Graph neural network architectures for predicting the electrophilicity index: insights from 2D and 3D molecular graph representations

Abstract

The electrophilicity index (ω) is a fundamental quantum chemical descriptor that quantifies a molecule's ability to accept electrons, playing a critical role in assessing reactivity and stability, and guiding molecular design. In this study, we benchmark a diverse set of graph neural network (GNN) architectures for predicting ω, using data derived from the QM9 dataset of organic molecules. As ω is a global quantum-chemical property derived from frontier molecular orbital energies, typically sensitive to the local chemical structure, we compare 3D molecular geometry models (SchNet, ALIGNN, GemNet), which account for the full atomic structure, with connectivity-based 2D models (attentive FP, GCN, GraphSAGE, GIN, GINE, GATv2) that consider only molecular topology. The results indicate that ω depends not only on molecular topology but also on the complete 3D atomic arrangement, as reflected by the superior predictive accuracy of 3D models—particularly ALIGNN. Nevertheless, some 2D GNNs provide a computationally efficient alternative. Notably, GINE achieves more than an order-of-magnitude reduction in training time compared to ALIGNN, while exhibiting only about a one percent-level decrease in predictive accuracy.