Anal. Chem. 2003, 75, 4389-4396
Cross-Reactive Metal Ion Sensor Array in a Micro Titer Plate Format Torsten Mayr,† Christian Igel,‡ Gregor Liebsch,† Ingo Klimant,§ and Otto S. Wolfbeis*,†
Institute of Analytical Chemistry, Chemo- and Biosensors, University of Regensburg, D-93040 Regensburg, Germany, Institut of Neuroinformatics, Ruhr University Bochum, D-44780 Bochum, Germany, and Institute of Analytical Chemistry, University of Technology Graz, A-8010 Graz, Austria
Sensor arrays enable the simultaneous determination and assessment of multiple chemical information. Existing sensing schemes employ a variety of chemical interaction strategies and include the use of conductive polymers, metal oxide field effect transistors, surface acoustic wave devices, catalytic (tin oxide), and electrochemical or optical sensors.1 The proper assembly of arrays from the multitude of existing sensors, for example, for environmental contaminants, is a demanding task, since practically all sensors are different in terms of sensing scheme and instru-
mentation. An ideal sensor array preferably is based on a single technology in order to minimize the instrumental effort. Bearing this in mind, we recently reported on a novel array concept that is based on (a) a uniform analytical protocol, (b) the use of fluorescent indicators with similar excitation and emission wavelengths, (c) integration into micro titer plate (MTP) technology (in order to form disposable sensor arrays), (d) the use of chemical imaging (which allows the reading of all parameters simultaneously), and (e) the use of inexpensive semiconductor and optical components.2 This combination paves the way for analyzing aqueous solutions for a large number of parameters using a single optoelectronic system and within a short time. The new scheme does not require highly sophisticated steps in preparation and signal processing, but rather makes use of the widely accepted MTP technique. This is in contrast to previous reports3 in which arrays based on microbeads contained in micro machined cavities are described. The scheme parallels a report (using different materials) to “see” odors by making use of sensor spots whose coloration is affected by vapors of certain odorants.4 Specifically, the scheme applies commercially available fluorescent indicators that can be excited in the range between 400 and 500 nm, emit light with a maximum at >500 nm, and display a decay time in the nanosecond range. The indicators are dispersed into thin films of a water-soluble polymer and placed on the bottom of the well of an MTP. Both the polymer and the indicator dissolve after adding the aqueous sample to the well. The interaction of analyte ion and indicator result in a fluorescence intensity related to the concentration of the analyte. The physical immobilization of the indicators in the wells of a MTP enables an easy preparation of arrays, since no chemical modification of the indicator is required. The arrays thus prepared can be conveniently stored for a long time. In this contribution, we describe a cross-reactive sensor array based on a novel concept and an improved imaging setup. In contrast to conventional sensor schemes that make use of specific interaction between the analyte and the receptor, this sensing strategy employs an array of unspecific sensors.5-8 This was also inspired by the excellent performance of biological olfactory
* To whom correspondence should be addressed. E-mail: otto.
[email protected]. † University of Regensburg. ‡ Ruhr University Bochum. § University of Technology Graz. (1) Albert, K. J.; Lewis, N. S.; Schauer, C. L.; Sotzing, G. A; Stitzer, S. E.; Vaid, T. P.; Walt, D. R. Chem. Rev. 2000, 100, 2595-2626.
(2) Mayr, T.; Liebsch, G.; Klimant, I.; Wolfbeis, O. S. Analyst 2002, 127, 201203. (3) Goodey, A; Lavigne, J. J.; Savoy, S. M.; Rodriguez, M. D.; Curey, T.; Tsao, A.; Simmons, G.; Wright, J.; Yoo, S.-J.; Sohn, Y. E.; Anslyn, V.; Shear, J. B.; Neikirk, D. P.; McDevitt, J. T. J. Am. Chem. Soc. 2001, 123, 2559-2570. (4) Rakow, N. A.; Suslick, K. S. Nature 2000, 406, 710-713.
A cross-reactive array in a micro titer plate (MTP) format is described that is based on a versatile and highly flexible scheme. It makes use of rather unspecific metal ions probes having almost identical fluorescence spectra, thus enabling (a) interrogation at identical analytical wavelengths, and (b) imaging of the probes contained in the wells of the MTP using a CCD camera and an array of blue-light-emitting diodes as a light source. The unselective response of the indicators in the presence of mixtures of five divalent cations generates a characteristic pattern that was analyzed by chemometric tools. The fluorescence intensity of the indicators was transferred into a timedependent parameter applying a scheme called dual lifetime referencing. In this method, the fluorescence decay profile of the indicator is referenced against the phosphorescence of an inert reference dye added to the system. The intrinsically referenced measurements also were performed using blue LEDs as light sources and a CCD camera without intensifiers as the detector. The best performance was observed if each well was excited by a single LED. The assembly allows the detection of dye concentrations in the nanomoles-per-liter range without amplification and the acquisition of 96 wells simultaneously. The pictures obtained form the basis for evaluation by pattern recognition algorithms. Support vector machines are capable of predicting the presence of significant concentrations of metal ions with high accuracy.
10.1021/ac020774t CCC: $25.00 Published on Web 08/01/2003
© 2003 American Chemical Society
Analytical Chemistry, Vol. 75, No. 17, September 1, 2003 4389
systems that use a large array of cross-reactive receptor cells that are not highly selective. The receptors signal the presence of an odor and create neuronal response patterns believed to be used to identify odors.9,10 In the past decade, this scheme principle was transferred to a manifold of approaches for vapors and liquids, so-called (opto)electronic noses or electronic tongues.1 Heavy metal ions were determined by potentiometric schemes,11,12 and also via a fiber optic system using specific fluorescent indicators but where varying wavelengths were needed for the detection of light intensity.13 Identification of the analyte is achieved in these systems by the recognition of distinct response patterns using various models for data evaluation. Here we present a cross-reactive array for the determination of mixtures of calcium(II), copper(II), nickel(II), cadmium(II), and zinc(II) ions. Generally, the fluorometric determination of heavy metal ions is interfered by other metal ions as a result of the lack of specific indicators. The approach reported here takes advantage of the poor probe selectivity forming cross-reactive arrays of unselective indicators. Selectivity is achieved by analyzing the obtained response pattern using support vector machines. In addition, we introduce a powerful novel setup capable of reading an entire 96-well MTP in one step within few microseconds and the detection of fluorophores in the nanomolar concentration range without any amplification. This was achieved by illumination of each single well with one LED. EXPERIMENTAL SECTION Chemicals and Solutions. The fluorescent probes Oregon Green BAPTA-5N, Phen Green FL, Newport Green, FluoZin-1, BTC-5N, Fluo-5N, and carboxyfluorescein were obtained from Molecular Probes Europe BV (Leiden, The Netherlands). Calcein and Lucifer Yellow CH were from Fluka (Buchs, Switzerland). The phosphorescent reference beads consisting of polyacrylonitrile (PAN) containing ruthenium(II)-tris(4,7-diphenyl)-1,10phenanthroline were a gift from Chromeon GmbH (Regensburg, Germany). They have been described in detail.5 Inorganic salts of analytical reagent grade were from Merck (Darmstadt, Germany); imidazol buffer, from Sigma (Vienna, Austria); and black micro titer plates (96 wells) with transparent bottom, from Greiner (Frickenhausen, Germany). Stock standard solutions of 1000 µmol L-1 metal ion concentration were prepared by dissolving the respective amount of the nitrate salt in 5 mM imidazole buffer solution of pH 7.0. Working solutions were prepared containing 10 and 100 µmol L-1 of the respective metal ion by diluting the stock with imidazole buffer. (5) Klimant, I.; Huber; C.; Liebsch, G.; Neurauter, G.; Stangelmayer, A.; Wolfbeis, O. S. In New Trends in Fluorescence Spectroscopy; Valeur, B. J., Brochon, C., Eds.; Springer Verlag: Berlin, 2001; Chapter 13. (6) Liebsch, G.; Klimant, I.; Krause, C.; Wolfbeis, O. S. Anal. Chem. 2001, 73, 4354-4363. (7) Persaud, K.; Dodd, G. Nature 1982, 299, 352-355. (8) Lundstro ¨m, I.; Erlandsson, R.; Frykman, U.; Hedborg, E.; Spetz, A.; Sundgren, H.; Welin, S.; Winquist, F. Nature 1991, 352, 47-52. (9) Lancet, D. Nature 1991, 351, 275-276. (10) Shepherd, G. Neuron 1994, 13, 771-790 (11) Vlasov, Y.; Legin, A. V.; Rudnitskaya, A. Sens. Actuators 1997, B44, 532537. (12) Kukla, A. L.; Starodub, N. F.; Kanjuk, N. K.; Shirshov, Y. M. Sens. Actuators 1999, B57, 213-218. (13) Prestel, H.; Gahr, A.; Niessner, R. Fresenius’ J. Anal. Chem. 2000, 368, 182-191.
4390 Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
Table 1. Indicator Concentrations and Effect of Ions on the Fluorescence, If Present, in 100 µmol L-1 Ion Concentrationa indicator Fluo-Zin1 BTC-5N calcein Lucifer Yellow Phen Green Newport Green Oregon Green BAPTA-5N Fluo-5N a
ion concn (µmol L-1)
Ca2+
Cu2+
2.5 5 1 5 5 1 1
+
-
2.5
+
-
-
Ni2+
Zn2+
Cd2+
+ -
+ +
+
+
+
+ +
+
+ indicates quenching; - indicates enhancement of fluorescence.
Preparation of Cross-Reactive Arrays. Ion mixtures were prepared by filling the wells of MTPs with 10 µL of solutions containing 10, 100, or 1000 µmol L-1 metal ion using a robotic system (from Hamilton, Darmstadt, Germany). Each column contained a different mixture. Finally, the volume was increased to 90 µL with buffer. In this way, 204 different ion mixtures were prepared. Arrays were obtained by adding 10 µL of solutions containing the fluorescent probes and suspensions of the fluorescent reference beads to the ion mixture at the same time using the eightchannel electronic pipet. Arrays of eight elements were arranged in columns. The concentrations of the fluorescent indicators and the nanobeads in the ion mixtures after dilution are given Table 1. The concentrations of the ions were 0, 1, 10, or 100 µmol L-1, respectively, after addition of indicator solution and bead suspension. The MTPs were immediately imaged after adding the indicator/bead solution. All experiments were performed at 22 ( 2 °C. Fluorescence Spectra. Excitation and emission spectra of the indicators and the reference particles were recorded with a Varian Cary Eclipse spectrofluorometer possessing a Schwarzschild collection optics and equipped with an MTP accessory, a 12-W xenon flash lamp as a light source (pulsed at 80 Hz), and a redsensitive photomultiplier tube (R777, Hamamatsu). The pulse width was 2 µs, and the pulse energy typically was 75 kW. Imaging Setup. Time-resolved imaging of the emission intensity was performed with a gated CCD camera and a pulsed blue LED as a light source, as described by Liebsch et al.,6 using a modified setup. An array of 96 LEDs was used in order to illuminate every well of the MTP by one LED. The LEDs (λmax 470 nm, NSPB, Nichia, Nu¨rnberg, Germany) were arranged to fit the wells of a 96-well MTP as shown in Figure 1A. This new LED array, developed in house, offers high flexibility, and the LEDs may be exchanged easily by others of different color. The array was covered with a combination of (a) a dichroic medium blue filter (Linos Photonics, Go¨ttingen, Germany) and (b) a shortpass filter type BG12 (Schott, Mainz, Germany) for excluding the red fraction of the LED emission. Pulses of blue excitation light hit the MTP containing the respective probes and samples in their wells. The geometry initially used2,6 does not allow, however, imaging of the area of a whole 96-well MTP simultaneously, but only a segment of 4 × 5 wells. Therefore, the emitted light was collected
Figure 1. (A) Picture of an array of 96 LEDs located under the wells of a 96-well micro titer plate (MTP). Each LED illuminates one well. (B) Fiber-optic adapter for imaging of MTPs; it reduces the imaged area to fit the standard optics of the imaging setup and enables the simultaneous evaluation of 96 wells at the same time. (C) Photograph of the combination of LED array (A) and the MTP fiber-optic adapter (B). (D) Schematic of the complete setup. Light from the LEDs passes an excitation filter and hits the wells of the micro titer plate. Light emitted by the fluorescent indicators passes an emission filter and is detected by the CCD camera.
and guided (via optical fibers of 3 mm i.d.) to an area of approximately 40 × 80 mm (between the two metal plates shown in Figure 1B). Emitted light then passed a KV 550 filter (Schott, Mainz, Germany) and was detected by the CCD camera (SensiMod, PCO, Kelheim, Germany). The camera has a black/white CCD chip with 640 × 480 pixels (307 200 pixels, VGA resolution) and a 12-bit resolution, equivalent to 4096 Gy-scale values. The CCD chip can be gated directly with a minimal trigger time of 100 ns; additional image intensification is not required. The camera and the light source are triggered by a pulse generator (DG535, Scientific Instruments, West Palm Beach), and the image data is transferred to a personal computer. Hardware and image processing are controlled by software developed in-house.15 Figure 1C shows the combination of adapter, MTP, and LED array. The complete setup is illustrated in Figure 1D.
Dual Lifetime Referencing (DLR). Most measurements were performed using dual lifetime referencing (DLR), a spectroscopic scheme that converts a fluorescence intensity information into a time-dependent parameter.5,6 A phosphorescent dye contained in poly(acrylonitrile) particles was added to the sample containing the respective fluorescent indicator. For a successful application of the scheme, it is mandatory that the excitation and emission spectra of the phosphor (in the nanobeads) overlap the respective spectra of the indicator. Both luminescences are excited simultaneously, and two images are acquired at different time gates by the camera. The first image is recorded during excitation and reflects the luminescence signal of both the fluorophorescent (14) Gadella, T. W. J., Jr.; Van Hoek, A.; Visser, A. J. W. G. J. Fluoresc. 1997, 7, 35-43. (15) Liebsch, G.; Klimant, I.; Frank, B.; Holst, G.; Wolfbeis, O. S. Appl. Spectrosc. 2000, 54, 548-559.
Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
4391
indicator dye and the phosphorescent reference dye. The second image is acquired after switching off the light source and is solely caused by the long-lived phosphorescent dye. Since the intensity of the fluorophore contains the information on the respective analyte, whereas phosphorescence is inert to it, the ratio of the images displays a referenced intensity distribution that reflects the concentration of the analyte at each picture element (pixel). Data Analysis. We used different methods of machine training to analyze the intensity distributions obtained. Both feed-forward multilayer perceptron neural networks16 and support vector machines17-19 (SVMs) were trained to predict, from a response pattern, whether particular metal ions are contained in the corresponding sample. Because the SVMs gave better results, we only describe their application. Support Vector Machines. SVMs have become a standard machine learning tool for classification. In essence, they are based on the following scheme: the input is mapped into a highdimensional feature space. This mapping is done implicitly by means of so-called kernel functions. The data points are linearly separated in that feature space. An SVM separates the points by the hyperplane with maximum soft margin with respect to the training data, that is, in a way that promises good generalization behavior. In the simplest case, the SVM maximizes the minimum distance of the examples to the separating hyperplane. A detailed description of SVMs can be found in introductory books.18,19 We used 1-norm soft-margin C-SVMs implemented in SVMlight.20 Let D ) {(x1, y1), ..., (xm, ym)} be a training data set with n-dimensional input patterns xi ∈ Rn and corresponding target classes yi ∈ {-1, 1}. The (dual) quadratic optimization problem solved by C-SVMs is minimizing m
W(r) ) -
∑
Ri +
i)1
1
m
∑ y y R R K(x , x )
2 i,j)1
i j i j
i
j
(1)
with respect to R ∈ Rm subject to m
∑ yR ) 0 j j
and 0 e Ri e C
(2)
i)1
for all i ) 1, ..., m. The constant hyperparameter C > 0 controls the regularization in that it governs the tradeoff between minimal error on the training data set and maximizing the soft margin. We used radial-basis function kernels of the type
K(x, z) ) exp(-γ||x - z||2)
(3)
with a bandwidth parameter γ > 0. After optimization, the decision function is given by (16) Bishop, C. M. Neural Networks for Pattern Recognition; Oxford University Press: Oxford, U.K., 1995. (17) Vapnik, V.; Statistical Learning Theory; Wiley: New York, 1998. (18) Christianini, N.; Shawe-Taylor, J.; An Introduction to Support Vector Machines; Cambridge University Press: Cambridge, U.K., 2000. (19) Scho¨lkopf, B.; Smola, A. J. Learning with Kernels - Support Vector Machines, Regularization, Optimization, and Beyond; MIT Press: Cambridge, MA, 2002. (20) Joachims, T. Making large-scale SVM learning practical. In Advances in Kernel Methods - Support Vector Learning; Scho ¨lkopf, B., Burges, C. J. C., Smola, A. J., eds; MIT Press: Cambridge, MA, 1999; pp 169-184.
4392
Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
m
f(x) ) sgn(
∑ y R K(x , x) + b) i i
i
(4)
i)1
m yjRjK(xj, xi) ) 1 for any where b ∈ R is chosen such that yi ∑j)1 i with 0 < Ri < C. The data points xi with Ri > 0 are called support vectors. Data Set and Preprocessing. Response patterns of m ) 204 different ion mixtures were available for training the models. The solutions contained variations of calcium(II), copper(II), nickel(II), zinc(II), and cadmium(II) ions in concentrations of 0, 1, 10, and 100 µmol L-1, respectively. We trained classifiers for the prediction of whether the amount of a particular ion is larger than a given threshold. We distinguished three tasks, namely, to classify whether the concentration exceeds either 0, 1, or 10 µmol L-1. For example, if a classifier is trained for the prediction of a copper concentration of >1 µmol L-1, then the training data {(x1, y1), ..., (xm, ym)} for the SVM were organized such that the inputs xi corresponded to the 5-dimensional response patterns (with each input dimension normalized to unit variance). We then set yi ) 1 if the concentration of copper in the corresponding sample was larger than 1 µmol L-1, and yi ) -1 otherwise. Model Selection. To obtain generalizing models, the SVM hyperparameters γ and C have to be adjusted. Therefore, we have performed a grid search for each combination of the five ions and the three threshold values. In each grid search, the leave-one-out (“LOO”) error (i.e., the m-fold cross-validation error) RLOO is computed for each pairing of γ ∈ {2-8, 2-7, ..., 21} and C ∈ {2-5, 2-4, ..., 24}. The LOO error is a very reliable estimate of the model performance after training, that is, its generalization behavior. For each data point (xi, yi) ∈ D an SVM fi is built using the training data set D\{(x1, y1)} . Then this SVM is tested whether it classifies x correctly by computing the loss term ei ) |fi(xi) - yi|. The LOO error is finally given by averaging over all data points, i.e., m RLOO ) 1/|D| ∑i)1 ei. Usually, RLOO is too time-consuming to be computed. However, in our application, the efficiency of SVM training and the size of the respective data set allowed for this rigorous evaluation. Hence, all results in this article refer to the minimum LOO error over the choices for the parameters γ and C.
RESULTS AND DISCUSSION Method. Fluorescence intensity is a parameter widely used in fluorometry, but it is a poor parameter with respect to quantitative imaging, because it depends on a number of variables other than the concentration of the species to be assayed. These include the variation of light intensity over the area to be imaged and inhomogeneities of the dye distribution. The most effective way to eliminate such adverse effects consists of the measurement of decay time (τ). Since τ it is an intrinsically referenced parameter, it is superior to intensity in that it is independent of local concentrations of fluorescent probes and fluctuations in light source intensity. On the other hand, lifetime imaging requires more sophisticated and expensive instrumentation,14 especially for approaches employing nanosecond-decaying fluorescent probes. Nevertheless, a recently developed scheme allows one to generate intrinsically referenced readouts by measuring these short-
decaying indicators using a microsecond-resolving system with comparatively simple instrumentation.5,6 This scheme, referred to as time domain dual lifetime referencing (t-DLR), converts the intensity information in a time-dependent parameter. Specifically, a phosphorescent dye (contained in polymeric nanobeads) is added to the sample containing the fluorescent indicator. Since the excitation and emission spectra of the phosphor overlap, the respective spectra of the indicator and indicator and reference dyes can be excited simultaneously, and two images can be taken at different time gates, one recorded in the excitation period (Aex), during which the light source is on, the other in the decay period (Aem), when the light source is off. Consequently, the first image reflects the luminescence signal of both the fluorescent indicator dye and the phosphorescent reference dye. The second image, in contrast, is solely caused by the long-lived phosphorescent dye. Since the intensity of the fluorophore contains the information on the respective analyte, and phosphorescence is inert to it, the ratio of the images displays a referenced intensity distribution that reflects the analyte at each pixel. The ratio can be described by the following relationship,
P)
Aex AREF-exc + AIND ) Aem AREF-em
Figure 2. Excitation and emission spectra of a solution of carboxyfluorescein buffered to pH 10 (solid lines) and of the reference particles (dashed lines). The dark gray and hatched areas reflect the optical properties of the imaging setup. The gray area gives the spectrum of the excitation light of a LED combined with a BG12 and a dichroic filter, while the hatched area displays the transmission characteristics of the emission filter and simultaneously represents the emission signal detected by the camera.
(5)
where the image Aex represents the sum of both luminescences AREF-exc and AIND, which are the signal intensities of the reference and the indicator in the excitation window, respectively. The second image Aem is equal to AREF-em, which is the signal intensity of the reference in the emission window. Performance of the Imaging Setup. The setup previously reported was adequate to demonstrate time-resolved imaging of sensor arrays in MTPs but lacks sensitivity.2 As a result, high indicator concentrations were required in order to compensate for a relatively inhomogeneous light field. Since the indicator/ analyte interaction is a complex reaction and follows the massaction law, indicator reactions in the lower micromoles-per-liter range are desired for the determination of heavy metal ions in concentration ranges set by standards and guidelines recommended by various institutions.21,22 The setup originally designed for evaluation of highly indicator-loaded sensor foils6 was modified so as to obtain excitation light of higher intensity. This was achieved by illuminating each well of a 96-well MTP with a single LED. Fluorescence is usually measured in a rectangular arrangement of excitation and emission light, which minimizes the amount of scattered excitation light.22This is in contrast to the arrangement chosen here, in which light source and detector are in line. Therefore, excitation and emission filters with appropriate transmission characteristics are required. The optimal performance was obtained by a combination of short-pass filters (medium blue and BG12) along with a long-pass filter (KV550). The spectral properties of the filtered light of the blue LED and of the transmission characteristics of the emission filter are shown in Figure 2. (21) World Health Organization. Guidelines for drinking water quality, 2nd ed.; WHO: Geneva, Switzerland, 1993. (22) European Commission. Water Quality Directive 98/83/EU. European Commission: Brussels; see: http://europa.eu.int/comm/environment/ water/water-drink/index_en.html, 1998.
Figure 3. Images illustrating the performance of the imaging system: (A) intensity picture of carboxyfluorescein solutions (of pH 4 and 10) contained in the wells of micro titer plates at a resolution of 160 × 120 pixels, (B) the same picture at a resolution of 80 × 60 pixels, and (C) t-DLR referenced picture after adding phosphorescent reference beads.
The performance of the setup was tested first by imaging MTPs containing solutions of carboxyfluorescein buffered to pH 4 and pH 10, respectively. Carboxyfluorescein was chosen as the model fluorophore because (a) many ion probes are derivatives of this compound, (b) it has a high quantum yield (0.92), and (c) its use is widespread. The excitation and emission spectra of carboxyfluorescein and dyed reference beads are given in Figure 2, as well. Both dyes can be excited using the LED/filter combination described before, and the emission spectra strongly overlap the transmission spectrum of the long-pass emission filter. Figure 3A and B shows the intensity pictures of solutions containing carboxyfluorescein in concentrations from 1 to 1000 nmol L-1 buffered to pH 4 and pH 10 (no reference beads added). The pictures were recorded with two resolutions of the CCD chip (160 × 120 and 80 × 60 pixels), respectively. Figure 3A demonstrates that a clear differentiation can be made between 10 nmol L-1 carboxyfluoresceins of pH 4 and pH 10 with negligible Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
4393
noise. The limit of detection was found to be 1 nmol L-1, as shown in Figure 3B. Although the fluorescence was not excited and detected at the spectral maximums, a detection in the lower nanomolar range is possible. This is remarkable, because the fluorescence signal was obtained without any amplification and is attributed to the high intensity of the excitation light. The intensity distribution of one well in Figure 3A and B shows inhomogeneities caused by the heterogeneous lightfield typical of LEDs. Figure 3C displays the resulting grayscale pictures after addition of the reference beads loaded with Ru(dpp) and applying t-DLR. The homogeneity of the gray distribution for each spot indicates that the inhomogeneities have been referenced out. The grayscales of the spots in Figure 3C indicate that a differentiation can be made between 10 nmol L-1 concentrations of carboxyfluorescein at pH 4 and pH 10, and differentiation of 1 nmol L-1 failed. Choice of Indicator. In contrast to common sensing schemes (in which poor selectivity is a severe drawback), the cross-reactive approach makes use of unspecific indicators. The main requirements to be fulfilled by indicators for use in a cross-reactive array are (a) a response to at least two target ions and (b) a highly different response. Furthermore, indicators had to be selected for the specific arrangement used here that meet the following criteria: (a) excitability by the blue LEDs, (b) decay times in the nanosecond range, (c) high quantum yields, and (d) good water solubility. FluoZin-1, BTC-5N, calcein, Lucifer Yellow, Phen Green, Newport Green, Oregon Green BAPTA 5N, and Fluo-5N were found to fulfill many of these requirements. All of these indicators display excellent water solubility and adequate quantum yields, can be excited with blue LEDs (λmax 470 nm), and emit light with a maximum at >500 nm. Their fluorescence decay time is in the range of 2-6 ns. The unselective and highly different response to certain target ions is summarized in Table 1. The response is illustrated by different symbols, and this, in fact, already represents a certain pattern. Fluorescence can be enhanced or quenched, depending on the signaling mechanism and the type of metal ion. Normalized excitation and emission spectra of the respective indicators in the presence and absence of the target ions are shown in Figure 4. Choice of Reference Dye. The main criteria for a reference luminophore to be useful in t-DLR imaging are (a) a decay time in the microsecond range for the use of microsecond resolving imaging setup, (b) spectral properties that are not at all affected by the sample, (c) an excitation spectrum that overlaps those of the indicators, and (d) emission spectra that overlap those of the indicators. Phosphorescent nanobeads (20-60 nm) were selected as the reference standard. These particles consist of poly(acrylonitrile) (PAN) dyed with ruthenium(II)-tris(4,7-diphenyl)-1,10-phenanthroline [Ru(dpp)] which has a quantum yield of >0.3 and a luminescence decay time of ∼6 µs.24 The luminescence of Ru(dpp) is known to be quenched by oxygen and oxidative or reductive compounds,25,26 but if incorporated into the gas(23) Valeur, M. Molecular Fluorescence; Wiley-VCH: Weinheim, New York, 2002. (24) Lin, C. T.; Boettcher, W.; Chou, W.; Creutz, C.; Sutin, N. J. Am. Chem. Soc. 1976, 98, 6536-6544. (25) Klimant, I.; Wolfbeis, O. S. Anal. Chem. 1995, 67, 3160-3166. (26) Juris, A.; Balzani, V.; Barigelletti, F.; Campagna, S.; Belser, P.; von Zelewsky, A. Coord. Chem. Rev. 1988, 84, 85-277.
4394
Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
impermeable PAN, quenching is virtually absent.27 In additon, the luminescence of beads is not affected by the target ions. The excitation and emission spectra of the PAN beads are included in Figure 2. Response Characteristics. After adding the indicator dye and the beads (as a suspension in water) to wells containing ion mixtures, these were illuminated by LEDs, and the entire MTP was imaged. The interaction of the metal ions and the indicators results in fluorescenct signals that generate a characteristic pattern for each mixture. Figure 5A shows the resulting grayscale picture reflecting the intrinsically referenced luminescence intensity of the wells. The gray scale pictures obtained in this way reflect R values (see eq 1) obtained by the ratio of the two images recorded during one acquisition cycle. Each spot in the images represents a signal in the response pattern of the array. For data evaluation, the values of the signal were obtained by extracting a circular area from the middle of each spot. This cutout represents 34 single measurements. Moreover, new images were generated in order to visualize the response of the array in the presence of various ion mixtures. The averaged pseudo color picture of the signals of 12 different arrays is shown in Figure 5B. The coloration of the spots depends on the signal intensity, resulting in a characteristic response pattern of each ion mixture. Data Evaluation. In a first attempt, regression models were built to estimate the ion concentrations from the response pattern, but the available training data turned out to be too sparse. Thus, SVMs were trained to classify whether the concentration of a particular ion in a sample exceeded the three different threshold values 0, 1, and 10 µmol L-1. The accuracy of the SVMs is represented by the leave-one-out errors. These are shown in Table 2. Obviously, it is easier to detect higher concentrations (i.e., to distinguish 0 or 1 µmol L-1 from 10 or 100 µmol L-1) than separating zero from nonzero concentrations. For example, we obtained unsatisfactory results when trying to predict whether calcium(II) was present in the sample; here, the error rate of 42.7% is close to tossing a coin. However, a calcium(II) concentration over 10 µmol L-1 can be predicted with an error rate of 17.7%. An explanation for this can be found in the dynamic range of the response of the indicators, which lies for most cases between 10 and 100 µmol L-1. Therefore, concentrations of 0 and 1 µm L-1 are hard to distinguish. Furthermore, it turns out that the accuracy of the classification strongly depends on the particular ion. For all threshold values, we obtain best results for copper(II), followed by cadmium(II), nickel(II), and zinc(II). Calcium(II) was most difficult to classify throughout. For example, a concentration of copper(II) of >10 µmol L-1 is detected with an accuracy of 93.1%, compared to 82.4% for calcium(II). The results are to be seen in relation to the number of indicators showing response to a particular ion. Table 1 shows that calcium(II) interacts with two indicators, whereas copper(II) interacts with six, and the other three ions, with at least four indicators. Additionally, copper(II) is a strong quencher, and this can superimpose the fluorescence enhancement by other ions. Neglecting Ca2+, the average classification rates for concentrations (27) Ku ¨ rner, J. M.; Klimant, I.; Krause, C.; Preu, H.; Kunz, W.; Wolfbeis, O. S. Bioconjugate Chem. 2001, 12, 883-889.
Figure 4. Excitation and emission spectra of the indicators employed (A-H) in the presence and absence of Ca2+, Cu2+, Ni2+, Zn2+, and Cd2+. Emission spectra were recorded at an excitation wavelength of 470 nm; excitation spectra, at an emission wavelength of 530 nm. The gray area gives the spectrum of the excitation light derived from a filtered LED. The hatched area displays the transmission characteristics of the long-pass filter and simultaneously represents the emission signal detected by the camera.
>0, >1, and >10 µmol L-1 are high, namely, 77.5, 86.8, and 90.3%, respectively. The latter result is particularly satisfying. However, we assume that the classification performance will be further improved once more training patterns are available.
The inadequate prediction found in certain cases is not assumed to be the result of pipetting errors, because solutions were prepared by a robotic system, and both the indicator and the reference dye were added from the same stock solutions. Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
4395
Table 2. Leave-One-Out Prediction Errors of SVMs for Ion Concentrations Larger than 0, 1, and 10 µMol L-1a RLOO, % ion Ca2+ Cu2+ Ni2+ Zn2+ Cd2+ a
>0 µmol 42.7 17.7 23.0 28.9 20.6
L-1
>1 µmol L-1
>10 µmol L-1
32.8 7.4 14.7 17.7 13.2
17.7 6.9 10.3 13.2 8.3
The prediction accuracy of the classifiers is given by 1 - RLOO.
interfering agents and competing natural chelators. The physical immobilization technique also includes the possibility of implementing buffers or masking agents by simply dispersing these reagents into a water-soluble polymer.
Figure 5. (A) Gray scale pictures of a 96-well micro titer plate (containing 12 cross-reactive arrays) after addition of the ion mixtures given beneath. Each column represents one array consisting of eight elements (spots). Rows were filled with solutions of the indicators given on the right side. The uniformity of the distribution of the gray tones indicates that local intensity variations have been successfully referenced out by the t-DLR method. Acquisition of the image takes 800 ms. (B) Pseudo color picture of the response pattern of a 96well micro titer plate containing 12 cross-reactive arrays on exposure of the ion mixtures given beneath. One row represents one array consisting of eight elements (spots). The spots represent the average value of an extracted circular area of the spots in picture A.
Kinetic effects are not expected, as well, since the measurements were performed immediately after adding the indicator/reference, and because all wells were imaged simultaneously. Further Potential. Because of the versatility of the scheme, a much larger number of parameters may be detected by the design and integration of selective and unselective indicators that may even be physically immobilized on the bottoms of the wells of MTPs. The high flexible setup offers the use of a large variety of indicators, since the LED source can be easily exchanged. In addition, existing sensors for oxygen, pH, carbon dioxide, ammonia, nitrate, or glucose may be implemented. Finally, chemometric methods may be used for the elimination of potential
4396 Analytical Chemistry, Vol. 75, No. 17, September 1, 2003
CONCLUSION A cross-reactive array for the determination of metal ions was presented that makes use of eight different unselective indicators for five cations. The array was arranged in micro titer plates, which were imaged with a CCD camera. The pattern obtained was analyzed using support vector machines. The classification results turned out to be dependent on the particular ion. For the detection of concentrations of more than 10 µmol L-1 , classification rates between 82.35% and 93.14% could be achieved. We assume that more training patterns could even improve the classification performance. For further investigations, the analysis of the data by separation in statistical independent components seems to be promising with respect to the complex interaction of indicator and mixture of ions. A powerful novel imaging setup was introduced that allows the determination of fluorophores in the nanomolar range without any amplification. Its limits of detection are comparable to those of a conventional MTP reader. The method has been shown to be superior to other imaging methods, because inhomogeneities can be eliminated via dual luminophore referencing at a comparably small instrumental effort. This imaging scheme enables all data to be acquired within the shortest time, since a multitude of parameters can be collected in one picture. This is advantageous for numerous fields of applications, including analysis of water and environmental samples and the determination of calcium when studying cellular interactions or in high-throughput screening. The scheme also allows for the continuous monitoring of processes taking place in MTPs or small bioreactors. ACKNOWLEDGMENT The authors gratefully acknowledge the input of Carsten Winkel (Institute of Neuroinformatics, Ruhr-University Bochum) for data evaluation. The work was supported by the German Environmental Foundation (DBU), project number 6000/615. Received for review December 22, 2002. Accepted May 28, 2003. AC020774T
