De Novo Drug Design Using Unified Multilayer Simple Recurrent Unit Model

Abstract

Drug de novo design has developed rapidly and innovatively in recent years. In this way, novel drug structures can be designed, which has made outstanding contributions to the field of drug design. Improving the quality of generated drugs is the unifying goal of most published experimental papers, and the sheer volume of data provides a solid foundation for achieving this goal. However, as the amount of drug data becomes larger and larger, we have to consider the time cost of such experiments. Correspondingly, when we are studying a small number of drugs with specific properties, the fitting ability of the model becomes critical. As a result, some models perform wildly differently on different datasets. In view of SMILES (Simplified molecular input line entry system) experiment data of different data sets, this paper proposes a “Unified Multilayer SRU De novo drug design acceleration Model” (USD) based on multi-layer Simple Recurrent Unit (SRU). Aiming at the problem that the amount of data in deep learning has a greater impact on the effect of experimental training, this experiment trains the SMILES data in the DugBank database (small data amount) and ChEMBL (big data amount), and finally generates novel drug molecular data. Both the efficiency of molecular generation and the time cost of model training and data generation have been greatly improved. A series of comparative experiments have proved that USD has a good balance ability between drug generation quality and time cost, which proves that this new research direction has certain experimental and reference value.