Use of near-infrared spectroscopy for the classification of medicinal cannabis cultivars and the prediction of their cannabinoid and terpene contents

Cannabis sativa L. is used to treat a wide variety of medical conditions, in light of its beneficial pharmacological properties of its cannabinoids and terpenes. At present, the quantitative chemical analysis of these active compounds is achieved through the use of laborious, expensive, and time-consuming technologies, such as high-pressure liquid-chromatography- photodiode arrays, mass spectrometer detectors (HPLC-PDA or MS), or gas chromatography-mass spectroscopy (GC-MS). Hence, we aimed to develop a simple, accurate, fast, and cheap technique for the quantification of major cannabinoids and terpenes using Fourier transform near infra-red spectroscopy (FT-NIRS). FT-NIRS was coupled with multivariate classification and regression models, namely partial least square-discriminant analysis (PLS-DA) and partial least squares regression (PLS-R) models. The PLS-DA model yielded an absolute major class separation (high-THC, high-CBD, hybrid, and high-CBG) and perfect class prediction. Using only three latent variables (LVs), the cross-validation and prediction model errors indicated a low probability of over-fitting the data. In addition, the PLS-DA model enabled the classification of chemovars with genetic-chemical similarities. The classification of high-THCA chemovars was more sensitive and more specific than the classifications of the remaining chemovars. The prediction of cannabinoid and terpene concentrations by PLS-R yielded 11 robust models with high predictive capabilities (R²_CV and R²_pred > 0.8, RPD >2.5 and RPIQ >3, RMSECV/RMSEC ratio <1.2) and additional 15 models whose performance was acceptable for initial screening purposes (R²_CV > 0.7 and R²_pred < 0.8, RPD >2 and RPIQ <3, 1.2 < RMSECV/RMSEC ratio <2). Our results confirm that there is sufficient information in the FT-NIRS to develop cannabinoid and terpene prediction models and major-cultivar classification models.