Environmental applications of molecular graph learning: Graph neural network based prediction of partition coefficients

In cheminformatics, predicting molecular properties is crucial for enhancing material research, toxicity assessment, and drug discovery. This research investigates the use of graph neural networks (GNNs) for predicting molecular properties by examining three different architectures: graph convolutional networks (GCNs), graph isomorphism networks (GINs), and graph attention networks (GATs). Employing molecular graph information, these models are evaluated on the MUTEG dataset and measured against key metrics such as accuracy and area under the receiver operating characteristic curve (AUC). Our experimental findings show that GIN has the highest accuracy at 89.2%, exceeding GCN (87.5%) and GAT (88.3%). GIN also achieves the highest AUC of 0.89, whereas the AUCs of GCN and GAT are 0.84 and 0.86, respectively, indicating GIN’s enhanced ability to effectively model graph isomorphisms. We selected GIN for this study because of its proven theoretical and empirical strength in capturing graph-level representations, particularly in domains such as cheminformatics, where molecular structures are naturally modeled as graphs. These results highlight the efficacy of GNNs in predicting molecular properties and position GIN as a favored framework for tasks that demand accurate graph feature extraction. This study further plays a pivotal role in understanding the environmental fate and transport of chemical compounds. We used GIN to identify partition coefficients such as the octanol-water partition coefficient, air-water partition coefficient, and soil–water partition coefficient from the MoleculeNet dataset.