PepDist: a new framework for protein-peptide binding prediction based on learning peptide distance functions

Area of Science:

• Computational biology
• Immunoinformatics
• Machine learning

Background:

• Protein-peptide interactions are crucial for cellular functions, including immune responses via MHC-peptide complexes.
• Accurate prediction of protein-peptide binding is vital for vaccine and drug development.

Purpose of the Study:

• To develop a novel computational approach, PepDist, for predicting protein-peptide binding affinity.
• To improve upon existing methods by learning a generalized distance function for protein families.

Main Methods:

• PepDist utilizes a semi-supervised learning framework with partial information (equivalence constraints).
• The core method involves learning a single peptide-peptide distance function applicable to entire protein families (e.g., MHC class I).
• DistBoost, a semi-supervised distance learning algorithm, is employed.

Main Results:

• PepDist significantly outperforms state-of-the-art methods on MHC class I and II datasets.
• The approach demonstrates superior performance, especially for proteins with limited labeled peptide data.
• A webserver (http://www.pepdist.cs.huji.ac.il) is available for predicting peptide binding to 35 MHC class I alleles.

Conclusions:

• Learning a single distance function across a protein family enhances prediction accuracy compared to individual classifiers.
• The study highlights the critical importance of experimentally determined non-binder data for better model generalization.
• Publication and public availability of non-binding peptide data are recommended to advance computational prediction models.