CAT-Site: Predicting Protein Binding Sites Using a Convolutional Neural Network.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate identification of protein binding sites is crucial for computer-assisted drug design.
• Existing methods often rely on limited ligand data, potentially limiting generalization.
• Understanding binding sites aids in predicting drug efficacy and potential side effects.

Purpose of the Study:

• To develop a novel workflow for predicting ligand binding sites on protein surfaces.
• To generalize binding site prediction by utilizing all available ligands for similar proteins.
• To improve the accuracy and reliability of computational drug design tools.

Main Methods:

• Protein clustering and merging of ligands from PDBbind and sc-PDB databases.
• Application of a three-dimensional convolutional neural network (3D-CNN) considering protein spatial structure.
• Integration of ligandability predictions to define joint binding sites.

Main Results:

• Achieved sensitivity of 0.829, specificity of 0.98, and F1 score of 0.517.
• For 54% of large, pharmacologically relevant binding sites, the predicted center is within 4 Å of the actual center.
• The model demonstrates robust performance in identifying potential drug interaction regions.

Conclusions:

• The proposed workflow effectively predicts protein binding sites using a 3D-CNN.
• This method offers a generalized approach by leveraging comprehensive ligand data.
• The findings contribute to more accurate and efficient drug design processes.