Background: Molecular interactions are central to numerous challenges in chemistry and the life sciences. Whether in solute–solvent dissolution, adverse drug–drug interactions, or protein complex formation, understanding the fundamental mechanisms of intermolecular interactions is essential for advancing molecular-level modeling.

Current machine learning strategies typically approach this problem through three paradigms: Embedding Merging, where molecules are encoded independently and then merged; Feature fusion, where features are combined via attention or other fusion mechanisms; and Merged molecular graphs, where solute and solvent atoms are unified into a single graph to explicitly model atomic-level interactions.

Among these, merged molecular graph-based approaches have shown strong performance and enhanced interpretability due to their ability to explicitly encode intermolecular interactions. A representative model in this category is MMGNN, which connects all possible atom pairs in a fused graph and applies attention mechanisms to prioritize critical interactions. While effective in improving the prediction of solvation free energy (ΔG_solv), MMGNN suffers from rapidly growing computational complexity as molecular size increases, limiting its scalability and general applicability.

Method: To address these limitations, we introduce a novel framework: the Molecular Merged Hypergraph Neural Network (MMHNN). MMHNN innovatively incorporates a predefined set of molecular subgraphs, replacing each with a supernode to construct a compact hypergraph. This architectural change substantially reduces computational overhead while preserving essential molecular interactions.

In addition, MMHNN explicitly models non-interacting or repulsive atomic pairs by introducing a mechanism rooted in Graph Information Bottleneck (GIB) theory. This component enhances the semantic interpretability of both nodes and edges in the fused molecular graph, thereby improving the transparency and explainability of predictions.

Through extensive experiments on multiple solute–solvent benchmark datasets, MMHNN not only demonstrates significantly improved predictive performance and efficiency over existing methods but also offers clearer interpretability of molecular interactions, paving the way for efficient and scalable modeling of intermolecular relationships.

Furthermore, to evaluate the model's generalization capability across diverse solute–solvent systems, the authors conducted systematic generalization analyses on different solvent environments and solute scaffolds, validating the robustness and effectiveness of the proposed model under distributional shifts.

In addition, a quantitative analysis of generalization errors across different solute–solvent systems reveals that the model exhibits greater error sensitivity for larger molecules and varies with specific atomic element types. Similarly, distinct distributional differences were observed among solute scaffolds, indicating that distribution shifts between training and testing data can significantly impact model performance.

Finally, it is evident that the hypergraph-based molecular fusion framework significantly reduces computational time and memory consumption compared to fully connected molecular fusion graphs, while simultaneously delivering superior predictive performance.

The complete study is accessible via DOI: 10.34133/research.0740