Multi-objective de novo protein design

从头开始的蛋白质设计包括产生折叠成所需结构的蛋白质序列。这一努力涵盖了医疗保健和精细化工行业的众多应用：制定和优化结合药物靶标的蛋白质疗法，产生可引发针对特定抗原的免疫反应的人造蛋白质，开发工业应用的酶催化剂，或构建能够更有效地输送药物的基于蛋白质的纳米结构。虽然蛋白质设计领域在过去十年中取得了重大进展，但它面临的一个突出挑战是对产生的设计的物理化学性质的有限控制，这总是导致设计-制造-测试实验周期的多次迭代。这个博士项目的目标是通过开发深度学习算法来设计具有特定物理化学性质的蛋白质来填补这一知识空白。该项目属于EPSRC数字医疗保健研究领域，更广泛地属于人工智能研究主题。该项目包括三个独立的工作包。-蛋白质性质优化算法。第一个工作包将夏季轮换开发的物业预测模型扩展为物业工程平台。性质预测算法将从热力学稳定性扩展到其他性质，如溶解度和翻译后修饰。候选人将探索多目标优化算法，以实现对蛋白质物理化学性质的有针对性的改进。该包包括跨不同深度学习体系结构的实验，从最初的图形卷积神经网络组合到包含神经注意力的其他体系结构。-从头开始蛋白质设计。该项目的下一阶段将解决同时为蛋白质生成一个新的、物理上可信的三维结构以及编码这些特征的氨基酸序列的问题。候选人将从目前在扩散模型和幻觉网络方面的现有文献开始开发生成模型，并以预测的物理化学特性(如稳定性和溶解性)为条件，以确保生成的设计对于蛋白质生产设施的开发是最佳的。这位候选人和她的主管目前正在与布里斯托尔大学的伍尔夫森实验室进行谈判，后者热衷于提供数据并通过实验验证生成的设计。然而，该项目的成功并不依赖于这种合作。-多目标蛋白质设计。该项目的最后部分将侧重于多目标蛋白质设计，即设计满足各种条件的蛋白质序列。候选人将结合她对蛋白质性质预测的理解和从头开始的蛋白质设计来产生人造蛋白质，如疫苗和抗体，这些人造蛋白质在免疫原性、热稳定性、毒性、半衰期等多个性质上都能通过性能分析。该项目的这一部分将整合候选人在前几年开发的机器学习架构，以及牛津蛋白质信息学集团生态系统内生成的其他模型。

De novo protein design involves generating protein sequences that fold to a desired structure. This endeavour encompasses numerous applications across the healthcare and the fine chemistry industries: formulating and optimising protein therapeutics that bind drug targets, generating artificial proteins that elicit immune response against a specific antigen, developing enzyme catalysts for industrial applications, or constructing protein-based nanostructures that can deliver drugs more efficiently. While the field of protein design has experienced significant progress in the past decade, one salient challenge it has faced is the limited control over the physicochemical properties of the generated designs, which invariably leads to multiple iterations of the design-make-test experimental cycle. The objective of this doctoral project is to fill this knowledge gap by developing deep learning algorithms to design proteins with specific physicochemical properties. This project falls within the EPSRC digital healthcare research area, and more generally under the artificial and intelligence research theme.The project comprises three independent work packages.- Protein property optimization algorithms. The first work package will expand the property prediction models developed in the summer rotation into a property engineering platform. The property prediction algorithms will be extended from thermodynamic stability into other properties such as solubility and post-translational modifications. The candidate will explore multi-objective optimization algorithms to enable targeted improvement of the physicochemical properties of a protein. This package includes experimentation across different deep learning architectures, starting from the initial combination of graph convolutional neural networks to other architectures incorporating neural attention.- De novo protein design. The next stage of the project will address the problem of simultaneously generating a new, physically plausible three-dimensional structure for a protein, and an amino acid sequence that encodes these features. The candidate will develop generative models, starting from the currently existing literature in diffusion models and hallucination networks, and condition them on predicted physicochemical properties such as stability and solubility to ensure that the generated designs are optimal for development at a protein production facility. The candidate and her supervisors are currently in conversations with the Woolfson Lab in the University of Bristol who are keen to provide data and validate generated designs experimentally. However, the project's success is not dependent on this collaboration.- Multi-objective protein design. The last part of the project will focus on multi-objective protein design i.e. designing protein sequences that fulfil a variety of conditions. The candidate will combine her understanding of protein property prediction and de novo protein design to generate artificial proteins, like vaccines and antibodies, that pass a performance profile on multiple properties like immunogenicity, thermostability, toxicity, half-life etc. This part of the project will integrate the machine learning architectures developed by the candidate in previous years, as well as other models generated within the ecosystem of the Oxford Protein Informatics Group.