Accurate predictions of drugs aqueous solubility via deep learning tools

In recent years, increasingly more data-driven approaches have been successfully applied in various kinds of properties predictions for medicine industry, with considerable accuracy; and such a progress in methodology has also facilitated the routine procedures of computational pharmaceutics. In this study, we proposed one novel and efficient quantitative structure-property relationship (QSPR) model for molecular properties predictions within the framework of deep learning neural network (DNN), using molecular descriptors calculated by Mordred. The key characteristic of this developed neural network lies in the fact that it could extract molecular information via flexibly utilising the metrics dependent descriptors, and therefore, it can be expected for various kinds of properties predictions, which are tightly associated with the molecular structures. We tried this approach in aqueous solubility predictions for drug-like molecules, and found that it performed well. The key factors determining the model's performance were also discussed in details. We believe such a useful yet convenient tool will enhance the efficiency of drug screenings, and even contribute to the development of modern pharmaceutics. Additionally, a systematic yet large-scale comparison between this novel DNN model and another proposed graph convolutional network (GCN) model was also presented in this study; and we hope to provide valuable reference to support research of computational pharmaceutics, focusing on properties predictions.