GeoPep: A Geometry-Aware Masked Language Model for Protein-Peptide Binding Site Prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein-peptide interactions are crucial in biological processes but challenging to predict computationally.
• Existing methods struggle with peptide flexibility and limited structural data for training structure-aware models.

Purpose of the Study:

• To develop a novel framework, GeoPep, for accurate peptide binding site prediction.
• To leverage transfer learning from large protein foundation models to overcome data limitations.

Main Methods:

• GeoPep utilizes transfer learning from ESM3, a multimodal protein foundation model, fine-tuning its representations for protein-peptide interactions.
• Integrates a Kolmogorov-Arnold Network (KAN) for complex nonlinear approximation.
• Employs distance-based loss functions incorporating 3D structural information for enhanced prediction.

Main Results:

• GeoPep significantly outperforms existing methods in protein-peptide binding site prediction.
• The framework effectively captures sparse and heterogeneous binding patterns.
• Demonstrates the utility of transfer learning for addressing data scarcity in specialized biological prediction tasks.

Conclusions:

• GeoPep presents a significant advancement in predicting peptide binding sites.
• The integration of foundation model transfer learning and advanced neural network architectures offers a powerful approach for biological sequence and structure prediction.
• This work paves the way for improved understanding and manipulation of protein-peptide interactions.