AOP-DRL: A deep representation learning framework for the computational prediction of antioxidant peptides

Molecular oxygen is vital for life but can generate reactive oxygen species (ROS), leading to oxidative stress and biomolecular damage. Antioxidant peptides mitigate ROS but face scalability challenges in traditional identification methods. To address this, we developed antioxidant peptides deep representation learning (AOP-DRL), a deep learning framework that combines protein language models with hierarchical convolutional networks for high-throughput prediction. The training dataset from the publicly available original study includes experimentally validated antioxidant sequences and negative controls orthogonally verified via public repositories under standard redox proteomics data partitioning protocols. Our architecture overcomes limitations in handling variable peptide lengths and capturing nonlinear residue interactions. Benchmarking experiments reveal that AOP-DRL outperforms state-of-the-art models in terms of accuracy while generalizing well across diverse datasets. Our model achieved accuracy improvements of 7.26 %, 2.57 %, 2.59 %, and 4.20 % on the prepartitioned subsets P60, P70, P80, and P90, respectively, compared with the mean accuracy of previously specialized models for antioxidant peptides. This approach provides a cost-effective, scalable alternative to wet-lab methods, accelerating the discovery of therapeutic and nutraceutical peptides. The adaptability of the framework also suggests broader applications in bioactive peptide prediction for personalized medicine and functional foods.