1H NMR-Based Metabolomics and Machine Learning Reveal Candidate Metabolic Markers Associated with Coffea arabica Genealogical Groups

The genetic diversity of Coffea arabica L. plays a strategic role in the development of cultivars with differentiated agronomic and sensory attributes. However, the chemical discrimination of genealogical groups still represents a challenge due to the strong environmental influence on bean metabolism. In this study, 1H NMR-based metabolomics associated with supervised chemometric models was applied to discriminate three genealogical groups of C. arabica (Bourbon, Mundo Novo, and Timor Hybrid) using 38 green coffee samples harvested during the 2020 and 2021 crop seasons. Spectra were processed by bucketing and analyzed using PCA, PLS-DA, Elastic Net, Random Forest, and SVM-RBF. PCA explained 52.91% of the total variability in the first two components but did not promote clear separation among groups. Among the supervised classifiers, PLS-DA showed the best balance between predictive performance, statistical stability, and chemical interpretability, achieving a balanced accuracy of 89.4% under repeated cross-validation. The five-component model presented an overall error rate of 12.1% after repeated 5-fold cross-validation with 100 simulations. VIP scores and loadings identified discriminant metabolites for each group: lipids, amino acids, and organic acids for Bourbon; amino acids and organic acids for Mundo Novo; and chlorogenic acids for Timor Hybrid. These results highlight the potential of 1H NMR metabolomics as a complementary tool for varietal discrimination and support for breeding programs.