pLM4Alg: Protein Language Model-Based Predictors for Allergenic Proteins and Peptides

Area of Science:

• Bioinformatics
• Computational Biology
• Allergy Research

Background:

• Allergy prevalence necessitates advanced bioinformatics tools for allergen identification.
• Pretrained protein language models (pLMs) show promise in predicting protein structure and function.
• Existing pLMs have not been applied to allergenic protein or peptide prediction.

Purpose of the Study:

• Develop robust models for allergenic protein/peptide prediction using pLMs.
• Assess model performance via fine-tuning with a convolutional neural network.
• Introduce a user-friendly web server for accessible prediction.

Main Methods:

• Utilized five pretrained protein language models (pLMs) of varying sizes.
• Fine-tuned pLMs with a convolutional neural network for allergenic prediction.
• Evaluated model performance using accuracy, MCC, and AUC metrics.

Main Results:

• Achieved state-of-the-art performance with accuracy (93.4-95.1%), MCC (0.869-0.902), and AUC (0.981-0.990).
• pLM4Alg models demonstrated capability in handling residue-missed and nonstandard amino acid sequences.
• A user-friendly web server was established for public access.

Conclusions:

• pLM4Alg represents a significant advancement in machine learning-based allergenic prediction.
• The models are expected to become leading tools for identifying allergenic peptides and proteins.
• Integration with other predictors can further accelerate allergy research.