Rethinking Retrosynthesis: Curriculum Learning Reshapes Transformer-Based Small-Molecule Reaction Prediction.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Chemical Synthesis

Background:

• Retrosynthesis prediction is crucial for drug discovery and materials science.
• Current models struggle with generalizing to rare or complex chemical reactions.
• Uniform training approaches fail to capture the nuances of chemical transformations.

Purpose of the Study:

• To introduce a curriculum learning (CL) framework for enhancing retrosynthesis prediction.
• To systematically control reaction difficulty during model training for improved chemical generalization.
• To address failure modes in models encountering rare or underrepresented reactions.

Main Methods:

• Developed a difficulty-aware curriculum learning framework.
• Introduced reactions in a chemically informed progression based on synthetic accessibility, ring complexity, and molecular size.
• Applied the CL framework to transformer-based architectures (ChemBERTa + DistilGPT2, ReactionT5v2, BART).

Main Results:

• Substantial performance gains across all tested architectures.
• CL improved BART model's top-1 accuracy from 27.0% to 75.9% without chemical pretraining.
• Significant improvements observed in low-data regimes and under scaffold-based/structurally dissimilar splits.
• Achieved gains without auxiliary labels, templates, or reaction class supervision.

Conclusions:

• The curriculum learning framework effectively enhances retrosynthesis prediction accuracy and robustness.
• This method offers a promising approach for planning synthetic routes for diverse chemical compounds.
• The framework shows potential for applications in pharmaceutical intermediates, catalysts, polymers, and functional materials.