Data undersampling models for the efficient rule-based retrosynthetic planning.

Area of Science:

• Artificial Intelligence in Chemistry
• Organic Synthesis
• Machine Learning for Chemical Discovery

Background:

• Computer-aided retrosynthetic planning is crucial for autonomous robotic chemists.
• Class imbalance in synthetic datasets hinders the prediction accuracy of retrosynthetic paths.
• AI development is significantly impacted by data imbalance issues.

Purpose of the Study:

• To investigate the impact of undersampling models on imbalanced reaction datasets for retrosynthetic planning.
• To improve the prediction performance of retrosynthetic templates for organic molecules.
• To enhance the accuracy of AI-driven chemical synthesis prediction.

Main Methods:

• Application of undersampling techniques to imbalanced reaction datasets.
• Clustering of molecular structures based on similarity (random and dissimilarity).
• Development and evaluation of a target-oriented undersampling model.

Main Results:

• Undersampling improved top-1 and top-10 prediction accuracies by 13.8% and 8.8% respectively, using similarity clustering.
• A target-oriented undersampling model yielded further improvements of 9.3% (top-1) and 4.2% (top-10) in accuracy.
• Demonstrated the critical role of dataset statistical distribution and sampling strategies.

Conclusions:

• Undersampling is an effective strategy to mitigate class imbalance in chemical synthesis datasets.
• Improved retrosynthetic prediction accuracy facilitates more reliable autonomous chemical synthesis.
• A deep understanding of training data is essential for advancing AI in chemistry.