Towards designing improved cancer immunotherapy targets with a peptide-MHC-I presentation model, HLApollo

Area of Science:

• Oncology
• Immunology
• Computational Biology

Background:

• Cancer immunotherapy success drives demand for personalized cancer vaccines.
• Accurate prediction of antigen presentation on MHC class I (MHC-I) is critical for effective T-cell responses against cancer.
• Computational methods are needed to model peptide-MHC-I (pMHC-I) interactions.

Purpose of the Study:

• Introduce HLApollo, a novel transformer-based model for pMHC-I presentation prediction.
• Enhance the accuracy and scope of computational pMHC-I binding predictions.
• Improve the development of personalized cancer vaccines and immunotherapies.

Main Methods:

• Developed HLApollo, a transformer model utilizing peptide and MHC sequences.
• Implemented end-to-end sequence processing and deconvolution of multi-allelic data.
• Employed a negative-set switching strategy for unlabelled ligandome data.
• Integrated protein features from protein language models.

Main Results:

• Achieved a 12.65% increase in average precision (AP) on ligandome data.
• Demonstrated a 4.1% AP increase on immunogenicity test data compared to existing models.
• Showed improved performance by incorporating protein language model features.
• Validated pan-allelic generalization, including rare alleles.

Conclusions:

• HLApollo significantly advances pMHC-I presentation prediction accuracy.
• The model's ability to generalize across alleles is crucial for clinical applications.
• HLApollo offers a powerful tool for developing next-generation personalized cancer vaccines.