Artificial Neural Networks Screening for JWH Synthetic Cannabinoids: a Comparative Analysis Regarding their Specificity and Accuracy

This study evaluates the impact of the dataset size and of the number of molecular descriptors selected to build Artificial Neural Networks (ANN) screening for JWH synthetic cannabinoids. The aim is to determine how to most economically use the available data on these illicit drugs and still avoid overfitting. The results indicate a proportional decrease in the number of inputs in terms of memory requirements, processing speed, and numerical precision by fitting a model with the same database of designer drugs and the same test set for each different sized training dataset (having 100, 50 and 25 samples respectively). The results indicate that the model trained with 100 samples performs nearly as well as the reference ANN system (built with 150 samples), but only modest results are recorded for training sets consisting of 50 or 25 samples.