A Retrieval-Based Molecular Style Transformation Optimization Model

Molecular optimization endeavors to enhance molecular properties through the alteration of intricate molecular structures, has important roles in drug discovery. In the past years, a lot of computational models have been proposed for molecular optimization. However, there is still great room for existing models to improve their performance, which is due to the following two facts. First, the majority of existing models are based on basic molecular descriptors, such as molecular weights and property values. Second, most of them are lack of the guidance from high-quality molecules. To address this issue, in this paper we propose a retrieval-based molecular style transformation model, which is terms as RMST. In the model, distinctive attribute features of high-property molecules are integrated with SMILES sequence data to navigate the optimization pathway. Besides, the model uses an attention-driven retrieval module to extract molecular attributes, which are fused with the SMILES sequence information to enrich the representation of molecules. Experimental results on two representative optimization tasks show that the proposed model is superior to three baseline models.