Synthetic cannabinoid receptor agonists (SCRAs) are a prominent danger to public health. Emerging SCRAs are most often highly active at the CB1 cannabinoid receptor. This high activity imposes serious health threats, illustrated by intoxications with SCRAs presenting at emergency departments (EDs). The rapid emergence of novel analogs makes the detection of these new derivatives challenging. However, there is a strong need for continuous monitoring of these compounds to adapt legislations and ensure public health. The EDs of some hospitals, located in relevant areas regarding drug abuse, serve as some kind of ‘sentinels’, allowing to keep guard on circulating, potentially highly dangerous SCRAs. An example is the ED of Guy's and St-Thomas’ Hospital in central London, which is ideally positioned to keep track of the latest changes on the illicit drug market. In the context of screening ED patient samples, the ideal assay is easy-to-perform and easily implementable. To speed up the work process and reduce the workload in the case of activity-based screening, we assessed whether machine learning could be of any help to the expert in deciding the eventual outcome of the screening assay. Following up on the success of a prior large-scale screening of serum samples for the presence of SCRA activity (Cannaert et al., Clinical Chemistry, 2019, 65, 347-349), we set out to screen a new large set (968 samples, inclusion from February 2019 to February 2020 inclusive), with the following aims: 1) assess the performance of the assay to using biological samples potentially containing newer circulating SCRAs, 2) exploration of a more structured way of manual scoring, and 3) exploration of computer-based scoring.

Serum samples were subjected to activity-based SCRA screening using a cell-based bioluminescence CB1 reporter assay and to High Resolution Mass Spectrometry (HRMS)-based analysis for confirmation and identification. Both strategies were run independently and were performed blind-coded. Screening results from the bioassay (obtained through an improved scoring system by the expert) were compared with analytical (HRMS) results (considered as the ‘gold standard’). Additionally, a machine learning model was built to evaluate the possibility of computer-assisted sample scoring based on the bioassay read-out (receptor-activation profiles). Two cross-validation settings were employed to evaluate the performance of potential machine learning models and to assess the robustness of the machine learning approach.

The activity-based bioassay yielded a sensitivity of 94.6% and a specificity of 98.5%. Within the sample set six different SCRAs and/or their metabolites were detected: 4F-MDMB-BINACA, 5F-MDMB-PICA, 5F-MDMB-BINACA, MDMB-CHMICA, MDMB-4en-PICA and MDMB-4en-PINACA. The sensitivity and specificity obtained by manual scoring of screening results could be matched by the machine learning model through different model settings. The performance of the activity-based screening method is in concordance with previous studies. The panel of identified SCRAs is largely distinct from the panel identified from April to December 2016, exemplifying the well-known phenomenon of market dynamics and, importantly, also underscoring the universal nature of activity-based screening. The successfully designed machine learning model could automatically discriminate positive from negative samples and enables optimal balancing of sensitivity and specificity based on model settings (pre-set threshold).

Depending on the desired sensitivity/specificity and the corresponding settings applied within the model, we can conclude that machine learning is an adequate alternative for manual scoring by the expert (e.g. sensitivity and specificity of 94% at a 0.055 threshold). Automation of this scoring process results in significant time saving and reduction of the workload (Janssens et al., Clinical Chemistry, 2022, in press, https://doi.org/10.1093/clinchem/hvac027).