GAPS: a geometric attention-based network for peptide binding site identification by the transfer learning approach

Area of Science:

• Computational Biology
• Drug Discovery
• Structural Bioinformatics

Background:

• Protein-peptide interactions (PPepIs) are crucial for cellular functions and drug design.
• Accurate identification of protein-peptide binding sites is essential for understanding PPepI mechanisms.
• Existing experimental and computational methods for binding site identification have limitations in accuracy, generality, or efficiency.

Purpose of the Study:

• To develop a novel computational model for accurate protein-peptide binding site identification.
• To address the drawbacks of current computational tools for PPepI research.
• To enhance the understanding of molecular interactions in biological systems.

Main Methods:

• Developed a geometric attention-based network named GAPS.
• Utilized geometric feature engineering for atom representations.
• Incorporated multiple attention mechanisms to update biological features.
• Implemented transfer learning from protein-protein binding site data.

Main Results:

• GAPS achieved state-of-the-art performance and robustness in protein-peptide binding site identification.
• The model demonstrated exceptional performance in predicting apo protein-peptide, protein-cyclic peptide, and AlphaFold-predicted binding sites.
• Transfer learning effectively leveraged protein-protein interaction data to improve PPepI recognition.

Conclusions:

• GAPS is a powerful, versatile, and stable method for diverse binding site predictions.
• The model offers a significant advancement over existing computational approaches.
• This work facilitates more efficient and accurate research into protein-peptide interactions for drug discovery.