Enhancing Retrosynthesis Prediction with Distillation Learning

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Synthetic Chemistry

Background:

• Single-step retrosynthesis prediction is vital for chemical synthesis pathway planning.
• Current methods exhibit performance gaps between high- and low-resource reaction classes, limiting overall effectiveness.
• Addressing these disparities is crucial for advancing automated synthesis design.

Purpose of the Study:

• To introduce novel distillation learning strategies to mitigate performance disparities in retrosynthesis prediction.
• To enhance the accuracy and robustness of template-free retrosynthesis prediction models.
• To improve the generalizability of predictive models across diverse chemical reaction classes.

Main Methods:

• Developed Retrosynthetic Mutual Distillation (Retro-MD) using dual sampling temperatures and cross-model knowledge transfer.
• Developed Retrosynthetic Self-Distillation (Retro-SD) employing a fixed temperature and iterative self-distillation.
• Applied these strategies to Transformer-based models for template-free retrosynthesis prediction.

Main Results:

• Achieved state-of-the-art performance among template-free retrosynthesis prediction approaches.
• Demonstrated significant improvements in prediction accuracy, particularly for low-resource reaction classes.
• Ablation studies confirmed the effectiveness of reaction-class-aware task partitioning.

Conclusions:

• Retro-MD and Retro-SD effectively bridge the performance gap in retrosynthesis prediction.
• Distillation learning offers a powerful approach to enhance chemical reaction prediction models.
• The proposed methods advance the capabilities of automated retrosynthesis and pathway planning.