Prediction of the aqueous solubility of small molecules was very challenging. Despite the effort of having the solubility contest back in 2008 and with the various machine learning algorithm availability, in silico models were given about only 0.65 of R2 value in prediction power to such basic properties of a molecule. Recent graph based neural network methods have been found successful and widely applied in prediction of chemical properties, with dramatic improvements of model performance in several endpoints. In this presentation, a graph convolution network (GCN) approach for water solubility prediction is demonstrated. In this model, the information from the structures was obtained and trained directly, preventing the information loss in data transformation. The experimental result indicates that the automatically feature extraction make a better performance than traditional methods. As a result, the R2 values of our models can be improved over 0.90, with the Delaney's solubility data set. Comparing to the traditional machine learning methods, the GCN has significantly improved the prediction ability of aqueous solubility.

Graph based neural network model for predicting aqueous solubility

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.