An interpretable CNN-transformer parallel model for unveiling osteogenic peptides and its application in chicken embryo egg hydrolysate

Addressing the critical concern of bone health in aging populations, a deep learning model termed PepOSX-AI: Osteogenesis was developed for the efficient prediction of food-derived osteogenic peptides (OPs) in this study, which integrated a parallel CNN-Transformer architecture with five physicochemical features. Based on ablation experiment and SHapley Additive exPlanation (SHAP) value analysis, hydrophobicity was revealed to be a critical physicochemical determinant. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.922 and accuracy of 83.0% after parameter optimization. By combining the CNN module's local feature extraction with the Transformer's self-attention, an interpretable analysis of the structure-activity relationship of OPs was performed, thereby providing insights into the underlying osteogenic mechanisms. Within this framework, peptides LL and LR from chicken embryo hydrolysate, exhibiting osteogenic activity, were successfully screened and validated through LC-MS/MS identification and model prediction. In conclusion, PepOSX-AI: Osteogenesis demonstrates promise for high-throughput screening of food-derived OPs.