AOXP: An Integrated Framework for Antioxidant Peptide Prediction Leveraging LLM-Based Semantic Features and Optimized Physicochemical Ensembles

Oxidative stress is a critical driver of various severe diseases, positioning antioxidant peptides (APs) as promising therapeutic agents. However, their discovery is constrained by the limitations of low-throughput wet-lab assays and insufficient computational prediction tools. To address this, we developed Antioxidant Peptides eXplorer (AOXP), a deep learning framework that leverages multiscale feature integration for high-precision AP prediction. Unlike existing models that often rely on isolated sequence patterns or manual feature descriptors, AOXP employs a hybrid encoding strategy that synergistically fuses deep, context-aware semantic representations from the protein large language model (ESM2) with explicit physicochemical and sequence distribution features. We constructed an optimized MLP architecture based on a refined 2400-dimensional feature vector, which was systematically benchmarked among five architectures and four feature selection strategies. Comprehensive evaluation demonstrates that AOXP achieves competitive performance across evaluation data sets, achieving a benchmark AUC of 0.736 on test_set1_drop and a sensitivity of 0.878 on the positive-only independent Test set 2, significantly outperforming current mainstream computational predictors in both sensitivity and robustness. Integrated Gradients analysis further quantified feature contributions, revealing that ESM2-derived semantics are the most influential features for model prediction (by feature number), followed by hydrophobicity, polarity, charge, and solvent accessibility. To demonstrate its therapeutic utility, AOXP was applied to screen a large peptide library targeting the Keap1-Nrf2 oxidative response pathway. A computationally predicted candidate peptide, MGRRHRHVTDEDDEENLGGDDDDGDDDVYMYPADMHPDERHAYREAV, was identified and validated through molecular dynamics simulations, exhibiting superior binding affinity for the Keap1 Kelch domain compared to the native ligand. AOXP provides a robust and efficient framework for the discovery of functional antioxidant peptides (available at: http://aidd.bioai-global.com/AOXP/).