RPSubAlign: a novel sequence-based molecular representation method for retrosynthesis prediction with improved validity and robustness

• Published in: Briefings in bioinformatics Vol. 26; no. 3
• Main Authors: Hu, Yuting, Hu, Feng, Zhang, Hongwen, Xu, Hongling, Gao, Jixiang, Deng, Wenshuai, Tian, Zijing, Hu, Qiaoyu, Li, Honglin, Diao, Yanyan
• Format: Journal Article
• Language: English
• Published: England Oxford University Press 01.05.2025; Oxford Publishing Limited (England)
• Subjects: Algorithms; Chemical synthesis; Computational Biology - methods; Data augmentation; Drug Discovery - methods; Predictions; Problem Solving Protocol; Representations; Robustness; Route planning; Software; Stability augmentation; Validity; deep learning; molecular representation; computer-aided synthesis planning
• ISSN: 1467-5463; 1477-4054; 1477-4054
• DOI: 10.1093/bib/bbaf257

Abstract Retrosynthetic route planning is essential for designing efficient pathways to synthesize complex molecules, serving as a cornerstone in drug discovery and organic synthesis. Sequence-based models have become a predominant approach in retrosynthetic route planning, yet its validity and robustness remain limited by the challenges in molecular representation methods. Current methods typically treat reactants and products as independent molecules, overlooking structural relationships crucial for accurate synthesis predictions. Herein, we introduce RPSubAlign, a molecular sequence representation method specifically tailored for retrosynthetic tasks, which aligns common substructures between reactants and products to enhance the validity and robustness of sequence-based models. Compared with conventional random and root-alignment representations, RPSubAlign achieves better performance on the USPTO-50K and USPTO-MIT datasets, improving up to a 34.8% increase in Top-N accuracy (with Self-Referencing Embedded Strings representation) and demonstrating enhanced stability across various data augmentation scenarios. RPSubAlign significantly improves syntactic validity, reaching 86.64% on USPTO-50K and 96.45% on USPTO-MIT (with Simplified Molecular Input Line Entry System representation), outperforming baseline methods. These results highlight RPSubAlign as a robust, effective approach for molecular characterization method for retrosynthesis predictions. Code for RPSubAlign is available at https://github.com/Aminoacid1226/RPSubAlign.