Improving Reaction Yield Prediction with Chemical Atom-Level Reaction Learning.

Area of Science:

• Computational Chemistry
• Chemical Synthesis
• Machine Learning

Background:

• Reaction yield is a key metric for chemical reaction efficiency.
• Predicting yields aids in optimizing synthetic pathways for drug development and materials science.
• Auxiliary molecules significantly influence reaction yields but are challenging to model computationally.

Purpose of the Study:

• To develop a novel computational framework for accurate reaction yield prediction.
• To explicitly model atom-level interactions between reactants and auxiliary molecules.
• To improve the fidelity of reaction modeling for predicting chemical outcomes.

Main Methods:

• Proposed a Chemical Atom-level Reaction Learning (CARL) framework.
• Utilized graph neural networks to model atom-level interactions.
• Evaluated performance on multiple benchmark datasets.

Main Results:

• CARL achieved state-of-the-art (SOTA) performance in yield prediction.
• The framework does not rely on handcrafted descriptors or domain-specific heuristics.
• Identified key substructures in reactants and auxiliary molecules influencing reaction outcomes.

Conclusions:

• CARL effectively models atom-level interactions for improved yield prediction.
• The framework offers a systematic approach to exploring larger reaction spaces.
• CARL enhances the design of synthetic pathways and accelerates chemical discovery.