Fluorescent Polymer Sensor Array for Detection and Discrimination of

Nov 11, 2010 - Corbin Company, Alexandria, Virginia 22314, United States. A fluorescent polymer sensor array (FPSA) was made from commercially availab...
4 downloads 0 Views 2MB Size
Download PDF
Anal. Chem. 2010, 82, 9917–9924

Fluorescent Polymer Sensor Array for Detection and Discrimination of Explosives in Water Marc D. Woodka* and Vincent P. Schnee RDECOM CERDEC Night Vision and Electronic Sensors Directorate, United States Army, Fort Belvoir, Virginia 22060, United States Michael P. Polcha Corbin Company, Alexandria, Virginia 22314, United States A fluorescent polymer sensor array (FPSA) was made from commercially available fluorescent polymers coated onto glass beads and was tested to assess the ability of the array to discriminate between different analytes in aqueous solution. The array was challenged with exposures to 17 different analytes, including the explosives trinitrotoluene (TNT), tetryl, and RDX, various explosiverelated compounds (ERCs), and nonexplosive electronwithdrawing compounds (EWCs). The array exhibited a natural selectivity toward EWCs, while the non-electronwithdrawing explosive 1,3,5-trinitroperhydro-1,3,5-triazine (RDX) produced no response. Response signatures were visualized by principal component analysis (PCA), and classified by linear discriminant analysis (LDA). RDX produced the same response signature as the sampled blanks and was classified accordingly. The array exhibited excellent discrimination toward all other compounds, with the exception of the isomers of nitrotoluene and aminodinitrotoluene. Of particular note was the ability of the array to discriminate between the three isomers of dinitrobenzene. The natural selectivity of the FPSA toward EWCs, plus the ability of the FPSA to discriminate between different EWCs, could be used to design a sensor with a low false alarm rate and an excellent ability to discriminate between explosives and explosive-related compounds. Broadly responsive array-based sensors have received significant attention in the recent literature for their ability to detect a variety of analytes. For vapor detection, so-called “electronic noses” have been fabricated with various sensing modalities including tin oxide sensors,1,2 conducting polymer3,4 or polymer composite5,6 * Author to whom correspondence should be addressed. E-mail: [email protected]. (1) Gardner, J. W.; Shurmer, H. V.; Corcoran, P. Sens. Actuators, B 1991, 4, 117–121. (2) Getino, J.; Horillo, M. C.; Gutierrez, J.; Ares, L.; Robla, J. I.; Garcia, C.; Sayago, I. Sens. Actuators, B 1997, 43, 200–205. (3) Freund, M. S.; Lewis, N. S. Proc. Natl. Acad. Sci. U.S.A. 1995, 92, 2652– 2656. (4) Janata, J.; Josowicz, M. Nat. Mater. 2003, 2, 19–24. (5) Lonergan, M. C.; Severin, E. J.; Doleman, B. J.; Beaber, S. A.; Grubbs, R. H.; Lewis, N. S. Chem. Mater. 1996, 8, 2298–2312. 10.1021/ac102504t  2010 American Chemical Society Published on Web 11/11/2010

chemically sensitive resistors, coated acoustic wave devices,7,8 dyeimpregnated polymers,9,10 functionalized metallic nanoparticles,11,12 and colorimetric methods.13,14 Similarly, for detection in water, “electronic tongues” have made use of sensing modalities such as functionalized gold nanoparticles,15 colorimetric optical devices,16-18 conductive polymers,19 and voltammetry.20 Reviews are available that cover the diversity of sensors used in both the nose-21,22 and tongue-23,24 based implementations. Analytes of particular interest in many defense- and securitybased applications include the high explosive trinitrotoluene (TNT) and explosive-related compounds (ERCs) such as dinitrotoluene, nitrotoluene, etc. Attempts at the detection of TNT and ERCs are routinely performed in the field to identify explosive hazards (e.g., bombs, land mines). Additionally, the presence of explosives in soils or wastewater at formerly used defense sites25

(6) Woodka, M. D.; Brunschwig, B. S.; Lewis, N. S. Langmuir 2007, 23, 13232– 13241. (7) Grate, J. W. Chem. Rev. 2000, 100, 2627–2648. (8) Grate, J. W.; Klusty, M. Anal. Chem. 1991, 63, 1719–1727. (9) White, J.; Kauer, J. S.; Dickinson, T. A.; Walt, D. R. Anal. Chem. 1996, 68, 2191–2202. (10) Li, D.; Mills, C. A.; Cooper, J. M. Sens. Actuators, B 2003, 92, 73–80. (11) Wohltjen, H.; Snow, A. W. Anal. Chem. 1998, 70, 2856–2859. (12) Zamborini, F. P.; Leopold, M. C.; Hicks, J. F.; Kulesza, P. J.; Malik, M. A.; Murray, R. W. J. Am. Chem. Soc. 2002, 124, 8958–8964. (13) Janzen, M. C.; Ponder, J. B.; Bailey, D. P.; Ingison, C. K.; Suslick, K. S. Anal. Chem. 2006, 78, 3591–3600. (14) Lim, S. H.; Musto, C. J.; Park, E.; Zhong, W.; Suslick, K. S. Org. Lett. 2008, 10, 4405–4408. (15) Cooper, J. S.; Raguse, B.; Chow, E.; Hubble, L.; Muller, K. H.; Wieczorek, L. Anal. Chem. 2010, 82, 3788–3795. (16) Folmer-Andersen, J. F.; Kitamura, M.; Anslyn, E. V. J. Am. Chem. Soc. 2006, 128, 2652–5653. (17) Palacios, M. A.; Nishiyabu, R.; Marquez, M.; Anzenbacher, P. J. Am. Chem. Soc. 2007, 129, 7538–7544. (18) Zhang, C.; Suslick, K. S. J. Am. Chem. Soc. 2005, 127, 11548–11549. (19) Riul, A.; Gallardo Soto, A. M.; Mello, S. V.; Bone, S.; Taylor, D. M.; Mattoso, L. H. C. Synth. Met. 2003, 2, 109–116. (20) Winquist, F.; Wide, P.; Lundstrom, I. Anal. Chim. Acta 1997, 357, 21–31. (21) Albert, K. J.; Lewis, N. S.; Schauer, C. L.; Sotzing, G. A.; Stitzel, S. E.; Vaid, T. P.; Walt, D. R. Chem. Rev. 2000, 100, 2595–2626. (22) Rock, F.; Barsan, N.; Weimar, U. Chem. Rev. 2008, 108, 705–725. (23) Ciosek, P.; Wroblewski, W. Analyst 2007, 132, 963–978. (24) Vlasov, Y. G.; Legin, A. V.; Ridnitskaya, A. M. Russ. J. Gen. Chem. 2008, 78, 2532–2544. (25) Lubbert, R. F.; Chu, T. J. Fed. Facil. Environ. J. 2007, 11, 5–18.

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

9917

or Federal waste sites26 poses environmental threats. The ability to detect explosives is critical to the ability to address these concerns. The chemical detection of explosives has been performed with many of the nose- and tongue-based sensing modalities listed above.27-36 The use of fluorescent polymers is particularly attractive due to their inherent sensitivity toward explosives.32,33,37 This natural sensitivity results from the strong electron-withdrawing nature of TNT and related ERCs,32,37 coupled with the molecular wire-based sensing scheme employed by fluorescent polymers.38 The transduction mechanism is an electron-transfer reaction between an excited-state electron in the polymer and an electron-withdrawing explosive compound such as TNT. The reaction is observed as an attenuation in the fluorescence of the polymer.32,33 Successful efforts utilizing this sensing scheme have focused on the use of a novel pentiptycene-derived poly(phenyleneethylene) fluorescent polymer, which exhibits increased fluorescence lifetimes and enhanced sensitivities/response times toward TNT.32,33 However, the natural sensitivity of the polymer toward alternative electron-withdrawing compounds32 (EWCs) could lead to uncertainty in the identification of compounds. Thus, a given quench in the fluorescence signal could be due to the presence of the strong quencher TNT at a low concentration, or the quench may be due to the presence of a weaker quenching nonexplosive/nonERC compound at a higher concentration. This could lead to false alarms in the field application of such a system. In this work, we describe the use of a sensor array fabricated from commercially available fluorescent polymers. The array was evaluated with respect to its ability to discriminate between various EWCs continuously sampled in aqueous solution over a 13 h period. Sensor diversity was obtained by utilizing polymers with differing functional repeat units along the polymer backbone. Responses from EWCs consisting of explosives and explosive hydrolysis products, and nonexplosive compounds, were evaluated by principal component analysis (PCA) and linear discriminant analysis (LDA). Included in the sampled analytes were three isomer groups. It is demonstrated that this nonspecific fluorescent polymer sensor array has the ability to operate over extended continuous periods of use and possesses an inherent ability to discriminate between EWCs, including chemically similar nitroaromatic compounds and the isomers of dinitrobenzene. (26) Riley, R. G.; Zachara, J. M.; Wobber, F. J. Chemical Contaminants on DOE Lands and Selection of Contaminant Mixtures for Subsurface Science Research. U.S. Department of Energy, 1992. (27) Albert, K. J.; Myrick, M. L.; Brown, S. B.; James, D. L.; Milanovich, F. P.; Walt, D. R. Environ. Sci. Technol. 2001, 35, 3193–3200. (28) Albert, K. J.; Walt, D. R. Anal. Chem. 2000, 72, 1947–1955. (29) Fu, X.; Benson, R. F.; Wang, J.; Fries, D. Sens. Actuators, B 2005, 106, 296–301. (30) Muralidharan, G.; Wig, A.; Pinnaduwage, L. A.; Hedden, D.; Thundat, T.; Lareau, R. T. Ultramicroscopy 2003, 97, 433–439. (31) Pinnaduwage, L. A.; Boiadjiev, V.; Hawk, J. E.; Thundat, T. Appl. Phys. Lett. 2003, 83, 1471–1473. (32) Yang, J. S.; Swager, T. M. J. Am. Chem. Soc. 1998, 120, 11864–11876. (33) Yang, J. S.; Swager, T. M. J. Am. Chem. Soc. 1998, 120, 5321–5322. (34) Yinon, J. Anal. Chem. 2003, 75, 98A–105A. (35) Harper, R. J.; Dock, M. L. Proc. SPIE 2007, 6540, 65400V. (36) Nguyen, H. H.; Li, X.; Wang, N.; Wang, Z. Y.; Ma, J.; Bock, W. J.; Ma, D. Macromolecules 2009, 42, 921–926. (37) Thomas, S. W.; Joly, G. D.; Swager, T. M. Chem. Rev. 2007, 107, 1339– 1386. (38) Swager, T. M. Acc. Chem. Res. 1998, 31, 201–207.

9918

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

Table 1. Fluorescent Polymers and Analyte Sample Vials Used in This Worka polymer

name

P1

poly[2-methoxy-5-(3′,7′-dimethyloctyloxy)-1,4phenylenevinylene] poly[(p-phenylenevinylene)-alt-(2-methoxy-5(2-ethylhexyloxy)-p-phenylenevinylene)] poly[(9,9-di-n-octylfluorenyl-2,7-diyl)-alt(benzo[2,1,3]thiadiazol-4,8-diyl)] poly[(9,9-dihexylfluoren-2,7-diyl)-co-(anthracen-9,10-diyl)] poly[2,5-bis(octyloxy)-1,4-phenylenevinylene]

P2 P3 P4 P5

analyte

name

TNT tetryl 2-am-DNT 4-am-DNT 1,2-DNB 1,3-DNB 1,4-DNB DQ DQ, 20 µM DQ, 200 µM 3-NT 4-NT TCl-1,4-benzQ

2,4,6-trinitrotoluene 2,4,6-trinitrophenylmethylnitramine 2-amino-4,6-dinitrotoluene 4-amino-2,6-dinitrotoluene 1,2-dinitrobenzene 1,3-dinitrobenzene 1,4-dinitrobenzene duroquinone (2,3,5,6-tetramethyl-1,4-benzoquinone) duroquinone, 20 µM duroquinone, 200 µM 3-nitrotoluene 4-nitrotoluene tetrachloro-1,4-benzoquinone (2,3,5,6-tetrachloro-1,4-benzoquinone) 1-chloro-2,4-dinitrobenzene 1,3,5-trinitroperhydro-1,3,5-triazine clean water, vial 1 clean water, vial 2

1-Cl-2,4-DNB RDX blank 1 blank 2 a

All analytes were delivered at 50 µM unless otherwise noted.

EXPERIMENTAL SECTION Materials. The names and molecular structures of the fluorescent polymers, purchased from Sigma-Aldrich, are given in Table 1 and Figure 1, respectively. 1,2-Dinitrobenzene, 1,3dinitrobenzene, 1,4-dinitrobenzene, duroquinone, 3-nitrotoluene, 4-nitrotoluene, tetrachloro-1,4-benzoquinone, and 1-chloro-2,4dinitrobenzene were purchased from Sigma-Aldrich in their pure chemical states. Acid-washed/silanized glass beads (250 µm diameter) and glass wool were purchased from Sigma-Aldrich. Trinitrotoluene, tetryl, 2-amino-4,6-dinitrotoluene, 4-amino-2,6dinitrotoluene, and 1,3,5-trinitroperhydro-1,3,5-triazine (RDX) were purchased from Cerilliant as 1000 µg mL-1 solutions in acetonitrile. All chemicals were used as received. Tap water was used with all of its impurities, dissolved oxygen, etc., without further treatment. See Table 1 for analyte abbreviations. Sensor Array. The sensor array was produced by placing polymer-coated glass beads into clear tubing, which allowed the coated beads to be illuminated and the resulting fluorescence to be monitored. The coating solutions for the glass beads were 1 mg/mL of polymers P1-P5, individually, in chloroform. Polymercoated beads were prepared by pushing a plug of glass wool into the bottom of a 2 mL Pasteur pipet. Glass beads (200 mg) were then placed into the plugged pipettes, and 1 mL of a 1 mg/mL polymer solution was allowed to freely pass over the glass beads, driven by gravity. A low flow rate of dry compressed air was flowed over the coated beads for 3 min to evaporate the remaining chloroform, and the dried coated beads were collected. This process was repeated for each of the polymer solutions P1-P5. This process was similar to the method used to coat dye onto beads for vapor detection.28,39

Figure 1. Structures of fluorescent polymers used in this work. See Table 1 for polymer names.

Figure 2. Sensor array configuration.

The sensor array was fabricated by packing the polymer-coated beads into 13 cm of transparent Chemfluor 367 scientific-grade tubing (0.125 in. o.d., 0.062 in. i.d.). A plug of glass wool was pushed into one end, followed by the pouring of 45 mg of P1coated beads into the tube. A small plug of glass wool was then pushed into the tube to serve as a sensor boundary indicator. This process was repeated for P2-, P3-, P4-, and P5-coated glass beads, followed by a larger plug of glass wool to retain the beads in the tubing. The tubing was then connected in-line to the sampling system. To obviate bubble formation in the sensor, a 75 psi back pressure regulator (IDEX Health Sciences) was connected to the outlet end of the tube. Fluorescence Measurements. Figure 2 depicts the configuration used for fluorescence monitoring. Custom-made dual probe Y-shaped couplers (Thorlabs, Inc., 600 µm) were used to measure the fluorescence of each sensor band. One forked end received light from a 410 nm light-emitting diode (LED) (Optek (39) Dickinson, T. A.; Michael, K. L.; Kauer, J. S.; Walt, D. R. Anal. Chem. 1999, 71, 2192–2198.

Technology), which was directed onto one polymer band. The other forked end fed directly into a spectrometer configured for fluorescence (Jaz spectrometer, Ocean Optics) for continuous monitoring of the signal. No filters or lenses were used. This was repeated for each of the five polymer bands. In all cases, fluorescence peaks were well-separated from their 410 nm excitation peaks, ranging from 500 to 650 nm. To ensure that decreases in fluorescence intensity were due to the expected electrontransfer reaction rather than the absorption of excitation or fluorescent light by injection of contaminated water, UV-vis absorption spectra of analytes were collected. Consistent with previous reports on TNT in water,40 analytes possessed absorption peaks in the UV range, with no absorption occurring in the visible (400-700 nm) range. Fluorescence measurements were collected at approximately 1 Hz. Sample Generation and Water Sampling. To generate contaminated samples, all pure analytes from Table 1 were made into 1000 µg/mL stock solutions in acetone. From the stock solutions, 50 µM aqueous samples were made by adding the appropriate amount of the stock solutions into a clean amber vial. The acetone solvent was then allowed to evaporate, leaving behind trace analyte contamination. Water was then added (20 mL), and the vials were sonicated at 40 °C for 20 min to form solutions. Duroquinone was additionally made at 20 and 200 µM, and two clean water blanks were included. This produced 17 analyte samples (nan ) 17) for random sampling. Samples were configured for automatic testing by use of computer-controlled Teflon-lined manifold sampling valves. Water was continuously passed through the sensor array over a 13 h period at 4 mL/min by use of a high-pressure gradient piston pump. During this period, 15 s pulses of the samples were drawn randomly. The number of exposures nexp for each of the analytes was 11, with sampling iterations consisting of 60 s of background water, 15 s of analyte, and 150 s of background water (total sampling time ) 225 s). The outlet water flow from the sensor was collected as hazardous aqueous waste and disposed of accordingly. Signal Processing and Analysis. Signal Extraction. The wavelength of maximum fluorescence intensity λmax was determined for each polymer. The average maximum fluorescence intensity I(t) was calculated as the mean of the fluorescence signal spanning 20 nm, from (λmax - 10) nm to (λmax + 10) nm. The percentage of fluorescence quenched IQ(t) was determined according to IQ(t) ) [1 - I(t)/IBL] × 100

(1)

where IBL was the mean fluorescence baseline intensity prior to analyte delivery, averaged from t ) 0 to t ) 10 s. For each analyte delivery, the IQ(t) response curve was used to calculate the maximum quench and area quenched to describe the response of each polymer. These descriptors were calculated within the time window of t ) 60-150 s of the 225 s analyte exposure. The maximum quench was calculated by finding the 10 consecutive IQ(t) measurements in this time window that, when averaged, yielded the largest mean quench. The area quenched was found by integrating the area under (40) Felt, D. R.; Larson, S. L.; Valente, E. J. Chemosphere 2002, 49, 287–295.

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

9919

Figure 3. Sensor responses showing (a) I(t) for each of the polymers throughout the 13 h sampling period; (b) IQ(t) response of the sensor array to 50 µM TNT, tetryl, 1,3-dinitrobenzene, and duroquinone in water and to a water blank; and (c) I(t) and IQ(t) responses for the sensor array upon exposure to 50 µM 1,3-dinitrobenzene in water, with 10 responses throughout the 13 h sampling period shown with each dose starting at t ) 0 s. See Table 1 for polymer designations.

the IQ(t) response curve from t ) 60 s to t ) 150 s. The array consisted of five polymer bands, thus ndesc ) 10 was the number of response descriptors used to describe the response vector r to each analyte. All processing and analyses were performed in MATLAB. Response Normalization. Response normalization has been shown to generate concentration-independent response signatures from sensors that respond linearly with concentration.41 Responses were therefore normalized according to ¯r )

r

(2)

∑r

2 i

i)1:10

where r is the 10-descriptor array response vector consisting of the maximum quench and area quenched for each of the polymers, the ri are components of r, and jr is the normalized response vector.42 Discrimination and Classification: Principal Component Analysis and Linear Discriminant Analysis. The ability of the sensor array to discriminate between the different test analytes was explored by principal component analysis (PCA) and linear discriminant analysis (LDA). PCA was performed by calculating the eigenvalue decomposition of the correlation matrix of the total response matrix R (nexp × ndesc), where nexp ) nannrep was the total number of exposures and ndesc ) 10. Response matrix R was meancentered and autoscaled and was projected onto the calculated eigenvectors to produce the principal components (PCs). The eigenvalues provided information on the percentage of variance contained in each of these PCs.43,44 LDA was performed by use of the classify function in MATLAB. A leave-one-out classification scheme was employed where each (41) Severin, E. J.; Doleman, B. J.; Lewis, N. S. Anal. Chem. 2000, 72, 658– 668. (42) Brereton, R. G. Chemometrics: Data Analysis for the Laboratory and Chemical Plant; John Wiley & Sons Ltd.: West Sussex, England, 2003.

9920

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

“unknown” 10-descriptor response vector was fit to nexp - 1 “known” samples and classified accordingly.43,44 PCA and LDA were performed on unnormalized and normalized response vectors. RESULTS AND DISCUSSION Sensor Lifetime. Figure 3a displays I(t) throughout the entire 13 h sampling period for the five polymer sensors. Each of the small drops in fluorescence represents a quenched signal in response to an analyte. The strong initial drop in fluorescence intensity was likely due to the washing off of loosely bound polymer from the glass beads. Because of this strong initial drop in I(t), the first exposure to each of the vials was discarded, leaving 10 exposures per sample for analysis. The initial drop was followed by a gradual decay in fluorescence signal due to photobleaching of the polymer,45 likely due to the presence of high concentrations of oxygen in the untreated (i.e., not bubbled with N2) water.46 In contrast to the remainder of the sensors, sensor P3 exhibited strong fluorescence intensity throughout the 13 h period with minimal degradation in intensity. Throughout the entire 13 h period, all polymers continued to respond. Array Response. Figure 3b displays the quenched fluorescence intensity IQ(t) response versus time for the sensor array during 15 s exposures to 50 µM TNT, tetryl, 1,3-DNB, and duroquinone in water and to a clean water blank. The 15 s exposures were initiated at t ) 60 s. Due to the volume of the sampling system, a 15 s delay was observed between exposure initiation and the arrival of the sample to the sensor array. The array response pattern differs for the different compounds. For (43) Duda, R. O.; Hart, P. E. Pattern Recognition and Scene Analysis; John Wiley & Sons: New York, 1973. (44) Johnson, R. A.; Wichern, D. W. Applied Multivariate Statistical Analysis, 5th ed.; Prentice Hall: Upper Saddle River, NJ, 2002. (45) Vanden Bout, D. A.; Yip, W. T.; Hu, D.; Fu, D. K.; Swager, T. M.; Barbara, P. F. Science 1997, 277, 1074–1077. (46) Wilkinson, F.; McGarvey, D. J.; Olea, A. F. J. Phys. Chem. 1994, 98, 3762– 3769.

Figure 4. (a) Area quenched and (b) maximum quench for polymers exposed to analytes in water at 50 µM (unless noted otherwise). Error bars denote (1 standard deviation in the response. See Table 1 for polymer designations.

example, among the polymers in the array, the strongest quenching response toward DQ is seen by P4, while P4 exhibits the weakest quenching response toward TNT and tetryl. These differences result from the physics governing the fluorescencequenching response: FQ ∝ CKb[exp(-∆G°)2]

(3)

where FQ is the amount of fluorescence quenched, C is the concentration of the analyte, Kb is the binding constant, and ∆G° is the overall free energy change for the electron-transfer reaction.32 For a given analyte, Kb and ∆G° are different for each of the polymers. The response kinetics are also unique for the different polymers in the array. Polymers P1 and P5 begin to respond simultaneously upon exposure to TNT. However, P5 exhibits a rapid sorption/desorption response, while P1 shows a more gradual desorption of TNT from the sensor. This is in contrast to the response to DQ, where P2 and P4 exhibit slower desorption kinetics while P1 displays a rapid sorption/desorption response. The use of the maximum quench and area quenched response descriptors collectively captures these different types of response information. Figure 4 displays the average maximum quench and the average area quenched for the sensor array upon exposure to the 17 sample vials. Averages of the 10 exposures to each vial are shown. Unique array response patterns are observed in response to most of the different analytes. Of particular interest is the difference in the response pattern for exposure to the isomers of dinitrobenzene: for 1,2-, 1,3-, and 1,4-DNB, the response of P3 changes significantly relative to the responses of the other four polymers. This relative change can be exploited as an identifying characteristic among these isomers. Additionally, the response pattern to the nonquencher RDX is identical with the responses to the two blanks. This natural selectivity of the sensors toward EWCs is advantageous because it minimizes the number

of compounds in the field that can cause response interference and/or produce false positives. Response Reproducibility. Figure 3c displays the responses of the sensors along the array to the 10 exposures of 50 µM 1,3dinitrobenzene. The left column displays the raw fluorescence signal intensity I(t), while the right column displays the quenched fluorescence signal IQ(t). Changes in I(t) response signals reflected the various levels of fluorescence decay observed in Figure 3a. When these signals were converted into percent quenched IQ(t), the responses of P3 converged into a highly reproducible response signal with respect to both response magnitude and response profile. The responses of P1, P2, P4, and P5 exhibited a similar response profile when converted into IQ(t); however, the response magnitudes showed differing degrees of variability. The ability to reproducibly extract response signatures from the array is demonstrated in Figure 4 in the form of error bars. Response variance was directly related to the drift observed in Figure 3a. For example, sensor P3, which exhibited the least decay in fluorescence signal, exhibited the least variance in response. Array Discrimination: Principal Component Analysis. PCA takes data in n dimensions and reorients the axis the data are represented along. The axes are chosen such that PC1 contains most of the response variance from the data, PC2 is orthogonal to PC1 and contains the second most response variance, etc. Reexpressing the original data along a truncated number of these principal components thus elucidates in two or three dimensions, which can be easily visualized, the natural response clustering of higher-dimensional data.42-44 Figure 5a displays PCA performed on the raw response descriptors. For clarity, only a portion of the analytes is shown. A tight cluster is observed for the blank vials and RDX, which exhibited no response. The explosives TNT and tetryl are clustered uniquely, as are a number of other compounds. Consistent with response patterns from Figure 4, separation is observed for the isomers of dinitrobenzene (1,2-DNB, 1,3-DNB, Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

9921

Figure 5. Principal component analysis (PCA) performed on (a) raw and (b) normalized response data; (c) legend. PCs 1-3 contain (a) 79%, 12%, and 5% and (b) 40%, 24%, and 18% of the variance, respectively.

and 1,4-DNB). However, the isomers of aminodinitrotoluene (2am-DNT and 4-am-DNT) are clustered together. Thus, the array has the ability to discriminate among certain isomer groups but cannot necessarily discriminate among all isomer groups. Additionally, linear signature trajectories are observed for certain analytes (e.g., DQ and 1-Cl-2,4-DNB). These patterns result from the drift observed in fluorescence signal intensity (Figure 3a). Figure 5b displays PCA for the normalized response descriptors. The normalization procedure (eq 2) placed all responses on the same response magnitude, minimizing the influence of response magnitude on the array signature. The nonresponders (blanks, RDX) were not shown, because the normalization scaled up the noise of their nonresponding baselines, causing their responses to scatter widely throughout the plot. For the majority of the analytes that elicited sensor responses, the normalization minimized the effect of the baseline drift observed in Figure 3a, creating more tightly clustered groupings of the analytes. When the raw response signatures were used, four sensors exhibited varying degrees of drift (P1, P2, P4, and P5; Figure 3a). Higher-order PCs would naturally contain a significant amount of this response drift, which is reflected in Figure 5a. The normalization procedure, however, minimized the effect of the response drift on response signature. In the extreme case of all sensors exhibiting the same amount of response drift, the normalization procedure would create an identical response signature throughout all phases of the drift. In the present case where four sensors (P1, P2, P4, and P5) exhibited some level of 9922

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

response drift and one sensor (P3) exhibited no response drift, the effect of the drift on the response signature was significantly lessened, leading to the improved response clustering observed in Figure 5b. Array Discrimination: Linear Discriminant Analysis. General. Table 2 displays the confusion matrix for analyte classification using LDA. The row denotes what was actually delivered to the array, and the column denotes what fraction of the 10 exposures was classified into each of the analyte classes. Thus, each row sums to 1, and perfect classification would yield an identity matrix with 1s along the diagonal. The top and bottom number (parentheses) are for classification with the raw and normalized response vector, respectively. False Positives/False Negatives: Explosives. No false negatives occurred for the classification of explosive analytes: the responses of TNT and tetryl were classified correctly 100% of the time with the raw response vectors, and tetryl was misclassified as TNT 1/10 times with the normalized response vector. Additionally, use of the raw response vectors for the nonresponders (two blanks and RDX) yielded random classification within nonresponding groups only. For example, RDX was classified 1/10, 5/10, and 4/10 times as RDX, blank 1, and blank 2, respectively. This suggests that the algorithm was not overtrained. Additionally, zero false positives occurred for any of the nonresponders with the raw response vector. For the normalized response vectors, occasional false positives were recorded for these nonresponders; for example, blank 1 was classified 1/10 times as tetryl and 2-amDNT, respectively. This was because the normalization procedure inflated the noise of the nonresponding signal to be on the same response magnitude as responding analytes, leading to occasional spurious classifications as a responding analyte. This could be addressed by implementing an algorithm that classifies a response within the training library only if a critical response magnitude is exceeded by the sensors. Structural Isomers. Of particular note was the classification of the isomers of DNB. With the raw response vectors, these analytes were correctly identified 100% of the time. A portion of that discrimination resulted from the differences in overall response magnitudes of the different analytes (Figure 4b). Classification with the normalized response vectors, which used response signatures of equal response magnitude, maintained a high level of classification among these isomers, correctly classifying 1,2-, 1,3-, and 1,4-DNB exposures 6/10, 10/10, and 9/10 times, respectively. Isomer classification is among the most difficult tasks for chemical identification, and standard analytical techniques such as GC/MS often struggle to discriminate within isomer groups.47 In contrast to the DNB isomers, the sensor array was not able to discriminate between the isomers 2-am-DNT and 4-am-DNT nor between the isomers 3-NT and 4-NT. For these isomers, classification was randomly assigned within the respective isomer class. Figure 4 shows that polymer P3 provided the discriminating information for the DNB isomers. The polymeric structure of P3 (Figure 1) thus contains chemistry capable of discriminating between the similar structural isomers of DNB but not the other isomer groups. This suggests that there may be alternative fluorescent polymers that could discriminate between larger (47) Armstrong, D. W.; DeMond, W.; Alak, A.; Hinze, W.; Riehl, T. E.; Bui, K. H. Anal. Chem. 1985, 57, 234–237.

Table 2. Confusion Matrix for LDA Classification among the 17 Randomly Delivered Sample Vialsa

a Rows represent the chemical delivered to the sensor array, and columns represent the chemical prediction of the sensor array using a leaveone-out classification scheme with linear discriminant analysis. Raw (top number) and normalized (bottom number, in parentheses) array response analyses are shown. Numbers are shown as fraction predicted; perfect classification would be 1s along the diagonal. All concentrations were 50 µM, unless otherwise noted. Nondiagonal zeroes have been deleted for clarity.

numbers of isomer groups, which could be included in a sensor array to generate a greater ability to discriminate between different analytes. Signature Concentration Dependences. To determine the effect of concentration on response signature, DQ was sampled at 20, 50, and 200 µM. Responses to 20 µM DQ appear to be below the detection limit of some of the sensors in the array (Figure 4). Above 20 µM DQ, the different concentrations displayed similar response profiles but different response magnitudes. This yielded near-perfect classification of DQ at the appropriate concentration when the raw response vector was used. When the normalized response vectors were used, DQ was correctly classified as DQ in the majority of cases; however, the ability to correctly predict concentration was lost (Table 2). For example, with the normalized response vector, 50 µM DQ was classified as 50 µM DQ 5/10 times and as 200 µM DQ 5/10 times. Similar effects were observed for classification of 200 µM DQ. When sampled at concentrations above the detection limit of the sensors, normalized response signatures appeared independent of concentration over the limited concentration range tested here, and they could be independent of concentration over several orders of magnitude.

PRACTICAL IMPLICATIONS The ability to identify and discriminate among all compounds of interest, with no concern for potential interferents and their associated false positives, is desired for any analytical technique. In the field application of an explosives detection unit, there are likely to be several interferents that could cause false positives. The high explosive RDX is of obvious interest for explosives detection; however, the sensor array was not able to obtain a response toward it. While this eliminates the utility of this approach as a panacea for explosives detection, it also helps to significantly reduce the likelihood of false positives in the field. Field use could encounter the challenge of discriminating among EWC interferents; however, in general the number of interferents one must be concerned with is significantly lessened due to the natural selectivity of fluorescent polymers toward EWCs. This is in contrast to some of the more broadly responsive sensor array transduction modalities.5,7,18,21 The fluorescent polymer sensor array has demonstrated the ability to discriminate between chemically similar analytes, including the isomers of DNB. The ability to discriminate between such compounds will be useful for discriminating between explosives, explosive-related compounds, and background interferents. This will increase the utility of the sensor for field use, providing the Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

9923

ability to discriminate between various explosives as well as nonexplosive interferents that could otherwise cause a false alarm. The normalization procedure minimized the effect of response drift, which resulted from the different polymer fluorescence decay profiles. If all of the sensors exhibited an identical fluorescence intensity decay profile, the normalization procedure would perfectly cancel out the response drift resulting from the fluorescence decay. This could be done by choosing a fluorescent polymer to be used as the signal transducer and creating polymer blends with various nonfluorescent polymers to obtain sensor diversity. This would be similar to the use of polymer-carbon black composites, where the carbon black serves as the signal transducer and nonconductive polymers generate chemical diversity along the array.5,6 The fluorescent polymer for signal transduction could be chosen on the basis of fluorescence stability/lifetime (P3), area quenched during exposure to TNT (P1), or any other desired response property. For the concentration-dependent DQ response data that were collected, the response signatures were consistent above the detection limit of the sensors in the array and over a limited concentration range. The relative responses of the sensors remained the same, and only their response magnitudes changed. This is because the amount of fluorescence quenched, FQ, is proportional to the concentration of analyte C (eq 3). This concentration dependence could facilitate response signature classification, could minimize the required training for the sensor array, and could be exploited for the identification of mixtures. CONCLUSIONS A fluorescent polymer sensor array was made from commercially available polymers coated onto glass beads and was used

9924

Analytical Chemistry, Vol. 82, No. 23, December 1, 2010

to detect analytes in aqueous solution. The array exhibited a natural selectivity for electron-withdrawing compounds (EWCs), while the non-electron-withdrawing explosive RDX produced no response. Discrimination among the different EWCs, including the explosives TNT and tetryl, explosive-related compounds, and the nonexplosive isomers of dinitrobenzene, was demonstrated by principal component analysis (PCA) and linear discriminant analysis (LDA). The beads exhibited continuous responses over a 13 h sampling period, with various degrees of fluorescence decay and response drift over time. A normalization procedure minimized the effect of sensor drift on response signature and created concentration-independent response signatures. The natural selectivity of the array toward EWCs, combined with the ability of the array to easily discriminate among different EWCs over extended periods of continuous operation, can be exploited to design a detector with the ability to discriminate between explosives, explosive-related compounds, and nonexplosives.

ACKNOWLEDGMENT The authors thank the Department of Defense Unexploded Ordnance Center of Excellence (DoD UXOCOE) for financial support. The National Research Council is acknowledged for a postdoctoral fellowship to V.P.S.

Received for review September 21, 2010. Accepted October 21, 2010. AC102504T