UC Berkeley

We report machine-learning (ML) models that predict with high accuracy the enantioselectivity of ketone hydrogenation with Noyori-type diphosphine-diamine catalysts. To mitigate bias in published data, we combined a data set curated from the literature with >1,000 examples from our own high-throughput experimentation. We demonstrate that 3D graph neural networks, trained on 3D representations derived from the enantiodetermining transition state, outperform models based on 2D graphs, molecular fingerprints, and descriptors derived from semiempirical quantum calculations. These results indicate that models trained on 3D structural representations that reflect catalyst-substrate interactions relevant to enantioselectivity can accurately extrapolate to structures distinct from those in the training set.