Graph-pMHC: graph neural network approach to MHC class II peptide presentation and antibody immunogenicity

Area of Science:

• Immunology
• Computational Biology
• Machine Learning

Background:

• Peptide-MHC class II (pMHCII) presentation is crucial for adaptive immunity but can cause anti-drug responses.
• Advances in mass spectrometry and machine learning have improved pMHCII presentation modeling.
• Accurate pMHCII prediction is vital for understanding immune responses and drug efficacy.

Purpose of the Study:

• To develop a novel graph neural network approach, graph-pMHC, for predicting pMHCII presentation.
• To improve the accuracy of pMHCII presentation modeling.
• To create a method for assessing antibody drug immunogenicity risk.

Main Methods:

• Developed graph-pMHC, a graph neural network utilizing AlphaFold2-multimer for adjacency matrices.
• Employed graph enumeration to address peptide-MHC binding groove alignment.
• Created an antibody drug immunogenicity dataset from clinical trial data.

Main Results:

• Graph-pMHC significantly outperformed existing methods, including NetMHCIIpan-4.0, with a +20.17% absolute average precision improvement.
• The developed method increased ROC-AUC by 2.57% for predicting antibody drug immunogenicity compared to OASis filtering alone.

Conclusions:

• Graph-pMHC offers a superior approach for predicting pMHCII presentation.
• The model effectively aids in evaluating the immunogenicity risk of antibody drugs.
• This work advances computational immunology and drug development strategies.