Comparative Evaluation of Graph Neural Networks for Molecular Lipophilicity Prediction

Abstract

Predicting molecular properties is a central challenge in cheminformatics and plays a critical role in drug discovery and materials science. Lipophilicity, which describes the partitioning of a molecule between lipid and aqueous environments, is a key physicochemical property influencing drug behavior. Although quantum mechanical methods can provide accurate predictions, they are often computationally demanding. In contrast, large experimental databases can be leveraged to train data-driven models such as graph neural networks (GNNs), which learn molecular representations directly from chemical structures. In this study, we compare four GNN architectures—graph convolutional networks (GCN), graph isomorphism networks (GIN), Graph Sample and AggregatE (GraphSAGE), and Attentive Fingerprint (Attentive FP)—for predicting lipophilicity using both regression and classification approaches. The models are evaluated on a benchmark dataset under consistent preprocessing, optimization, and training conditions. Our results show that the Attentive FP model consistently outperforms the other architectures, yielding improved predictive accuracy across both tasks. We further examine the impact of outlier removal during preprocessing, observing reduced prediction errors and improved robustness. These findings provide insight into the capabilities of representative GNN models for molecular lipophilicity prediction.