Quantitative analysis of thiram based on SERS and PLSR combined with wavenumber selection

Abstract

Surfaced-enhanced Raman spectroscopy (SERS) is a novel analytical technology mainly for quick, simple and ultrasensitive analysis. In this study, SERS and partial least squares regression (PLSR) were used for the quantitative analysis of thiram combined with a wavenumber selection method based on the adaptive genetic algorithm (AGA). The conventional genetic algorithm (CGA) was also used for wavenumber selection and the effects of the two were contrasted. Moreover, the impact on individual analytical results was evaluated of different numbers of subintervals included in the wavenumber selection. We achieved the best results (root mean square error of cross-validation = 0.2957 μM, R = 0.9995) by PLSR using an AGA with 5-subintervals. The experiments indicated that the proposed AGA-based wavenumber selection was superior to the CGA and the effect of multi-subintervals on wavenumber selection was better than that of a single subinterval.

---

1 Introduction

Thiram, a nonsystemic dithiocarbamate fungicide, is widely used as a protectant on foliage and fruits to control a variety of fungal diseases, such as rust, scab, anthracnosis, botrytis, fusarium, monilia and so on.¹ The residue of thiram is harmful to aquatic organisms, birds, and mammals.² Thus, the quantitative and qualitative analysis of thiram in environmental materials is important. Although the conventional analytical methods such as ELISA, HPLC, LC-MS/MS are accurate, they are time consuming, laborious, costly and are not suitable for fast detection or analysis.^3–5

Surface-enhanced Raman spectroscopy (SERS) is a novel analytical method providing ultrasensitive, simple and rapid detection of organic chemicals.⁶ SERS overcomes the low sensitivity of traditional Raman spectroscopy by tremendous enhancement of scattering signals via absorption of analytes onto the rough surface of noble metals (gold, silver and so on). Besides, SERS technology has the excellent characteristics of high speed, low cost and its ability to record spectra of analytes without pretreatment. Therefore, SERS is suitable for fast and accurate analysis of thiram.

While of long standing fundamental importance, SERS as a quantitative analytical tool has only recently begun to be used.^7–9 Chemometrics methods such as principal component regression (PCR), partial least squares regression (PLSR) and support vector machine regression (SVR) are generally employed to process data in quantitative analysis. Meanwhile, some studies have proved that proper wavelength or wavenumber selection can improve the results before using the chemometrics methods.^10–13 In general, the methods used for wavenumber selection can be divided into manual and automated procedures. The manual methods are based on the selection of the spectra near characteristic peaks of the analytes; the automatic methods mainly depend on optimization algorithms. In comparison, the latter methods are carried out without human intervention, and are simple and easy to operate. It has been proved that the optimization algorithms such as the genetic algorithm (GA),^14,15 simulated annealing (SA),¹⁶ and ant colony optimization (ACO)¹⁷ can effectively select appropriate wavenumber regions. Since the GA is simple and easy to realize, it was used for wavenumber selection in this research. In the method of GA-based wavenumber selection, a combination of spectral variables or subintervals is assigned to a chromosome. Through continuous crossover, mutation and selection, chromosomes of high fitness value are obtained with high predictive accuracy of the corresponding regression models. As the rate of mutation and crossover is fixed in CGA, the algorithm usually has the shortcomings of easily falling into local optimization and destruction of the excellent individuals.¹⁸ One possible way of avoiding this problem is the use of the adaptive genetic algorithm AGA so that the rate of mutation and crossover can be adjusted by the fitness value of the individuals. The corresponding crossover and mutation rates of the individuals with high fitness value are low in AGA, hence they can be saved and optimized further with the AGA. But for individuals with a low fitness value, it is highly probable that they are eliminated because of their high mutation rate. The selection unit of wavenumber selection can be divided into spectral variables and spectral subintervals.¹⁹ The resultant analysis results tend to be better in the latter for taking into account the high correlation of the spectral variables.

In this study, SERS and PLSR were used for the quantitative analysis of thiram combined with AGA-based wavenumber selection methods. Different GA-based wavenumber selection methods (AGA, CGA) were compared by 10-fold cross-validation on the basis of the correlation coefficient (Rp) and the root mean square error of cross-validation (RMSECV) of the corresponding regression models. Furthermore, the effects of the single subinterval and multi-subintervals on wavenumber selection were also assessed. The details are shown in below.

2 Materials and methods

2.1 Samples preparation

Ag nitrate, and trisodium citrate were purchased from Sinopharm Chemical Reagent Co., Ltd. (Shanghai, China). All the chemicals used were of analytical grade or better and were used without further purification. Preparation of Ag nanoparticles: silver nanoparticles were synthesized by adding 10 mL 10⁻² mol L⁻¹ Ag nitrate to 90 mL distilled water in a 250 mL three-necked flask.²⁰ The solution was heated to boiling and then 3 mL 1% trisodium citrate solution was added followed by heating for 40 min. After cooling to room temperature, the solution was finally centrifuged to obtain a sol solution of the nanoparticles. The concentration of the thick Ag colloidal suspension was approximately 1 M. Commercial thiram was obtained from YuanDa (Beijing, China). It was dissolved in ethanol (99%, V/V) to obtain solutions of concentrations of 0.1, 0.5, 1, 5, 10, and 50 μM.

2.2 SERS measurement

Silver nanoparticles were selected as SERS-active substrates because of their excellent performance in enhancing the signal intensity.²¹

Ultrasonic dispersed Ag nanoparticles, were placed into a centrifuge tube. Then thiram solutions of different concentrations (the same volume) were added to the Ag sol, followed by mixing using ultrasonic dispersion for 10 min.²² Then, the mixed solution was dropped onto a silicon wafer which acted as the enhancing substrate. After evaporation of the solvent at room temperature the SERS spectra were recorded. The spectra were acquired using a LabRAM HR800 Raman spectrometer equipped with a 532 nm diode laser source. All the spectra were recorded with a 50× microscope objective and 200 mW laser power. For each original spectrum, 10 scans were performed at sampling points of 710 over the range of 400–1600 cm⁻¹ with an integration time of 10 s. For each sample, 10 original spectra were recorded taking into account the location of the sampling points. And 100 spectra were collected with sampling once a week. A single original spectrum was not suitable for quantitative analysis because of the poor reproducibility of the SERS signal. To solve this problem, the mean of 30 spectra, randomly selected from 100 original spectra, was used for the analysis spectroscopy and 50 analysis spectra were acquired by cyclic operation 50 times. The baseline of each spectrum, which was caused by the fluorescence background, was approximated by a seventh-order polynomial fit and subtracted using NGSLabSpec software (Jobin Yvon).

2.3 Data analysis methods

2.3.1 PLS regression (PLSR). PLSR is an algorithm for building linear regression models that has been applied in several disciplines such as agriculture, chemistry, economics and medicine.^23,24 The core of PLSR is a dimension reduction technique that operates under the assumption of a basic latent decomposition of the response matrix (Y) and predictor matrix (X):

X = TP^T + E, Y = TQ^T + F (1)

where T is a matrix that produces k linear combinations (scores); P and Q are matrices of coefficients (loadings); E and F are matrices of residual errors. To specify the latent component matrix T such that T = XW, PLSR requires columns of W = (w₁, w₂,…, w_k) to be determined from successive optimization problems. The criterion for finding the k^th direction vector w_k for Y is formulated aswhere is the covariance of X. So far, the relationship between X and Y is derived.

2.3.2 GA-based wavenumber selection method. GA mimics the theory of natural biological evolution in which individuals with a high fitness can prevail and survive to the next generation, and the best individuals can be adapted by crossover and mutation.^19,25 The GA-based wavenumber selection method is based on the framework of GA. Each individual defined by a chromosome represents a set of spectral subintervals. The objective of the method is to search for a combination of subintervals (wavenumber regions) which are developed for the better model.

First, chromosomes are randomly generated to form the initial population. Then, to find the best combination of the wavenumber subintervals, the calculations for the fitness values, selection of chromosomes, crossover, and mutation are repeated until the number of iterations exceeds the predefined limit. In this study, one subinterval in a spectrum is represented by two integer values. The first value denotes the starting wavenumber of a subinterval, and the second value denotes the finishing wavenumber. If the range of the represented region exceeds the full spectral range or two regions overlap, the excess or redundant parts are ignored.

Two objectives are considered when designing the fitness function of chromosomes. One is to maximize R_p, and the other is to minimize RMSECV. So the fitness function was designed as: R_p/(1 + RMSECV). And the fitness values of chromosomes were calculated by PLSR using the SERS spectra of the selected wavenumber regions. Other operators were set as follows: the number of iterations was 500; the size of the population was 20. The number of subintervals in each chromosome was 1 (single subinterval) or 5 (multi-subintervals). The crossover rate was 0.6 and the mutation rate was 0. 1 in CGA; but for AGA, the crossover rate (P_c) and the mutation rate (P_m) were set as following:

where f_max, f̄, f_min are the maximum fitness value, the average fitness value and the minimum fitness value of the individuals in each generation. And p_c1 > p_c2 > p_c3, p_m1 > p_m2 > p_m3 were assigned by the value of 0.0–1.0. The corresponding rate of crossover and mutation of the individuals with low fitness is high in eqn (3), the individuals can be washed out during the iteration. For the individuals with high fitness, the rate of crossover and mutation is small, and the individuals are inherited and further optimized. But in CGA, the fixed rate of crossover and mutation may cause the individuals with low fitness not to be washed out quickly; the individuals with high fitness are not more likely to be inherited, and the better individual is not conducive to being obtained.

2.3.3 Experiments. A flowchart detailing the development of PLSR models combined with wavenumber selection is shown in Fig. 1. Firstly, 45 spectra selected from 50 analysis spectra were used to select the wavenumber with CGA-PLSR/AGA-PLSR every time. Then, models whose performance was evaluated on the basis of Rp and RMSECV were developed with the residual spectra of the selected wavenumber regions. The process was repeated until the entire spectral set (10-fold cross validation methods) had been traversed.

Fig. 1 Flowchart of the development of PLSR models with GA-based wavenumber selection.

All the computation and chemometric methods were implemented in MATLAB 2011b (The Mathworks Inc., Natick, MA, USA). The PLSR or GA algorithms were carried out by ourselves.

3 Results and discussion

3.1 SER spectra of thiram

The SERS spectra recorded for thiram solutions of different concentrations (0.1, 0.5, 1, 5, 10, and 50 μM) are shown in Fig. 2. The peak at 1386 cm⁻¹ is the strongest and is due to the CN stretching mode and symmetric CH₃ deformation mode. The other peaks, such as the peak at 440 cm⁻¹, which is due to the CH₃NC deformation and C=S stretching modes, the peak at 561 cm⁻¹, which is associated with the S=S stretching mode, the peak at 925 cm⁻¹, which is due to the stretching CH₃N and C=S modes, are all clearly identifiable. From Fig. 2, we can also find that the CN stretching mode and rocking CH₃ mode occurs at 1150 cm⁻¹; the CN stretching mode occurs at 1444 cm⁻¹ as well as the rocking CH₃ mode.²⁶ At the same time, the CN stretching mode and the deformation, rocking mode of CH₃ can be observed at 1514 cm⁻¹. All the intensities of these peaks are stronger while the concentrations are higher. The phenomenon partially indicates that quantitative analysis of SERS is possible.

Fig. 2 SERS spectra for different concentrations of thiram.

Spectral wavenumber regions for the establishment of the regression model have a significant influence on the performance of the mode, and the conventional methods directly select the spectra near characteristic peaks. Although all characteristic information of the spectra is used to develop the model in these methods, introduction of some information has a negative effect on the results and the non-characteristic information rather than a positive effect. Thus, the optimization algorithms (CGA, AGA) were used to identify which few wavenumber regions to use in this study.

3.2 Quantitative analysis of thiram

3.2.1 Results of wavenumber selection. CAG and AGA with different subintervals were used to select the spectral wavenumber for building of the regression model. The results obtained through the 10-fold cross-validation methods are showed in Fig. 3. The vertical axis of the figure represents the selection frequency of each wavenumber. As shown in the figure, selected spectral ranges are mostly near the SERS characteristic peaks of thiram. While the selected wavenumber range in CAG1/AGA1 is wide, some irrelevant information can be easily introduced to interfere with the subsequent analysis. Due to the excellent flexibility, the algorithm with 5-subintervals can select the spectra which are more suitable for the quantitative analysis. Furthermore, the ranges of wavenumber selected by the AGA are more focused on the characteristic peaks of thiram than that of the CGA. To evaluate the effect of the different algorithms, the spectra with a selection ratio value of above 0.70 were used to build the quantitative analysis model of thiram concentrations.

Fig. 3 Selection ratio of spectral wavenumber selected by CGA/AGA; (a) CGA with single subinterval (CGA1), (b) AGA with single subinterval (AGA1), (c) CGA with multi-subintervals (CGA5), (d) AGA with multi-subintervals (AGA5).

3.2.2 Quantitative analysis models. The performance of the modes which were developed by PLSR with different spectral ranges are shown in Table 1. The RMSECV and R of the model based on full region spectra were 0.4666 and 0.9934, respectively. The results were worse than those for the models with GA-based wavenumber selection, which illustrates that wavenumber selection is necessary before regression analysis. And the experimental results show that the CAG5/AGA5 are superior to the CAG1/AGA1 thus verifying the previous discussion. In addition, the corresponding model of AGA5 (RMSECV = 0.2957 μM, R = 0.9995) is better than the CGA5 (RMSECV = 0.3642 μM, R = 0.9986). This is because the excellent selection results could easily be saved and optimized further and the bad results were eliminated with big probability through adjustment of the rate of crossover and mutation with the fitness value of the individuals in AGA5.

Table 1 Summary of regression results for SERS spectra

Data set	Method	RMSECV	R
a Without wavenumber selection algorithms.
SER spectra of 400–1600 cm^−1		0.4666	0.9934
	CGA1	0.4597	0.9953
	AGA1	0.3908	0.9977
	CGA5	0.3642	0.9986
	AGA5	0.2957	0.9995

---

Fig. 4 shows the PCA plot of the PLSR scores of the spectra selected by AGA5/CGA5. It is found that the PLSR scores of the samples of the same concentration are more centred in AGA5, which means that the deviation between the samples of the same class is reduced. Thus the performance of the corresponding model is improved. The predictive error of the model which was built with the selected spectral wavenumbers in AGA5 is shown as Fig. 5. The analysis results are generally accurate, but the relative deviation for the low-concentration solution (0.1 μM) is too big to appear to be a mistake.

Fig. 4 PCA scores plot of PLSR latent variables of the thiram spectra selected by CGA5/AGA5; (a) CGA5, (b) AGA5.

Fig. 5 The predictive error of the model developed by PLSR with AGA5-based wavenumber selection.

3.3 Reliability of SERS for quantitative analysis

SERS is widely used as a rapid and accurate technology for both qualitative and quantitative analyses of trace amounts of contaminants. When SERS are measured, they can be affected by laser frequency, the surface of the substrate and the molecular structures of the analytes.⁶ So it is very challenging to apply SERS for quantitative analysis because of the difficulty in obtaining reproducible spectra. To avoid the problem, the experimental conditions and the preparation of SERS-active substrates were strictly controlled, and each analysis spectrum was the mean of 30 original spectra. Through such processing, the SERS deviation of samples of the same concentration became small and were available for quantitative analysis. In future, we believe SERS for quantitative analyses will certainly be used in more and more research fields with the aid of the rapid development of microscopy Raman spectrometers and substrates.

4 Conclusions

Firstly, the experiments prove that the SERS spectra of thiram which are acquired under controlled experimental conditions and statistical methods are available for quantitative analysis. Then it is known that the wavenumber selection prior to the analysis is necessary by contrasting the performance of the corresponding models. The CGA/AGA–based wavenumber selection methods were used to select spectral wavenumbers in this study. As the local optimum can be effectively avoided and the excellent individuals can easily be saved and optimized further in AGA, the effect of wavenumber selection is better. Additionally, the influence of methods with different subintervals on the wavenumber selection was also evaluated. And the methods with 5-subintervals were found to be more suitable in this aspect because of their flexibility. The best analysis results were obtained using the model based on PLSR combined with AGA5 (RMSECV = 0.2957 uM, R = 0.9995). Finally, we hope that the results obtained by us will help further chemometric investigations (wavenumber selection) and investigations in the sphere of vibrational spectroscopy (Near infrared, Mid-infrared and Raman) of multi-component systems.

Acknowledgements

This study was supported by the National High Technology Research and Development Program of China (no. SS2013AA100302).

References

1. S. Walia, R. K. Sharma and B. S. Parmar, Bull. Environ. Contam. Toxicol., 2009, 83, 363–368.
2. C. Cereser, S. Boget, P. Parvaz and A. Revol, Toxicology, 2001, 162, 89–101.
3. L. Queffelec, F. Boisde, J. P. Larue, J. P. Haelters, B. Corbel, D. Thouvenot and P. Nodet, J. Agric. Food Chem., 2001, 49, 1675–1680.
4. C. Fernández, A. J. Reviejo, L. M. Polo and J. M. Pingarrbn, Talanta, 1996, 43, 1341–1348.
5. S. B. Ekroth, B. Ohlin and B. G. Osterdahl, J. Agric. Food Chem., 1998, 46, 5302–5304.
6. F. Zhai, Y. Huang, C. Li, X. Wang and K. Lai, J. Agric. Food Chem., 2011, 59, 10023–10027.
7. W. E. Smith, K. Faulds and D. Graham, Top. Appl. Phys., 103, 381–396.
8. C. L. Haynes, C. R. Yonzon, X. Zhang and R. P. V. Duyne, J. Raman Spectrosc., 2005, 36, 471–484.
9. S. C. Pinzaru, I. Pavel, N. Leopold and W. Kiefer, J. Raman Spectrosc., 2004, 35, 338–346.
10. C. H. Spiegelman, M. J. Mcshane, M. J. Goetz, M. Motamedi, Q. L. Yue and G. L. Cote, Anal. Chem., 1998, 70, 35–44.
11. B. Nadler and R. R. Coifman, J. Chemom., 2005, 19, 107–118.
12. W. Fan, H. Li, Y. Shan, H. Lv, H. Zhang and Y. Liang, Anal. Methods, 2011, 3, 1872–1876.
13. G. Nasser and D. Shahsavani, Anal. Methods, 2012, 4, 3733–3738.
14. R. Leardi and L. Nørgaard, J. Chemom., 2004, 18, 486–497.
15. X. Zou, J. Zhao, X. Huang and Y. Li, Chemom. Intell. Lab. Syst., 2007, 87, 43–51.
16. U. Hörchner and J. H. Kalivas, Anal. Chim. Acta, 1995, 311, 1–13.
17. M. Shamsipur, V. Zare-Shahabadi, B. Hemmateenejad and M. Akhond, J. Chemom., 2006, 20, 146–157.
18. R. Balamurugan, C. V. Ramakrishnan and N. Singh, Appl. Soft Comput., 2008, 8, 1607–1624.
19. M. Arakawa, Y. Yamashita and K. Funatsu, J. Chemom., 2011, 25, 10–19.
20. P. Li, X. Zhou, H. Liu, L. Yang and J. Liu, J. Raman Spectrosc., 2013, 44, 999–1003.
21. A. Michota and J. Bukowska, J. Raman Spectrosc., 2003, 34, 21–25.
22. K. Qian, L. Yang, Z. Li and J. Liu, J. Raman Spectrosc., 2013, 44, 21–28.
23. K. C. Weber, K. M. Honório, A. T. Bruni, A. D. Andricopulo and A. B. da Silva, Struct. Chem., 2006, 17, 307–313.
24. T. Mehmood, K. H. Liland, L. Snipen and S. Sæbø, Chemom. Intell. Lab. Syst., 2012, 118, 62–69.
25. W. Patrick, R. Pell and E. Comas, Chemom. Intell. Lab. Syst., 2009, 98, 108–114.
26. B. Saute and R. Narayanan, Analyst, 2011, 136, 527–532.

---