Accurate PROTAC-targeted degradation prediction with DegradeMaster.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Proteolysis-targeting chimeras (PROTACs) offer a novel approach to degrade disease-causing proteins, including previously
• undruggable
• targets.
• Current deep learning methods for predicting PROTAC efficacy often overlook crucial 3D structural information or rely on limited labeled datasets.

Purpose of the Study:

• To develop a more accurate computational method for predicting PROTAC-targeted protein degradation.
• To address the limitations of existing deep learning models by incorporating 3D spatial information and leveraging unlabeled data.

Main Methods:

• Developed DegradeMaster, a semisupervised E(3)-equivariant graph neural network-based predictor.
• Incorporated 3D geometric constraints using an E(3)-equivariant graph encoder.
• Utilized a memory-based pseudolabeling strategy to enhance training with unlabeled data.
• Designed a mutual attention pooling module for interpretable graph representations.

Main Results:

• DegradeMaster significantly outperformed state-of-the-art baselines, improving AUROC by 10.5%.
• Achieved high accuracy in predicting degradability for specific PROTAC candidates (e.g., 88.33% for VZ185 on BRD9).
• Demonstrated the importance of structural information in warhead and E3 ligand regions through attention weight visualization.

Conclusions:

• DegradeMaster shows significant potential for accelerating the discovery of novel PROTAC compounds.
• The model's ability to interpret functional PROTAC components highlights the value of 3D structural data in prediction.
• This work provides a powerful tool for advancing targeted protein degradation strategies.