AI-Driven Discovery\nof Asymmetric Pauson–Khand\nReactions: A New Toolbox in a Synthetic Chemist’s Treasure

Enantioselective\ncatalytic reactions have a significant impact\non chemical synthesis, and they are important components in an experimental\nchemist’s toolbox. However, development of asymmetric catalysts\noften relies on the chemical intuition and experience of a synthetic\nchemist, making the process both time-consuming and resource-intensive.\nThe machine-learning-assisted reaction discovery can serve as a very\nefficient platform for obtaining high-performing catalysts in a time-economical\nmanner without extensive experimentation. Herein, we report a data-driven\nand machine learning method for reliably predicting enantiomeric excess\n(�) of 211 asymmetric Pauson–Khand reactions (PKR 1–PKR\n211) between a variety of 45 unique 1,6-enyne substrates and 12 unique\naxially chiral biaryl ligands in the presence of different reaction\nconditions like varying CO gas pressure, temperature, and solvent\npolarity. Four different machine learning algorithms have been studied:\nextreme gradient boosting (XGBoost), random forest (RF), light gradient\nboosting machine (LGBM), and neural network (NN). A fivefold cross\nvalidation method was applied to our k-means SMOTE-augmented data\nset to obtain the optimized hyperparameters for the training set,\nand subsequently, these parameters were used in the test data set.\nIn the case of the out-of-box set, the XGBoost method is found to\nbe superior among all four machine learning methods investigated.\nOur out-of-box samples contain a total of 12 unique asymmetric Pauson–Khand\nreactions (PKR 212–PKR 223) arising from three new 1,3-benzodioxole-based\nSEGPHOS catalysts, which were never included in the training set.\nThe XGBoost algorithm shows an impressive root mean square error (RMSE)\nof 7.06 (±1.11) in predicting �. The XGBoost-predicted �\nvalues match reasonably well with the experimental results. The absolute\ndifference between the experimental and XGBoost-calculated � values\nranges from 0.9 to 7.6 for the majority of the out-of-box Pauson–Khand\nreactions. The reactions with fluoro-substituted-SEGPHOS ligand <b>L14</b> shows smaller deviations from the experimental � values\ncompared to the reactions with <b>L13</b> and <b>L15</b> catalysts where the benzodioxole units do not have fluorine atoms.\nFinally, we have discovered a library of 3357 lead reactions with\nexcellent � (≥99) by engaging the experimentally unknown\ncombinations of the catalysts, substrates, and reaction conditions.\nThe axially chiral biaryl catalysts and enyne substrates present in\nthe library are synthetically accessible. The ligand space in the\nlibrary is dominated by the presence of tol-BINAP and the DTBM-OMe-BIPHEP\nligands. The substrate space is predominantly occupied by NTs-tethered,\nO-tethered, NBn-tethered, and C(CO<sub>2</sub>Me)<sub>2</sub>-tethered\n1,6-enynes that have an H or methyl functional group present in the\nalkyne unit. Our newly discovered library assists a synthetic chemist\nto develop a highly enantioselective PKR by starting with <i>a priori</i> knowledge without extensive trial-and-error experimentation.