Contrastive learning of protein representations with graph neural networks for structural and functional annotations

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Protein annotation is crucial but limited, despite vast sequence data growth.
• Existing computational tools often neglect 3D structure, relying solely on sequence data.
• High-quality predicted protein structures are increasingly available (e.g., AlphaFold).

Purpose of the Study:

• To develop a general deep learning framework, PenLight, for protein structural and functional annotation.
• To leverage both 3D protein structure and advanced language model representations.
• To improve annotation accuracy and coverage compared to existing methods.

Main Methods:

• Developed PenLight, a graph neural network (GNN) framework.
• Integrated 3D protein structure data and protein language model representations.
• Employed contrastive learning for GNN training to capture semantic similarities.

Main Results:

• PenLight demonstrated superior prediction accuracy and coverage on benchmark tasks.
• Achieved better performance than state-of-the-art methods in structural classification and functional annotation.
• Learned protein representations capturing similarities beyond sequence identity.

Conclusions:

• PenLight offers a powerful new approach for protein annotation.
• Integrating 3D structure and semantic similarity learning enhances predictive capabilities.
• This framework advances computational methods for understanding the protein universe.