Using AI based modelling to drive the engineering of biology

本提案旨在应用人工智能方法来推动生物学工程。最终的愿景是创建一个自动驾驶实验室，人工智能可以选择下一个实验，解释数据并迭代实现给定的设计目标。这将能够以可持续的方式快速合理地设计微生物，用于生产化学品、化合物和食品（例如蛋白质），从而解决净零的需求。我们认为，真正利用人工智能的力量将促进（人工智能）对生物系统的学习，因此未来的设计和优化任务将建立在现有知识库的基础上。与将生物设计问题视为孤立的优化任务相比，这将是一个重大的进步，因为没有学习可以从一个任务推进到下一个任务。在这里，我们将实施由国际合作者开发的人工代谢网络（AMN）方法(1)。这在一个创新的机器学习（ML）框架中编码了大肠杆菌的机制代谢模型，该框架能够以传统的通量平衡分析（FBA）方法无法实现的方式进行学习，FBA是过去30年来代谢模型的支柱。我们寻求将我们的生物系统自动化构建实验平台(2)与人工智能指导方法集成在一起，以模拟工程大肠杆菌细胞代谢。经典FBA建模的一个关键限制是，它对工程系统（如基因敲除（KOs））的定量预测能力较差，因此在生物系统的合理正向工程中只有有限的效用。它在识别可能需要考虑的基因方面具有价值，但在基因编辑的定量预测方面存在不足。它也不兼容当前的深度学习技术。我们的合作者开发了AMN(1)，这是一种混合方法，它将FBA作为白盒机制层嵌入到一种能够训练黑盒中性网络（NN）输入层的形式。可训练性已经被证明提供了一种很好的方法来拟合来自遗传ko的数据，这些数据不能用经典的FBA充分建模，从而证明混合AMN方法可以成为预测工程系统行为的有力工具。

This proposal seeks to apply AI methodology to drive the engineering of biology. The ultimate vision is to create a self-driving lab, where AI can select the next experiment, interpret data and iterate towards a given design objective. This will enable rapid and rational engineering of microbes for the production of chemicals, compounds and food (e.g. protein) in a sustainable manner that addresses the needs of net zero.We consider that a true harnessing of the power of AI will facilitate the learning (by the AI) of the biological system, so that future design and optimisation tasks will build on the existing knowledge base. This will represent a significant advance compared to treating biodesign problems as isolated optimisation tasks where no learning can be carried forward from one task to the next. Here we will implement the Artificial Metabolic Network (AMN) methodology developed by our international collaborators (1). This encodes the mechanistic metabolic model of E. coli in an innovative Machine Learning (ML) framework that enables learning in way that is not possible with the classical approach to Flux Balance Analysis (FBA), which has been the mainstay of metabolic models for the past 3 decades.We seek to integrate our experimental platform for automated building of biological systems (2) with an AI guided approach to modelling engineered E. coli cell metabolism. A key limitation of classic FBA modelling is that it has poor quantitative predictors of engineered systems, such as gene knockouts (KOs) and thus has only limited utility in the rational forward engineering of biological systems. It has value in identifying likely genes to consider but falls short in quantitative predictions of gene edits. It is also not compatible with current deep learning technologies. Our collaborators have developed AMN (1), a hybrid method that embeds FBA as a white-box mechanistic layer in a form that enables training of a black-box Neutral Net (NN) input layer. The trainability has been demonstrated to provide an excellent approach to fit data derived from genetic KOs that could not be adequately modelled with classical FBA, thus demonstrating that the hybrid AMN approach could be a powerful tool in predicting the behaviour of engineered systems.