Rapid prediction of 3-isobutyl-2-methoxypyrazine (IBMP) in Chilean grape extracts with absorbance-transmittance excitation emission matrix (A-TEEM) spectroscopy

Detecting methoxypyrazines such as 3-isobutyl-2-methoxypyrazine (IBMP) in wine grapes constitutes a significant challenge due to the extremely low concentrations at which they impart orthonasal sensory effects. Associated with vegetal or 'green' aromas in wine, IBMP accumulates in immature grapes and diminishes with maturity. Such aromas can be viewed as detrimental to the quality of red wines, with Cabernet-Sauvignon being a particularly vulnerable variety. The sensory threshold for red wine matrices is recognised to range from 2 to 16 ng/L. Hence, detection conventionally requires expensive, time-consuming analysis by gas chromatography with mass spectrometry. Seeking a fast (i.e., < 10 min per sample) and inexpensive assay, this study investigated the ability to empirically predict IBMP in red winegrape varieties using HORIBA's patented absorbance-transmittance and fluorescence excitation-emission matrix (A-TEEM) spectroscopy. Cabernet-Sauvignon, Carménère, Merlot, and Cabernet franc grapes were sourced and extracted during harvest season over a period of 4 years from > 70 vineyards throughout Chile. A-TEEM spectral data represented the fusion of the UV-Vis absorbance and unfolded fluorescence EEM to yield more than 10,000 variables per sample. Model optimisation included a recursive partial least squares variable selection for the calibration set, and IBMP test set predictions were evaluated by multivariate regression using a randomly selected calibration/validation set of 1816/584 samples (including replication). Importantly, both cross-validation (CV) and test set prediction were based on exclusive sample repetition classes to eliminate replicate trapping. Support vector machine (SVM) regression yielded root mean square errors for calibration, CV, and test set prediction of < 0.4 ng/kg, with an R2 = 0.879. The regression results were supplemented by SVM discriminant analysis based on a pass/fail threshold of 2 ng/kg, yielding a Matthews correlation coefficient of 0.970 with no misclassified samples in the < 2 ng/kg class. This work highlights the potential of A-TEEM with machine learning data analysis to provide an inexpensive, rapid, and accurate screening method for IBMP in extracts of susceptible red winegrape varieties at or below the typical wine sensory threshold.