DeepMHCI: an anchor position-aware deep interaction model for accurate MHC-I peptide binding affinity prediction

Area of Science:

• Immunological bioinformatics
• Computational biology
• Cancer immunotherapy

Background:

• Predicting MHC-I peptide binding is crucial for personalized cancer vaccines.
• Current deep learning methods struggle with non-9-mer peptides due to input representation limitations.
• Accurate prediction requires understanding anchor positions in MHC binding motifs.

Purpose of the Study:

• Develop a high-performance deep model for MHC-I peptide binding affinity prediction.
• Improve accuracy, especially for variable-length peptides (non-9-mers).
• Enhance the identification of neoantigens for therapeutic cancer vaccines.

Main Methods:

• Developed DeepMHCI, an anchor position-aware deep learning model.
• Incorporated a position-wise gated layer for controlled peptide information flow.
• Utilized a residual binding interaction convolution layer to model peptide-MHC interactions.

Main Results:

• DeepMHCI demonstrated superior performance across four benchmark datasets.
• The model excelled in predicting binding affinity for non-9-mer peptides.
• Experimental validation included cross-validation, independent testing, and external vaccine/epitope identification.

Conclusions:

• DeepMHCI offers a significant advancement in MHC-I peptide binding prediction.
• The anchor position-aware approach effectively addresses limitations of previous methods.
• The model provides a valuable tool for neoantigen discovery and personalized vaccine development.