CholBindNet as an interpretable neural network for cholesterol-binding site classification

Cholesterol is a key modulator of membrane protein structure and function, yet predicting cholesterol-binding sites remains challenging due to its non-druglike physicochemical properties. Here, we curated more than 800 high-resolution transmembrane protein structures containing cholesterol, and developed an interpretable, atom-based graph neural network, called CholBindNet. A positive-unlabeled (PU) training strategy was employed to address the scarcity of negative samples due to the promiscuous nature of cholesterol binding. We show that CholBindNet substantially outperforms existing machine learning models trained on general ligand-binding datasets, including AlphaFold3, P2Rank, and DiffDock. The performance and generalizability of the model on unseen membrane proteins were further demonstrated by rapidly assessing cholesterol-binding sites in the PIEZO2 ion channel against all-atom molecular dynamics (MD) simulations conducted on Anton3 supercomputer. Additionally, strong model interpretability was achieved for CholBindNet through atom-level feature encoding, Grad-CAM visualization, and attention-based scoring analysis. Overall, CholBindNet provides an efficient and scalable approach for classifying and ranking cholesterol-binding sites on membrane proteins, achieving performance comparable to computationally expensive MD simulations while offering rich biophysical insights into the atomic-level spatial patterns beyond amino-acid sequence. This work lays the foundation for future deep-learning models targeting membrane protein drug-binding sites and cholesterol-modulated therapeutics. Cholesterol is a key modulator of membrane protein structure and function, yet predicting cholesterol-binding sites remains challenging due to the relatively limited data on cholesterol-binding pockets compared to drug-like ligand binding sites. Here, the authors report CholBindNet, an interpretable, atom-based graph neural network that relies on a positive-unlabeled (PU) training strategy, demonstrating its superior performance over existing models and assessing strong, moderate, and weak cholesterol-binding sites in the PIEZO2 ion channel against all-atom molecular dynamics simulations.