Scaffold-Directed Face Selectivity Machine-Learned from Vectors of Non-covalent Interactions

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Scaffold-directed reactions are crucial in synthetic chemistry.
• Understanding non-covalent interactions is key to controlling these reactions.
• Traditional methods for predicting reaction outcomes have limitations.

Purpose of the Study:

• To develop a novel method for vectorizing and Machine Learning (ML) non-covalent interactions.
• To improve the prediction accuracy of reaction outcomes in synthetic chemistry.
• To assess the role of mechanistic knowledge in ML model performance.

Main Methods:

• Vectorization of non-covalent interactions.
• Training ML models on vectorized interaction data.
• Evaluating model performance on Michael additions and Diels-Alder cycloadditions.

Main Results:

• ML models achieved ~90% accuracy in predicting the correct face of approach.
• This accuracy surpasses traditional ML descriptors, energetic calculations, and expert intuition.
• Models lacking mechanistic knowledge showed limited ability to generalize to new reactions.

Conclusions:

• The developed ML approach effectively captures non-covalent interactions for reaction prediction.
• Incorporating mechanistic knowledge is vital for robust and transferable ML models in chemistry.
• This method offers a significant advancement over existing predictive techniques in synthetic chemistry.