GraphMHC: Neoantigen prediction model applying the graph neural network to molecular structure

Area of Science:

• Computational biology
• Immunoinformatics
• Cancer research

Background:

• Neoantigens are crucial biomarkers for predicting patient response to immune checkpoint inhibition.
• Current prediction models using deep neural networks lack interpretability regarding biomolecular interactions.
• Accurate prediction of neoantigen-MHC binding is essential for personalized cancer therapy.

Purpose of the Study:

• To develop an interpretable graph neural network model, GraphMHC, for simulating neoantigen-MHC protein binding.
• To improve the accuracy of predicting neoantigen binding affinity and its clinical relevance.
• To provide a feature-intrinsic method for modeling biomolecular interactions in cancer immunology.

Main Methods:

• Utilized graph neural networks (GNNs) to model molecular structures of amino acid sequences from the Immune Epitope Database (IEDB).
• Converted amino acid sequences into graph representations capturing atomic information and inter-atomic connections.
• Employed stacked graph attention and convolution layers for binding classification.

Main Results:

• GraphMHC achieved high prediction performance with an Area Under the Receiver Operating Characteristic Curve (AUC) of 92.2%.
• The model surpassed baseline prediction models in accuracy.
• Application to The Cancer Genome Atlas (TCGA) melanoma data revealed significant differences in stromal score and borderline differences in overall survival between high and low neoantigen load groups, a distinction missed by the baseline model.

Conclusions:

• GraphMHC offers the first feature-intrinsic GNN-based method for modeling neoantigen-MHC binding using molecular structures.
• The model provides highly accurate, interpretable insights into neoantigen binding.
• GraphMHC has the potential to significantly improve prognosis prediction for cancer patients undergoing immune checkpoint inhibitor therapy.