Interpretable side reaction detection via atom representations derived from multi-tier reaction perspectives

Recent deep learning-based methods have advanced chemical retrosynthesis. However, they neglect side reactions, strong competitors to desired synthesis reactions that reduce the yield of anticipated product molecules. To identify side-reaction sites, we introduce an atom representation learning framework, SideReact, which explores synthesis reactions from a top-down chemical perspective (i.e., from reaction to motif, to property, and to quantum w.r.t. atoms). Its multi-tier atom encoder (MtAE) extracts top-down observations as atom representations, including reaction paradigms, motif reactivities and co-reactivities, atomic identities and bonding circumstances, and atoms' electron-attracting abilities. Moreover, its hierarchical cross-attention module (HCA) captures the chemical hierarchy among encoding tiers in a bottom-up way. Especially, HCA enables an interpretable path through the hierarchy to reveal how to find reaction sites among atoms in reactants. The comparison with ten state-of-the-art methods demonstrates the superiority of SideReact in both side reactions and desired reactions. Furthermore, a real case study, including ten synthetic examples in textbooks, a complex synthetic reaction of Atorvastatin with side-reaction sites, and a synthetic reaction for a drug-like molecule, substantiates the effectiveness of SideReact. More importantly, analyses demonstrate how its MtAE works and how its HCA unveils the interpretable inference of side-reaction sites, highlighting its potential in retrosynthesis planning.