CholBindNet: Interpretable Neural Networks for Cholesterol Binding Site Prediction.

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Cholesterol is crucial for membrane protein function, but predicting its binding sites is difficult.
• Existing deep-learning models struggle with membrane proteins, especially for cholesterol binding.

Purpose of the Study:

• To develop an accurate and efficient deep-learning framework for predicting cholesterol binding sites on transmembrane proteins.
• To provide a scalable alternative to computationally intensive molecular dynamics simulations.

Main Methods:

• Curated over 800 high-resolution transmembrane protein structures with cholesterol.
• Developed CholBindNet, an atom-based deep-learning framework with four model architectures.
• Employed a Positive-Unlabeled (PU) training strategy to handle scarce negative samples.

Main Results:

• CholBindNet significantly outperformed existing deep-learning models for ligand-binding prediction.
• The model accurately ranked cholesterol-binding sites (strong, median, weak) in the PIEZO2 ion channel, matching molecular dynamics simulations.
• Achieved high model interpretability through atom-level features, Grad-CAM, and attention analysis.

Conclusions:

• CholBindNet offers an efficient and scalable method for predicting cholesterol binding sites on membrane proteins.
• The framework provides mechanistic biophysical insights, aiding drug design for membrane protein targets.
• This work facilitates the study of lipid regulation in membrane proteins and the development of cholesterol-modulated therapeutics.