may suffer from developability issues - such as high immunogenicity, instability, self-association, high viscosity, polyspecificity, or poor expression - which may make it untenable as a therapeutic. It is therefore important to recognise and flag any issues as early as possible in the drug discovery pipeline. Many such issues arise due to structure-based biophysical properties of the antibody. Current tools exist which are able to calculate these properties and flag such developability issues, such as the Therapeutic Antibody Profiler (TAP). However, there is currently no tool capable of multi-objective structural optimisation of these flagged candidates, removing developability issues while retaining antigen binding properties. Development of such a tool would prove invaluable in preventing high affinity antibodies that are flagged as undevelopable from being discarded from the drug discovery pipeline. This would save money, help diversify lead therapeutics, and potentially unlocking new targets. This project aims to create such a tool, capable of suggesting mutations to overcome developability issues for an antibody, while also retaining binding affinity for its target epitope. For a multi-optimisation problem like this, a machine learning (ML) model would be well suited. The ML model developed will take advantage of state-of-the-art deep learning advances, such as equivariant graph neural networks and transformers. Advances in the accuracy of computational structural modelling of antibodies (e.g. ABlooper) allows for a structure-aware approach to be taken. Rules based limitations on the mutational space can also be integrated into the model to prevent a reduction in antigen binding affinity. The project falls within the EPSRC chemical biology and biological chemistry theme. The project is based at the Oxford Protein Informatics Group, Department of Statistics in collaboration with AstraZeneca UK, with academic supervision by Professor Charlotte Deane, and industrial supervision by Dr Rebecca Croassdale-Wood. AstraZeneca are providing antibody synthesis and assaying capabilities, allowing experimental validation of model predictions.