CoBiSyn: A Bidirectional Search Framework for Chemical Synthesis Planning

ICLR 2026 Conference Submission16488 Authors

Keywords: AI for Science, Machine Learning

TL;DR: CoBiSyn accelerates synthesis planning through coordinated bidirectional search, producing shorter and higher-quality pathways.

Abstract: Artificial Intelligence is increasingly advancing scientific discovery, with chemistry being a key application domain. Synthesis planning, which aims to identify feasible reaction pathways connecting target molecules to available starting materials, is a fundamental task in organic synthesis and drug discovery. Prior work typically relies on backward search, iteratively applying single-step retrosynthesis models, which neglects information from the starting materials and often leads to inefficient exploration and redundant reactions. In this paper, we propose CoBiSyn (Coordinated Bidirectional Synthesis Planning), a framework that alternates between "backward decomposition'' and "forward construction'', while coordinating these two directions through shared frontier information. To support this process, we introduce a conditional embedding projection mechanism and a learned asymmetric synthetic distance, which together provide local and global cost estimates to steer the search. The experiments on multiple benchmark datasets demonstrate that CoBiSyn significantly improves the efficiency and quality for synthesis planning, compared to existing approaches.

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

Submission Number: 16488