Data Augmentation and Pretraining for Template-Based Retrosynthetic Prediction in Computer-Aided Synthesis Planning

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Organic Synthesis

Background:

• Machine learning models for reaction template recommendation often prioritize frequent templates, excluding potentially valuable transformations.
• Current models focus on high accuracy for one-to-one mapping, limiting the scope of recommended reactions.

Purpose of the Study:

• To augment machine learning algorithms for improved reaction template recommendation in computer-aided synthesis planning.
• To increase the recall of template applicability and the diversity of predicted precursors.

Main Methods:

• Augmenting open-access organic reaction datasets with calculated template applicability.
• Pretraining a template-relevance neural network on the augmented dataset.

Main Results:

• Reported an increase in template applicability recall.
• Observed an increase in the diversity of predicted precursors.
• Demonstrated improved top-1 accuracy, particularly for rare templates, even on small datasets.

Conclusions:

• Data augmentation and pretraining enhance neural networks to recognize a broader set of theoretically successful reaction templates.
• These strategies are effective for improving reaction template recommendation, especially in scenarios with limited data.