Alg-MFDL: A multi-feature deep learning framework for allergenic proteins prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Immunology

Background:

• Global incidence of allergies is rising, impacting public health.
• Identifying allergens is crucial for prevention but traditional lab methods are slow and costly.
• There is a need for efficient computational methods to predict allergenic proteins.

Purpose of the Study:

• To develop a novel computational predictor, Alg-MFDL, for identifying allergenic proteins.
• To integrate pre-trained protein language models (PLMs) and handcrafted features for enhanced protein representation.
• To establish a reliable and efficient tool for allergen prediction.

Main Methods:

• Compared eight pre-trained PLMs (ProtTrans, ESM-2) and selected optimal models.
• Evaluated three handcrafted features and their combinations to find the best feature set.
• Fused protein representations and trained a convolutional neural network (CNN).
• Validated Alg-MFDL on benchmark datasets.

Main Results:

• Alg-MFDL achieved high performance: 0.973 accuracy, 0.996 precision, 0.951 sensitivity, and 0.973 F1-score.
• The model outperformed most current state-of-the-art methods across key metrics.
• The integration of PLMs and handcrafted features led to a more complete protein representation.

Conclusions:

• Alg-MFDL is a highly accurate and efficient tool for predicting allergenic proteins.
• The developed model surpasses existing methods in allergen identification.
• Alg-MFDL offers a valuable computational approach for allergy research and prevention.