Improving Chemical Reaction Prediction with Unlabeled Data.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Predicting organic chemical reaction products is crucial, particularly for novel reactants.
• Traditional machine learning models require extensive high-quality labeled data, limiting their application.
• Developing methods to leverage unlabeled data is essential for enhancing model performance.

Purpose of the Study:

• To propose a novel method combining semi-supervised learning and graph convolutional neural networks for chemical reaction prediction.
• To improve the accuracy of predicting products in organic chemical reactions by utilizing unlabeled datasets.
• To address the limitations of data dependency in traditional machine learning models for chemical synthesis.

Main Methods:

• A Mean Teacher Weisfeiler-Lehman Network was developed to identify reaction centers.
• A candidate product set was constructed based on predicted reaction centers.
• An Improved Weisfeiler-Lehman Difference Network was employed to rank the candidate products.

Main Results:

• The proposed framework demonstrated improved top-5 accuracy by 0.7% using 35k unlabeled data alongside 400k labeled data.
• A significant performance gain of 1.8% was observed with 80k labeled and 35k unlabeled data, showcasing the benefit of unlabeled data.
• The method's performance scales positively with an increased proportion of unlabeled data.

Conclusions:

• The integration of semi-supervised learning with graph convolutional neural networks offers a powerful approach for chemical reaction prediction.
• Leveraging unlabeled data significantly enhances prediction accuracy, particularly in scenarios with limited labeled datasets.
• This method provides a valuable tool for accelerating chemical discovery and synthesis planning.