DFL-MHC: MHC identification model based on dual-stage training and multi-view feature fusion

Area of Science:

• Immunology
• Bioinformatics
• Computational Biology

Background:

• The major histocompatibility complex (MHC) is critical for adaptive immunity, influencing antigen presentation, immune surveillance, and disease susceptibility.
• Current MHC identification methods often rely on manually engineered features or single protein language models (PLMs), limiting their ability to capture comprehensive information.
• Existing models struggle to model deep semantic dependencies within sequences due to conventional machine learning algorithms or simple classifiers.

Purpose of the Study:

• To develop an advanced MHC identification model that overcomes the limitations of existing methods.
• To introduce a novel framework, DFL-MHC, that integrates multi-sequence and multi-model views through dual-stage training and multi-view feature fusion.
• To enhance the accuracy and reliability of MHC identification for immunological research and clinical applications.

Main Methods:

• A dual-stage training strategy combining feature extraction and feature modeling.
• A cross-sequence and cross-model multi-view scheme for feature extraction, fusing information from two PLMs applied to truncated protein sequences.
• Dimensionality reduction to obtain an optimal feature subset, followed by a bi-directional long short-term memory (BiLSTM) network with an attention mechanism for deep semantic dependency modeling.

Main Results:

• The DFL-MHC framework effectively captures complementary information across different sequence lengths and PLMs.
• The BiLSTM network with attention successfully models long-range and deep semantic dependencies within sequences.
• DFL-MHC demonstrates superior performance in MHC identification compared to existing methods.

Conclusions:

• Multi-view feature fusion and dual-stage training are effective strategies for accurate and reliable MHC identification.
• The DFL-MHC model offers a significant advancement in MHC identification, benefiting immunological research and clinical practice.
• This approach highlights the potential of integrating diverse data views and advanced deep learning architectures for complex biological sequence analysis.