RetroMPA: A Molecular Property-Aware Auxiliary Framework for Enhancing Retrosynthesis Prediction

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Data-driven deep learning models for retrosynthesis lack explicit chemical knowledge integration.
• Existing models learn reaction patterns from data but struggle with incorporating prior chemical expertise.

Purpose of the Study:

• To introduce RetroMPA, a molecular property-aware module to enhance retrosynthesis predictions.
• To provide a model-agnostic framework for integrating chemical knowledge into existing retrosynthesis algorithms.

Main Methods:

• RetroMPA acts as a posthoc enhancement module, filtering and optimizing predictions from various algorithms.
• It operates at the molecular level using a property-aware latent embedding space.
• The framework is designed as a plug-and-play solution, requiring no modifications to original model architectures.

Main Results:

• RetroMPA improved top-1 accuracy by an average of 5.50% across eight models on USPTO-50K.
• Performance gains of approximately 2.03% were observed on the large-scale USPTO-Full dataset.
• Wet-lab experiments validated novel substrate combinations for Suzuki-Miyaura coupling, Bucherer reaction, and Friedel-Crafts acylation.

Conclusions:

• RetroMPA effectively injects chemical knowledge into retrosynthesis pipelines, enhancing predictive accuracy.
• The module's model-agnostic and plug-and-play nature offers broad applicability in computational chemistry.
• Experimental validation suggests RetroMPA's potential beyond data fitting, aiding in novel synthetic route discovery.