TRAP

TRAP predicts CD8+ T-cell epitopes presented by HLA-I using deep learning and transfer learning to estimate peptide immunogenicity and T-cell recognition potential.

Key Features:

• Deep Learning Architecture: Employs a deep learning framework with transfer learning to learn immunogenicity-associated patterns from limited datasets and mitigate HLA bias for pathology-specific CD8+ T-cell epitope prediction.
• MHC-I Binding Information Utilization: Integrates MHC-I binding data to generate context-specific predictions of which peptides are likely presented by HLA-I and recognized by CD8+ T-cells.
• Low-Confidence Prediction Detection: Implements a mechanism to identify and abstain on low-confidence predictions for peptides that deviate substantially from training data.
• RSAT Metric Development: Computes RSAT (relative similarity to autoantigens and tumour-associated antigens) to complement "dissimilarity to self" approaches and refine immunogenicity estimation.
• Limited-data Generalization: Extracts immunogenicity-associated properties from limited data on emerging pathogens to enable translation of predictions to related species.
• Imbalanced Dataset and Multi-length/Species Robustness: Reduces loss of potential epitopes in imbalanced datasets and provides immunogenicity estimates across peptide lengths and species.

Scientific Applications:

• Cancer neoantigen discovery: Applied to identify epitopes from glioblastoma patient data to prioritize putative tumor-associated CD8+ T-cell targets.
• Infectious disease epitope identification: Used to identify immunogenic peptides from SARS-CoV-2 and to extract properties useful for emerging pathogen analysis across related species.
• Comparative algorithm evaluation: Demonstrated superior performance relative to existing algorithms in both cancer and infectious disease contexts.
• Immunogenicity estimation for translational research: Provides peptide-level immunogenicity estimates to inform immunotherapy and vaccine research across peptide lengths and species.

Methodology:

Applies deep learning with transfer learning, integrates MHC-I binding information, implements low-confidence detection/abstention, and computes the RSAT metric (relative similarity to autoantigens and tumour-associated antigens).