DeepSeqPan, a novel deep convolutional neural network model for pan-specific class I HLA-peptide binding affinity...

Area of Science:

• Immunoinformatics
• Computational Biology
• Genomics

Background:

• Human leukocyte antigen (HLA)-peptide interactions are crucial for immune responses.
• Accurate prediction of HLA-binding peptides is vital for peptide drug discovery.
• Current models often rely on pseudo-sequence encoding derived from limited structural data.

Purpose of the Study:

• To develop a novel deep convolutional neural network (DCNN) model for predicting HLA-peptide binding.
• To overcome limitations of existing models by learning HLA sequence and binding context directly from data.
• To achieve state-of-the-art performance without requiring HLA structural information.

Main Methods:

• Proposed a Deep Convolutional Neural Network (DCNN) model named DeepSeqPan.
• Developed a binding context extraction layer within the DCNN architecture.
• Implemented dual outputs for both binding affinity and binding probability predictions.
• Trained and evaluated the model on public benchmark datasets.

Main Results:

• DeepSeqPan achieved state-of-the-art performance across numerous HLA alleles.
• The model demonstrated strong generalization capabilities on unseen data.
• Performance was achieved without utilizing HLA structural information during training.
• The model requires only raw peptide and HLA sequences for prediction.

Conclusions:

• The proposed DCNN model, DeepSeqPan, offers a powerful and flexible approach for HLA-peptide binding prediction.
• The method's ability to learn directly from sequence data enhances its applicability to diverse HLA alleles, including those lacking structural information.
• The model's architecture and performance pave the way for broader applications in other protein-binding prediction tasks, such as protein-DNA and protein-RNA interactions.