AllerTrans: An Improved Protein Allergenicity Prediction Model Using Deep Learning

Abstract

Recognizing the potential allergenicity of proteins is essential for ensuring their safety. Allergens are a major concern in determining protein safety, especially with the increasing use of recombinant proteins in new medical products. These proteins need careful allergenicity assessment to guarantee their safety. However, traditional laboratory testing for allergenicity is expensive and time-consuming. To address this challenge, bioinformatics offers efficient and cost-effective alternatives for predicting protein allergenicity. In this study, we developed an enhanced deep-learning model to predict the potential allergenicity of proteins based on their primary structure represented as protein sequences. Our approach utilizes two protein language models, to extract distinct feature vectors for each sequence, which are then input into a deep neural network model for classification. Each feature vector represents a specific aspect of the protein sequence, and combining them enhances the final result and balances the model's sensitivity and specificity. The model classifies proteins into allergenic or non-allergenic classes. Our proposed model demonstrates admissible improvement across all evaluation metrics compared to the AlgPred 2.0 model, achieving a sensitivity of 97.91%, specificity of 97.69%, accuracy of 97.80%, and an impressive area under the ROC curve of 99% on the AlgPred 2.0 dataset using standard five-fold cross-validation.