Prediction of Reaction Yield for Buchwald-Hartwig Cross-coupling Reactions Using Deep Learning.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Chemical reaction yield is crucial for optimizing reaction conditions.
• Existing machine learning models for reaction yield prediction using high-throughput experiment (HTE) data lack practical predictive ability.
• The Buchwald-Hartwig cross-coupling reaction is a significant area for yield optimization.

Purpose of the Study:

• To develop an improved machine learning model for predicting chemical reaction yield.
• To enhance the predictive performance of message passing neural network (MPNN) models.
• To evaluate the utility of Mol2Vec feature vectors as atom embeddings in MPNNs for chemical yield prediction.

Main Methods:

• Development of a message passing neural network (MPNN) model.
• Utilizing Mol2Vec feature vectors, pre-trained on a large compound database, as initial atom embeddings.
• Application and evaluation on the Buchwald-Hartwig cross-coupling high-throughput experiment (HTE) data set.
• Comparison with five previously reported predictive models.

Main Results:

• The proposed MPNN model with Mol2Vec embeddings demonstrated superior predictive ability compared to five previous models on three out of five data sets.
• Visualization using a self-attention mechanism highlighted the effectiveness of Mol2Vec embeddings.
• Mol2Vec embeddings outperformed other atom embedding methods, including simpler representations.

Conclusions:

• The proposed MPNN model incorporating Mol2Vec embeddings represents a significant advancement in predicting chemical reaction yield.
• Mol2Vec is a highly effective atom embedding strategy for enhancing the performance of MPNNs in chemical applications.
• This approach offers a more practical and accurate method for reaction condition optimization.