NED-HGT: a node-edge decoupled heterogeneous graph transformer for accurate solubility prediction

Accurate prediction of molecular solubility and lipophilicity is critical for drug discovery, influencing candidate screening, formulation development, and pharmacokinetic evaluation. Prevailing homogeneous molecular graph representations often fail to capture the nuanced physicochemical roles of chemical bonds and their spatial arrangements, which are critical determinants of solvation free energy. Here, we propose a Node-Edge Decoupling Heterogeneous Graph Transformer (NED-HGT) for predicting solubility properties of drug-like molecules. Specifically, we construct a heterogeneous molecular graph where atoms and chemical bonds are explicitly modeled as distinct nodes, with four edge types capturing inter-atomic distances, bonding connectivity, atom-bond membership, and bond-bond angles. To fully exploit this structural information, we develop a node-edge decoupled heterogeneous graph transformer that independently learns atom-centric and bond-centric representations, fused via an attention-based gated recurrent unit module. Concurrently, 2D molecular descriptors are extracted via RDKit and processed through a convolutional neural network to capture complementary physicochemical features. The two channels are integrated to generate a comprehensive molecular representation for solubility prediction. Evaluated on five public benchmarks, NED-HGT achieves competitive or state-of-the-art predictive performance across all datasets under random splits. Under the more challenging scaffold-aware splits, it ranks first on four datasets (ESOL, FreeSolv, Lipophilicity, and CASR-1) and secures a close second-place result on AqSolDB. Ablation experiments confirm the synergistic contributions of both channels, while case studies provide qualitative illustrations of the model’s predictive behavior on structurally diverse molecules. The source code is publicly available at https://github.com/ShiLab-GitHub/NED-HGT.