OrgAIcat: Leveraging Literature Data for Enantioselectivity Prediction and Rapid Optimization in Organocatalysis

Abstract

Published reaction literature is heterogeneous and biased, yet it remains the largest practical record of asymmetric catalysis. Here we ask whether carefully curated literature data can nevertheless support useful enantioselectivity prediction in organocatalysis. We present OrgAIcat, a machine-learning workflow built on iSynth, a manually curated dataset of over 22,000 Aldol reactions and Michael additions collected from 844 publications. Using R-SPOC, a chemically informed reaction representation that combines catalyst, transformation, and condition descriptors without quantum-mechanical calculations, OrgAIcat achieves R² values of 0.716 and 0.774 and mean absolute errors (MAE) of 0.282 and 0.265 kcal/mol for the Aldol reactions and Michael additions, respectively. External tests on recent literature and in-house reactions show that the model performs well in related organocatalytic systems. In a human-guided closed-loop workflow, OrgAIcat was used to prioritize catalysts for a challenging Aldol reaction, after which Bayesian optimization improved the enantiomeric excess (ee) from 41% to 96% in 12 experiments. These results indicate that curated literature data, despite their biases, can provide a practical basis for enantioselectivity prediction, catalyst prioritization, and rapid optimization in organocatalysis.