Predicting MHC class I binder: existing approaches and a novel recurrent neural network solution

Area of Science:

• Immunoinformatics
• Computational Biology
• Genomics

Background:

• Major Histocompatibility Complex (MHC) is crucial for understanding complex human diseases.
• Numerous computational tools exist for predicting MHC class I binders, but they have limitations.
• Existing tools typically require fixed peptide sequence lengths, hindering comprehensive analysis.

Purpose of the Study:

• To review and evaluate existing MHC class I binding prediction tools.
• To address the limitation of fixed peptide sequence lengths in current predictors.
• To develop a novel, accurate, and efficient predictor for variable-length MHC class I binders.

Main Methods:

• Comprehensive review of 27 MHC class I binding prediction tools.
• Evaluation of feature representation, prediction algorithms, and model training strategies.
• Development of a novel bilateral and variable long short-term memory (BVLSTM)-based approach (BVLSTM-MHC).

Main Results:

• Identified fixed peptide sequence length as a common limitation in existing tools.
• BVLSTM-MHC demonstrated superior performance compared to 10 mainstream predictors across six of eight metrics.
• Achieved the best performance on an independent validation dataset.

Conclusions:

• BVLSTM-MHC is the first variable-length MHC class I binding predictor.
• The developed model offers enhanced accuracy and efficiency for MHC class I binder prediction.
• A web server is available for predicting MHC class I binders in multiple species.