AllerTrans: a deep learning method for predicting the allergenicity of protein sequences.

Area of Science:

• Bioinformatics
• Computational Biology
• Allergenicity Assessment

Background:

• Allergens pose significant risks in medical products, necessitating robust safety assessments.
• Traditional allergenicity testing is costly and time-consuming.
• Bioinformatics and deep learning offer efficient alternatives for predicting protein allergenicity.

Purpose of the Study:

• To develop a superior computational model for predicting protein allergenicity.
• To enhance the safety evaluation of recombinant proteins in medical applications.

Main Methods:

• Developed an enhanced deep learning model utilizing two protein language models (pLMs) to extract feature vectors from protein sequences.
• Integrated feature vectors into a deep neural network (DNN) for classification.
• Employed ensemble modeling to combine top-performing models, balancing sensitivity and specificity.

Main Results:

• The proposed model achieved high performance: 97.91% sensitivity, 97.69% specificity, 97.80% accuracy, and 99% AUC.
• Demonstrated improved prediction capabilities compared to existing methods.
• Validated using standard 2-fold cross-validation.

Conclusions:

• The AllerTrans model provides a powerful and efficient tool for predicting protein allergenicity.
• This approach significantly improves the safety assessment of proteins in medical products.
• The model is accessible as a public web-based prediction tool.