A comprehensive analysis of optimizers in message passing neural networks for molecular property prediction task

Molecular property prediction is fundamental to cheminformatics, particularly in drug discovery and materials science. While Graph Neural Networks (GNNs) have emerged as powerful tools for molecular classification, the choice of optimizer significantly influences model performance. However, there is a lack of systematic studies evaluating the impact of different optimizers on molecular GNNs. To address this gap, this research investigates the effects of eight widely used optimizers on the performance of a message-passing neural network (MPNN) for binary molecular classification. Using benchmark datasets, we analyze training stability, convergence behavior, and classification accuracy. Our results reveal that adaptive gradient-based optimizers outperform traditional methods in terms of convergence stability and predictive accuracy. These findings highlight the critical role of optimizer selection in molecular classification and provide insights into optimization strategies that enhance model reliability and performance.