Improving LogP Prediction of Chemical Compounds for Drug Development by a Novel Graph CNN Design

Computational prediction of the properties of drug molecules has become increasingly valuable, as it can reduce both the costs and time associated with drug development. One key property is LogP, which measures how easily a molecule can pass through lipid membranes and determines how well a drug will be absorbed, distributed, and transported throughout the body. There are various prediction methods, such as quantum calculation, classical machine learning models, and emerging deep learning techniques. However, they are still not accurate enough to replace the experimental determination of LogP. In this paper, we developed a computational method to predict the LogP values with greater accuracy than existing methods. Our approach employs a graph convolutional neural network (GCNN) with triple graph convolutional (GC) layers, utilizing the DeepChem library of Python modules adapted from a previously reported deep neural network (DNN). The model is trained on a carefully curated, publicly available dataset of 13,785 compounds, learning to recognize patterns between molecular structures and LogP values by passing the data through a series of graph convolutional layers. The model is tested on a dataset of 1,350 molecules against the original DNN and two benchmark methods (support vector regression model and random forest). Our model achieved a mean absolute error is 0.509 LogP units, while the other methods ranged from 0.550 to 1.734 LogP units. These results suggest that the proposed model offers superior prediction accuracy over existing ones and warrants further developments.