MHCAttnNet

MHCAttnNet predicts binding affinity between major histocompatibility complex (MHC) alleles and peptides to identify high-affinity peptide–MHC interactions relevant for personalized cancer vaccine design.

Key Features:

• Attention-based deep neural model: Employs an attention mechanism within a deep neural network to capture and emphasize relevant amino acid subsequences in peptides and MHC alleles.
• Bi-directional Long Short-Term Memory (Bi-LSTM) encoder: Uses a Bi-LSTM encoder to handle variable-length peptide sequences and improve generalizability across diverse peptides and MHC alleles.
• Performance metrics: Achieves an AUC-PRC of 94.18% for class I MHC alleles, reports a higher F1-score than comparator models, and covers a broader range of class II MHC alleles.
• Attention mechanism visualization: Generates attention heatmaps over amino acids to highlight important subsequences and trigrams within peptide and MHC allele sequences.
• Reduction in experimental load: Reduces candidate subsequences from 9,251 possible trigrams to approximately 258 significant trigrams to limit the number of subsequences requiring experimental testing.

Scientific Applications:

• Personalized cancer vaccine design: Identifies peptides with high-affinity binding to patient-specific MHC alleles to support design and synthesis of personalized cancer vaccines.
• Target prioritization for immunotherapy: Prioritizes peptide–MHC candidates for experimental validation to streamline selection of targets that enhance immune recognition of cancer cells.

Methodology:

Trains a deep neural network with an attention mechanism and a Bi-LSTM encoder on datasets of peptide and MHC allele sequences, learns subsequence patterns and trigrams relevant for binding affinity, and produces attention heatmaps while selecting approximately 258 significant trigrams from 9,251 candidates.