Whilst cannabis is known for having a strong and distinct odour, there is considerable diversity in the language used to describe cannabis odour [10]. With the recent acceptance of cannabis as a recreational drug and medicinal product, plant breeding has enabled the development of many new varieties, which often have exotic odour-inspired names and rich odour descriptions. However, cannabis naming is unregulated [47], and there is significant chemical and genetic diversity between strains of the same name [48, 49], which has resulted in a reputation within the industry of products having inaccurate descriptions or misleadingly suggestive names [50–52].

A significant demographic of medicinal cannabis users are the paediatric and geriatric patient populations, which have specific considerations for drug design and usability [53–56]. Paediatrics in particular, are sensitive to organoleptic properties such as taste and odour, with rejection occurring from unpalatable medications [57–60]. Cannabis oils are reported to have unpleasant organoleptic properties for both adults and children, which has resulted in cannabis capsule product developments [60, 61], however swallowing can be difficult in paediatrics making this unavailable to those patients [62]. For liquid oral formulations such as Epidiolex, strawberry flavouring is added [63], however no such alternative is available for raw flower products. With the inconsistencies in product labelling and descriptions, there is a risk of these special patient groups receiving medicinal products with different sensory properties to what they expect, which may result in rejection and thereby negatively affect their therapeutic outcomes. Accordingly, there is a need to increase the consistency and accuracy of cannabis labelling and descriptions to improve patient experiences and therapeutic outcomes.

The simplest measure of odour strength is OAV which is the ratio of the concentration of a compound in a sample relative to its ODT. However, OAV doesn’t inform on how that chemical will be perceived by an individual or, if present in a mixture, how other compounds are contributing to the overall odour. Odour intensity and vector addition are proposed as improved methods for determining odour perception, particularly for mixtures. As OAV is an input to OI calculation, ODT for each compound present in a mixture is required in order to perform odour vector addition of OI. Accordingly, as cannabis and other natural products contain many volatiles which have not been tested by sensory analysis to determine their ODT, there is a need to predict ODT from in silico measures to enable vector modelling of profiles containing these compounds.

Polynomial regression was utilised to generate a model for the prediction of ln ODT from physicochemical measures, which had an R² = 0.6892 and a 10-fold R² = 0.6484 (Fig 2, Table S5 in S1 File). The similarity between the two R² measures indicates comparable accuracy for both the training and the test sets across the fold splits, supporting that the model is not overfit [64]. The model includes 69 total 1^st, 2^nd, and 3^rd order terms (Table S5 in S1 File), which are all significant (P < 0.05). Of these 69 terms, 49 include terms relating to the presence of functional groups, and 23 contain terms relating to molecular size. The inclusion of these properties for the prediction of ODT is consistent with previous studies which identified that ODT is impacted by the related measures of carbon chain length and functional group presence [41, 42, 65]. The model also included terms with measures relating to molecular properties (e.g. Log P), and types of bonds present (e.g. rotatable bonds), which may indicate their association with ODT perception and could therefore be the targets for further studies on physicochemical properties impacting ODT.

Previous studies have presented predictive models for ODT, however, these tend to utilise small data sets. Polster and Schieberle [65] present an ODT model from 44 compounds with a R² of 0.89, Rodríguez, Teixeira [66] identified a model from 121 compounds with R² of 0.769, Anker, Jurs [67] and Junkes, Arruda [68] present models from 49 compounds with R² of 0.863 and 0.765, respectively, whilst Pal, Mitra [44] utilised 42 compounds to develop a model with R² of 0.809 and an R² of 0.813 when applied to a test set of 11 compounds. Abraham, Sánchez-Moreno [43] also modelled a log transformation of ODT against two datasets containing 206 and 353 compounds to produce models with R² of 0.748 and 0.759, respectively. These works support the use of modelling of ODT from chemical properties, however, the data sets utilised in these previous models are small compared to the 10,000 identified volatile odour chemicals [69]. Therefore, as statistical models are limited in their application to compounds with properties similar to those within the modelling dataset, and the relatively small size of these models means they only account for a small proportion of the diversity of properties across all volatile compounds, these models are expected to have low generalisability for predicting ODT in new compounds. By contrast, the model presented herein utilised data from 1,274 chemicals, which is a substantial expansion in the breadth of chemical properties modelled (Table S6 in S1 File) and therefore has greater generalisability to new compounds. Furthermore, the accuracy of the model herein was found to be relatively similar to the accuracy of empirical determination of ODT through human participant sensory testing. The ratios for the ODT data variance and the model variance (defined in method 2.3 herein) were found to be on average 2.36 and 2.21, respectively (Table S7 and Fig S1 in S1 File), and were not significantly different (p = 0.771), indicating that the predicted values fall within a similar range of the reported ODT values to their median. Human olfactory sensitivity is known to be highly variable [70], and for ODT specifically, variations of up to 4 or 5 magnitudes are reported between studies of the same chemicals due to variations in sensory methodology [71]. Accordingly, the model herein which has achieved prediction of comparable accuracy could be considered as a suitable substitute for ODT in the absence of sensory testing.

It was hypothesised that odour vector modelling could be utilised to generate OI profiles for the highly complex volatile mixtures present in cannabis flowers, and that these profiles would be more informative to the overall odour than their associated terpene profiles. Additionally, whilst current odour vector modelling approaches utilise empirically solved values for the angles between compound OI vectors, it was hypothesised herein that a calculated estimation could be used in place of these values. Finally, the incorporation of OD ranking into OI calculation was hypothesised to account for the relative strengths of ODs and thereby improve the predictive power of the resulting OI profile.

Cannabis terpene and OI profiles were explored for the prediction of cannabis SD by LR modelling. Of the 13 SD categories modelled, terpene data was able to produce models for 10 SDs (Equations S10–S19 in S1 File) whilst OI profiles could produce models for 12 SDs (Equations S20–S31 in S1 File). The performance of each model was assessed for its ability to accurately assign SDs to each sample (Fig 3). Six SD categories were able to be modelled by OI significantly (p < 0.05) better than terpene (herbal, sweet, berry, tree fruit, roasted, and mellow), 5 SD categories were modelled equally successfully for both OI and terpene data (mineral, earthy, tropical, warming, and floral), 1 SD category was modelled by terpene significantly better than OI data (citrus), and 1 SD could not be modelled by either terpene or OI data (animal). When comparing the overall performance of the two modelling strategies, it was identified that OI data produced models with an average of 70.33% accuracy and models produced from volatile data had an average of 48.39% accuracy, which was a significant (p = 0.031) difference of 21.94%.

The results indicate that odour vector modelling produces profiles which are more informative for predicting overall cannabis odour (SD) than profiles of terpene data produced by routine testing within the Californian cannabis industry. The transformation of these relatively easy-to-produce terpene profiles, into OI profiles which are more informative for odour prediction is a promising development in achieving more consistent labelling and product descriptions within the cannabis industry. Noting that these routine profiles contain only 18 volatiles of the more than 200 which contribute to the cannabis odour [72], increasing the predictive capability of these (if possible) may require considerable expansion of the breadth of chemical profiling. This increased volatile profiling would result in greater costs to manufacturers and possibly consumers and patients. Improving on the current practices is therefore the easiest and cheapest way to confer benefits to the industry and the patients. However, as some of the SDs were not able to be modelled well by either the terpene or the OI data sets (floral and animal), expansion of the volatile profiling or cannabis sample data may still be desirable to achieve more accurate modelling of these SDs. As the expansion of the volatile profiles also adds data to the corresponding OI profiles, and expansion of cannabis sample data provides more data points for the modelling, these strategies may increase the accuracy of either the terpene or the OI modelling strategies. Furthermore, the application of this strategy to explore statistical relationships between volatile and sensory profiles is not limited to the cannabis industry. Industries wherein volatile chemical profiles and sensory data are both readily collected—such as perfume or food industries—could potentially utilise this strategy to improve predictive capabilities and eventually guide alterations to volatile profiles to achieve targeted changes to sensory profiles.