Hydrophobicity identifies false positives and false negatives in peptide-MHC binding

Area of Science:

• Immunology and Bioinformatics
• Computational Biology and Biochemistry

Background:

• Major Histocompatibility Complex (MHC) Class I molecules present peptides to T-cells for immune surveillance.
• Computational tools like NetMHC-4.0 and NetMHCpan-4.1 predict peptide-MHC binding but lack biochemical property considerations.
• Hydrophobicity is a key biochemical factor in peptide-MHC interactions, yet is not explicitly modeled in current prediction tools.

Purpose of the Study:

• To investigate the correlation between peptide hydrophobicity and predicted binding affinity to specific MHC Class I variants (HLA-A*0201, HLA-B*2705, HLA-B*0801).
• To identify and analyze biases in MHC-binding prediction tools regarding peptide hydrophobicity.
• To explore the impact of these biases on applications like vaccine development and neoantigen discovery.

Main Methods:

• Analysis of peptide binding predictions from NetMHC-4.0 and NetMHCpan-4.1.
• Investigation using both the original training datasets and a sample of the human proteome.
• Application of Machine Learning metrics to identify sources of prediction bias (false positives/negatives).

Main Results:

• NetMHC-4.0 demonstrated a statistically significant bias towards predicting highly hydrophobic peptides as strong binders for HLA-A*0201 and HLA-B*2705.
• This bias was observed in both the training data and the human proteome sample.
• The bias was characterized by an excess of hydrophobic false positives and a deficit of hydrophilic false negatives.

Conclusions:

• Current neural network-based MHC-binding prediction tools exhibit a significant hydrophobicity bias, particularly for certain HLA types.
• Retraining these models with biochemical features like hydrophobicity and improved datasets could enhance prediction accuracy.
• Improved prediction accuracy holds potential for advancing vaccine design and neoantigen identification.