RetroAux: Boosting Retrosynthesis via Molecular Property-Aware Learning

Mianzhi Liu, Fan Xiao, Zhiliang Yu, Huayang Huang, Yuke Li, Yi Yang, Yu Wu

Keywords: Retrosynthesis Prediction, Molecular Property Awareness, Knowledge-Constrained Learning, Molecular Representation Learning, Autoregressive Molecular Generation, Chemical Semantic Space

TL;DR: Boosting Retrosynthesis via Molecular Property-Aware Learning

Abstract: Retrosynthesis remains a critical task in both drug discovery and organic synthesis. Current methodologies in this field predominantly rely on purely data-driven paradigms, where models are expected to autonomously learn reaction patterns from extensive retrosynthesis datasets without incorporating established chemical knowledge. To address this limitation, we introduce RetroAux, a framework injected with molecular property knowledge to enhance existing approaches to achieve significant performance enhancements. Specifically, throughout the entire retrosynthesis pipeline from reaction rule acquisition to reactant molecule generation, our methodology systematically integrates molecular property knowledge (e.g., functional groups, chirality) as both chemical priors and foundations, thereby enhancing retrosynthesis prediction reliability. Our knowledge-driven framework can be seamlessly integrated with multiple existing data-driven methods and improve their performance stably. Experimental results demonstrate that it enhances various existing data-driven retrosynthesis models with average top-1 accuracy improvements of 2.22% without retraining origin models, signifying a paradigm evolution in retrosynthesis from purely data-driven approaches to knowledge-driven methodologies.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 1346