Chemical Graph-Based Transformer Models for Yield Prediction of High-Throughput Cross-Coupling Reaction Datasets.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Chemical Reaction Engineering

Background:

• Chemical reaction yield is crucial for optimizing reaction conditions.
• Data-driven models using high-throughput experimentation are emerging for yield prediction.
• Accurate yield prediction aids in efficient chemical synthesis and process development.

Purpose of the Study:

• To propose a novel neural network architecture for predicting chemical reaction yield.
• To leverage chemical graph representations of reaction components for enhanced prediction.
• To compare the performance of the proposed model against existing state-of-the-art methods.

Main Methods:

• Developed a Message Passing Neural Network (MPNN) combined with a Transformer encoder (MPNN-Transformer).
• Represented reaction components as molecular matrices and incorporated compound role embeddings.
• Evaluated model performance on Buchwald-Hartwig cross-coupling (BHC) and Suzuki-Miyaura cross-coupling (SMC) datasets.

Main Results:

• The MPNN-Transformer model achieved high prediction accuracy on BHC datasets.
• Demonstrated strong performance on extrapolation-oriented SMC datasets.
• Accuracy improved with larger training dataset sizes and showed limitations with purely structural similarity.

Conclusions:

• The MPNN-Transformer architecture is effective for chemical reaction yield prediction.
• The model shows promise for optimizing cross-coupling reactions.
• Data-driven yield prediction has limitations, particularly concerning chemical structural similarity.