All cultivars analyzed in this study were high THC-containing varieties (>10%) as determined by high performance liquid chromatography (HPLC) (data not shown). Cultivar names, which were analyzed at least five times (n≥5) in our laboratory over the data collection period, were selected to develop the classification scheme. This resulted in a data set containing 233 samples with 30 different cultivar names. The weight loss upon drying was between 6% and 14% indicating that the samples were not excessively dry and had similar moisture content. Typically, samples were submitted for testing a few days before they were made available to patients and usually represented samples from different batches purchased by the dispensary from various producers (personal communication with dispensary owners). Cannabinol (CBN), the primary degradation product of THC was <0.1% in all samples. The carboxylic acid form of THC, THC-acid, which is biosynthesized by the plant and converted into THC by decarboxylation,^17,36 dominated in all samples. Although there was no way of determining the age or exact storage conditions of the plant material, taken together, these observations suggest that the plant material was dried and stored properly.

As mentioned by Hazekamp, certain terpenoids in cannabis, mainly certain sesquiterpenoids, are difficult to identify due to poor resolution and lack of reference materials.³⁰ For the purposes of classification, we chose to build a database containing only unequivocally identified terpenoids for which we had an authentic reference standard whose peak identification was confirmed by GC-MS (Table 1). It is worth noting that previous studies identified cis-ocimene at a relative retention time (RRT) of 0.40.^28,30,39 In this study, we were able to reidentify this peak as trans-ocimene due to the availability of a reference standard that contained a known mixture of cis-ocimene and trans-ocimene isomers. Isomers could be distinguished based on known elution order⁴¹ and mass spectra (Supplementary Fig. S1). The quantitative levels of the 16 terpenoids in the different cultivars are shown in Table 2.

Initially each cultivar's quantitative terpenoid data was analyzed by HCA (Fig. 1). Based on the HCA, the cultivar names could be broken down into five groups. First, a myrcene-dominant group made up of the Purple Cream, Grape Ape, Purple Princess, Blue Dream, Strawberry Haze, Godfather, and Purple Urkle cultivars. A second terpinolene-dominant group composed of the Jack Herer and Trainwreck Cultivars. Another third group composed of the cultivars named as Crown Og, Skywalker Og Kush, Og Kush, Gas, Tahoe Og Kush, Triple O, Gelato, and Miami White Kush dominated in myrcene and limonene. Distinguishing characteristics among these cultivars were the relatively higher levels of the monoterpenoid alcohols α-terpineol, endo-fenchyl-alcohol, and linalool. A fourth group of cultivars were dominated by β-caryophyllene, which included Blue Cookies, Girl Scout Cookies, Animal Cookies, Thin Mints, Fortune Cookies, Sherbert, Chemdog, and Gorilla Glue #4 (gorilla glue). A fifth group composed of Master Kush, Bubba Kush, Mr. Nice, and Sour Diesel tended to dominate in myrcene, limonene, or β-caryophyllene. One distinguishing characteristic of this group was the relatively higher levels of α-bisabolol in Master Kush, Bubba Kush, and Sour Diesel. We termed the five groups myrcene (first), terpinolene (second), Og Kush (third), caryophyllene (fourth), and bisabolol (fifth), respectively.

To explore the above categorization, data were submitted to PCA (Fig. 2). Figure 2 shows the scores plot of Principal component 1 (PC1) versus PC2 with samples labeled according to the five groups identified in the HCA. PC1 explained 46.5% of variance, and PC2 explained 22.9% of variance in the data (total 69.4%). If PC3 is included, 85% of variance is explained (Supplementary Fig. S2). The higher percentage of variance explained in our model, compared with previous studies,^29,30 is most likely from avoiding unnecessary variance from cannabinoid content as well as reasonable quantitative consistency of the terpenoid levels of the samples tested. The terpinolene category clearly clustered by itself along the negative PC1 and positive PC2 axes. The myrcene category also clustered mainly by itself along the positive PC1 axis. The Og Kush group partially overlapped with the caryophyllene, and bisabolol group along the negative PC1 axis, whereas very little overlap was observed along the positive PC1 axis. The loadings plot confirmed that the distinguishing characteristics of the terpinolene and myrcene groups were their respective dominant terpenoids, whereas relatively higher levels of limonene, β-caryophyllene, humulene, and linalool were characteristics of the Og Kush, β-caryophyllene, and α-bisabolol groups.

To further discriminate cultivars within the major groups identified by HCA and PCA, metabolite patterns of the cultivars were analyzed with supervised multivariate data analysis techniques. PLS-DA is a multivariate regression method that can be used to evaluate the differences between defined classes in a data matrix. OPLS-DA is an extension of PLS-DA that employs orthogonal signal correction, and was used here to identify variables (terpenoids) important for discriminating the cultivar classes. In OPLS-DA, the X-axis displays the maximum variation between classes, whereas the Y-axis displays maximum variation within classes.⁴²

The approach taken with OPLS-DA to aid in explaining the differences between cultivars within their respective groups is illustrated in detail with the β-caryophyllene group (Fig. 3). If all cultivars are assigned as their own class, clustering is observed among Blue Cookies, Girl Scout Cookies, Animal Cookies, Thin Mints, and Fortune Cookies, whereas Gorilla Glue #4 and Sherbert clustered more to themselves. Chemdog samples displayed a large within-class variation (Fig. 3a). Inspection of the quantitative data reveals that Blue Cookies, Girl Scout Cookies, Animal Cookies, Thin Mints, and Fortune Cookies shared the characteristic of the next most abundant terpenoids after β-caryophyllene; namely limonene and humulene in a ∼1:1 ratio. These cultivars were thus assigned to a class termed “Cookie” due to their similarity in nomenclature. Gorilla glue contrasted with the cookie cultivars in that limonene and myrcene (∼1:1) were the next most abundant terpenoids after β-caryophyllene. Sherbert contained the highest levels of limonene and higher levels of the alcohol-substituted monoterpenoid, linalool, followed by endo-fenchyl-alcohol, and α-terpineol.

When analyzed again, with cookie as a defined class, better clustering was observed within the cookie class, although it was still difficult to observe the distinction between Gorilla Glue, Sherbert, and Chemdog (Fig. 3b). The larger within-class variation of the Chemdog samples is most likely a result of the high variation in myrcene levels. Chemdog was removed from the data set and the remaining samples were analyzed again by OPLS-DA. Distinct clustering was thus observed between the Cookie, Gorilla Glue, and Sherbert classes (Fig. 3c). The S-plot shows significant metabolites contributing to the variation across the defined classes (Fig. 3d). Metabolites in the upper right and lower left corners of the S-plot (i.e., with the highest positive and lowest negative p-value versus p[corr]-values) were the most significant variables responsible for the discrimination between classes. The S-plot confirmed that the main terpenoids distinguishing the classes within the β-caryophyllene group are relatively higher levels of humulene in the cookies category, relatively higher levels of limonene and alcohol substituted terpenoids in the sherbert class, and relatively higher levels of myrcene in the Gorilla Glue class.

Inspection of the quantitative data within the myrcene group suggested that four classes could potentially be distinguished. Blue dream contained higher levels of α-pinene and β-pinene compared with all other cultivars in this study. The ratio between α-pinene, β-pinene, and myrcene was consistent at ∼2:1:4 throughout the samples of Blue Dream analyzed. Godfather contained on average the highest levels of myrcene followed by α-pinene and limonene. Godfather also contained lower amounts of β-caryophyllene and humulene relative to other cultivars in the myrcene group. Cultivars named Purple Cream, Grape Ape, Purple Princess, and Purple Urkle dominated in myrcene followed by α-pinene in a ∼4:1 ratio followed by trans-ocimene. Therefore, Purple Cream, Grape Ape, Purple Princess, and Purple Urkle were assigned to a class termed “Purple” due to their shared morphological characteristic of purple-tinged flower buds. Strawberry Haze had the highest levels of trans-ocimene followed by limonene. When analyzed by OPLS-DA, the samples in each of the assigned classes within the myrcene group cluster, and clusters are separated along the first principal component (Fig. 4a). The S-plot confirmed that trans-ocimene and linalool were the most important distinguishing terpenoids for the Strawberry Haze and Purple categories, whereas α-pinene and β-pinene were the most important for distinguishing the Blue Dream category (Fig. 4b).

Within the Og Kush group, Crown Og, Gas, Og Kush, Skywalker Og Kush, Superman Og Kush, and Tahoe Og Kush cultivars all tended to dominate in myrcene with similar or slightly lower levels of limonene. These cultivars were assigned to a class termed “Og Kush” due to similarity in nomenclature. Miami White Kush and Triple O dominated in limonene compared with myrcene in a ∼2:1 ratio and were assigned to a class termed “Limonene Og Kush.” Gelato dominated in limonene, but contained even lower amounts of myrcene. Gelato also tended to contain the highest levels of linalool, endo-fenchyl-alcohol, α-terpineol, and geranyl acetate compared with other cultivars. OPLS-DA analysis of these classes revealed a continuum of terpenoid profiles with some overlap between Og Kush and Limonene Og Kush, as well as between Limonene Og Kush and Gelato (Fig. 5a). The S-plot confirmed that the main source of variation within the Og Kush group is whether the chemotype tends to dominate in myrcene or limonene (Fig. 5b).

Jack Herer and Trainwreck cultivars were distinguishable in a previous study²⁹ and bear strong resemblance to the Bedrocan cultivar available through Dutch pharmacies as well as other accessions described as sativa dominant.^28,39 Within the terpinolene group, Jack Herer mainly differed from Trainwreck due to lower levels of myrcene and limonene. Another important characteristic of the terpinolene chemotype is the presence of α-phellandrene and 3-carene. As discussed in Hazekamp 2016,³⁰ β-phellandrene overlaps with limonene. GC-MS analysis of representative terpinolene chemotype samples revealed that the peak at RRT 0.377 is most likely composed of β-phellandrene and limonene. In representative samples of cultivars that lack α-phellandrene and terpinolene, the limonene peak does not seem to contain β-phellandrene (Supplementary Fig. S3). Although the quantitative analysis failed to detect 3-carene and α-terpinene above detection limits in Jack Herer samples by GC-FID, the presence of these compounds was confirmed by GC-MS analysis of a representative Jack Herer sample (Supplementary Fig. S1B). Since the terpinolene chemotype had such a distinctive profile no further analysis was necessary.

In the bisabolol group Master and Bubba Kush tended to dominate in limonene, β-caryophyllene, and myrcene in a ∼2:2:1 ratio. Mr. Nice had a ∼1:1:1 ratio of myrcene, limonene, and β-caryophyllene. Initial OPLS-DA analysis with each cultivar as class resulted in poor clustering for Sour Diesel likely due to high variation in the sample set and lack of other distinguishing characteristics (Supplementary Fig. S4). If Bubba Kush and Master Kush were combined into a class termed “Kush” and analyzed by OPLS-DA with Mr. Nice discrimination is observed (Fig. 6a). The S-plot confirmed that Mr. Nice was distinguishable due to its higher levels of myrcene, whereas the Kush class had more humulene and alcohol-substituted terpenoids, mainly α-bisabolol (Fig. 6b).

Based on the overall terpenoid profiles, HCA, PCA, and OPLS-DA, a scheme for classifying the cultivars was developed based on dominant terpenoids, ratios of major terpenoids, presence of unique terpenoids, or other characteristic terpenoids (Table 3). To test this classification approach, different PLS-DA models were analyzed with different sample sets and class assignments (Table 4). Sour Diesel and Chemdog samples were not included in PLS-DA models. In model 1, cultivar names were assigned as classes. In model 2, class names from Table 3 were used as classes. In model 3, samples from the data collection period that contained three to four replicates and could be assigned to the classes in Table 3 were added (Table 3 italics). This included an additional 57 samples representing 17 additional cultivar names (Supplementary Table S1). In model 4, only classes that contained >15 samples (from the sample set in model 3) per class were used. This included the Blue Dream, Cookie, Og Kush, Limonene Og Kush, Purple, and Terpinolene classes.

The PLS-DA models were subjected to cross-validation and permutation tests.⁴³ For cross-validation five components were searched and leave-one-out cross-validation was used. Prediction accuracy, R² (sum of squares captured by model) and Q² (cross-validated R²) were plotted and Q² used as performance measure. The closer Q² is to 1, the better the goodness of prediction of the PLS-DA model. The best performance of the models (as indicated by the red asterisk in the plots) was achieved using all five components (Supplementary Figs. S5–S8, top). From models 1 to 2, Q² increased substantially with only minor increase in Q² between models 2 and 3, whereas model 4 has the highest Q² value (Table 4). Permutation tests were performed on each model to determine whether discrimination between classes were statistically significant, or due to random noise.⁴³ Two thousand permutations were performed, during each of which a five-component PLS-DA model was built between the terpenoid data and permuted class labels. Histograms show permutation test scores, and compare it to the performance based on the defined classifications (red arrow) (Supplementary Figs. S5–S8, bottom). The further the separation distance, based on a sum of squares between/sum of squares (B/W) ratio, between the observed statistic and the distribution resulting from permuted data, the more significant the discrimination.⁴⁴ The separation difference increased from model 1 to 2 from 2 to 3 and from 3 to 4 (Supplementary Figs. S5–S8, bottom). The PLS-DA results confirm that classification based on terpenoid classes outperforms classification based on cultivar name. When new samples were added to the sample set from models 1 and 2, a slightly improved model 3 resulted. However, the highest performing model was achieved with classes that contain ≥15 samples (model 4). These results suggest that better predictive models are constructed from terpenoid profiles with more representative samples.