ProteinMAE: masked autoencoder for protein surface self-supervised learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein functions are dictated by surface properties, crucial for tasks like protein design and interaction prediction.
• Current deep learning methods for protein surface analysis are hindered by limited experimental data (label scarcity).
• Self-supervised learning (SSL) has shown promise in overcoming data limitations in other fields.

Purpose of the Study:

• To introduce ProteinMAE, a novel self-supervised framework for protein surface representation.
• To address the challenge of label scarcity in learning-based protein surface analysis.
• To develop a computationally efficient method for pretraining protein surface models.

Main Methods:

• Developed an efficient network architecture for protein surface representation.
• Utilized a large corpus of unlabeled protein data for self-supervised pretraining (ProteinMAE).
• Fine-tuned the pretrained model on downstream tasks: binding site identification, protein pocket classification, and protein-protein interaction prediction.

Main Results:

• ProteinMAE significantly improved performance across all evaluated downstream tasks.
• The method achieved competitive results compared to state-of-the-art approaches.
• The ProteinMAE network demonstrated substantial computational advantages, requiring less than 10% of the memory cost of previous methods.

Conclusions:

• ProteinMAE effectively mitigates label scarcity in protein surface representation through self-supervised learning.
• The framework offers a powerful and computationally efficient alternative for analyzing protein surfaces.
• This approach holds significant potential for advancing various applications in structural biology and drug discovery.