Improving chemical reaction yield prediction using pre-trained graph neural networks.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Graph neural networks (GNNs) are effective for predicting chemical reaction yields.
• Performance of GNNs degrades with insufficient or non-diverse training datasets.

Purpose of the Study:

• To investigate the effectiveness of pre-training GNNs for chemical reaction yield prediction.
• To introduce a novel GNN pre-training method to enhance prediction performance.

Main Methods:

• Pre-trained GNNs on a large molecular database using a novel pre-text task.
• Calculated molecular descriptors and reduced dimensionality using principal component analysis (PCA).
• Assigned principal component scores as pseudo-labels for GNN pre-training.
• Fine-tuned the pre-trained GNN on reaction yield datasets.

Main Results:

• Demonstrated the effectiveness of GNN pre-training through experimental evaluation.
• The proposed method showed improved performance in chemical reaction yield prediction.
• Successfully addressed the issue of limited training data for GNNs.

Conclusions:

• GNN pre-training is a viable strategy to improve chemical reaction yield prediction accuracy.
• The novel pre-training approach using PCA-derived pseudo-labels is effective.
• This method offers a solution for scenarios with limited chemical reaction data.