Creating an A.I. tool for chemical reaction design is challenging, especially for complex catalytic reactions. Researchers at the Institute for Chemical Reaction Design and Discovery and the Max Planck Institute für Kohlenforschung demonstrated a machine learning method using advanced, efficient 2D chemical descriptors that accurately predict highly selective asymmetric catalysts—without quantum chemical computations.
Molecules are usually represented as a collection of descriptors, often consisting of small parts or molecule fragments. These are easier for A.I. to process and can be arranged and rearranged to construct different molecules. But computationally cheaper 2D descriptors don’t accurately represent complex catalyst structures. Researchers developed Circular Substructure 2D descriptors to represent cyclic and branched hydrocarbon structures, common in catalysis.
A fully trained model virtually tested 190 catalysts not part of the training data. The A.I. model could predict highly selective catalysts after only having been trained on the data of catalysts with moderate selectivity, extrapolating beyond training data. This high selectivity is especially crucial for the design of new medicines.