Statistical comparison and uncertainty analysis of graph neural networks and machine learning models for molecular property prediction in drug discovery

Accurate prediction of molecular properties is essential for accelerating drug discovery. While graph neural networks (GNNs) have emerged as a powerful tool for this task, they have not been systematically benchmarked against traditional machine learning methods, particularly regarding the crucial aspects of predictive accuracy, interpretability, and uncertainty. To systematically evaluate state-of-the-art GNN architectures against classical machine learning methods for predicting key physicochemical properties. This study provides a multi-faceted comparison of model performance, statistical robustness, prediction uncertainty, and chemical interpretability. We implemented and compared seven models: Graph Attention Networks (GAT), Graph Convolutional Networks (GCN), Graph Isomorphism Networks (GIN), SimpleGNN, Support Vector Regression (SVR), Random Forest, and ElasticNet. These were evaluated across three MoleculeNet datasets: ESOL (aqueous solubility), FreeSolv (hydration free energy), and Lipophilicity (partition coefficient). The evaluation framework included rigorous statistical testing, bootstrap-based uncertainty quantification, and analysis of GAT attention mechanisms for chemical insight. GAT consistently achieved superior performance, with test RMSE values of 0.1863 (ESOL), 0.1953 (FreeSolv), and 0.4922 (Lipophilicity), outperforming traditional methods by a significant margin. GNNs demonstrated substantial advantages over classical approaches, which showed considerably higher prediction errors. Furthermore, GAT provided the most reliable predictions with the lowest uncertainty and generated chemically relevant insights through its attention mechanism, successfully identifying key functional groups driving molecular properties. This systematic evaluation provides compelling evidence for the superiority of GNNs, particularly GAT, over traditional machine learning for molecular property prediction. GAT’s high accuracy, combined with its robust uncertainty quantification and chemical interpretability, establishes it as a preferred computational approach for pharmaceutical research and development.