Classification of Protein-Binding Sites Using a Spherical Convolutional Neural Network.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Protein-binding site analysis is crucial for drug discovery applications like hit finding and repurposing.
• Classifying binding sites has been a long-standing challenge, with numerous methods developed over 30 years.
• Advancements in machine learning and structural data availability enable deep learning for binding site comparison.

Purpose of the Study:

• To develop and validate a deep learning model for protein-binding site representation and comparison.
• To assess the model's capability in similarity and classification tasks.
• To evaluate the model's potential for drug discovery, including lead hopping.

Main Methods:

• Utilized a spherical convolutional neural network based on the DeepSphere architecture.
• Trained the model on TOUGH-C1 and TOUGH-M1 datasets.
• Validated the model using the ProSPECCTs datasets and tested on kinase families.

Main Results:

• The model demonstrated strong performance in protein-binding site similarity and classification.
• The model successfully learned and differentiated physicochemical properties of binding sites.
• Effective clustering of kinase subfamilies was achieved, highlighting the model's practical application.

Conclusions:

• The developed deep learning method offers a powerful tool for analyzing protein-binding sites.
• The model shows significant promise for advancing drug discovery processes, particularly in lead hopping.
• This approach enables direct binding site information utilization for identifying novel drug candidates.