Comparative evaluation of vibrational spectroscopy and deep learning for hierarchical authentication of collagen-based biological matrices

The authentication of highly similar biological matrices remains a major challenge in analytical chemistry because weak and overlapping spectroscopic signatures often limit reliable classification. Here, we establish a comparative analytical framework to evaluate the performance of vibrational spectroscopy coupled with conventional chemometric and deep-learning approaches using donkey-hide gelatin (DHG) as a representative collagen-rich model system. In five-fold cross-validation of 60 authentic and counterfeit samples, linear discriminant analysis achieved accuracies of 81.7%, 96.7%, and 100% using SERS, FT-IR, and NIR spectra, respectively; the first two NIR principal components accounted for 90.0% of the spectral variance. NIR also enabled complete cross-validation classification of four animal-derived gelatins. For the more closely related E1-E4 commercial products, AggMapNet achieved approximately 99% accuracy, compared with approximately 80-90% for conventional models. For three E4 commercial grades, 157 of 160 samples were correctly classified, corresponding to 98.1% overall accuracy, and balanced accuracy of 98.0%. These findings show that conventional chemometric analysis effectively resolves broader differences in authenticity and animal source, whereas AggMap-assisted representation learning enables finer discrimination among closely related commercial products and product grades.