Accurate structure-activity relationship prediction of antioxidant peptides using a multimodal deep learning...

Area of Science:

• Biochemistry and Bioinformatics
• Computational Chemistry
• Peptide Science

Background:

• Antioxidant peptides (AOPs) show promise for treating oxidative stress diseases and have applications in functional foods and cosmetics.
• Developing accurate predictive models for AOPs is crucial for their efficient discovery and application.

Purpose of the Study:

• To create a comprehensive quantitative structure-activity relationship (QSAR) model for predicting AOP activity.
• To utilize a generative model for designing novel AOP candidates.
• To enhance the accuracy, robustness, and interpretability of AOP prediction.

Main Methods:

• Integrated 6 sequence-based structure representations with stacking ensemble neural networks (CNNs, BiLSTM, Transformer).
• Employed a generative model for novel AOP candidate design.
• Utilized SHAP analysis for model interpretability.

Main Results:

• Achieved high predictive metrics (accuracy, AUROC, AUPRC > 0.90; MCC > 0.80) with stacking models using one-hot encoding.
• Identified key amino acid residues influencing antioxidant activity (positive: Pro, Leu, Ala, Tyr, Gly; negative: Met, Cys, Trp, Asn, Thr).
• Computationally identified 604 high-confidence AOPs.

Conclusions:

• The multimodal deep learning framework significantly improves AOP prediction accuracy, robustness, and interpretability.
• This approach enables efficient discovery of high-potential AOPs.
• Provides a powerful pipeline for accelerating peptide discovery in pharmaceutical and functional applications.