Transferable enantioselectivity models from sparse data.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Catalysis

Background:

• Optimizing enantioselectivity in new reactions is difficult, especially with limited data on catalyst-substrate interactions.
• Existing statistical models struggle with mechanistically complex transformations and sparse datasets.

Purpose of the Study:

• To develop a novel descriptor generation strategy for predicting catalyst performance in enantioselective reactions.
• To enable modeling of reactions with diverse ligand and substrate types, addressing data scarcity.

Main Methods:

• Generated descriptors accounting for changes in the enantiodetermining step based on catalyst/substrate identity.
• Collected data on enantioselective nickel-catalyzed C(sp3)-couplings.
• Trained statistical models using features from proposed transition states and intermediates.

Main Results:

• Developed models applicable to unseen ligands and reaction partners.
• Successfully optimized poorly performing examples within a substrate scope.
• Demonstrated a strategy for quantitatively transferring knowledge from sparse data to new chemical spaces.

Conclusions:

• The new descriptor strategy effectively models complex enantioselective reactions.
• This approach streamlines catalyst and reaction development by enabling prediction across diverse chemical spaces.
• Facilitates knowledge transfer from limited data to novel applications in asymmetric catalysis.