The increasing prevalence of food and environmental allergies has intensified the demand for reliable computational approaches for protein allergen prediction. Recent advances in protein language models have enabled end-to-end sequence representation learning without reliance on handcrafted features. In this study, ProtBERT was optimized and calibrated for protein allergen prediction using curated datasets of allergenic and non-allergenic proteins. ProtBERT embeddings were coupled with a lightweight neural classification head and optimized through learning-rate tuning, dropout regularization, partial fine-tuning of upper transformer layers, and class-balanced loss functions. Model calibration and decision-threshold optimization were further applied to enhance probabilistic reliability. Following optimization, the ProtBERT-based model achieved an accuracy of 0.9392, F1-score of 0.9339, and ROC-AUC of 0.9757, demonstrating strong discriminative capability directly from raw amino acid sequences. For contextual evaluation, ProtBERT performance was compared with physicochemical feature-based machine learning approaches, which served as reference baselines. Despite differences in modeling paradigms, ProtBERT offers advantages in scalability, reduced feature engineering, and compatibility with foundation-model frameworks. These findings indicate that optimized and calibrated ProtBERT models provide a robust sequence-based framework for protein allergen prediction and represent a promising foundation for future hybrid and large-scale computational allergology systems.<\/p>\n