Chemistry-informed deep learning model for predicting stereoselectivity and absolute configuration in asymmetric...

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Catalysis

Background:

• Asymmetric hydrogenation of olefins is crucial for synthesizing chiral molecules.
• Existing machine learning models struggle with predicting stereoselectivity and absolute configurations for complex reactions, especially those with multiple prochiral sites.
• Current models often rely on predefined descriptors and have limited substrate-catalyst applicability.

Purpose of the Study:

• To develop a novel deep learning model, Chemistry-Informed Asymmetric Hydrogenation Network (ChemAHNet), for predicting outcomes in olefin asymmetric hydrogenation.
• To overcome the limitations of existing models in terms of substrate-catalyst scope and prediction accuracy for reactions with multiple prochiral sites.
• To enable concurrent prediction of stereoselectivity and absolute configuration.

Main Methods:

• Developed ChemAHNet, a deep learning model incorporating reaction mechanism principles.
• Utilized three structure-aware modules within the deep learning architecture.
• Employed simplified molecular-input line-entry system (SMILES) strings as input, capturing atomic-level spatial and electronic interactions.
• Defined the activation energy barrier ($\mathrm{\Delta \Delta }{G}^{‡}$) through catalyst-olefin interactions.

Main Results:

• ChemAHNet accurately predicts the absolute configuration of major enantiomers across a wide range of catalysts and substrates.
• The model successfully defines the $\mathrm{\Delta \Delta }{G}^{‡}$ of asymmetric hydrogenation.
• Achieved concurrent prediction of stereoselectivity and absolute configuration.
• Demonstrated applicability beyond olefin asymmetric hydrogenation to other asymmetric catalytic reactions.

Conclusions:

• ChemAHNet offers a robust and accurate method for predicting outcomes in asymmetric hydrogenation, surpassing previous machine learning approaches.
• The model's mechanism-informed design and ability to process SMILES inputs enable it to capture complex chemical interactions.
• ChemAHNet provides a powerful tool for target-directed molecular engineering and catalyst design in asymmetric synthesis.