Abstract
This work focuses on organic reaction prediction with deep learning, with the recently introduced fragSMILES representation - which encodes molecular substructures and chirality, enabling compact and expressive molecular representation in a textual form. In a systematic comparison with well-established molecular notations - simplified molecular input line entry system (SMILES), self-referencing embedded strings (SELFIES), sequential attachment-based fragment embedding (SAFE) and tree-based SMILES (t-SMILES) - fragSMILES achieved the highest performance across forward- and retro-synthesis prediction, with superior recognition of stereochemical reaction information. Moreover, fragSMILES enhances the capacity to capture stereochemical complexity - a key challenge in synthesis planning. Our results demonstrate that chirality-aware and fragment-level representations can advance current computer-assisted synthesis planning efforts.