Graph neural networks have become essential for molecular property prediction in drug discovery and materials design. However, graph neural networks face a critical challenge called oversmoothing: after multiple message passing layers, node features become overly similar, causing atoms to lose their unique physicochemical identity. This severely limits model explainability – researchers cannot identify which atoms or substructures drive a prediction. Existing solutions designed for general graph data often break molecular topological structures, which is unacceptable for molecules where real chemical meaning would be lost.
In a study published in ENG. Chem. Eng., researchers at East China University of Science and Technology propose AdapGNN, a framework specially designed for molecule property prediction. The core idea is simple yet effective: during each message passing step, original atom features are merged with features from the previous layer. To further emphasize crucial parts of a molecule, a weight projection module generates node specific weights based on local chemical environments. Two projection types are defined: L₁ using only the atom’s own features, and L₂ incorporating first order neighbor information via one message passing step.
Because existing benchmark datasets lack ground truth atom importance, the team also created MolExplain, a dataset of nearly 300,000 small molecules with molecular weight below 300, covering 12 distinct tasks. Each task includes the input molecule, a target descriptor value, and a binary atom mask indicating which atoms truly contribute. Properties range from simple atom counts to complex molecular descriptors like logP, topological polar surface area, and atomic surface area, with transformation functions that can be constant, linear, or nonlinear.
Experiments compared AdapGNN integrated with two base graph neural networks, namely AttentiveFP and graph convolution network, against the standalone models on predictive accuracy and explainability. AdapGNN achieved lower root mean squared error than baselines on 10 of 12 tasks, with especially large improvements for molecule energy where the error was reduced from 1.54 to 0.64 for AttentiveFP, and from 2.67 to 0.74 for graph convolution network. For explainability, measured by the area under the ROC curve for atom level classification of contributing versus non contributing atoms, AdapGNN consistently scored above 75 percent and often much higher, reaching up to 99.5 percent for carbon atom count and 99.7 percent for topological surface area, while baseline models sometimes performed near random at 50 percent. The muted root mean squared error metric also confirmed AdapGNN’s superior explainability. Visualizations showed that AdapGNN correctly focuses on non carbon atoms for non carbon count tasks and on in ring atoms for in ring count tasks, whereas baseline models often assign importance broadly across all atoms.
The choice between L₁ and L₂ projections depends on the task: L₁ works better for atom type dependent properties, while L₂ better captures local bonding environments. This framework provides a practical path toward more interpretable and trustworthy graph neural network models for molecular property prediction.

DOI
10.1007/s11705-026-2659-1