DeepAttentionPan

DeepAttentionPan predicts peptide binding affinity to Major Histocompatibility Complex class I (MHC-I) proteins using convolutional neural networks combined with attention mechanisms to provide interpretable and accurate predictions across polymorphic HLA alleles.

Key Features:

• Attention Mechanism Integration: Incorporates attention mechanisms to highlight peptide positional contributions to MHC-I binding affinity.
• Convolutional Neural Networks and Ensemble: Utilizes convolutional neural networks with an ensemble of 20 trained networks to deliver flexible and stable prediction performance.
• State-of-the-Art Performance on IEDB Benchmark: Demonstrates state-of-the-art results on 21 test allele datasets from the IEDB weekly benchmark dataset.
• Transfer Learning Capability: Supports transfer learning for allele-specific fine-tuning when sample sizes are limited.
• Interpretability via Positional Attention Weights: Generates peptide positional attention weights that provide mechanistic insights aligned with experimentally verified binding motifs and anchors.

Scientific Applications:

• Vaccine Design: Predicts peptide–MHC interactions to identify candidate binders across diverse HLA alleles for therapeutic vaccine development.
• Immune Response Characterization: Facilitates characterization of immune responses by identifying key peptide positions that influence MHC-I binding.

Methodology:

End-to-end deep learning approach employing convolutional neural networks combined with attention mechanisms, an ensemble of 20 trained networks, transfer learning for allele-specific fine-tuning, evaluation on the IEDB weekly benchmark dataset, and analysis of peptide positional attention weights.