Investigating the human and nonobese diabetic mouse MHC class II immunopeptidome using protein language modeling

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Major histocompatibility complex class II (MHCII)-peptide identification is crucial for evaluating immunoregulatory therapeutics.
• Accurate MHCII-peptide presentation prediction has significant biopharmaceutical applications, including drug safety assessment and immunomodulatory drug discovery.
• Large datasets on immune responses and MHC-associated peptide proteomics have been compiled, alongside advances in deep learning for sequence data analysis.

Purpose of the Study:

• To train and evaluate a compact transformer model (AEGIS) for MHCII-peptide presentation prediction using human and mouse data.
• To assess the model's performance against existing deep learning algorithms and investigate the impact of cross-organism dataset integration.
• To demonstrate the model's capability in predicting peptides in a preclinical type 1 diabetes model.

Main Methods:

• A compact transformer model, AEGIS, was trained on human and mouse MHCII immunopeptidome data.
• Model performance was evaluated on peptide presentation prediction, including comparisons with existing deep learning methods.
• Variants of the model were trained with and without MHCII information, and the impact of including I-Ag7 presented peptides was assessed.

Main Results:

• The AEGIS transformer model achieved performance comparable to existing deep learning algorithms for MHCII-peptide prediction.
• Combining datasets from multiple organisms enhanced the model's predictive performance.
• The inclusion of I-Ag7 presented peptides enabled accurate in silico prediction of presented peptides in a preclinical type 1 diabetes model for the first time.

Conclusions:

• The AEGIS model provides accurate MHCII-peptide presentation prediction, performing on par with state-of-the-art deep learning methods.
• Cross-organism data integration improves model performance, highlighting the value of diverse datasets.
• The model's success in predicting peptides in a preclinical type 1 diabetes model offers promising therapeutic applications for immune-related diseases.