Chemical discrimination of wild and cultivated Dendrobium devonianum by ATR-FTIR spectroscopy and machine learning with targeted metabolomic analysis.

Wild and cultivated Dendrobium devonianum are visually similar, but their chemical differences complicate raw-material authentication. Recent spectroscopic studies of Dendrobium have focused primarily on species identification and geographical traceability, with limited chemical interpretation of wild-cultivated discrimination. ATR-FTIR spectroscopy and machine learning were applied to 145 samples (74 wild and 71 cultivated), while targeted metabolomics was performed using six pooled composite samples (three independently prepared pools per group). Spectral differences were concentrated in the 1800-900 cm-1 fingerprint region, although unsupervised analyses showed substantial overlap between groups. Three of 36 preprocessing-model combinations achieved 100% accuracy in an initial 43-sample test set. However, repeated stratified five-fold cross-validation (10 repeats) yielded a highest mean balanced accuracy of 0.9522 ± 0.0496 for partial least-squares discriminant analysis of raw spectra, providing a more cautious assessment of discrimination performance. SHAP identified 1058-1066 cm-1 as a key region associated with carbohydrate and glycosidic vibrations. Targeted metabolomics revealed 22 differential metabolites spanning sugars, amino acid-related compounds, and flavonoids. Among them, d-mannose, maltose, glucose, l-rhamnose, and glycosylated flavonoids showed associations with discriminative infrared variables. Compared with classification-focused studies, this framework connects spectral discrimination with complementary metabolite-level evidence, supporting more interpretable source authentication while highlighting the need for independent external validation.