Science: successful machine learning challenges cross coupling reaction

Scientists have long been able to use computers to help explore the chemical world, discover new synthetic pathways and predict reaction results However, productivity prediction software often makes mistakes, because the data collected by many organizations over the years are often inconsistent and incomplete with the reality For example, reactions that don't work are usually not recorded Now we can accurately predict the yield of cross coupling reaction by a computer program The key of the algorithm is that the training data is extracted from thousands of small-scale reactions "This is just a small step, and the ultimate goal is to be able to predict the reaction performance of the new substrate without experiments." Abigail Doyle of Princeton University explained that she led the work with Spencer Dreher of Merck The team created a customized database of nearly 5000 Buchwald Hartwig coupling reactions - palladium catalyzed reactions that form a bond between carbon and nitrogen Isoxazole, a heterocycle known to inhibit cross coupling, was added to each reaction Although the difficulty is increased, the Princeton Merck team can use these data to train the algorithm and finally predict the yield correctly in a small error, close to the experimental results After inputting the data from 5000 cross coupling experiments into the program, the yield of such reactions can be predicted (source: Science) Since it takes months or even years for anthropologists to conduct 5000 experiments, Doyle and Dreher use Merck's high-throughput experimental platform to perform 1500 nanomolar reactions in one day Then the random forest algorithm is used to predict the results of 3000 reactions and the calculated parameters of each reagent, such as homo and LUMO energy Forest algorithm learns by building decision tree For each problem, the program adds a new branch The output is the average of thousands of decision trees In order to test the prediction accuracy of the algorithm, the team conducted 230 experiments Although the predicted results are not completely consistent with the actual experimental results, the error only fluctuates in a small range However, it is still challenging to find out the reasoning process behind the algorithm prediction As the authors say, models can be difficult to explain Although the reasoning result of this "black box" method is quite gratifying, if we can't understand the principle behind it, the reliability of the prediction result of the algorithm will be greatly reduced The important information reflected in this paper is that if there are enough prediction factors, the prediction of reaction yield is feasible The team will continue to train their algorithms to handle more complex compounds In previous studies, the substrates are all planar structures, which will be replaced by three-dimensional structures in the future Such differences will bring additional challenges, and the application prospects will be broader Paper doi: 10.1126/science.aar5169 corresponding author: Abigail g Doyle Spencer D Dreher