Protein-Protein Interaction Prediction for Targeted Protein Degradation

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Protein-protein interactions (PPIs) are crucial for biological functions and disease understanding.
• Experimental PPI site detection is costly and time-consuming.
• Current machine learning methods for PPI prediction are limited by sequence-based approaches and not applied to targeted protein degradation.

Purpose of the Study:

• To develop a novel deep learning architecture for predicting protein-protein interaction sites and interactions using surface representations.
• To evaluate the model's performance on established datasets and assess its generalization capabilities.
• To apply the model to predict PPIs relevant for targeted protein degradation drug development.

Main Methods:

• Developed a deep learning architecture based on graph representation learning.
• Utilized protein surface representations as input for the model.
• Evaluated performance using AUROC scores on the MaSIF dataset and a new, diverse PPI dataset.
• Assessed accuracy on ternary complex data for targeted protein degradation applications.

Main Results:

• The model achieved state-of-the-art performance on the MaSIF dataset.
• Demonstrated strong generalization capabilities on a new, diverse dataset of protein interactions.
• Showed high accuracy in predicting PPIs relevant for targeted protein degradation, including ternary complex data.

Conclusions:

• The proposed deep learning model accurately predicts protein-protein interaction sites and interactions from surface representations.
• The model's ability to generalize and predict PPIs for targeted protein degradation makes it a valuable tool for drug discovery.
• This approach offers a computationally efficient alternative to experimental methods for screening protein pairs in drug development.