Explainable AI for secure and accurate prediction of bacteriophage virion proteins using NLP descriptors and transformer-guided ideal proximity matrix reconstruction

Abstract

The biological functions of bacteria are significantly impacted by bacteriophage virion proteins (BVPs), which are bacterial viruses. BVPs play a major role in phage therapy and genetic engineering. Secure and accurate identification of these proteins is essential for understanding phage-host interactions and for bioinformatics and medical applications. However, ensuring privacy and robustness in computational models is challenging, especially when handling complex biological data. Previous works relied on wet-lab experiments, had limited scalability, incomplete feature coverage, and low generalization ability. In this study, we introduce a privacy-preserving and adversarial-robust deep learning framework. It integrates natural language processing (NLP) descriptors with transformer-guided ideal proximity matrix reconstruction to capture rich information from protein sequences. For post-hoc interpretability, we use SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). These techniques increase openness and confidence in predictions. SHAP analyzes the dataset to identify the most significant proximity-based and NLP-derived descriptors at global and class levels. LIME provides instance-specific explanations, emphasizing local decision boundaries for particular predictions. The proposed model achieved 95.75 % and 90.27 % accuracy on the training and independent datasets, respectively. We calculated statistical measures, such as Chi-Square and P-value, for each dataset to demonstrate reliability. Our model improves predictive outcomes, transparency, and security. The empirical results validate its outstanding performance compared to existing models, while preserving security and explainable AI. This makes it suitable and reliable for real-world applications in proteomics and bioinformatics.