SaltySought: Machine learning and deep neural approaches for identifying and explaining salty peptides

Abstract

Salty peptides remain underexplored in taste-active peptide research, with most studies focusing on umami and sweet peptides. Existing research primarily examines individual peptides rather than adopting a systematic approach. To address this, we propose SaltySought, an online platform integrating machine learning and deep learning for salty peptide classification. Our framework employs high-dimensional feature extraction (ESM, AAC, CTD, GAAC) to comprehensively represent peptide sequences. Feature selection was conducted using Random Forest, XGBoost, and Extra Trees, refining the most informative features before model training. The selected features were then used to systematically evaluate six models, with CNN1D achieving the best performance (AUC-ROC 0.88). Finally, SHAP analysis was applied to interpret model decisions, revealing that short peptide length, PRAM900101, PONP930101, and solvent accessibility jointly determine saltiness perception. These findings support the development of low-sodium functional foods by integrating AI-driven peptide identification, deep learning embeddings, and feature explainability techniques, enhancing food formulation strategies.