Congratulations to C-CAS members Simone Gallarati, Abby Doyle, and Matt Sigman, as well as Doyle lab member Erin Bucci, for the publication of their recent manuscript, "Transferable enantioselectivity models from sparse data," in Nature.
Project Background
When optimizing a new asymmetric reaction, machine learning models may be used to accelerate the identification of selective ligands, however the scarcity and sparsity of data is a key obstacle in training such models (e.g., only limited information on catalyst–substrate interactions is available). We have developed a computational strategy to generate features for ML models that are inexpensive yet highly descriptive of the reaction mechanism. This allowed us to train models on few, diverse reactions involving distinct ligand and substrate types, which have helped us optimize unseen transformations.
Project Outcome
Using our modeling workflow, we were able to train models on Ni-catalyzed C(sp3)-coupling reactions involving distinct ligand and substrate types. The models suggested that changes in the enantiodetermining step are likely occurring with different catalyst and substrate classes, a behavior that often limits the applicability of ML algorithms. We found these models to afford accurate predictions on a related, yet distinct reaction class involving unseen ligand and substrate types, highlighting how our workflow may offer a higher degree of transferability.
Impact and Benefits
The computational workflow, which we have made publicly accessible, is applicable to any reaction for which some mechanistic information is available. Our work demonstrated how this approach helps accelerate catalyst discovery and reaction optimization for new transformations by leveraging existing, yet sparse data on related reactions, decreasing the number of expensive and time-consuming experiments that must be carried out. We expect this approach to help establish other “transfer learning” strategies in asymmetric catalysis, further expanding the field of ML-assisted ligand design.