Improving MHC class I antigen-processing predictions using representation learning and cleavage site-specific...

Area of Science:

• Immunoinformatics
• Computational Biology
• Machine Learning

Background:

• Peptide processing and presentation by Major Histocompatibility Complex (MHC) class I molecules are critical for adaptive immunity.
• Accurate prediction of this process is essential for developing effective vaccines and targeted therapies.
• Existing computational tools have limitations in predicting peptide processing efficiency.

Purpose of the Study:

• To introduce MHCrank, a novel deep-learning model for predicting peptide processing probability for MHC class I presentation.
• To evaluate MHCrank's performance against established baseline methods.

Main Methods:

• Developed a deep-learning model incorporating cleavage site-specific kernels and learned amino acid embeddings.
• Utilized visualization of site-specific amino acid enrichment patterns.
• Analyzed cosine similarity of learned embeddings against known physiochemical properties.

Main Results:

• MHCrank significantly outperformed MHCflurry and netMHCpan in predictive performance.
• Visualizations revealed biologically relevant amino acid enrichments at key positions.
• Learned embeddings showed high correlation with experimentally validated physiochemical properties crucial for peptide processing.

Conclusions:

• MHCrank offers superior performance in predicting peptide processing for MHC class I presentation compared to current methods.
• The model's insights into amino acid preferences and physiochemical correlations enhance understanding of the processing mechanism.
• MHCrank holds significant potential for advancing drug discovery and vaccine design.