Feeding Machine Learning Algorithms with Mechanistic Data to Predict Outcomes of Copper-Catalyzed Couplings

项目摘要 将芳基卤与氧和氮亲核试剂结合形成C-N和C-O的联合收割机交叉偶联反应 键是合成医学相关分子的重要工具。这些反应通常由以下物质催化： 过渡金属络合物，如钯或铜。铜催化的交叉偶联反应具有 与钯介导的对应物相比具有几个优点，但与铜相关的缺点 催化作用往往超过这些好处。大多数Cu催化的交叉偶联方法需要高负载的 催化剂和高温以及铜催化剂通常不能实现芳基氯的交叉偶联。 为了解决这些问题，增加铜催化剂对于C-N和C-O交叉偶联的活性的配体已经被发现。 已经发现，酰胺配体可以产生一些最具活性的催化剂。一系列 Ma的出版物表明，在某些情况下，这种催化剂可以反应10，000次，并且可以 交叉偶联芳基氯，尽管在高温（120 °C）和催化剂负载量（5-10摩尔%）下。然而，在这方面， 确定促进给定底物对交叉偶联的反应条件可能是困难的 己经使用双酰胺配体以高产率实现底物的不同组合的偶联。 在这里，我们建议使用机械研究，以及机器学习（ML），以促进 铜盐与N-二甲酰胺配体介导的C-O交叉偶联反应条件的确定 并促进改进的配体和方法的开发。我们假设， 理解可以用来构建改进的ML模型，可以使用100-1000点的数据集 从而有效地预测反应产率。一旦建立，能够预测产量的ML模型可以用于评估 计算机模拟配体生成用于芳基氯交叉偶联的催化剂的潜力，或预测 反应条件以实现偶联配偶体的新组合的高产率。 我们的研究策略如下。首先，我们将阐明C-O交叉偶联反应的机理 催化的铜盐与异戊酰胺配体，并确定如何配体结构影响反应 机制获得的机械见解将用于识别或开发机器学习的输入 模型（特征）。我们将使用高通量实验工具进行960个C-O偶联反应， 各种芳基卤化物、亲核试剂和N，N-二甲基酰胺配体。数据集将被用来比较我们的手- 选择的特征与通过ML算法通过其预测C-O耦合产率的能力选择的特征。的 将在三个不同的ML优化任务中进行比较。我们的机械研究将建立 首次研究了铜盐与草酰胺配体催化的C-O交叉偶联反应机理， 为研究其他C-O和C-N交叉偶联反应奠定了基础 系统我们的ML研究将建立方法，使合理的小数据集来预测反应产率。

PROJECT SUMMARY Cross-coupling reactions that combine aryl halides with oxygen and nitrogen nucleophiles to form C-N and C-O bonds are vital tools for the synthesis of medicinally-relevant molecules. These reactions are often catalyzed by complexes of transition metals, such as palladium or copper. Copper-catalyzed cross-coupling reactions have several advantages over their palladium-mediated counterparts, but the disadvantages associated with copper catalysis often outweigh these benefits. Most Cu-catalyzed cross-coupling methods require high loadings of catalyst and high temperatures, and copper catalysts are often unable to effect cross-coupling of aryl chlorides. To address these issues, ligands that increase activity of copper catalysts for C-N and C-O cross-couplings have been sought, and oxalamide ligands have been shown to generate some of the most active catalysts. A series of publications by Ma have shown that such catalysts can, in some cases, react with 10,000 turnovers, and can cross-couple aryl chlorides, albeit at high temperature (120 °C) and loading of catalyst (5-10 mol %). However, identifying reaction conditions to promote cross-coupling of a given pair of substrates can be difficult – 12 different oxalamide ligands have been used to achieve couplings of different combinations of substrates in high yield. Herein, we propose to use mechanistic research, together with machine learning (ML), to facilitate the identification of reaction conditions for C-O cross-coupling reactions mediated by Cu salts with oxalamide ligands and to facilitate the development of improved ligands and methods. We hypothesize that mechanistic understanding can be used to build improved ML models that can use data sets on the order of 100-1000 points to predict reaction yield effectively. Once built, an ML model capable of predicting yield can be used to evaluate in-silico the potential of a ligand to generate a catalyst for the cross-coupling of aryl chlorides, or to predict reaction conditions to achieve high yield for a new combination of coupling partners. Our research strategy is as follows. First, we will elucidate the mechanism of C-O cross-coupling reactions catalyzed by copper salts with oxalamide ligands, and determine how ligand structure influences the reaction mechanism. The mechanistic insights gained will be used to identify or develop input for a machine learning model (features). We will use high-throughput experimentation tools to carry out 960 C-O coupling reactions with a variety of aryl halides, nucleophiles, and oxalamide ligands. The data set will be used to compare our hand- selected features with features selected by a ML algorithm by their ability to predict C-O coupling yield. The comparison will be made across three different ML optimization tasks. Our mechanistic studies will establish for the first time the mechanism of a C-O cross-coupling reaction catalyzed by Cu salts with oxalamide ligands, laying the foundation for the investigation of other C-O and C-N cross-coupling reactions using the same catalyst system. Our ML studies will establish methods that enable reasonably small datasets to predict reaction yield.