DeepNetBim

DeepNetBim predicts HLA-peptide interactions and their immunogenic potential using network-based deep learning to support epitope binding prediction and neoantigen identification.

Key Features:

• Network-based representation: Models HLA molecules and peptides as nodes in a weighted HLA-peptide binding network to capture complex interaction propensities.
• Integration of binding and immunogenic data: Combines quantitative class I HLA-peptide binding data with qualitative immunogenic data from T cell activation, MHC binding, and MHC ligand elution assays sourced from the Immune Epitope Database (IEDB).
• Deep learning architecture: Employs a convolutional neural network (CNN) enhanced with an attention mechanism.
• Network centrality features: Integrates network centrality metrics into the model to improve prediction of both binding and immunogenicity versus models lacking these features or using shuffled networks.
• Performance metrics: Reported AUC of 93.74% for HLA-peptide binding prediction, outperformed 11 state-of-the-art models, and showed improved positive predictive value (PPV) and increased neoantigen recognition when filtering negative immunogenic predictions.

Scientific Applications:

• Pan-specific epitope prediction: Enables prediction of epitopes across multiple HLA alleles using integrated binding and immunogenicity information.
• Vaccine and immunotherapy design: Informs selection of candidate peptides with favorable binding and immunogenicity profiles for vaccine and therapeutic development.
• Neoantigen identification for personalized cancer immunotherapy: Enhances identification and prioritization of patient-specific tumor neoantigens by combining binding and immunogenicity predictions.

Methodology:

Constructs weighted HLA-peptide binding and immunogenic networks, integrates these networks into a CNN with attention, and extracts features from network attributes of both binding and immunogenic models, with combined dual-model predictions used to improve performance.