DeepNetBim: deep learning model for predicting HLA-epitope interactions based on network analysis by harnessing...

Area of Science:

• Computational immunology
• Bioinformatics
• Machine learning in drug discovery

Background:

• Epitope prediction is crucial for cancer immunology and immunotherapy, aiding vaccine development.
• Current computational methods often fail to identify immunogenic epitopes, limiting clinical applications.
• Understanding both binding and immunogenic potential of human leukocyte antigen (HLA)-peptide pairs is essential for effective vaccines.

Purpose of the Study:

• To develop a novel network-based deep learning method, DeepNetBim, for predicting HLA-peptide interactions.
• To integrate binding and immunogenic data within a network framework to improve epitope prediction accuracy.
• To assess the clinical utility of DeepNetBim in identifying potent and immunogenic epitopes for vaccine design.

Main Methods:

• Retrieved quantitative HLA-peptide binding and qualitative immunogenic data from the Immune Epitope Database.
• Constructed weighted HLA-peptide binding and immunogenic networks.
• Integrated network features into a deep learning algorithm combining convolutional neural networks and attention mechanisms.

Main Results:

• DeepNetBim significantly outperformed existing models in HLA-peptide binding prediction, achieving an AUC score of 93.74%.
• Integration of network centrality metrics enhanced both binding and immunogenicity predictions.
• The combined model demonstrated improved performance in neoantigen identification, increasing positive predictive value and neoantigen recognition.

Conclusions:

• DeepNetBim, a network-based deep learning tool, effectively predicts HLA-peptide interactions by extracting network attributes.
• The combined binding and immunogenic models show superior performance compared to individual models and other state-of-the-art methods.
• DeepNetBim holds significant potential for clinical applications in epitope-based vaccine development.