The electrostatic landscape of MHC-peptide binding revealed using inception networks

Area of Science:

• Biophysical Sciences
• Computational Biology
• Immunoinformatics

Background:

• Macromolecular recognition and protein-protein interactions are crucial in biophysics but complex to model.
• Peptide presentation by polymorphic Major Histocompatibility Complex class I (MHC-I) molecules presents a significant modeling challenge.
• Accurate prediction of peptide-MHC interactions is vital for understanding immune responses.

Purpose of the Study:

• To develop a novel deep learning model for predicting peptide binding motifs across a wide range of MHC-I alleles.
• To integrate molecular electrostatics into a deep convolutional neural network for enhanced prediction accuracy.
• To apply the predictive model to analyze associations with HIV disease progression and immune checkpoint inhibitor response.

Main Methods:

• Developed HLA-Inception, a deep biophysical convolutional neural network.
• Integrated molecular electrostatics to capture non-bonded interactions.
• Trained and validated the model across 5,821 human leukocyte antigen (HLA) alleles.

Main Results:

• HLA-Inception accurately predicts peptide binding motifs for numerous HLA alleles.
• Model predictions show strong correlation with experimental peptide binding and presentation data.
• Demonstrated utility in analyzing HLA associations with HIV progression and treatment response.

Conclusions:

• Deep learning models integrating biophysical properties can effectively predict peptide-MHC binding motifs.
• HLA-Inception offers a powerful tool for immunoinformatics and personalized medicine.
• The model facilitates deeper understanding of immune recognition and its clinical implications.