RetroRanker: leveraging reaction changes to improve retrosynthesis prediction through re-ranking

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Chemical Synthesis

Background:

• Data-driven methods are advancing retrosynthesis, but suffer from frequency bias, leading to suboptimal predictions.
• Frequency bias causes chemically unreasonable reactants to be ranked highly due to training data distribution.
• Existing template-based approaches may miss valid reactants among low-ranked, low-confidence predictions.

Purpose of the Study:

• To introduce RetroRanker, a novel graph neural network-based ranking model.
• To mitigate frequency bias in existing retrosynthesis prediction models through re-ranking.
• To improve the accuracy and reliability of retrosynthesis predictions.

Main Methods:

• Developed RetroRanker, a graph neural network model for re-ranking retrosynthesis predictions.
• RetroRanker evaluates potential reaction changes for predicted reactants to assess chemical reasonableness.
• The model re-ranks predictions, lowering the rank of chemically implausible outcomes.

Main Results:

• RetroRanker demonstrated performance improvements on most state-of-the-art retrosynthesis models.
• Re-ranked results on public retrosynthesis benchmarks showed significant enhancements.
• Preliminary studies suggest RetroRanker can also boost multi-step retrosynthesis performance.

Conclusions:

• RetroRanker effectively reduces frequency bias in retrosynthesis predictions.
• The model enhances the accuracy of chemical synthesis planning tools.
• RetroRanker shows promise for both single-step and multi-step retrosynthesis applications.