Identifying the reaction centers of molecule based on dual-view representation

In the process of drug retrosynthesis, identifying the reaction centers where the chemical reactions occurring is an important fundamental issue in semi-templated models. However, few publications pay their attention on this research point currently and most of them only simply employ GNN-based or Transformer-based approaches to detect the reaction centers in a drug, overlooking the significance of the mutual interactions between atoms and bonds, which has been shown to play crucial roles in the chemical properties. Besides, previous works mainly focus on atom-level representation learning, ignoring the molecular fragment information that also influences the nature of a drug. To this end, we propose a dual-view molecular representation learning model called DVMR-CL (Dual-View Molecular Representation Contrastive Learning) to learn the molecular representations. More specially, the proposed method captures the molecular fragment information by learning dual-view molecular embeddings, leveraging cross-attention mechanism to address the aforementioned problem. Experimental results demonstrate that the proposed DVMR-CL significantly improves the accuracy of reaction center recognition compared with other baselines. The source code underlying this article is freely available at https://github.com/jingwang-0415/DVMR-CL.