Machine Learning and Molecular Modelling: A Synergistic Approach to Rapid Reactivity Prediction

新型化学反应的计算设计被认为是计算有机化学和生物化学的“圣杯”之一。在药物设计、毒理学和化学合成中，准确、快速地预测化学反应的计算方法将为耗时的实验方法提供具有成本效益的替代方案，在某些情况下还可以替代动物试验。非常重要的是基于机制的预测模型，因为与计算性“黑箱”模型相比，它们更有可能得到普遍接受，因为计算性“黑箱”模型无法深入了解预测的方式和原因。提供这样的洞察力对于模型在毒理学和药物设计中获得监管接受尤为重要。然而，目前没有一种计算方法(分子建模或机器学习(ML))提供快速、准确的预测与清晰的机制洞察力的组合；为了追求其他特征，必须牺牲这些理想的一个或多个特征。该项目将开发一种新颖的、协同的分子建模和ML方法，用于快速、高精度和基于机理的反应性预测，用于毒理学、药物设计和化学合成，从而帮助实现“圣杯”。我们将在大数据集(约10,000种化合物)上训练和验证ML模型，这些模型可以纠正从快速分子建模技术获得的能量障碍，而不是从极其缓慢的高精度方法获得的能量障碍。因此，我们对反应建模的协同方法将是从这些快速分子建模技术中获得机理上的洞察力，并使用我们的ML模型来获得快速和准确的反应障碍。C-N键形成反应的模型将被开发用于共价药物设计(靶向赖氨酸)、毒理学(预测突变和呼吸敏化)和药物合成规划。为了展示我们的协同方法的广泛效用，我们将使用它来合理化生物和综合相关系统的实验反应性数据，对于这些系统，目前的建模方法的使用将缓慢得令人望而却步。预测不需要超级计算机，即使在笔记本电脑上也是可能的，这将代表着反应模型的范式转变。

The computational design of new chemical reactions is regarded as one of the "Holy Grails" of computational organic chemistry and biochemistry. Accurate and fast computational approaches to predicting chemical reactivity would provide cost-effective alternatives to time-consuming experimental approaches, and in some cases animal testing, in drug design, toxicology and chemical synthesis. Of great importance are mechanism-based prediction models because they are much more likely to reach general acceptance compared to computational "black-box" models which offer no insight into how and why predictions are made. Providing such insight is especially important for models to gain regulatory acceptance in toxicology and drug design. However, no current computational approach (molecular modelling or machine learning (ML)) to reactivity prediction offers the combination of fast, accurate predictions with clear mechanistic insight; one or more of these desirable characteristics must be sacrificed in pursuit of the others. This project will develop a novel, synergistic molecular modelling and ML approach to rapid, high-accuracy and mechanism-based reactivity prediction for use in toxicology, drug design and chemical synthesis and thus help realise the "Holy Grail".We will train and validate ML models on large datasets (~10,000 compounds) that can correct energy barriers obtained from rapid molecular modelling techniques to those derived from prohibitively slow, high-accuracy methods. Our synergistic approach to reaction modelling will thus be to derive mechanistic insight from these rapid molecular modelling techniques and use our ML models to obtain fast and accurate reaction barriers. Models for C-N bond-forming reactions will be developed for use in covalent drug design (targeting lysine), toxicology (predicting mutagenicity and respiratory sensitisation) and pharmaceutical drug synthesis planning. To demonstrate the broad utility of our synergistic approach, we will use it to rationalise experimental reactivity data of biologically and synthetically relevant systems for which the use of current modelling approaches would be prohibitively slow. Rather than requiring a supercomputer, predictions will be possible even on a laptop which will represent a paradigm shift in reaction modelling.