Principal Component Analysis Coupled with Artificial Neural Networks—A Combined Technique Classifying Small Molecular Structures Using a Concatenated Spectral Database

Abstract: In this paper we present several expert systems that predict the class identity of the modeled compounds, based on a preprocessed spectral database. The expert systems were built using Artificial Neural Networks (ANN) and are designed to predict if an unknown compound has the toxicological activity of amphetamines (stimulant and hallucinogen), or whether it is a nonamphetamine. In attempts to circumvent the laws controlling drugs of abuse, new chemical structures are very frequently introduced on the black market. They are obtained by slightly modifying the controlled molecular structures by adding or changing substituents at various positions on the banned molecules. As a result, no substance similar to those forming a prohibited class may be used nowadays, even if it has not been specifically listed. Therefore, reliable, fast and accessible systems capable of modeling and then identifying similarities at molecular level, are highly needed for epidemiological, clinical, and forensic purposes. In order to obtain the expert systems, we have preprocessed a concatenated spectral database, representing the GC-FTIR (gas chromatography-Fourier transform infrared spectrometry) and GC-MS (gas chromatography-mass spectrometry) spectra of 103 forensic compounds. The database was used as input for a Principal Component Analysis (PCA). The scores of the forensic compounds on the main principal components (PCs) were then used as inputs for the ANN systems. We have built eight PC-ANN systems (principal component analysis coupled with artificial neural network) with a different number of input variables: 15 PCs, 16 PCs, 17 PCs, 18 PCs, 19 PCs, 20 PCs, 21 PCs and 22 PCs. The best expert system was found to be the ANN network built with 18 PCs, which accounts for an explained variance of 77%. This expert system has the best sensitivity (a rate of classification C = 100% and a rate of true positives TP = 100%), as well as a good selectivity (a rate of true negatives TN = 92.77%). A comparative analysis of the validation results of all expert systems is presented, and the input variables with the highest discrimination power are discussed.