Fragment-based t-SMILES for de novo molecular generation

At present, sequence-based and graph-based models are two of popular used molecular generative models. In this study, we introduce a general-purposed, fragment-based, hierarchical molecular representation named t-SMILES (tree-based SMILES) which describes molecules using a SMILES-type string obtained by doing breadth first search (BFS) on full binary molecular tree formed from fragmented molecular graph. The proposed t-SMILES combines the advantages of graph model paying more attention to molecular topology structure and language model possessing powerful learning ability. Experiments with feature tree rooted JTVAE and chemical reaction-based BRICS molecular decomposing algorithms using sequence-based autoregressive generation models on three popular molecule datasets including Zinc, QM9 and ChEMBL datasets indicate that t-SMILES based models significantly outperform previously proposed fragment-based models and being competitive with classical SMILES based and graph-based approaches. Most importantly, we proposed a new perspective for fragment based molecular designing. Hence, SOTA powerful sequence-based solutions could be easily applied for fragment based molecular tasks.