Investigation of Transition-Metal Catalysed C-H Activation Using Quantum Chemistry and Machine Learning

直接官能化C-H键以形成新的C-C键（或C-N、C-O等）在过去的二十年里，特别是在药物发现方面，受到了极大的关注。这类广泛的反应是有吸引力的，因为它能够实现更短和更有效的合成，避免在进行它们的偶联之前使用卤化物/胺预官能化的分子。C-H活化反应通常需要使用过渡金属（TM）催化剂（例如，Pd，Pt）。该反应建议通过在C-H键处插入金属来进行，尽管替代机制是可能的。然而，即使使用TM催化剂，C-H活化也需要苛刻的反应条件，这是由于其高的键离解能，这降低了敏感官能团的耐受性。此外，通常存在差的区域选择性和立体选择性。该项目的目标是双重的，第一个是用量子化学方法自动生成这些反应的反应路径。第二个目标是通过机器学习（ML）方法预测给定反应物的反应产物。计算研究可以帮助阐明这些反应的机制，从而提高它们的反应性和选择性。以前的大部分工作都是通过人工探索势能面来寻找过渡态（TS），这很耗时，需要专家用户。这种分析的自动化将减少用于建模的资源，并增加其对合成开发的影响。这种类型的反应的自动化研究的努力，最近才被报道与模型系统。该项目旨在利用C-H活化反应的实用性。我们将探索使用半经验（SE）方法来获得TS猜测，然后使用基于ML的校正来增强该猜测，以获得接近DFT优化的TS结构的几何形状。这种SE-ML组合方法已被应用于溶剂化能、生成热等的预测，并且显示出比纯ML模型更快地收敛并且需要更少的训练数据。我们还将致力于扩展这种方法的通用性，以预测任何金属催化剂的C-H活化（即，Ru、Fe等）。第二个探索的途径将是使用ML方法来预测反应产物，特别是C-H活化反应的区域/立体选择性，给定的反应物。在这里，我们将使用IBM Rxn平台作为起点，该平台已被证明在第一次预测已知反应结果时达到90%的准确度。这种ML方法由Schwaller等人介绍，由神经网络语言转换器模型组成，该模型将反应物的SMILES字符串“转换”为产物SMILES。尽管可以使用ML模型来预测一般反应，但它们对C-H活化反应的有效性尚未得到广泛测试。我们打算研究基于语言转换器的ML模型是否可以预测C-H激活的区域选择性和立体选择性。我们还将实现基于3D结构的ML模型（例如，基于分子静电势）并测试它在预测选择性方面是否更成功。该项目属于EPSRC研究领域“计算和理论化学”的福尔斯。但它也涉及反应机理、药物化学、过渡金属催化、DFT和机器学习等领域。该项目符合EPSRC在该领域的目标，包括与制药部门和软件开发的合作。后者将有利于改进计算工作流程和研究与更广泛的科学背景相关的催化机制。我们很高兴有阿斯利康作为我们的工业合作伙伴。他们通过建议哪些调查路线可能更好地与制药行业的合成和药物发现目标保持一致来提供帮助。

Direct functionalisation of C-H bonds to form new C-C bonds (or C-N, C-O etc.) has received significant interest in the past two decades, particularly in drug discovery. This broad class of reactions is attractive since it enables shorter and more efficient syntheses, avoiding the use of halide/amine pre-functionalised molecules prior to performing their coupling. C-H activation reactions usually require the use of a transition metal (TM) catalyst (e.g., Pd, Pt). The reaction is suggested to proceed via metal insertion at the C-H bond, although alternative mechanisms are possible. However, even with TM catalysts, C-H activation requires harsh reaction conditions due to its high bond dissociation energy, which reduces tolerance of sensitive functional groups. Additionally, there is often poor regio- and stereo-selectivity. The aim of this project is twofold, the first is the automation of reaction path generation for these reactions with quantum chemical methods. The second goal is the prediction of reaction products for given reactants via machine learning (ML) approaches. Computational studies can help elucidate the mechanism of these reactions and thereby allow improving their reactivity and selectivity. Most of the previous work has been done via manual exploration of potential energy surfaces to find transition states (TS), which is time consuming and requires expert users. Automating such analyses would reduce the resources spent on modelling and increase their impact on synthetic development. Efforts to automate the study of this type of reactions are only recently being reported with model systems. This project aims to leverage the utility for C-H activation reactions. We will explore the use semi-empirical (SE) methods to obtain TS guesses and then augment that with ML-based correction to obtain a geometry close to the DFT-optimised TS structure. This combined SE-ML approach has been applied to the prediction of solvation energy, heats of formation etc., and shown to converge faster and require less training data than pure ML models. We will also aim to extend the generalisability of this approach to predict C-H activation with any metal catalyst (i.e., Ru, Fe etc.). The second avenue to explore will be the use of ML methods to predict the reaction products, especially the regio-/stereo-selectivity of C-H activation reaction, given the reactants. Here we will employ the IBM Rxn platform as a starting point, which has been shown to achieve 90% accuracy on the first prediction against the known reaction outcome. This ML approach, introduced by Schwaller et al, consists of a neural network language-transformer model which "translates" SMILES strings of reactants into the product SMILES. Even though ML models to predict general reactions are available, their efficacy for C-H activation reactions has not been extensively tested. We intend to investigate whether language-transformer based ML models can predict C-H activation regioselectivity and stereoselectivity. We will also implement 3D structure-based ML models (e.g., based on molecular electrostatic potential) and test whether it is more successful in predicting selectivity. This project falls within the EPSRC research area "computational and theoretical chemistry". But it also touches upon areas of reaction mechanism, medicinal chemistry, transition metal catalysis, DFT and machine learning. This project aligns with the goals of EPSRC in that area, including collaboration with pharmaceutical sector and software development. The latter will be beneficial to improving computational workflows and the study of catalytic mechanisms which are relevant in the broader scientific context. We are very happy to have AstraZeneca as our industrial collaborators. They assist by suggesting which routes of inquiry might align better with the synthetic and drug discovery goals of pharmaceutical industries.