FTIR microspectroscopy coupled with variable selection methods for the identification of flunitrazepam in necrophagous flies

Abstract

The detection and identification of a drug in a corpse through the analysis of fly larvae feeding on the body by spectroscopic techniques promises to be of great value, because of their sensitivity, promptness, low cost and simplicity. Therefore, the purpose of this study was to develop a method based on Fourier-transform infrared (FTIR) microscopy to identify and discriminate flunitrazepam in necrophagous flies (Chrysomya megacephala, Chrysomya albiceps and Cochliomyia macellaria) as a non-invasive and non-destructive technique. Thirty-two Wistar mice were divided into two groups of sixteen and supplemented in two categories: group 1 – ethanol; and group 2 – standard flunitrazepam at a dose of 2 mg kg⁻¹. Spectra from the larvae samples were analyzed by principal component analysis-linear discriminant analysis (PCA-LDA), and variable selection techniques such as successive projection algorithm (SPA-LDA) and genetic algorithm (GA-LDA) to determine if control versus flunitrazepam could be segregated. In addition, the multivariate classification accuracy results were tested based on sensitivity, specificity, positive (or precision) and negative predictive values, Youden index, and positive and negative likelihood ratios. For control vs. flunitrazepam category, the sensitivity and specificity levels, using 46 wavenumbers by SPA-LDA, gave relatively good accuracy (up to 82.3% control vs. flunitrazepam). The resulting GA-LDA model also successfully classified both classes with respect to the main biochemical alterations induced by flunitrazepam using only 40 wavenumbers (up to 88.2% control vs. flunitrazepam). Compared to classical methods, this new approach could represent an alternative and an innovative tool for faster and cheaper evaluation in entomotoxicology.

---

Introduction

Entomotoxicology ranks among the newest branches of forensic entomology that deals with the identification and quantitation of drugs and other toxins in carrion-feeding arthropods in decomposing tissues, and the study of drug-induced changes in arthropod growth with respect to the estimate of the post-mortem interval by entomological methods.^1,2 The use of necrophagous species as a matrix for qualitative compound (drugs,^3,4 metals^5,6 and pesticides^7,8) detection in insect tissues has been generally accepted by forensic toxicologists. Although there has been recent progress in the detection of toxic substances in intact insects, there are some limitations such as insufficient knowledge of insect development and activity, proper use and validation of analytical procedures and lack of a general consensus concerning experimental set-up and sampling.

Several analytical drug detection/quantification procedures have been used for the analysis of insect tissues. These include radioimmunoassay,^9,10 gas chromatography/mass spectrometry,^11,12 and high-performance liquid chromatography-mass spectrometry,^13,14 coupled with classic extraction techniques such as protein precipitation, liquid–liquid extraction or solid phase extraction. However, these techniques carried out in an entomotoxicological context are expensive, invasive, destructive, involve numerical preparation steps and most of the time they require pools of specimens to detect any present drug. As a consequence, there has been increased interest in the use of alternative, new methods for detection and identification of a drug being present in a corpse.

The detection and identification of a drug present in a corpse through the analysis of fly larvae feeding on the body by spectroscopic techniques promises to be of great value, because of their sensitivity, promptness, low cost and simplicity. Recently, we used near-infrared spectroscopy (NIRS) as a novel non-destructive method for the identification of flunitrazepam in Chrysomya megacephala (Fabricius) (Diptera: Calliphoridae) larvae, puparia and adults,¹⁵ with the resulting study successfully detecting biochemical alterations for the insect. NIRS is characterized by broad overlapping spectral peaks produced by the overtones and combinations of infrared vibrational modes. While these results are encouraging, other databases of vibrational spectra for entomotoxicological results must be established. Fourier-transform infrared (FTIR) spectroscopy is one technique with potential applications in the field of entomotoxicology. FTIR spectroscopy has the ability to rapidly generate a “biochemical cell-fingerprint” of the material under analysis.¹⁶ IR spectroscopy is also characterized by a minimum of sample handling, requiring no extractions and is non-destructive.¹⁷ Unlike conventional techniques used in the analysis of insect tissues, IR spectroscopy yields a precise representation of all the chemical bonds present in a sample and offers the opportunity to very quickly observe all metabolic modifications induced by a specific compound.¹⁸ In this context, it could be interesting to develop a strategy based on IR spectroscopy for detection and identification of a drug in a corpse.

Also, the use of appropriate chemometric tools for multivariate calibration and classification is largely responsible for advancing spectroscopic techniques, for instance, IR and NIR. Computational approaches [e.g., principal component analysis (PCA),¹⁹ linear discriminant analysis (LDA),²⁰ genetic algorithm (GA)²¹ and successive projections algorithm (SPA)²²] permit the processing of large amounts of spectroscopic data variables that subsequently require data reduction approaches in order to identify sources of variance across spectra and for inter-class variation to be identified.

Not only is the choice and development of computational approaches important to ensure reliable drug detection and quantification using spectroscopic techniques, but also multivariate classification quality features such as sensitivity, specificity, positive (or precision) and negative predictive values, Youden index, and positive and negative likelihood ratios should be calculated to ensure the validity of the results in accordance with international guidelines.²³ When reviewing entomotoxicological publications, the methods tend to lack proper validation.

Herein, we have attempted to evaluate the potential of a novel non-destructive method based on attenuated total reflection (ATR)-FTIR microspectroscopy for identification of flunitrazepam in 32 larvae. Flunitrazepam is the N-methyl-2′-fluoro analogue of nitrazepam and is available in a number of western European countries for use as a hypnotic (Rohypnol and Noriel®) and anesthetic (Narcozep®) agent. The detection of flunitrazepam, the most frequently abused pharmaceutical drug in the world, in necrophagous flies (Chrysomya megacephala, Chrysomya albiceps and Cochliomyia macellaria) as a non-invasive and non-destructive technique does not appear to be well documented. In our study, sample preparation, spectroscopic measurement, data preprocessing, feature selection and analytical validation were addressed. To our knowledge, there is no reported use of FTIR microscopy for the detection and identification of a drug being present in a corpse.

Materials and methods

Thirty-two Wistar mice (Rattus norvegicus) with an average weight of 300 g were divided into two groups of sixteen and supplemented in the following way: group 1 – ethanol; and group 2 – standard flunitrazepam at a dose of 2 mg kg⁻¹. One hour after the supplementation, the mice were sacrificed, individually housed in fly traps and distributed at eight points along a track of a nearby forest. On the third and fourth days, 10 larvae were collected from each mouse and analyzed with ATR microspectroscopy. The average weight of one larva was estimated at 80 mg. All experiments were performed in compliance with the relevant laws and institutional guidelines, where the ethics committee of the Federal University of Rio Grande do Norte (UFRN/Brazil) have approved the experiments [Research and Ethics Committee (REC) approval no. 044/2013].

IR spectra [n = 264, 32 larvae (Chrysomya megacephala, Chrysomya albiceps and Cochliomyia macellaria) and eight random points] were collected from larvae (placing larvae individually on their backs on the plate) using a Bruker Lumus FTIR spectrometer with motorized ATR crystal (Bruker Optics Ltd, Coventry, UK). Prior to analyzing each specimen, the diamond crystal within the spectrometer was washed and a background spectrum was obtained to account for atmospheric composition.

The data import, pre-treatment and construction of chemometric classification models (PCA-LDA, SPA-LDA and GA-LDA) were implemented in MATLAB R2014a software (http://www.mathworks.com). Raw spectra were pre-processed by cutting between 1800 and 900 cm⁻¹ (235 wavenumbers at 3.84 cm⁻¹ spectral resolution) and baseline-corrected. For PCA-LDA, SPA-LDA and GA-LDA models, the samples were divided into training (70%), validation (15%) and prediction sets (15%) by applying the classic Kennard–Stone (KS) uniform sampling algorithm to the IR spectra. The KS algorithm was applied separately to each class to extract a representative set of objects from a given dataset by maximizing the minimal Euclidean distance between already selected objects and the remaining objects. The training samples were used in the modelling procedure (including variable selection for LDA), whereas the prediction set was only used in the final evaluation of the classification. The optimum number of variables for SPA-LDA and GA-LDA was determined with an average risk G of LDA misclassification. Such a cost function is calculated in the validation set as:

where g_n is defined aswhere I(n) is the index of the true class for the nth validation object x_n.

In this definition, the numerator is the squared Mahalanobis distance between object x_n (of class index I_n) and the sample mean m_I(n) of its true class. The denominator in eqn (2) corresponds to the squared Mahalanobis distance between object x_n and the center of the closest wrong class.

The GA routine was carried out during 40 generations with 80 chromosomes each. Crossover and mutation probabilities were set to 60% and 10%, respectively. Moreover, the algorithm was repeated three times, starting from different random initial populations. The best solution (in terms of the fitness value) resulting from the three realizations of the GA was employed. For this study, LDA scores, loadings, and discriminant function (DF) values were obtained for the specimen.

Sensitivity (the confidence that a positive result for a sample of the label class is obtained), specificity (the confidence that a negative result for a sample of non-label class is obtained), positive predictive value (PPV; measures the proportion of correctly assigned positive examples and its value varies between 0 and 1), negative predictive value (NPV; measures the proportion of correctly assigned negative examples and its value varies between 0 and 1), Youden's index (YOU; evaluates the classifier's ability to avoid failure), and the likelihood ratios (LR+) (represents the ratio between the probability of predicting an example as positive when it truly is positive, and the probability of predicting an example as positive when it actually is not positive) and (LR−) (represents the ratio between the probability of predicting an example as negative when it is actually positive, and the probability of predicting an example as negative when it truly is negative) were calculated as important quality standards in test evaluation. The quality metrics used in this study for evaluating the classification results can be calculated using the following equations:

YOU = SENS − (1 − SPEC) (7)

where FN is defined as false negative, FP as false positive, TP as true positive and TN as true negative. SENS is defined as sensitivity and SPEC as specificity.

Results and discussion

In total, n = 264 spectra were acquired. The average IR spectrum for each original class (control, black line; flunitrazepam, red line) in the range 900–1800 cm⁻¹ after baseline correction is shown in Fig. 1. As can be seen, discriminating between the two categories of specimens on the basis of IR measurements is not straightforward, owing to the complexity of the spectra. Although all spectra had a similar shape in the regions 1150–1190 cm⁻¹, 1470–1490 cm⁻¹ and 1505–1520 cm⁻¹, the spectra were shifted downwards.

Fig. 1 Average spectrum for each original class (control, black line; flunitrazepam, red line).

Distinguishing these categories only by spectral observation is difficult; so, to identify markers, it is necessary to apply computational analysis (PCA-LDA, and variable selection techniques such as SPA-LDA and GA-LDA). The optimum numbers of PCs for PCA and variables for SPA-LDA and GA-LDA were determined by power versus cost calculation using the minimum cost function G. These were adopted to systematically classify normal vs. flunitrazepam based on IR spectra. Further, comparisons were made between rates, interpretability and training times.

Fig. 2 is a 2-D PCA-LDA Fisher scores plot of the derived spectral points from each category. We can see that the Fisher scores do not show good segregation. The PCA-LDA model was built using five PCs (93% variance in the data).

Fig. 2 Discriminant function versus samples calculated by using the PCA-LDA model from two categories (control and flunitrazepam).

Then, SPA-LDA and GA-LDA were applied to the dataset to obtain the optimum number of variables by the minimum cost function G. Fisher scores for both models (SPA and GA) were obtained, and this improved the segregation between categories when compared with PCA-LDA.

SPA-LDA resulted in the selection of 46 variables (Table 1). Using solely 46 selected wavenumbers, the Fisher scores were calculated for both categories of the dataset, as shown in Fig. 3. As can be seen in Fig. 3, there is a greater effect of homogeneity among classes, with a little misclassification being obtained. Examination of the selected wavenumbers following SPA-LDA showed that the main biochemical alterations discriminating control vs. flunitrazepam were lipid, protein, nucleic acid, carbohydrate and, to a lesser extent, DNA vibrations. Several selected wavenumbers appear to be of particular interest, namely, the variables at 1315 cm⁻¹, 1389 cm⁻¹, 1505 cm⁻¹ and 1550 cm⁻¹, associated with amide III of proteins, COO⁻ symmetric stretch in fatty acids, amide II of proteins and C–O stretching of predominantly a-sheet of amide II, respectively.

Table 1 Variables for SPA-LDA and GA-LDA determined from the minimum cost function G used to achieve classification of control and flunitrazepam for a given validation dataset

Chemometric analysis	Wavenumber (cm^−1) selected
SPA-LDA	1300	1304	1308	1315	1323	1327	1331	1339
	1345	1350	1355	1361	1367	1373	1378	1382
	1389	1396	1402	1409	1415	1420	1425	1430
	1434	1441	1445	1449	1454	1458	1462	1467
	1473	1478	1486	1492	1499	1505	1510	1517
	1522	1527	1534	1540	1545	1550
GA-LDA	1304	1318	1319	1326	1327	1329	1331	1334
	1335	1342	1345	1352	1356	1369	1371	1382
	1386	1390	1391	1395	1403	1405	1421	1439
	1441	1445	1447	1448	1450	1451	1471	1472
	1477	1484	1486	1488	1500	1501	1505	1527

---

Fig. 3 Discriminant function versus samples calculated by the SPA-LDA model from two categories (control and flunitrazepam) using 46 wavenumbers selected.

GA-LDA was applied to the dataset (control vs. flunitrazepam) and resulted in the selection of 40 variables (Table 1). Fig. 4 is a scores plot that shows that GA-LDA generates the best segregation between the two categories. Several selected wavenumbers (GA-LDA) appear to be of particular interest, namely, the variables at 1334 cm⁻¹ and 1527 cm⁻¹, representing the amide III from proteins and CH bending and CH₂ wagging, respectively. These findings suggest that FTIR microscopy is a very promising technique for the non-destructive identification of flunitrazepam in Chrysomya megacephala, Chrysomya albiceps and Cochliomyia macellaria specimens. In addition, this finding is significant due to its potential for translation into entomotoxicology practice. Currently, no spectroscopic techniques are the gold standard in the detection procedures for the analysis of intact insects for toxicology decision-making.

Fig. 4 Discriminant function versus samples calculated by the GA-LDA model from two categories (control and flunitrazepam) using 40 wavenumbers selected.

Classification rates were determined by using the best models. Table 2 presents the validation results for the optimized model (PCA-LDA, SPA-LDA and GA-LDA) of each category. According to results of sensitivity shown in Table 2, it is possible to see that the sensitivity rate from PCA-LDA, SPA-LDA and GA-LDA achieved scores of 64.7%, 70.5% and 64.7% for the control category, respectively, showing that the control category can be relatively well classified by these methods. For the flunitrazepam category, the sensitivity values from PCA-LDA, SPA-LDA and GA-LDA models were 64.7%, 82.3% and 88.2%, respectively. Furthermore, the specificity as shown in Table 2 for both categories suggests that following SPA-LDA and GA-SPA, an improved accuracy in comparison with PCA-LDA was obtained. In general, distinguishing between the control and flunitrazepam categories was more successful when using GA-LDA, demonstrating that ATR-FTIR microspectroscopy in conjunction with powerful chemometric approaches has the potential to detect and identify drugs present in a corpse.

Table 2 Values of quality performance features from three classification methods (PCA-LDA, SPA-LDA and GA-LDA) by FTIR microspectroscopy for each category

Stage performance features	PCA-LDA	SPA-LDA	GA-LDA
Control
Sensitivity (%)	64.7	70.5	64.7
Specificity (%)	76.4	70.5	64.7
Positive predictive value (PPV)	73.3	70.5	64.7
Negative predictive value (NPV)	68.4	70.5	64.7
Youden index (YOU)	41.1	41.1	29.4
Positive likelihood ratio (LR+)	2.7	2.4	1.8
Negative likelihood ratio (LR−)	0.4	0.4	0.5
Flunitrazepam
Sensitivity (%)	64.7	82.3	88.2
Specificity (%)	64.7	82.3	88.2
Positive predictive value (PPV)	64.7	82.3	88.2
Negative predictive value (NPV)	64.7	82.3	88.2
Youden index (YOU)	29.4	64.7	76.4
Positive likelihood ratio (LR+)	1.8	4.6	7.5
Negative likelihood ratio (LR−)	0.5	0.2	0.1

---

Conclusions

The results of this study show that ATR-FTIR microspectroscopy coupled with variable selection techniques (SPA or GA) of necrophagous flies may be an alternative approach for the detection and identification of flunitrazepam. We report a fast, clean and non-destructive methodology involving minimal sample preparation to categorize the specimens. In a case study for flunitrazepam samples, the resulting GA-LDA model successfully detected the biochemical alteration based on 40 wavenumbers and produced 82.3% and 88.2% sensitivity and specificity accuracies. For this dataset investigated, these selected wavenumbers (SPA-LDA and GA-LDA) appear to be of particular interest for the detection and identification of flunitrazepam. This is required to robustly validate the classification and biomarker extraction models for necrophagous flies and identification of flunitrazepam. Although the determination of abused drugs in insects is usually provided by gold standards (solid phase extraction chromatography methods), the proposed methodology can be applied to new drugs or necrophagous insects where the processing time and reagent costs required are a major advantage. This method was thoroughly validated in accordance with international guidelines, being considered suitable for use as an official methodology for entomological contexts. Further validation of these approaches exploiting other biospectroscopy techniques and using larger and architecturally more robust datasets is required.

Acknowledgements

The authors would like to thank PPGQ-UFRN for scientific support. K. M. G. Lima acknowledges the CNPq/Capes project (grants 070/2012 and 442087/2014-4), FAPERN (PPP 005/2012) for financial support. We are grateful to Fabio Godoy (Bruker Optics Ltd) for excellent technical assistance for this study with the Bruker Lumus FTIR spectrometer.

References

1. A. Tracqui, C. Keyser-Tracqui, P. Kintz and B. Ludes, Int. J. Leg. Med., 2004, 118, 194–196.
2. M. Gosselin, S. M. R. Wille, M. D. M. R. Fernandez, V. Di Fazio, N. Samyn, G. de Boeck and B. Bourel, Forensic Sci. Int., 2011, 208, 1–9.
3. C. P. Campobasso, M. Gherardi, M. Caligara, L. Sironi and F. Introna, Int. J. Leg. Med., 2004, 118, 210–214.
4. B. Levine, M. Golle and J. E. Smialek, Am. J. Forensic Med. Pathol., 2000, 21, 59–61.
5. E. M. Roeterdink, I. R. Dadour and R. J. Watling, Int. J. Leg. Med., 2004, 118, 63–70.
6. L. Lagoo, L. S. Schaeffer, D. W. Szymanski and R. W. Smith, J. Forensic Sci., 2010, 55, 624–632.
7. R. A. Rashid, K. Osman, M. I. Ismail, R. M. Zuha and R. A. Hassan, Trop. Biomed., 2008, 25, 184–190.
8. X. Liu, Y. Shi, H. Wang and R. Zhang, Forensic Sci. Int., 2009, 192, 14–18.
9. B. Bourel, G. Tournel, V. Hedouin, M. Deveaux and M. L. Goff, Forensic Sci. Int., 2001, 120, 127–131.
10. V. Hedouin, B. Bourel, L. Martin-Bouyer, A. Becart, G. Tournel, M. Deveaux and D. Gosset, J. Forensic Sci., 1999, 44, 351–353.
11. C. P. Campobasso, M. Gherardi, M. Caligara, L. Sironi and F. Introna, Int. J. Leg. Med., 2004, 118, 210–214.
12. D. T. Manhoff, I. Hood, F. Caputo, J. Perry, S. Rosen and H. G. Mirchandani, J. Forensic Sci., 1991, 36, 1732–1735.
13. D. W. Sadler, J. Richardson, S. Haigh, G. Bruce and D. J. Pounder, Am. J. Forensic Med. Pathol., 1997, 18, 397–403.
14. D. W. Sadler, L. Robertson, G. Brown, C. Fuke and D. J. Pounder, J. Forensic Sci., 1997, 42, 481–485.
15. J. S. Oliveira, T. C. Baia, R. A. Gama and K. M. G. Lima, Microchem. J., 2014, 115, 39–46.
16. M. J. Baker, J. Trevisan, P. Bassan, R. Bhargava, H. J. Butler, K. M. Dorling, P. R. Fielden, S. W. Fogarty, N. J. Fullwood, K. A. Heys, C. Hughes, P. Lasch, P. L. Martin-Hirsch, B. Obinaju, G. D. Sockalingum, J. Sulé-Suso, R. J. Strong, M. J. Walsh, B. R. Wood, P. Gardner and F. L. Martin, Nat. Protoc., 2014, 9, 1771–1791.
17. A. L. Mitchell, K. B. Gajjar, G. Theophilou, F. L. Martin and P. L. Martin-Hirsch, J. Biophotonics, 2014, 7, 153–165.
18. A. Derenne, M. Verdonck and E. Goormaghtigh, Analyst, 2012, 137, 3255–3264.
19. M. Cavagna, R. Dell'Anna, F. Monti, F. Rossi and S. Torriani, J. Agric. Food Chem., 2010, 58, 39–45.
20. L. F. Chiu, P. Y. Huang, W. F. Chiang, T. Y. Wong, S. H. Lin, Y. C. Lee and D. Bin Shieh, Anal. Bioanal. Chem., 2013, 405, 1995–2007.
21. M. Ferrand, B. Huquet, S. Barbey, F. Barillet, F. Faucon, H. Larroque, O. Leray, J. M. Trommenschlager and M. Brochard, Chemom. Intell. Lab. Syst., 2011, 106, 183–189.
22. A. S. Marques, M. C. N. de Melo, T. A. Cidral and K. M. G. de Lima, J. Microbiol. Methods, 2013, 98, 26–30.
23. M. Thompson, S. L. R. Ellison and R. Wood, Pure Appl. Chem., 2002, 74, 835–855.

---