Graph-Theoretic Models for the Prediction of Molecular Measurements

Researchers Anna Niane and Prudence Djagba have published a study demonstrating that a systematically enhanced classical graph-theoretic model can compete with modern deep learning for molecular property prediction. The work began by testing a baseline model based on the D(G)-ζ(G) polynomial indices, which achieved a low average R² of 0.24 across five diverse MoleculeNet datasets (BACE, ESOL, LogP, etc.), confirming its poor generalization. To address this, the authors proposed a rigorous enhancement framework.

This framework progressively incorporated Ridge regularization, additional graph descriptors, physicochemical properties, ensemble learning with Gradient Boosting, Lasso for feature selection, and a hybrid approach combining topological indices with Morgan fingerprints (a common molecular representation). The result was a dramatic performance boost, raising the average best R² to 0.79, representing improvements of 165% to 274% per dataset, all statistically significant (p < 0.001).

In a direct, controlled comparison, these enhanced classical models matched or outperformed a Graph Convolutional Network (GCN)—a standard deep learning architecture for molecules—on all five benchmark tasks. The models also proved competitive against a recent, more advanced hybrid GNN+PGM model. The entire pipeline requires no GPU, completes training in under five minutes, and relies solely on open-source software, making it a highly accessible and efficient alternative for researchers, particularly in resource-constrained environments.