Prediction of sauvignon blanc quality gradings with static headspace−gas chromatography−ion mobility spectrometry (SHS−GC−IMS) and machine learning

AIM: The main goal of the current study is the development of a cost-effective and easy-to-use method suitable for use in the laboratory of commercial wineries to analyze wine aroma. Additionally, this study attempted to establish a prediction model for wine quality gradings based on their aroma, which could reveal the important aroma compounds that correlate well with different grades of perceived quality METHODS: Parameters of the SHS−GC−IMS instrument were first optimized to acquire the most desirable chromatographic resolution and signal intensities. Method stability was then exhibited by repeatability and reproducibility. Subsequently, compound identification was conducted. After method development, a total of 143 end-ferment wine samples of three different quality gradings from vintage 2020 were analyzed with the SHS−GC−IMS instrument. Six machine learning methods were employed to process the results and construct a quality prediction model. Techniques that aim to explain the model to extract useful insights were also applied. RESULTS: The SHS−GC−IMS method was able to detect 23 compounds among 65 peaks, mostly esters and higher alcohols, using the current instrumentation. Several identified compounds, including methyl acetate, ethyl formate, and amyl acetate, have seldomly been reported in Sauvignon Blanc wines before. The method also indicated decent repeatability and reproducibility, both of which were below 10%. The quality prediction model was successfully established using artificial neural network (ANN) based on all peaks regardless of their identity. The model returned a highly satisfactory prediction accuracy of 95.4% using 10-fold cross-validation. SHapley Additive exPlanations (SHAP) values was used to delineate the prediction mechanism of the model. SHAP values revealed that isoamyl acetate, ethyl decanoate, ethyl octanoate and 1-hexanol were positively linked to better quality, whereas hexyl acetate, isoamyl alcohol, and 1-butanol could lower the quality grading. CONCLUSIONS: This study has successfully developed a method alternative to GC−MS based instruments for the non-targeted screening of wine volatile compounds. With its simple design featuring a headspace sampling unit, highly simplified sample preparation, and nitrogen being the only gas supply, the instrument has shown outstanding practicality desired by commercial winery laboratories. The powerful prediction model and the insights extracted by SHAP values could serve as a starting point for winemakers to investigate the effects of winemaking operations on the expression of the volatiles shown to correlate with higher gradings, to enhance the quality of wines. The findings of this study have been published as an original research article in the Journal of Agricultural and Food Chemistry: J. Agric. Food Chem. 2021, 69(10), 3255−3265.

Authors: Wenyao Zhu – School of Chemical Sciences, University of Auckland, Private Bag 92019, Auckland, New Zealand; Drylands Winery, Constellation Brands NZ, Blenheim 7273, New Zealand., Frank BENKWITZ, Drylands Winery, Constellation Brands NZ, Blenheim 7273, New Zealand. Paul A. KILMARTIN, School of Chemical Sciences, University of Auckland, Private Bag 92019, Auckland, New Zealand.


Keywords: Sauvignon blanc, static headspace−gas chromatography−ion mobility spectrometry (SHS−GC−IMS), quality grading, machine learning, artificial neural network (ANN), model explanation

Share via
Copy link
Powered by Social Snap