Retro-MTGR: Molecule Retrosynthesis Prediction via Multi-Task Graph Representation Learning

Abstract It is a vital bridging step to infer appropriate synthesis reaction routes (i.e., retrosynthesis) of newly-designed molecules. Unlike classical experience-based retrosynthesis approaches, artificial intelligence enables a cheap and fast retrosynthesis approach. Template-based models, limited in known synthesis templates, leverage substructure searching to infer candidate reaction centers (i.e., bonds). In contrast, both translation-based models (TransMs) and discriminative methods (DiscMs) are free to synthesis templates. TransM regards retrosynthesis as a translation from the target molecule to its reactants by generative algorithms. DiscM, directly inspired by chemical synthesis, performs reaction center recognition and leaving group identification in turn. Nevertheless, TransMs are redundant and weakly interpretable, while existing DiscMs neglect the associations between reaction centers and leaving groups. To address these issues, this paper elaborates a novel discriminative Multi-Task Graph Representation learning model of Retrosynthesis prediction (Retro-MTGR). It solves two major supervised discriminative tasks (i.e., the reaction center recognition and the leaving group identification respectively), and an auxiliary self-supervised task (i.e., atom embedding enhancer) simultaneously. The comparison with various state-of-the-art methods first demonstrates the superiority of Retro-MTGR. Then, the ablation studies reveal how its crucial components contribute to the prediction respectively, including the atom embedding enhancer, bond energies, and the leaving group co-occurrence graph. More importantly, comprehensive investigations validate its chemical interpretability by answering two questions: why a bond can be the reaction center or not, and what leaving groups are appropriate to given synthons. The answers demonstrate that Retro-MTGR can reflect five underlying chemical synthesis rules by characterizing molecule structures alone. Finally, two case studies demonstrate that the inferred retrosynthesis routes by Retro-MTGR are significantly consistent with those achieved by performed chemical synthesis assays. It’s anticipated that our Retro-MTGR can provide prior guidance for real retrosynthesis route planning. The code and data underlying this article are freely available at https://github.com/zpczaizheli/Retro-MTGR.