EpiMII: Structure-Aware Graph Neural Networks for MHC-II Epitope Generation

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Major histocompatibility complex class II (MHC-II) neoantigens are crucial for cancer immunotherapy, influencing T cell responses.
• Current methods for predicting and designing functional neoantigens are limited by accuracy and the scarcity of experimental data.
• Developing novel, highly immunogenic neoantigens is essential for advancing cancer vaccines and personalized treatments.

Purpose of the Study:

• To develop a structure-aware graph neural network model, EpiMII, for the de novo design of MHC-II epitopes.
• To generate mimotopes that maintain T cell specificity while improving MHC-II binding affinity and immunogenicity.
• To overcome limitations in existing neoantigen prediction tools through a structure-guided approach.

Main Methods:

• EpiMII utilizes an inverse folding strategy integrating 3D structural information to design MHC-II epitopes.
• The model was trained on a large dataset of 142,934 homology-modeled MHC-II epitope structures.
• Performance was benchmarked against existing tools using sequence recovery rates on held-out and crystallized epitope datasets.

Main Results:

• EpiMII achieved a 66.7% sequence recovery rate on a held-out test set and 79.0% on crystallized epitopes, outperforming ProteinMPNN.
• In a hepatocellular carcinoma model, EpiMII-designed epitopes activated CD4+ T cells and induced cytokine secretion (IFN-γ, TNF-α) in vitro.
• One designed epitope (P4) significantly reduced tumor volume in mice, demonstrating in vivo efficacy.

Conclusions:

• EpiMII is a powerful tool for structure-guided neoantigen discovery, enabling the de novo design of immunogenic MHC-II epitopes.
• The model's ability to enhance T cell activation and reduce tumor growth has significant implications for cancer vaccine development.
• EpiMII advances personalized immunotherapy by facilitating the design of targeted and effective neoantigen-based treatments.