A Directed Message Passing Neural Network Model for Predicting True Solubility in Drug Discovery

This study presents a novel directed message-passing neural network (DMPNN) model for predicting true solubility (logS) in drug discovery. Traditional methods such as high-throughput screening and QSAR modelsexemplified by the RuleofFivehave historically guided early discovery efforts but often fall short in handling modern high-dimensional, complex chemical datasets. Recent advances in integrating machine learning, including support vector machines, random forests, and deep learning, have improved prediction accuracy. However, high-dimensional data and accurate error estimation remain significant challenges. To address these issues, the proposed model leverages a directed message-passing mechanism that explicitly captures bond-directional interactions and complex non-linear relationships between atoms in molecular graphs. The model processes SMILES strings by first converting them into molecular graphs, featuring atoms and bonds via established cheminformatics techniques, and then iteratively refining bond representations through directed message passing. A final readout function aggregates atomic embeddings into a global fingerprint that feeds into a multilayer perception for solubility prediction. Tested on benchmark datasets from DeepChem, the DMPNN achieved an RMSE of 8.030, an MAE of 2.841, and an R of 0.989, demonstrating robust performance and reliability in predicting solubility. These results suggest our model could speed up and simplify the drug discovery process.