APPLICATION OF THE MACHINE AND DEEP LEARNING METHODS FOR THE CLASSIFICATION OF CANNABINOID- AND CATHINONE-DERIVED COMPOUNDS

Abstract

Objective: New psychoactive substances (NPS) have been rapidly developed to avoid legal entanglement. In 2013–2018, the number of cathinonederived
compounds increased from 30 to 89. In 2016, of 56 NPS compounds, 21 were identified as cannabinoid-derived; only 43 were regulated in
the narcotics law. Artificial intelligence, such as machine and deep learning, is a method of data processing and object recognition, including human
poses and image classifications.
Methods: Herein, the machine and deep learning methods for cathinone- and cannabinoid-derived compound classification were compared using
pharmacophore modeling as the reference method. For classifying cathinone-derived compounds, the structure was transformed into fingerprints,
which was used as a learning parameter for the machine and deep learning methods. Contrarily, the physicochemical properties and fingerprint shape
were utilized as learning materials for the deep learning method to classify the cannabinoid-derived substances.
Results: Consequently, in the cathinone-derived compound classification, the deep learning method produced the accuracy and Cohen kappa values
of 0.9932 and 0.992, respectively. Furthermore, such values in the pharmacophore modeling method were higher than those in the machine learning
method (0.911 and 0.708 vs. 0.718 and 0.673, respectively). In the cannabinoid-derived compound classification, the deep learning method with the
fingerprint form had the highest accuracy and Cohen kappa values (0.9904 and 0.9876). Such values in this method with the descriptor form were
higher than those in the pharmacophore modeling method (0.8958 and 0.8622 vs. 0.68 and 0.396, respectively).
Conclusion: The deep learning method has the potential in the NPS classification.