GIHP: Graph convolutional neural network based interpretable pan-specific HLA-peptide binding affinity prediction.

Area of Science:

• Immunoinformatics
• Computational Biology
• Structural Biology

Background:

• Accurate prediction of Human Leukocyte Antigen (HLA)-peptide binding affinities is vital for vaccine development and immunotherapy design.
• Current sequence-based methods lack structural insights and model interpretability, hindering the identification of critical binding residues.
• Understanding HLA-peptide interactions is fundamental to deciphering adaptive immune responses.

Purpose of the Study:

• To develop an interpretable prediction method for HLA-peptide binding affinities that incorporates structural information.
• To enhance model interpretability for identifying key amino acids involved in HLA-peptide binding.
• To validate the clinical relevance of identified key residues in predicting immunotherapy outcomes.

Main Methods:

• Proposed GIHP, an interpretable graph convolutional neural network (GCNN) model.
• Represented HLA structure as an amino acid-level graph and peptide SMILE strings as atom-level graphs.
• Developed a novel visual explanation method, gradient weighted activation mapping (Grad-WAM), for identifying key binding residues.

Main Results:

• GIHP demonstrated superior prediction accuracy compared to state-of-the-art methods across multiple datasets.
• Identified key HLA-peptide binding residues were validated for their ability to distinguish immunotherapy patient survival groups.
• The identified residues successfully separated patient survival groups in breast, bladder, and pan-cancer datasets.

Conclusions:

• GIHP significantly improves the accuracy and interpretability of HLA-peptide binding predictions.
• The identified key residues hold potential for guiding personalized cancer immunotherapy strategies.
• This work provides a foundation for leveraging structural and interpretable models in immunoinformatics.