Direct classification of new psychoactive substances in seized blotter papers by ATR-FTIR and multivariate discriminant analysis

Microchemical Journal 133 (2017) 96–103

Contents lists available at ScienceDirect

Microchemical Journal journal homepage: www.elsevier.com/locate/microc

Direct classification of new psychoactive substances in seized blotter papers by ATR-FTIR and multivariate discriminant analysis☆ Leandro S.A. Pereira a, Fernanda L.C. Lisboa b,c, José Coelho Neto b,c, Frederico N. Valladão b, Marcelo M. Sena a,d,⁎ a

Departamento de Química, ICEx, Universidade Federal de Minas Gerais, 31270-901 Belo Horizonte, MG, Brazil Divisão de Laboratório, Instituto de Criminalística da Polícia Civil de Minas Gerais, 30180-060 Belo Horizonte, MG, Brazil Departamento de Física e Química, Pontifícia Universidade Católica de Minas Gerais, 30535-901 Belo Horizonte, MG, Brazil d Instituto Nacional de Ciência e Tecnologia em Bioanalítica, 13083-970 Campinas, SP, Brazil b c

a r t i c l e

i n f o

Article history: Received 15 November 2016 Received in revised form 10 February 2017 Accepted 17 March 2017 Available online 20 March 2017 Keywords: New psychoactive substances (NPS) Phenethylamines Direct analysis Partial least squares - discriminant analysis (PLS-DA) Blotter paper Forensic analysis

a b s t r a c t Due to the general increase in drug trafficking crime rates, a high amount of drug samples is continuosly seized and requires forensic analysis. In order to cover the demand for this great amount of samples in forensic investigations, non-destructive, fast and direct analysis methods are desirable. A new supervised classification method using PLS-DA (partial least squares discriminant analysis) and ATR-FTIR (attenuated total reflectance Fourier transform infrared spectroscopy) was developed to identify NPS (new psychoactive substances) drugs in blotter papers. A multivariate model was built to classify NBOMe, 2C-H, LSD, MAL (methallylescaline) and discriminate them of blank papers. A submodel was also built to discriminate 25B-NBOMe, 25C-NBOMe and 25I-NBOMe inside NBOMe class. Both models were validated through the estimate of specific figures of merit. The average of reliability rate (RLR) was 88.9%, accordance (ACC) was 91.1% and concordance (CON) was 86.1%. For the NBOMe submodel RLR was 82.2%, ACC was 100% and CON was 94.4%. The model presented high correct classification rates for all the classes, with the exception of LSD, possibly due to its lower concentration on seized blotters. The proposed method has potential to be used on blotter screening routine. The analysis is cost-effective, rapid, 2 min per sample, and utilizes ATR-FTIR, a technique whose use is increasing on forensic laboratories around the world. © 2017 Elsevier B.V. All rights reserved.

1. Introduction The new psychoactive substances (NPS) are quickly spreading through the world, and the detection and identification of these drugs are fundamental to control and confront this market [1]. Most of these substances have undergone small chemical substitutions from known illicit drugs. These small changes in their chemical structures take some NPS out of the category of illegal substances, allowing them to be marketed legally instead of prohibited. For this reason, they are also called “legal highs”. As NPS are emerging as a social problem, the prohibition or control of their use through scheduling has been a key step to tackle this drug problem. However, regulating NPS have proved to be difficult. These new drugs become less attractive to dealers as they are legally banned, and so, other “legal” drugs have emerged to replace the illegal ones [2, ☆ Selected paper from the 18th Brazilian Meeting on Analytical Chemistry, 18–21 September, 2016, Brazil ⁎ Corresponding author at: Departamento de Química, ICEx, Universidade Federal de Minas Gerais, 31270-901 Belo Horizonte, MG, Brazil. E-mail address: [email protected] (M.M. Sena).

http://dx.doi.org/10.1016/j.microc.2017.03.032 0026-265X/© 2017 Elsevier B.V. All rights reserved.

3]. The availability of different substituted drugs is broad, even if only the main classes of substituted NPS are considered, such as phenethylamines, tryptamines and cannabinoids [4,5]. The fast growth of forms originates additional problems at both the analytical and legislation levels [6]. At the analytical level, there is a need for developing rapid methods for chemical detection and characterization of new drugs. Specifically, forensic analysts must face the uncertainty about the identity of seized substances or products, and need to have access to pure standards and certified reference materials [7]. The phenethylamine class of drugs has been known for a long time [8]. Alexander and Ann Shulgin have given an important contribution to the knowledge about this class as they have described the synthesis and psychotropic effects of several substituted phenethylamines [9]. This class is named due to its basic structure, which comprises a phenyl group bounded to an amino group by a two carbons chain. Through this chemical backbone (Fig. S1A, Supplementary material) some favorable substitution sites are available. This class includes some drugs well known for their potential for abuse, such as methamphetamine [1,10], and others only recently introduced in the traffic. NBOMe series and 2C series are substituted phenethylamines recently regulated in some countries, because they have been openly

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103

sold as “legal highs” and have become a public health problem [11,12]. In the NBOMe series, a 2-methoxybenzyl group is added to the nitrogen of the 2C series (Fig. S1B). The 2C series (Fig. S1C) is composed of primary derivatives of 2,5-dimethoxyphenethylamine (2C-H). Another drug that has attracted attention due to its potent hallucinogenic effect is methallylescaline (MAL) (Fig. S1D). This synthetic analog of mescaline was reported to be seized for the first time in Sweden in 2013 [13]. MAL has also been seized in Brazil [14], but it is not yet officially regulated. These NPS are powerful agonists of serotonin receptors and potent hallucinogens with effects similar to LSD [15]. They are currently sold in blotter papers and drug dealers can mislead the users by selling them as LSD. Thus, users are consuming NPS without knowing the toxicological effects of the unknown drug. Several intoxication cases after ingestion of supposed LSD or “acid” blotters have been reported, including fatalities [11,15–17]. Forensic laboratories are dealing with increasingly large amounts of seized NPS blotter samples in recent years, but there are no currently reference color tests for these drugs [18]. The most common methods for detecting and characterizing NPS have been based on chromatographic techniques [17,19–22]. These methods have the disadvantages of requiring pretreatment steps, such as sample extraction, being destructive, expensive and time consuming. Other alternative methods have been based on mass spectrometry direct analysis [23–25]. Some papers have also characterized NPS by using spectroscopic techniques, such as Fourier transform infrared (FTIR), nuclear magnetic resonance (NMR) and Raman [17,22,26], but the relatively small number of samples analyzed did not allow to obtain broad and robust predictive models. For this aim, more sophisticated data analysis techniques based on multivariate statistics should be used. A particularly promising technique for qualitative and quantitative analysis of NPS is FTIR. Modern FTIR spectrophotometers provide rapid determinations with adequate signal-to-noise ratios, allowing direct and non-destructive analysis of solids or liquids when an attenuated total reflectance (ATR) accessory is available. Although FTIR is a very useful technique for extracting structural information from pure substances [27], IR spectra of complex real matrices cannot be satisfactorily analyzed only by simple spectral matching and univariate methods. Hence, the use of multivariate statistics is necessary. A previous paper [14] has analyzed some NPS by FTIR using spectral matching and discriminant analysis. However, this work has not developed a robust model by adopting systematic criteria for selecting training and test samples, by conducting a robust validation and by performing the spectral characterization of the model. Similarly, other recent paper [28] has applied near infrared spectroscopy (NIRS) and PCA for discriminating synthetic cannabinoids from phenethylamines, but no supervised classification model was developed. Supervised classification chemometric methods, mainly partial least squares discriminant analysis (PLS-DA), are the most appropriate alternative to build robust predictive models applied to discriminate forensic samples. They should be used within an entire multivariate strategy, including proper data preprocessing,

Fig. 1. ATR-FTIR spectra of 73 seized blotter samples and 21 paper samples.

representative criteria for splitting samples in training and test sets, estimate of specific figures of merit for method validation and identification of the most discriminant variables by inspecting model informative vectors. Recently, several papers have developed PLS-DA models for the discriminant analysis of forensic samples, such as drugs [29–31], explosives [32], fuels [33], soils [34], documents [35], adulterated food [36] and counterfeit beverages [37]. As quantitative methods, supervised classification methods can be submitted to a full analytical validation in order to obtain official recognition. However, multivariate qualitative validation has received only very recently attention in the literature [38–40]. Since the responses provided by qualitative methods are discrete, statistical tests and procedures used differ from the ones used in quantitative validation. The most common qualitative figures of merit (FOM) are false-positive rate (FPR), false-negative rate (FNR), sensitivity rate (SNR) and selectivity rate (SLR). The former two FOM are related to the trueness of the methods, while the latter two are related to their qualitative selectivity. All of these are easily calculated from the confusion matrix. Other more general FOM are used to evaluate trueness, such as the rate of correctly classified samples (%CC) [41] and the reliability rate (RLR) [39]. The former is calculated as the ratio between the sum of true-positives plus true-negatives and all the results, while the latter is more robust and is calculated with result rates of misclassified samples. The qualitative precision is estimated as accordance (ACC) at the repeatability level, and concordance (CON) at the level of intermediary precision or reproducibility [42]. Both of these FOM are calculated by combinatory possibilities of two concordant results. The aim of this paper was to develop rapid and non-destructive supervised classification methods for discriminating seized blotter samples from four different classes of synthetic drugs, NBOMe, 2C-H, LSD and MAL, and blank papers, based on PLS-DA and ATR-FTIR spectra. In a hierarchical strategy, a PLS-DA submodel was built in the sequence to differentiate three specific NBOMe molecules. All the developed methods were validated through the estimate of proper FOM. In

Table 1 Number of samples in each class for the training and test sets. Class Main model NBOMe 2C-H LSD1 LSD2 MAL Paper NBOMe sub-model 25B-NBOMe 25C-NBOMe 25I-NBOMe

Training set

Test set

17 6 7 6 13 14

8 3 3 3 7 7

3 5 8

2 3 4

97

Fig. 2. PCA model including all the samples. PC1 vs. PC2 scores.

98

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103

detector and a single-bounce Smart Orbit accessory module containing a diamond ATR crystal. Chromatographic reference analyses were carried out using an Agilent 7890A GC system with an Agilent 5975C mass spectrometry detector, for gas chromatography, and a Shimadzu LCMS-8030 (mass detector) UFLC class 20A, for liquid chromatography. Data was obtained in the software OMNIC 9 (Thermo Fisher Scientific, Waltham, MA, USA) and handled using MATLAB version 7.13 (The MathWorks, Natick, MA, USA). Chemometrics routines came from the PLS Toolbox, version 6.7.1 (Eigenvector Technologies, Manson, MA, USA). Spectral comparisons were performed based on the spectral libraries of OMNIC software, which included HR Comprehensive Forensic FT-IR Collection and HR Georgia State Forensic Drugs Library, available commercially from Thermo Fisher Scientific, and SWGDRUG IR library [44]. Fig. 3. Mean spectra of classes LSD1 and LSD2.

2.2. Samples and spectra acquisition addition, informative vectors, such as variable importance in projection (VIP) scores [43] and regression coefficients, were used for interpreting models and spectrally characterizing each type of drug. 2. Materials and methods 2.1. Apparatus and software Blotter sample spectra were recorded using a Thermo Nicolet iZ10 ATR-FTIR spectrophotometer with an EverGlo IR source, a DLaTGS IR

73 blotter samples seized in 2014 and 2015 in Minas Gerais State were used for the development and validation of the method. Samples were analyzed with the reference methods used in laboratory of the “Polícia Civil de Minas Gerais”, by GC–MS or LC-MS, which were based on the relevant literature [17,45]. The chromatographic results were used as reference class assignments for the construction of classification models. Seized blotter samples containing more than one drug were not used in the models. In the analyzed samples, six different drugs were found: 25B-NBOMe, 25C-NBOME, 25I-NBOMe, 2C-H, LSD and MAL. A

Fig. 4. Classification plots for the main model. Classes (A) NBOMe, (B) 2C-H, (C) LSD1, (D) LSD2, (E) MAL, and (F) paper. Horizontal line indicates the threshold value. Vertical line splits training and test samples.

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103 Table 2 Estimated FOM for the main model, as percentage, for each class in training and test sets. Training set Class

FPR

FNR

RLR

SNR

SLR

NBOMe 2C-H LSD1 LSD2 MAL Paper Mean

2.2 0.0 7.1 0.0 0.0 6.1 2.6

17.6 0.0 28.6 0.0 15.4 7.1 2.6

80.2 100 64.3 100 84.6 86.8 86.0

82.4 100 71.4 100 84.6 92.9 88.5

97.8 100 92.9 100 100 93.9 97.4

Test set Class

FPR

FNR

RLR

SNR

SLR

ACC

CON

NBOMe 2C-H LSD1 LSD2 MAL Paper Mean

0.0 7.1 7.1 0.0 0.0 4.2 3.1

25.0 0.0 0.0 0.0 14.3 0.0 10.7

75.0 92.9 92.9 100 85.7 95.8 91.2

75.0 100 100 100 85.7 100.0 89.3

100 92.9 92.9 100 100 95.8 96.9

100 100 46.7 100 100 100 91.11

100 100 33.3 83.3 100 100 86.1

“blank paper” class was created to simulate blank blotters with 21 samples representative of papers used in the NPS trafficking. Spectra were recorded directly from the artwork side of the blotter, in the wavenumber range from 4000 to 400 cm−1 with resolution of 4 cm−1, and as an average of 16 scans. Before recording each spectrum, a background correction was performed in order to minimize atmospheric interference and instrumental noise.

99

2.3. Modeling Data were preprocessed by standard normal variate (SNV) in order to remove non-linear baseline deviations (drifts) caused by multiplicative light scattering, which is present in reflectance spectra of solid samples [46]. A principal component analysis (PCA) model was previously built with mean centered data, aiming to observe the similarities between samples in a non-supervised manner. Then, a general PLS2-DA model was built. The three detected NBOMe molecules have similar spectra, since they differ only by a different halogen atom in the same chemical structure. Thus, the general PLS-DA model was not able to discriminate specific NBOMe, and these three drugs were grouped in a single class. LSD samples were separated in two classes (LSD1 and LSD2) because they presented two different patterns of spectral features, as will be discussed in the section 3.1. For the whole analysis, a hierarchical strategy was defined, consisting of two models. The main model was trained with 6 classes: NBOMe, 2C-H, LSD1, LSD2, MAL, and blank paper. The secondary model was built to discriminate each specific member of the NBOMe class: 25B-NBOMe, 25C-NBOMe, and 25I-NBOMe. For each class, samples were systematically separated as 2/3 for the training set and 1/3 for the test set, using the Kennard-Stone algorithm [47]. The number of samples in each class is shown in Table 1. Data used for building supervised classification methods should not be mean centered if the classes have different numbers of samples, since this leads the model to be weighted by the larger classes [48]. So, a class centroid centering preprocessing was preferred. For predictions, Bayesian threshold values were adopted [40,48]. The number of latent

Fig. 5. VIP scores for the main model. Classes (A) NBOMe, (B) 2C-H, (C) LSD1, (D) LSD2, (E) MAL, and (F) paper.

100

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103

variables (LV) was estimated by cross-validation, using venetian blinds with 10 splits for the main model and leave-one-out for the secondary NBOMe model. The chosen number of LV was based on the minimum cross-validation classification error (CVCE). 2.4. Validation Qualitative validation was performed by evaluating trueness, through FPR, FNR and RLR, selectivity, through SNR and SLR, and precision, through ACC and CON. FPR, FNR, RLR, SNR and SLR were estimated, for each class, in both training and test sets, and the average of each FOM was attributed to the overall model. ACC and CON were estimated using one sample of each class from the validation set. Six replicates of each sample were obtained for estimating ACC. The same samples were analyzed in other day by other analyst to estimate CON. 3. Results and discussion 3.1. Preliminary modeling ATR-FTIR spectra of all the samples, including the “blank paper” class, are displayed in Fig. 1. In addition, representative spectra of one sample of each drug class (the spectrum most similar to the average spectrum of each class) are shown in Fig. S2. A preliminary PLS-DA model was built modeling each type of drug as a single class. This model provided misclassification of a high number of NBOMe samples, which were attributed to other NBOMe classes. The spectra of the three types of NBOMe drugs are very similar and this model was not able to discriminate them in the much larger variance space that included all the others classes. Hence, a single class incorporating all the studied NBOMe was created in the main model. In the sequence, a hierarchical sub-model was also built to discriminate only each specific NBOMe type. Another problem was observed with the preliminary general model. Some LSD samples were discriminated with high predicted y values, being classified with high probability of belonging to their true class. However, the other part of LSD samples were predicted with low y values and misclassified as paper. In an attempt to understand the reasons of these results, a PCA model was built. The first two PC of this model accounted for 75.51% of the spectral variance. In the scores plot showed in Fig. 2, LSD samples were grouped in two different regions. The first part showed high positive scores on PC2 and the other part of the LSD samples presented lower negative or positive scores on both PC1 and PC2. In brief, PC2 discriminated almost half of the LSD samples from all the other drug samples. Then, a search in spectral libraries was carried out in an attempt to characterize these samples. The spectra of the samples with high positive scores on PC2 reasonably matched with the spectrum of a copolymer of styrene and methyl methacrylate. It was concluded that these samples contained this additive, which was not found in any of the other analyzed drug samples. In fact, PC2 modeled the presence of this copolymer rather than discriminating the LSD itself. This is an interesting side aspect of this work, since the presence of specific blotter additives that can be correlated with specific drugs represents valuable information for the investigations of NPS trafficking. Thus, for the new PLS-DA model LSD samples were split in two classes, LSD1 and LSD2, without and with this copolymer, respectively. Mean spectra of LSD1 and LSD2 classes are presented in Fig. 3. Major differences observed between these two classes of spectra were the narrow peaks at 698, 758, 1452, 1493, 1601 and 1728 cm−1, and the band between 2800 and 3100 cm−1. All those vibrations were attributed to the detected copolymer: the peaks at 1601 and 1728 cm−1 were attributed to the stretching of the C_C aromatic ring of styrene, and of the carbonyl group of methyl methacrylate, respectively; these peaks have specifically been used for the determination of this copolymer [49]; the peak at 698 cm−1 is characteristic of the aromatic C\\H out-of-plane bending of the styrene component [49]; the peaks at 758

Table 3 The most discriminant wavenumbers for each class, detected by VIP scores and reference spectra matching. In parenthesis, the signal contribution of the wavenumbers for each regression vector. Class

Discriminant wavenumbers (cm−1)

NBOMe 2C-H LSD1 LSD2 MAL Paper

1495 (+); 1224 (−); 1213 (+); 756 (+); 698 (−) 1503 (+); 1224 (+); 798 (+) 1224 (−); 1213 (+); 1130 (−); 873 (+) 2940–2840 (+); 1728 (+); 1213 (−); 1030 (−); 698 (+) 1590 (+), 1424 (+); 1340 (+); 1213 (+); 1130 (+); 1030 (−) 1244 (+); 1213 (−); 1130 (+); 1030 (+); 873 (−); 698 (+)

and 1452 cm− 1 were also attributed to the aromatic C_C bending, while the bands around 3000 cm−1 were associated to symmetric and asymmetric stretching of the aromatic C\\H [50]. This copolymer has been reported in printing papers to modify their wettability [51]. Considering that this type of copolymer is utilized to fix the ink to the paper and produce better resolution printing, the same property may be used to adjust the drug content in blotters. Hence, the difference between the two classes of LSD samples was attributed to distinct printings of the blotter papers.

Fig. 6. Classification plots for the NBOMe model. Classes (A) 25B-NBOMe, (B) 25C-NBOMe, and (C) 25I-NBOMe. Horizontal line indicates the threshold value. Vertical line splits training and test samples.

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103 Table 4 Estimated FOM for the NBOMe sub-model, as percentage, for each class in training set and test set. Training set Class

FPR

FNR

RLR

SNR

SLR

25B-NBOMe 25C-NBOMe 25I-NBOMe Mean

0 18.2 0 6.1

33.3 0 12.5 15.3

66.7 81.8 87.5 78.7

66.7 100 87.5 84.7

100 81.8 100 93.9

Test set Class

FPR

FNR

RLR

SNR

SLR

ACC

CON

25B-NBOMe 25C-NBOMe 25I-NBOMe Mean

0 0 20 6.7

0 33.3 0 11.1

100 66.7 80 82.2

100 66.7 100 88.9

100 100 80 93.3

100 100 100 100

100 83.3 100 94.4

3.2. Main model The general PLS-DA model was built with 9 LV, accounting for 99.5% of spectral variance (X) and 64.1% of class variance (Y). Predicted values for the samples of each class are presented in Fig. 4. After the PLS-DA model was built, a standard constraint (PLS Toolbox, v. 6.7.1) was

Fig. 7. VIP scores for the NBOMe model. Classes (A) 25B-NBOMe, (B) 25C-NBOMe, and (C) 25I-NBOMe.

101

imposed forcing each sample to belong to only one single class. Hence, on the basis of the Bayesian theorem each sample was classified as belonging to the class to which it showed higher probability. Then, the confusion matrix was obtained (Table S1, Supplementary material). If a new drug not included in the model development is predicted, it would be expected that this drug is not attributed to any of the classes (all the predicted y values below the respective thresholds). However, if this new drug has a chemical structure very similar to an already modeled NPS, it may be wrongly predicted as a false-positive. In any case, when the model is expanded to include samples of this new drug, it likely provides correct predictions. The method validation was performed by estimating the FOM presented in Table 2. The main model misclassified only 8 samples in the training set and 5 samples in the test set. The majority of these misclassified samples are related to LSD1 class, as false-positive or false-negative. This class was not well modeled and its RLR, 64.3%, is considered low. In Fig. 4C, which shows y values for LSD1 class, all samples presented low y values, even the own LSD1 samples, for which y values close to 1 would be expected. Furthermore, the y values for LSD1 class of several samples of NBOMe, MAL and paper classes were above the threshold, which could be wrongly predicted as belonging to this class if the model had not been constrained to single class predictions. This is another evidence that the LSD1 class was not well modeled. LSD is a very potent hallucinogen and its concentration in blotters is usually as low as dozens μg per dose [19,52], one or more orders of magnitude below other NPS. Hence, misclassifications of LSD1 samples as paper samples were expected, due to the limited sensitivity of the technique. Two samples of 25B-NBOMe class were misclassified into paper class and, by analogy, low drug contents could be expected in these samples. Only one paper sample (out of 21) was misclassified as containing drug. ACC and CON results were both 100%, except for LSD classes, showing that the model is very precise and the predictions are reproducible. Aside from LSD and the two samples of NBOMe in the test set, all other classes presented RLR above 80%. Considering that there are no regulations that define acceptable values for these figures of merit, these results can be considered reasonable. As a screening method, false positive results can be circumvented by the complementary analysis with a combination of various techniques required for identity confirmation, according to SWGDRUG recommendations [44]. Spectral interpretation of the models is an important aspect in the development of robust classification or calibration multivariate spectroscopic methods. The use of regression coefficients as the only informative vector for interpreting multivariate models has been criticized, since these vectors are dependent on the composition of the samples in the training set, on the implicit covariance of the components, and on the signal to noise ratio [53]. As a better alternative for this interpretation, VIP scores vectors have been preferred to identify the most discriminant variables in the spectra [43]. VIP scores higher than 1.0 are considered significantly discriminant for the models. However, VIP scores provide absolute values, not allowing relating the most important variables with the analyte or the interferences. Thus, model spectral interpretation may be improved by jointly inspecting VIP scores and regression coefficients. Fig. 5 shows the VIP scores for each class of the main model. Regression vectors are shown in Fig. S3. For spectral matching, reference spectra of LSD, styrene and acrylate (components of the copolymer found in samples of the LSD2 class), and blank paper (copier paper) were available in the spectral libraries of the OMNIC software [54]. 25B-NBOMe, 25C-NBOMe, 25I-NBOMe and 2C-H spectra can be found in the SWGDRUG IR spectral library [44]. MAL is a very recent drug in forensic research and no spectrum was found in either published works or spectral libraries. Then, a full picture blotter sample (5 × 5 doses) was extracted with methanol and purified. Purified extract content was confirmed to be MAL by GC–MS and LC-MS. A spectrum of this extract was recorded (Fig. S4) and used as reference. By observing the VIP scores (Fig. 5), the fingerprint region of the infrared spectra was considered important to all the classes. The highest

102

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103

VIP scores of each class were checked for the presence in the respective reference spectra. The most important VIP scores and their signal contribution for the regression vector of each class (Fig. S3) are listed in Table 3. Positive regression coefficients mean specific contribution for discriminating a drug (or paper) class. For all the classes, with the exception of LSD1, these variables/wavenumbers were identified in the reference spectra. The most discriminant signals for NBOMe and 2C-H classes were similar, with small shifts. Peaks at 1495 and 1503 cm−1 can be attributed to aromatic C_C\\C vibration [17], peaks at 1213 and 1224 cm−1 are related to aryl ethers [22], and at 756 and 798 cm−1 can be attributed to the out-of-plane C\\H bending of the aromatic ring [50]. For the LSD1 class, the highest VIP scores did not match the main peaks of the reference spectrum, and some variables were also discriminant for blank papers, such as 1213, 1130 and 873 cm− 1. The peak at 873 cm− 1 was associated to the carbonate out-of-plane bending, since calcium carbonate is a common addictive in the paper industry for filling and coating [55]. This strengthens the limitation of this model for classifying LSD samples, due to the low content of this drug in the blotters and the lack of sensitivity for these samples. For the same reasons, samples of the LSD2 class were correctly classified based on the presence of the styrene-methyl methacrylate copolymer. Spectral attributions for this substance were already discussed in the end of section 3.1 and most of the discriminant wavenumbers previously attributed. MAL samples were discriminate mainly due to the wavenumbers 1590 and 1130 cm−1, attributed to the N\\H bending (scissoring) and the C\\N stretching of primary amines, respectively, 1213 cm−1, associated to C\\O\\C stretching vibrations of aromatic ethers, and 1424 cm−1, attributed to C\\H in plane bending of aliphatic alkenes [50]. Finally, the “paper” class can be mainly discriminated by the peaks at 1213 and 1130 cm− 1, which were attributed to C\\H bending and C\\O\\C asymmetric stretching of cellulose. 3.3. NBOMe model A hierarchical submodel was built aiming to specifically discriminate the three NBOMe found in the samples. This model was built with 5 LV and accounted for 96.4% of the variance in X and 66.5% of the variance in Y. The classification plots for each class are presented in Fig. 6 and the confusion matrix is presented in Table S2 (Supplementary material). Estimated FOM for this model are presented in Table 4. This submodel is limited by the low number of available samples in each class, which makes it less robust and less reliable than the previous main model. As an example, a single sample of 25B-NBOMe was misclassified in the training set, which means 33.3% of FNR for this class. Nevertheless, the overall RLR for this model are considered reasonable, with values of 79% and 89% for the training and test sets, respectively. For this submodel, VIP scores vectors are presented in Fig. 7. Regression vectors are shown in Fig. S5. The similarity of the structures and the spectra of these three NBOMe leads to several important VIP scores in common for the three classes, such as 2917, 2849, 1503, 1492, 1217, 801, 765 and 756 cm− 1. The main discriminant wavenumbers for each class were: 1492 and 756 cm− 1 for 25B-NBOMe; 2917, 2849, 1505, 810 and 765 cm−1 for 25C-NBOMe; and 1217 and 765 cm−1 for 25I-NBOMe. 4. Conclusions A direct analysis method based on ATR-FTIR and PLS-DA was developed and validated for the rapid screening of NPS in blotter papers. A multivariate model classified samples of three classes, NBOMe, 2C-H and MAL, with relatively high efficiency, and was able to differentiate them from blank papers. However, this model was not effective for the classification of LSD samples due to the lower contents of this drug found in seized blotters, and to the limit of sensitivity of the technique.

The present method is simple, of low cost, non-destructive and fast, spending only 2 min per sample analysis, including background scan and cleaning, which makes it a potential candidate for the development of rapid screening routines in forensic laboratories. This is a promising approach that can be made more robust by increasing the number of samples, and be expanded by incorporating new classes of NPS. This last feature is very useful in this type of forensic analysis due to the dynamic nature of the NPS illegal market. ATR-FTIR spectrophotometers are commonly available in most of the official forensic laboratories in Brazil. Hence, this method could also be expanded and generalized for the simultaneous use in several laboratories located at different places by developing calibration/classification transfer models [56]. Conflict of interest None of the authors have any conflict of interest. Acknowledgement CAPES and CNPq for students sponsorships, CAPES for granting the project CAPES/Pró-Forenses N° 025/2014, and FIP/PUC Minas program. Appendix A. Supplementary data Chemical structures of the four analyzed drugs, representative spectra of each drug class, correlation coefficients for the main model and the submodel, ATR-FTIR spectrum of MAL, and confusion matrices for the main model and the submodel. Supplementary data associated with this article can be found in the online version, at http://dx.doi. org/10.1016/j.microc.2017.03.032. References [1] United Nations Office on Drugs and Crime (UNODC), World Drug Report, Vienna, Austria, 2015. [2] A. Ledberg, The interest in eight new psychoactive substances before and after scheduling, Drug Alcohol Depend. 152 (2015) 73–78. [3] D. de Boer, I. Bosman, A new trend in drugs-of-abuse; the 2C-series of phenethylamine designer drugs, Pharm. World Sci. 26 (2004) 110–113. [4] L.A. Johnson, R.L. Johnson, R.-B. Portier, Current “legal highs”, J. Emerg. Med. 44 (2013) 1108–1115. [5] M.E. Nelson, S.M. Bryant, S.E. Aks, Emerging drugs of abuse, Dis. Mon. 60 (2014) 110–132. [6] L.A. King, A.T. Kicman, A brief history of ‘new psychoactive substances’, Drug Test. Anal. 3 (2011) 401–403. [7] S.D. Brandt, L.A. King, M. Evans-Brown, The new drug phenomenon, Drug Test. Anal. 6 (2014) 587–597. [8] F.J. Carod-Artal, Hallucinogenic drugs in pre-Columbian Mesoamerican cultures, Neurologia 30 (2015) 42–49. [9] A. Shulgin, A. Shulgin, Pihkal: A Chemical Love Story, Transform Press, Berkeley, USA, 1991. [10] K. Tsujikawa, K. Kuwayama, H. Miyaguchi, T. Kanamori, Y.T. Iwata, H. Inoue, Chemical profiling of seized methamphetamine putatively synthesized from phenylacetic acid derivatives, Forensic Sci. Int. 227 (2013) 42–44. [11] V.B. Kueppers, C.T. Cooke, 25I-NBOMe related death in Australia: a case report, Forensic Sci. Int. 249 (2015) E15–E18. [12] B. Dean, S. Stellpflug, A. Burnett, K. Engebretsen, 2C or not 2C: phenethylamine designer drug review, J. Med. Toxicol. 9 (2013) 172–178. [13] European Monitoring Centre for Drugs and Drug Abuse (EMCDDA), EMCDDA– Europol, Annual Report on the Implementation of Council Decision 005/387/JHA, 2014, Publications Office of the European Union, Luxembourg, 2013. [14] J. Coelho Neto, Rapid detection of NBOME's and other NPS on blotter papers by direct ATR-FTIR spectrometry, Forensic Sci. Int. 252 (2015) 87–92. [15] J. Suzuki, M.A. Dekker, E.S. Valenti, F.A.A. Cruz, A.M. Correa, J.L. Poklis, A. Poklis, Toxicities associated with NBOMe ingestion - a novel class of potent hallucinogens: a review of the literature, Psychosomatics 56 (2015) 129–139. [16] J.L. Poklis, S.A. Raso, K.N. Alford, A. Poklis, M.R. Peace, Analysis of 25I-NBOMe, 25BNBOMe, 25C-NBOMe and other dimethoxyphenyl-N-(2-methoxyphenyl) methyl ethanamine derivatives on blotter paper, J. Anal. Toxicol. 39 (2015) 617–623. [17] D. Zuba, K. Sekula, A. Buczek, 25C-NBOMe - new potent hallucinogenic substance identified on the drug market, Forensic Sci. Int. 227 (2013) 7–14. [18] E. Cuypers, A.J. Bonneure, J. Tytgat, The use of presumptive color tests for new psychoactive substances, Drug Test. Anal. 8 (2016) 137–141. [19] P.A. Marinho, E.M. Alvarez Leite, Quantification of LSD in illicit samples by high performance liquid chromatography, Braz. J. Pharm. Sci. 46 (2010) 695–703.

L.S.A. Pereira et al. / Microchemical Journal 133 (2017) 96–103 [20] L. Burrai, M. Nieddu, M. Palomba, M.A. Pirisi, Identification and quantitation of 4bromo-2,5-dimethoxyamphetamine in seized blotters, Legal Med. 17 (2015) 56–59. [21] D.A.S. Robayo, W.F.G. Mendez, G.T. Ocampo, M.R. Moreano, Analysis of blotter papers employed in the commercialization of new hallucinogenic substances of the 2,5-dimethoxy-N-(2-methoxybenzyl) phenethylamine series seized in the city of Bogotá by applying gas chromatography coupled to a selective mass ion trap detector, J. Braz. Chem. Soc. 27 (2016) 992–997. [22] D. Zuba, K. Sekula, Analytical characterization of three hallucinogenic N-(2methoxy)benzyl derivatives of the 2C-series of phenethylamine drugs, Drug Test. Anal. 5 (2013) 634–645. [23] W. Romao, B.D. Sabino, M. Bueno, B.G. Vaz, A.C. Junior, A.O. Maldaner, E.V.R. de Castro, R.A. Lordeiro, C.C. Nascentes, M.N. Eberlin, R. Augusti, LSD and 9,10dihydro-LSD analyses in street drug blotter samples via easy ambient sonic-spray ionization massspectrometry (EASI-MS), J. Forensic Sci. 57 (2012) 1307–1312. [24] T.J. Kauppila, V. Arvola, M. Haapala, J. Pol, L. Aalberg, V. Saarela, S. Franssila, T. Kotiaho, R. Kostiainen, Direct analysis of illicit drugs by desorption atmospheric pressure photolonization, Rapid Commun. Mass Spectrom. 22 (2008) 979–985. [25] T.C. Carvalho, I.F. Oliveira, L.V. Tose, G. Vanini, J.B. Kill, A.C. Neto, L.F. Machado, J.C.L. Ambrosio, V. Lacerda, B.G. Vaz, W. Romao, Qualitative analysis of designer drugs by paper spray ionisation mass spectrometry (PSI-MS), Anal. Methods 8 (2016) 614–620. [26] R. Christie, E. Horan, J. Fox, C. O'Donnell, H.J. Byrne, S. McDermott, J. Power, P. Kavanagh, Discrimination of cathinone regioisomers, sold as ‘legal highs’ by Raman spectroscopy, Drug Test. Anal. 6 (2014) 651–657. [27] Scientific Working Group for the Analysis of Seized Drugs (SWGDRUG), Supplemental Document SD-2 for part IVB - Quality Assurance/Validation of Analytical Methods, 2009. [28] R. Risoluti, S. Materazzi, A. Gregori, L. Ripani, Early detection of emerging street drugs by near infrared spectroscopy and chemometrics, Talanta 153 (2016) 407–413. [29] N.V.S. Rodrigues, E.M. Cardoso, M.V.O. Andrade, C.L. Donnici, M.M. Sena, Analysis of seized cocaine samples by using chemometric methods and FTIR spectroscopy, J. Braz. Chem. Soc. 24 (2013) 507–517. [30] M.C.A. Marcelo, K.C. Mariotti, M.F. Ferrao, R.S. Ortiz, Profiling cocaine by ATR-FTIR, Forensic Sci. Int. 246 (2015) 65–71. [31] C.M. Liu, Z.D. Hua, Y.P. Bai, Y. Liu, Profiling and classification of illicit heroin by ICPMS analysis of inorganic elements, Forensic Sci. Int. 239 (2014) 37–43. [32] M.A.F. de la Ossa, C. Garcia-Ruiz, J.M. Amigo, Near infrared spectral imaging for the analysis of dynamite residues on human handprints, Talanta 130 (2014) 315–321. [33] N.C. da Silva, M.F. Pimentel, R.S. Honorato, M. Talhavini, A.O. Maldaner, F.A. Honorato, Classification of Brazilian and foreign gasolines adulterated with alcohol using infrared spectroscopy, Forensic Sci. Int. 253 (2015) 33–42. [34] M. Baron, J. Gonzalez-Rodriguez, R. Croxton, R. Gonzalez, R. Jimenez-Perez, Chemometric study on the forensic discrimination of soil types using their infrared spectral characteristics, Appl. Spectrosc. 65 (2011) 1151–1161. [35] F.D.L. Borba, R.S. Honorato, A. de Juan, Use of Raman spectroscopy and chemometrics to distinguish blue ballpoint pen inks, Forensic Sci. Int. 249 (2015) 73–82. [36] K.M. Nunes, M.V.O. Andrade, A.M.P. Santos, M.C. Lasmar, M.M. Sena, Detection and characterisation of frauds in bovine meat in natura by non-meat ingredient additions using data fusion of chemical parameters and ATR-FTIR spectroscopy, Food Chem. 205 (2016) 14–22.

103

[37] H.V. Pereira, V.S. Amador, M.M. Sena, R. Augusti, E. Piccin, Paper spray mass spectrometry and PLS-DA improved by variable selection for the forensic discrimination of beers, Anal. Chim. Acta 940 (2016) 104–112. [38] M.I. Lopez, M.P. Callao, I. Ruisanchez, A tutorial on the validation of qualitative methods: from the univariate to the multivariate approach, Anal. Chim. Acta 891 (2015) 62–72. [39] C.D. Gondim, O.A.M. Coelho, R.L. Alvarenga, R.G. Junqueira, S.V.C. de Souza, An appropriate and systematized procedure for validating qualitative methods: its application in the detection of sulfonamide residues in raw milk, Anal. Chim. Acta 830 (2014) 11–22. [40] B.G. Botelho, N. Reis, L.S. Oliveira, M.M. Sena, Development and analytical validation of a screening method for simultaneous detection of five adulterants in raw milk using mid-infrared spectroscopy and PLS-DA, Food Chem. 181 (2015) 31–37. [41] R.G. Brereton, Consequences of sample size, variable selection, and model validation and optimisation, for predicting classification ability from analytical data, TrAC Trends Anal. Chem. 25 (2006) 1103–1111. [42] S.D. Langton, R. Chevennement, N. Nagelkerke, B. Lombard, Analysing collaborative trials for qualitative microbiological methods: accordance and concordance, Int. J. Food Microbiol. 79 (2002) 175–181. [43] I.G. Chong, C.H. Jun, Performance of some variable selection methods when multicollinearity is present, Chemom. Intell. Lab. Syst. 78 (2005) 103–112. [44] Scientific Working Group for the Analysis of Seized Drugs (SWGDRUG), SWGDRUG IR Library v. 1.2, 2015. [45] United Nations Office on Drugs and Crime (UNODC), Guidance for the Validation of Analytical Methodology and Calibration of Equipment used for Testing of Illicit Drugs in Seized Materials and Biological Specimens, United Nations Publication, Vienna, Austria, 2009. [46] A. Rinnan, F. van den Berg, S.B. Engelsen, Review of the most common pre-processing techniques for near-infrared spectra, TrAC Trends Anal. Chem. 28 (2009) 1201–1222. [47] R.W. Kennard, L.A. Stone, Computer aided design of experiments, Technometrics 11 (1969) 137–148. [48] R.G. Brereton, G.R. Lloyd, Partial least squares discriminant analysis: taking the magic away, J. Chemom. 28 (2014) 213–225. [49] S. Mori, Compositional analysis and infrared-spectra of styrene methyl-methacrylate random copolymers, J. Appl. Polym. Sci. 38 (1989) 547–555. [50] D.L. Pavia, G.M. Lampman, G.S. Kriz, J.R. Vyvyan, Introduction to Spectroscopy, fourth ed. Cengage Learning, Stanford, USA, 2008. [51] M. Stankovska, J. Gigac, M. Letko, E. Opalena, The effect of surface sizing on paper wettability and on properties of inkjet prints, Wood Res. 59 (2014) 67–76. [52] T. Passie, J.H. Halpern, D.O. Stichtenoth, H.M. Emrich, A. Hintzen, The pharmacology of lysergic acid diethylamide: a review, CNS Neurosci. Ther. 14 (2008) 295–314. [53] C.D. Brown, R.L. Green, Critical factors limiting the interpretation of regression vectors in multivariate calibration, TrAC Trends Anal. Chem. 28 (2009) 506–514. [54] Thermo Fisher Scientific, Omnic 9.1.27, Waltham, Massachusetts, USA. [55] M.A. Hubbe, R.A. Gill, Fillers for papermaking: a review of their properties, usage practices, and their mechanistic role, Bioresources 11 (2016) 2886–2963. [56] T. Fearn, Standardisation and calibration transfer for near infrared instruments: a review, J. Near Infrared Spectrosc. 9 (2001) 229–244.