DapNet-HLA

DapNet-HLA predicts non-classical Human Leukocyte Antigen (HLA) binding sites using an adaptive dual-attention deep learning network to improve allele-specific binding prediction.

Key Features:

• Adaptive Dual-Attention Mechanism: An adaptive dual-attention network focuses on relevant sequence features to enhance predictive accuracy.
• Amino Acid Sequence Transformation: Amino acid sequences are converted into digital vectors via a lookup table and fused word embedding to mitigate one-hot sparsity, mapping each residue to a low-dimensional vector optimized during classifier training.
• Group Convolution Block (GCB): GCBs capture complex local patterns in sequence-derived feature maps.
• SENet Attention: Squeeze-and-excitation (SENet) modules recalibrate feature maps to increase the weight of informative features.
• BiLSTM: Bidirectional long short-term memory networks model dependencies in both forward and backward sequence directions.
• Bahdanau Attention Mechanism: A Bahdanau attention mechanism improves recurrent network focus on relevant input regions.
• Performance Metrics and Validation: Evaluated on five datasets with 5-fold cross-validation, showing an ACC improvement of 4.89% and an MCC increase of 0.0933 versus existing methods, alongside ROC and PR curve analysis with reported AUC fluctuations.

Scientific Applications:

• Cancer immunotherapy target identification: Prediction of non-classical HLA binding sites to support identification of immunotherapy targets in cancer.
• COVID-19 immune response analysis: Application to studies of immune responses in COVID-19 through predicted HLA–peptide interactions.
• Allergy research: Contribution to investigations of immune mechanisms relevant to allergic diseases via non-classical HLA binding predictions.

Methodology:

Amino acid sequences are vectorized using a lookup table and fused word embeddings into low-dimensional vectors optimized during classifier training; these features are processed by an adaptive dual-attention deep learning architecture incorporating Group Convolution Block, SENet, BiLSTM, and Bahdanau attention, and the model is assessed by 5-fold cross-validation with ROC/PR curve and AUC, ACC, and MCC reporting.