Multianalyte Single-Cell Analysis with Multiple Cell Lines Using a

Nov 1, 2007 - produces rich single-cell data and discriminates between single-cell responses from clonal populations stimulated with different agonist...
0 downloads 0 Views 466KB Size
Download PDF
Anal. Chem. 2007, 79, 9045-9053

Multianalyte Single-Cell Analysis with Multiple Cell Lines Using a Fiber-Optic Array Ragnhild D. Whitaker and David R. Walt*

Chemistry Department, Tufts University, 62 Talbot Avenue, Medford, Massachusetts 02155

A single-cell drug screening method is described that produces rich single-cell data and discriminates between single-cell responses from clonal populations stimulated with different agonists. Ligand-induced receptor activation is commonly detected by observing intracellular Ca2+ oscillations using high-throughput screening (HTS) methods. In most cases, HTS results in an average signal from several cells and is not sensitive enough to enable the identification of population outliers or population variance. In order to obtain this information, many individual cells must be analyzed simultaneously. We have developed a novel system using a specialized fiber-optic platform and have combined it with statistical analysis, to simultaneously analyze the dynamics of Ca2+ oscillations in a large number of single cells. Mammalian cells ectopically expressing different human GPCR receptors were stimulated, and Ca2+ changes in numerous single cells were recorded over time using a fluorescent microscope and a CCD camera. We determined the percentage of live cells in a population responding to stimuli, the distribution of responses within a population of clonal cells, and the number of outliers. By employing principal component analysis and K-nearest neighbor modeling, we classified the time-resolved Ca2+ traces of single cells as a function of the stimulus type with high certainty for a population of cells. This method is potentially a powerful tool for identifying new drug targets or for investigating the single-cell behavior of an existing target or known receptor. The development of this single-cell drug screening method is presented, and fluorescent and statistical analyses of single-cell dynamic responses are discussed. In recent years, an increased number of single-cell methods for detecting cellular events have been reported.1-4 Individual cells in a clonal population can exhibit heterogeneous behavior; however, the bulk response from thousands of cells does not capture this variation.5-7 One area where single-cell analyses are * To whom correspondence should be addressed. E-mail: david.walt@ tufts.edu. (1) Yamamura, S.; Kishi, H.; Tokimitsu, Y.; Kondo, S.; Honda, R.; Rao, S. R.; Omori, M.; Tamiya, E.; Muraguchi, A. Anal. Chem. 2005, 77, 8050-8056. (2) Rettig, J. R.; Folch, A. Anal. Chem. 2005, 77, 5628-5634. (3) Di Carlo, D.; Aghdam, N.; Lee, L. P. Anal. Chem. 2006, 78, 4925-4930. (4) Wheeler, A. R.; Throndset, W. R.; Whelan, R. J.; Leach, A. M.; Zare, R. N.; Liao, Y. H.; Farrell, K.; Manger, I. D.; Daridon, A. Anal. Chem. 2003, 75, 3581-3586. (5) Raser, J. M.; O’Shea, E. K. Science 2004, 304, 1811-1815. 10.1021/ac701744x CCC: $37.00 Published on Web 11/01/2007

© 2007 American Chemical Society

particularly valuable is the investigation of ligand-induced receptor activation, because the detailed nature of single-cell responses provides important information for understanding the full complexity of the receptor response8. We used ligand-induced activation of G-coupled protein receptors (GPCR) for single-cell analysis. GPCR receptors are the largest family of cell-surface receptors and are involved in a variety of physiological disorders, making them an important target for drug development.9 Multiple methods exist for detecting GPCR binding and activation, including immunological detection,10,11 radiometric detection,12,13 reporter gene incorporation,14 and patch clamp measurements.15 Optical detection methods such as fluorescence,16,17 luminescence,17 fluorescence resonance energy transfer,18,19 bioluminescence resonance energy transfer,20 and fluorescence correlation spectroscopy (FCS) have also been employed.21,22 In general, receptor activation is followed by a reaction cascade, leading to a change in cytoplasmic Ca2+.23,24 Detection of these Ca2+ oscillations is fast and convenient and is commonly used in drug development and high-throughput screening.23,24 Traditionally, these Ca2+ oscillations have been investigated using microtiter plates, providing a bulk result from many (6) Raser, J. M.; O’Shea, E. K. Science 2005, 309, 2010-2013. (7) Rao, C. V.; Wolf, D. M.; Arkin, A. P. Nature 2002, 420, 231-237. (8) Di, Carlo, D.; Lee, L. P. Anal. Chem. 2006, 78, 7918-7925. (9) Lefkowitz, R. J. J. Biol. Chem. 1998, 273, 18677-18680. (10) Milligan, G. Int. Congr. Ser. 2003, 1249, 15-25. (11) Zhang, Y.; Pool, C.; Sadler, K.; Yan, H.-p.; Edl, J.; Wang, X.; Boyd, J. G.; Tam, J. P. Biochemistry 2004, 43, 12575-12584. (12) Williams, C. Nat. Rev. Drug Discovery 2004, 3, 125-135. (13) Drew, J. E.; Barrett, P.; Conway, S.; Delagrange, P.; Morgan, P. J. Biochim. Biophys. Acta: Mol. Cell Res. 2002, 1592, 185-192. (14) Hirano, M.; Zang, L.; Oka, T.; Ito, Y.; Shimada, Y.; Nishimura, Y.; Tanaka, T. Biochem. Biophys. Res. Commun. 2006, 351, 185-191. (15) Pihl, J.; Sinclair, J.; Sahlin, E.; Karlsson, M.; Petterson, F.; Olofsson, J.; Orwar, O. Anal. Chem. 2005, 77, 3897-3903. (16) Levasseur, G.; Persuy, M.-A.; Grebert, D.; Remy, J.-J.; Salesse, R.; PajotAugy, E. Eur. J. Biochem. 2003, 270, 2905-2912. (17) Jacquier, V.; Pick, H.; Vogel, H. J. Neurochem. 2006, 97, 537-544. (18) Hoffmann, C.; Gaietta, G.; Buenemann, M.; Adams, S. R.; Oberdorff-Maass, S.; Behr, B.; Vilardaga, J.-P.; Tsien, R. Y.; Ellisman, M. H.; Lohse, M. J. Nat. Methods 2005, 2, 171-176. (19) Janetopoulos, C.; Jin, T.; Devreotes, P. Science 2001, 291, 2408-2410. (20) Turu, G.; Szidonya, L.; Gaborik, Z.; Buday, L.; Spaet, A.; Clark, A. J. L.; Hunyady, L. FEBS Lett. 2006, 580, 41-45. (21) Philip, F.; Sengupta, P.; Scarlata, S. J. Biol. Chem. 2007, 282, 19203-19216. (22) Vukojevic, V.; Pramanik, A.; Yakovleva, T.; Rigler, R.; Terenius, L.; Bakalkin, G. Cell. Mol. Life Sci. 2005, 62, 535-550. (23) Berridge, M. J. Adv. Second Messenger Phosphoprotein Res. 1992, 26, 211223. (24) Bootman, M. D.; Taylor, C. W.; Berridge, M. J. J. Biol. Chem. 1992, 267, 25113-25119.

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007 9045

cells.25 As mentioned above, this bulk response does not reflect the variation within a population and ignores information about the differences between clonal cells. We have developed an optical fiber-based method to detect and analyze real-time Ca2+ responses in single cells. This method enables the determination of variance in cellular populations, the percentage of cells in a population responding to a particular drug challenge, and the occurrence of outliers when cells are exposed to different agonists. Using clustering algorithms, we show that single-cell Ca2+ traces from a population of cells can unequivocally identify the type of drug used to stimulate these cells. EXPERIMENTAL SECTION Strains and Media. Chinese hamster ovarian (CHO) cells were stably transfected to express chemokine (C-C motif) receptor 5 (CCR5), formyl peptide receptor-like 2 (FPRL2), and ChemerinR23 (ChemR23) receptor and were a generous gift from Dr GilbertVassart(Universite´ LibredeBruxelles,Brussels,Belgium).26-28. Dulbecco’s modified medium (DMEM), fetal bovine serum (FBS), trypsin, L-glutamine, phosphate-buffered saline (PBS), Zeocin, penicillin-streptomycin-neomycin (PSN) antibiotic mixture, Oregon Green Bapta-1, Influx pinocytic cell-loading reagent, and Calcein AM were all purchased from Invitrogen (Carlsbad, CA). F2L peptide and WKYMVm-peptide were purchased from Phoenix Pharmaceuticals (Burlingame, CA), hChemerin, Mip1Alpha, Mip1Beta, and RANTES were purchased from R&D Systems Inc. (Minneapolis, MN). Sodium bicarbonate was purchased from Sigma Aldrich (St. Louis, MO). Hydrochloric acid was obtained from Fisher Scientific (Hampton, NH). The cells were grown in DMEM containing 10% FBS, 2 mM L-glutamine, 1.2 ng/mL NaCO3H, 100 µg/mL penicillin, 100 µg/ mL streptomycin, 200 µg/mL neomycin, and 250 µg/mL Zeocin at 37 °C, 5% CO2 in a Cole Parmer 39200-00 incubator (Sheldon Manufacturing, Cornelius, OR). The cell lines were subcultured every 2-3 days. The day before the fiber-optic experiments, the cells were trypsinized using trypsin/EDTA (0.05% trypsin) for 5 min at 37 °C, diluted in growth medium, and spun for 10 min at 1200 rpm in a Sorvall RC-3 centrifuge (Sorvall, Asheville, NC). Cells were then resuspended in growth medium to yield a concentration of ∼5 × 105 cells/mL before being placed on the fiber-optic array. Fiber-Based Living Cell Array Fabrication. A fiber bundle containing ∼2000 optical fibers, with each fiber 22 µm in diameter (Schott, Elmsford, NY), was polished using a series of lapping films (30, 12, 8, 3, 1, and 0.3 µm; Mark V Laboratories, East Granby, CT) and an Ultrapol fiber polisher (Ultra Tec Inc., Santa Ana, CA) and etched using 0.025 M HCl for 17 min to create wells 15 µm deep.29 The resulting fibers were subsequently sonicated in DI H2O for 45 s to remove residual debris from the etching (25) Hodder, P.; Mull, R.; Cassaday, J.; Berry, K.; Strulovici, B. J. Biomol. Screening 2004, 9, 417-426. (26) Blanpain, C.; Libert, F.; Vassart, G.; Parmentier, M. Recept. Channels 2002, 8, 19-31. (27) Migeotte, I.; Riboldi, E.; Franssen, J.-D.; Gregoire, F.; Loison, C.; Wittamer, V.; Detheux, M.; Robberecht, P.; Costagliola, S.; Vassart, G.; Sozzani, S.; Parmentier, M.; Communi, D. J. Exp. Med. 2005, 201, 83-93. (28) Wittamer, V.; Franssen, J.-D.; Vulcano, M.; Mirjolet, J.-F.; Le, Poul, E.; Migeotte, I.; Brezillon, S.; Tyldesley, R.; Blanpain, C.; Detheux, M.; Mantovani, A.; Sozzani, S.; Vassart, G.; Parmentier, M.; Communi, D. J. Exp. Med. 2003, 198, 977-985. (29) Taylor, L. C.; Walt, D. R. Anal. Biochem. 2000, 278, 132-142.

9046

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

process. The distal end of the fiber bundle was modified by attaching a 1-in. plastic sleeve (o.d. 1/16 in., i.d. 1/32 in.) (Smallparts Inc., Miami Lakes, FL) to act as a vessel for the cell suspension. Before placing the cells on the fibers, the fibers were coated with fibronectin (10 µg/mL) for 35 min and washed in PBS. A 30-µL cell suspension was then placed in the plastic vessel, and the cells were incubated vertically overnight at 37 °C in 5% CO2 to allow the cells to fully adhere to the wells. In this position, the cells settle in the wells preferentially due to gravitational forces.29 The day of the experiment, an Influx pinocytic cell-loading reagent was prepared according to the protocol30 and diluted with Oregon Green Bapta-1-dextran (10 000 MW 5 mg/mL). The dye was added to the loading reagent in a 1:4 ratio. The cells on the fiber were then loaded with the dextran-bound dye using a protocol adapted from the original loading protocol for the reagent.30 The loading protocol was as follows: the cells were incubated with the dye-loading reagent solution described above for 20 min at 37 °C, the solution was then removed, and the cells were incubated with a lysis medium (serum-free medium and DI H2O in a 3:2 ratio) for 2 min. After removal of the lysis medium, the cells were allowed to recover for 20 min in fresh growth medium. The loading procedure was then repeated. The loading medium allows the cell-impermeable dye to be loaded into the cells through polymer vesicles, which are taken into the cells through endocytosis. The vesicles then burst and release the dye in the cytosol when the cells are incubated with hypotonic lysis medium.30 The cells are only incubated with the lysis medium long enough for the polymer vesicles to burst without affecting the cell membrane. Fiber-Optic Imaging Experiments. Upon dye loading, the fiber was placed on an inverted microscope (model IX81, Olympus America, Melville, NY). The growth medium above the cells was removed and replaced with 5 µL of pure DMEM (serum-free medium). DMEM was also used to dilute the agonists used in the experiments. Three background images were taken using a CCD camera (Orca-ER, Hamamatsu, Japan) at wavelengths of 494 and 530 nm for excitation and emission, respectively, with a 900ms exposure (filter set 31003, Chroma Technology Corp., Rockingham, VT). The cells expressing the different receptors were exposed to different agonists; for the promiscuous receptors CCR5 and FPRL2, different types of agonists were presented at both high and low concentrations; for ChemR23, the chemerin agonist was presented at three different concentrations (Table 1). The ability to exhibit intracellular Ca2+ increase was tested in the cells by exposure to 80 µM ATP. The baseline and nonspecific Ca2+ oscillations were determined using serum-free medium instead of agonists. The cells were stimulated by adding 5 µL of the agonist, ATP, or a nonstimulating compound to the fiber bundle. An image was taken every 3 s for 4 min using the parameters described for the background images. The cells were washed three times in DMEM after exposure to a compound, and the process was repeated for the next compound. When several compounds were used, as for the promiscuous receptors CCR5 and FPRL2,27,31 the order for presenting the compounds was varied between fibers to counteract any bias in the measurement. After stimulating the cells with all the relevant compounds, an on-fiber assay was performed to determine if the cells were alive by (30) http://probes.invitrogen.com/media/pis/mp14400.pdf, 2001. (31) Oppermann, M. Cell. Signalling 2004, 16, 1201-1210.

Table 1. Receptors Expressed in Separate CHO Cell Lines, and Working Concentrations of Compounds Used in Real-Time Ca2+ Measurementsa receptor receptor/agonist

CCR5

Mip1Beta H ) 100 nM Mip1Beta L ) 10 nM Mip1Alpha H ) 250 nM Mip1Alpha L ) 25 nM RANTES H ) 100 nM RANTES L ) 10 nM F2L H ) 400 nM F2L L ) 40 nM W-peptide H ) 5 µM W-peptide L ) 500 nM hChemerin H ) 100 nM hChemerin M ) 50 nM hChemerin L ) 25 nM medium ATP ) 80 µM

+ + + + + +

FPRL2

ChemR23

+ + + + + + + +

+

+

a Plus signs indicate cells responded to the compound, with H, M, and L indicating high, medium, and low concentrations. The order of the compounds was varied from experiment to experiment. Different concentrations of the same compounds were added sequentially.

introducing 10 µL of 4 µM Calcein AM to the fiber bundle and incubating for 40 min. Calcein AM is cell permeable and is cleaved by esterases inside the cells to yield a fluorescent product if the cell is alive.32 Live cells were identified by detecting an increase in fluorescence over time using the same filter cube as used for the Oregon Green dye. Dead cells do not exhibit an increase in fluorescence and were omitted from the data analysis. Fiber experiments that did not exhibit any fluorescence increase when ATP was added, or did not exhibit a significant fluorescence increase when incubated with the Calcein AM dye, were excluded from further analysis. Image processing was performed using IP Lab Imaging analysis software (Scanalytics, Fairfax, VA). Microtiter Plate Experiments. The microtiter plate experiments were used to test the dye’s ability to detect Ca2+ spikes, receptor activation, and the cells’ ability to increase intracellular Ca2+. A Tecan Infinite 200 Series (Tecan, Zurich, Switzerland) microtiter plate reader was used for fluorescence measurement using wavelengths of 494 and 525 nm for the excitation and emission, respectively. In these experiments, the Oregon Green Bapta 1 dye was used in the acetoxymethyl ester form instead of the dextran-bound form used in the fiber-optic experiments. This substitution was due to the more facile loading of the dye in the acetoxymethyl ester form when experiments were conducted in microtiter plates. The cells were trypsinized using trypsin/EDTA (0.05% trypsin) for 5 min at 37 °C, diluted in growth medium, and spun for 10 min at 1200 rpm in a Sorvall RC-3 centrifuge. Cells were then resuspended in growth medium to yield a concentration of ∼1 × 106 cells/mL and seeded at 20 000 cells/well in a 96-well flat bottom transparent cell culture treated microtiter plate (Fisher Scientific). The cells were incubated overnight at 37 °C with 5% CO2; 100 µL of 10 µM Oregon Green Bapta-1 was added to the wells followed by 45-min incubation at room temperature. The cells were then washed in PBS, and 100 µL of PBS was added to (32) Oral, H. B.; George, A. J. T.; Haskard, D. O. Endothelium 1998, 6, 143151.

the wells. A total of 100 µL of agonist in PBS was then added to the cells using the injector mode on the microtiter plate reader. The agonist concentrations used for these experiments were the highest concentrations listed in Table 1. The fluorescence signal was measured every second for 2 min after injection. The microtiter plate reader was set to read each well three times, with 70-µs integration time, and a manual gain of 140. The cells were also tested for Ca2+ increase in response to ATP and serum-free medium using the same settings. RESULTS AND DISCUSSION Mammalian single-cell drug screening was performed by placing single cells in etched wells (22 µm in diameter and 15 µm deep) at the distal end of a fiber-optic bundle. Well size and cell concentration were used to ensure that only one cell was present in each well. We employed three different cells lines, ectopically expressing a human GPCR receptor, as model systems when developing this method. The three different cell lines were CHOcellsexpressingeitherCCR5,FPRL2,orChemR23receptors.26-28,31 Each cell line was determined to respond only to specific agonists (Table 1). Agonists at different concentrations were used to stimulate the cells (Table 1), and upon agonist stimulation, Ca2+ measurements were recorded as individual cell fluorescence intensities over time using a Ca2+-sensitive dye, a fluorescent microscope, and a CCD camera. We used image analysis software to extract time-resolved data depicting the Ca2+ modulations in single cells. The activation as a result of each agonist stimulus was determined, and the data were analyzed for outliers. We then investigated whether the fiber-optic method could provide agonistspecific results that were not obtainable when the bulk cell responses were analyzed. Promiscuous receptors, such as CCR5 and FPRL2 used here, are activated by more than one agonist. The Ca2+ transients resulting from stimulation of promiscuous receptors have previously been suggested to be agonist specific,23,33,34 and this specificity was investigated using both bulkand single-cell data from the fiber-optic arrays. In this work, we investigated whether mathematical modeling and pattern recognition algorithms could be used to analyze the kinetic Ca2+ traces to determine the nature of the stimulus used to activate the different receptors. Due to the stochastic nature of protein expression and the inherently large phenotypic variation in clonal populations, the cell response distribution can only be observed by employing single-cell analysis on a large number of clonal cells. The average results obtained from bulk analysis methods mask the variance and individual responses in clonal populations. Single-cell responses can also identify the response amplitude over several types of stimuli and identify outliers that are not visible in a bulk experiment.3,5,6,35-37 The method employed here observes individual cells upon exposure to different stimuli, which provides unique insight into how different agonists affect the same cell over time. Since the cells are stimulated serially within a short time (33) Savineau, J.-P.; Guibert, C.; Marthan, R. Lung Biol. Health Dis. 2005, 197, 167-183. (34) Sanchez-Bueno, A.; Cobbold, P. H. Biochem. J. 1993, 291, 169-172. (35) Stevens, J. L. Chem. Res. Toxicol. 2006, 19, 1393-1401. (36) Sigal, A.; Milo, R.; Cohen, A.; Geva-Zatorsky, N.; Klein, Y.; Liron, Y.; Rosenfeld, N.; Danon, T.; Perzov, N.; Alon, U. Nature 2006, 444, 643-646. (37) Cruz-Monteagudo, M.; Gonzalez-Diaz, H. Eur. J. Med. Chem. 2005, 40, 1030-1041.

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

9047

Figure 1. Example of single-cell results from one fiber array stimulated with three agonists. Background-subtracted intensity is on the left and normalized intensity is on the right. The images show a population of CCR5 cells stimulated with Mip1Beta (A), Mip1Alpha (B), and RANTES (C). Each trace on the graph represents one cell, and the intensity represents the Ca2+ concentration in the cell at each time point. The images show how the patterns change when the background-subtracted image is normalized. The normalization enables comparison of response patterns only and enables better comparison between fibers; 152 × 133 mm (266 × 266 DPI).

period, there is minimal variation in cellular protein- and receptor levels between stimulations. The spatial restrictions exercised on the cells confined to the wells have been shown to arrest cell proliferation and maintain them in a quiescent phase.38 Therefore, intercellular differences in protein levels due to the difference in cell cycle phase should not be significant.36,39 We monitored the receptor activation by acquiring an image every 3 s for 4 min starting immediately after the stimulus was added to the fiber. Ca2+ oscillations in cells are short-lived, and by taking images frequently, Ca2+ changes were captured with high temporal resolution. In all the analyses described here, only live cells were included; live cells were identified after the experiment was over using Calcein AM, a cell-permeable ester that is cleaved by cytoplasmic esterases, yielding a fluorescent product only in live cells. The background signal for each single (38) Chen, C. S.; Mrksich, M.; Huang, S.; Whitesides, G. M.; Ingber, D. E. Science 1997, 276, 1425-1428. (39) Sigal, A.; Milo, R.; Cohen, A.; Geva-Zatorsky, N.; Klein, Y.; Alaluf, I.; Swerdlin, N.; Perzov, N.; Danon, T.; Liron, Y.; Raveh, T.; Carpenter, A. E.; Lahav, G.; Alon, U. Nat. Methods 2006, 3, 525-531.

9048 Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

cell was determined by averaging three background images taken before the addition of any compounds. A cell was defined as a positive responder if the percent increase in fluorescence from one image to the next at any point was higher than the percent value of three times the standard deviation of the background for that cell. These increases from the previous images in the time series were used instead of the increase from an initial image (t0) in order to minimize and account for photobleaching. Photobleaching of the dye occurs throughout the experiment causing a downward drift in baseline levels,40 and this drift can vary from cell to cell. Table 1 shows the compounds and concentrations used to test the different cell lines and also shows the specific agonist that activated each cell line. The results from this receptor specificity testing were in accordance with previous publications.27,28,41 Figure 1 is an example of the single-cell data obtained (40) Bencic-Nagale, S.; Walt, D. R. Anal. Chem. 2005, 77, 6155-6162. (41) Blanpain, C.; Wittamer, V.; Vanderwinden, J.-M.; Boom, A.; Renneboog, B.; Lee, B.; Le, Poul, E.; El, Asmar, L.; Govaerts, C.; Vassart, G.; Doms, R. W.; Parmentier, M. J. Biol. Chem. 2001, 276, 23795-23804.

Table 2. Percent Responders and Standard Deviations (SD) for Each Receptor Responding to a Given Stimulusa CCR5 Receptor percent responders/SD

low responders medium responders high responders total responders

Mip1Beta

Mip1Alpha

RANTES

ATP

medium

20/10 13/8 1/1 34/7

23/7 13/9 1/1 38/8

19/8 33/10 6/9 58/20

7/4 9/5 69/7 84/6

4/5 0.3/0.6 0/0 4/6

FPRL2 Receptor percent responders/SD

low responders medium responders high responders total responders

F2L

W-peptide

ATP

medium

17/16 33/18 35/30 85/5

17/26 45/19 16/5 77/14

2/0 87 /14 8/9 97/5

0/0 0.3/0.4 0.1/0.2 0/1

ChemR23 Receptor percent responders/SD

low responders medium responders high responders total responders

Chemerin H

Chemerin M

Chemerin L

ATP

medium

12/21 28/13 44/41 84/25

34/23 6/4 5/9 43/12

7/3 5/3 0.5/0.7 12/2

12/15 78/17 0.5/0.7 90/10

9/8 3/3 0.5/0.9 13/4

a CCR5 and FPRL2 are given for the high concentration only. ChemR23 is given for three different concentrations. Abbreviations: Chemerin H, M, and L stands for Chemerin high, medium, and low concentrations. ATP stimulation was performed separately for the fibers with cells expressing CCR5 receptor.

from a fiber containing CCR5-expressing cells stimulated with three different agonists. Each live cell on the fiber is represented as a trace on the graph, and each time-point on each trace represents the Ca2+ concentration in the cell at that time. The figure shows both the background-subtracted data and the corresponding normalized data. The normalization process and the utility of the normalized data are described below. Control experiments included testing the cells for crossreaction with the cytokines used for the other cell lines, measuring baseline Ca2+ oscillations when only medium was added, and measuring cellular Ca2+ upon ATP stimulation. Cross-reactions were determined to be equal to or less than the response measured when adding just medium (data not shown), and therefore, these cross-reaction experiments were not included in the series of experiments performed on each fiber. Cross-reaction controls were not conducted on the fibers because many experiments were performed on each fiber, and in order to minimize the decrease in signal response due to photobleaching of the dye40 and possible desensitization of the receptor,9,42 the number of experiments on each fiber was kept to a minimum. ATP stimulation was performed on each fiber to assess the cells’ ability to increase intracellular Ca2+. ATP increases the intracellular Ca2+ through ATP receptors rather than through ligand binding of the cytokine receptors, so we expected to see all live cells exhibit an intracellular increase of Ca2+ upon ATP stimulation.28 (42) Freedman, N. J.; Lefkowitz, R. J. Recent Prog. Horm. Res. 1996, 51, 319353.

Determination of Single-Cell Receptor Activation. After extracting single-cell data from the acquired images, we determined the percentage of responders resulting from agonist stimulation and how this response varied with agonist concentration. The cells were divided into low (cutoff plus 0-1%), medium (cutoff plus 1-5%), and high (greater than cutoff plus 5%) responders. Table 2 shows the percentage of cells responding to each stimulus at high concentrations for the CCR5 and FPRL2 and all concentrations for ChemR23 receptors. Generally, it was observed that only some live cells in each population responded to the different stimuli and that ATP stimulation caused the highest total percentage of the population to respond. The only exception to this general observation was the ChemR23 receptor, where there was no significant difference in the total number of responders between high concentration of hChemerin and ATP. For CCR5 and FPRL2, a similar number of cells responded to different concentrations of stimuli, while for ChemR23 receptor, the total number of cells increased with increasing concentration of stimuli. These results demonstrate that, for these cells, only a subset of the cell population responds to stimuli, and in some cases (ChemR23 here), this subset varies with concentration. The single-cell analysis provides the ability to distinguish between two very different scenarios: a high response caused by a few high-responding cells or by many cells responding with similar intensities. The first scenario can be indicative of a ligand generating significant outliers from the general cell population, while the second scenario would be outlier free. Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

9049

Table 3. Outliers in Populations Expressing CCR5 (A), FPRL2 (B), and ChemR23 (C)a % of responding cells classified as outliers upper threshold - Std

% of responding cells classified as outliers lower threshold - Std

Mip1Beta H Mip1Beta L Mip1Alpha H Mip1Alpha L RANTES H RANTES L

(A) CCR5 9.0-3.4 6.3-3.0 6.6-5.4 4.0-4.0 8.8-3.8 5.8-5.1

4.9-0.3 7.6-5.7 4.6-1.7 2.7-2.6 3.3-3.0 3.2-2.1

F2L H F2L L W-peptide H W-peptide L

(B) FPRL2 5.2-2.1 9.5-14.9 6.8-4.5 3.2-2.5

8.2-7.3 1.5-1.5 3.9-3.2 1.4-1.5

(C) ChemR23 11.8-9.7 11.1-6.8 13.7-7.9

1.4-1.5 1.3-1.0 3.5-3.3

hChemerin H hChemerin M hChemerin L

a The outliers were calculated as being above or below the set threshold (1.5 IQD ( upper/lower quartile). The results are summarized from four different fibers. Abbreviations H, M, and L stand for high, medium, and low concentrations.

Single-Cell Outlier Detection. After investigating population responses to the various stimuli, outliers in the same cell population were examined. Using conventional statistical methods, we mapped the outliers caused by each agonist at different concentrations (Table 3). We used the interquartile distance (IQD) and defined outliers as cells responding outside (1.5 times IQD over the upper/lower quartile mark of the population.43 The CHO cells employed in these experiments were transfected with human receptors expressed ectopically, and to simulate these receptors, we used their natural ligands with the expectation that they would exhibit a low outlier percentage.27,28 As expected, the percentage of outliers from positively responding cells barely exceeded 10%. The cells expressing CCR5 receptors showed a higher number of outliers at a higher agonist concentration while the two other cell lines did not show a correlation between concentration and the percentage of outliers. This result demonstrates that a high concentration of agonist does not necessarily result in a higher percentage of outliers. Outlier detection is important to identify potentially toxic effects of both new and existing compounds. The single-cell method provides data that can uncover such outliers. Outlier detection is normally performed between research subjects (e.g., humans or animals) in a study and can expose adverse reactions in a minority of the sample population. These extreme responses can cause unwanted side effects and toxicity.44,45 Identifying outliers in a population of clonal cells can also uncover extreme effects, although the toxicity caused by outliers on the single-cell level is not known and could vary from receptor to receptor. A large percentage of outliers indicates significant differences in responses between the cells, and a bulk measure(43) Crawshaw, J.; Chambers, J. A concise course in A-levels Statistics, 2rd ed.; Stanley Thornes Ltd.: London, 1999. (44) Edwards, I. R.; Aronson, J. K. Lancet 2000, 356, 1255-1259. (45) Rostami-Hodjegan, A.; Tucker, G. T. Nat. Rev. Drug Discovery 2007, 6, 140148.

9050 Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

Figure 2. Example of PCA plot of bulk analysis of four fibers analyzed on 1 day. The PCA plot and the corresponding KNN analysis show that the traces from the different agonists did not cluster well, and the bulk Ca2+ traces were not specific for the agonist used to stimulate the cell population; 169 × 135 mm (96 × 96 DPI).

ment will not be an accurate reflection of this response diversity. A dose-dependent response determined exclusively on the basis of the bulk response could also be inaccurate.8,46 The outliers in cells with ectopically expressed receptors, such as the receptors used here, will not necessarily be identical to outliers seen with receptors naturally expressed on their original host cells,6 but high outlier numbers in transfected cells could still uncover unexpected drug reactions. Pattern Recognition Using Principal Component Analysis (PCA). The activation of different pathways and the specificity of Ca2+ transients when stimulating promiscuous receptors with different agonists have been described in the literature.23,27,31,33,34 We wanted to investigate whether the patterns of Ca2+ oscillations occurring as a result of receptor activation were specific for each agonist and if these different conditions could be distinguished using the kinetic responses of the Ca2+ traces from a population. To compare single-cell and bulk-cell analyses, each fiber was first analyzed as if it were a bulk-cell analysis system. The obtained bulk Ca2+ traces were tested for agonist specificity as described below. It was determined that these responses did not demonstrate agonist specificity, so we proceeded to analyze the single-cell responses with respect to specific response patterns for each agonist. When investigating the Ca2+ kinetics, the mean fluorescence intensity of the entire fiber array at each time point was used for the bulk analysis, and for the single-cell analysis, the fluorescence intensity for each cell at each time point was used. Only positively responding cells selected according to the abovementioned criteria were included in the single-cell analysis. The responses obtained when stimulating the cells with different agonists were clustered using the mathematical models described below. It was decided that, after having extracted all desired amplitude-based results from the data sets, the Ca2+ traces would be analyzed in the normalized form. By normalizing the responses, the cell responses are manipulated to have the same maximum amplitude. There are many reasons why clonal cells respond with different intensities such as the amount of dye loaded in each cell, cellular protein levels, and receptor expression (46) Valet, G. Drug Discovery Today 2006, 11, 785-791.

Figure 3. Example of PCA plot for cells expressing CCR5 stimulated with Mip1Alpha, Mip1Beta, and RANTES in high concentration. The time series were normalized, and three PCA values were calculated and plotted. Each data point represents one cell. The PCA plot to the left is from one fiber, while the PCA plot to the right is from pooling all the fibers on that day. KNN analyses from the same fibers are shown in Table 5; 127 × 68 mm (266 × 266 DPI).

levels.3,8 In order to focus on the actual response patterns of the cells rather than the response intensities, the normalized traces were used to determine if the Ca2+ responses were agonist specific. Figure 1, as mentioned above, illustrates how the patterns change when the background-subtracted cell traces are normalized. The patterns are exaggerated, and are no longer a good visual representation of how the cell population behaves, but are a mathematical representation of the response pattern.47The intensities over time for the bulk fiber response or for the single positively responding cells were normalized according to the following formula:

Ix ) In - I0 | Ix | Imax ) Max (| Ix |) Inormalized ) Ix/Imax where Ix is the background-corrected intensity obtained by subtracting the intensity at time t ) 0 (I0 at t0) from the intensities at every time point In in the time series. Imax is the maximum absolute value from the background-corrected time series (| Ix |). The normalized matrix of Ca2+ traces was imported into MATLAB (Mathworks, Natick, MA), and PCA was performed on the data set using the PLS Toolbox (Eigenvector Research, Inc., Wenatchee, WA). PCA analysis reduces the number of variables while retaining most of the variance in the data set. PCA is an efficient method for compressing data and provides excellent graphical representation. In order to cluster the different classes of data, the analysis weights the time points where there is more variance between the data series. The data are then recalculated into a number of principal components (PCs), where the first PC represents the largest variance in the data, the second PC represents less variance, and so on.47 Each data set we used (47) Bakken, G. A.; Jurs, P. C. J. Med. Chem. 2000, 43, 4534-4541.

contained the Ca2+ kinetic trace responses of the cell population to the different agonist stimuli, as well as their responses to ATP and medium. In most cases, single-cell data sets clustered based on three or four PCs, with three PCs on average reflecting 98% of the covariance in the dataset. Using the normalized Ca2+ traces and PCA analysis, data sets obtained when the cells were stimulated with different agonists were clustered. The bulk fiber responses showed no clustering on the PCA plot (Figure 2). For the single-cells data, Ca2+ responses from cells expressing CCR5 and FPRL2 stimulated with high concentrations of agonist clustered well, with most of the Ca2+ traces identified as the correct agonist by simply setting a cutoff value for each PC (Figure 3). Pattern Recognition Using K-Nearest Neighbor (KNN) Analysis. We next explored whether we could use the singlecell Ca2+ responses to correctly identify unknown compounds. For compound identification, we employed KNN classification in the Weka machine learning program (University of Waikato, Hamilton, New Zealand). KNN employs Euclidian distances at each time point to compare the time-series data. Again, we normalized each time series before using KNN. As mentioned above, this procedure eliminates the amplitude differences of the Ca2+ traces for each agonist (each class), ensuring that the clustering is based on the patterns of responses and not on the amplitude of responses. We used 10-fold cross-validation when performing the analysis. In 10-fold cross-validation, the data are partitioned into 10 blocks: 9 of the blocks in each data set are used to build the model, and then each single trace of the remaining block is fed into the model one by one as an unknown and assigned a class. This procedure is repeated 10 times, with a different block left out of the model each time, and so each trace is at one point fed into the model as an unknown. Class assignment for each time series in KNN is based on the closest similarity to a specific class. The overall class for the time series is assigned as the most common class over all the time points. The number of voting neighbors was K ) 3, meaning the trace was assigned by the Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

9051

Table 4. Example of KNN Classification When Bulk Results from Four Fibers with CCR5 Expressing Cells Were Analyzed Using KNN Algorithma classified as Actual Mip1Betay Mip1Alpha RANTES

Mip1Beta

Mip1Alpha

RANTES

1 4 3

3 0 1

0 0 0

correctly classified instances: 1, 10% incorrectly classified instances: 11, 90% a Most of the traces were identified as the same (Mip1Beta), Thus, the classification was unsuccessful. The results show that bulk data from the cells were not specific for each agonist.

Table 5. Examples of Confusion Matrices Obtained from the CCR5 (A, B) and FPRL2 (C) Cell Lines after Performing KNN Pattern Recognition on Ca2+ Traces Obtained Using High Concentration of Agonista (A) CCR5 Receptor, Single Cell, 1 Fiber classified as actual Mip1Beta Mip1Alpha RANTES

Mip1Beta

Mip1Alpha

RANTES

101 3 6

3 123 6

4 3 104

correctly classified instances: 328, 93% incorrectly classified instances: 25, 7% (B) CCR5 Receptor, Single Cell, 3 Fibers classified as

majority of its three nearest neighbors.48 Table 4 shows a confusion matrix obtained when traces from bulk-cell analysis were fed into the model. The majority of traces were not correctly classified; most of the traces were identified as resulting from the same type of stimuli. Sections A-C in Table 5 show examples of the confusion matrices obtained from single-cell analysis of CCR5 and FPRL2 cell lines. The KNN analyses performed on the cell traces showed that the single-cell results could unequivocally identify the agonist used to stimulate the cells. Even though between 7 and 12% of the cell traces were incorrectly classified by the model, the majority of the traces were correctly classified, providing a high level of certainty to the agonist identity. The pattern recognition of single-cell Ca2+ traces can provide correct classification of the agonist used to stimulate a cell population because the majority of single-cell Ca2+ traces are correctly classified. When the majority of single-cell traces are correctly classified, a classification error of the population as a whole is unlikely to be made. The confidence in these measurements arises from the high number of cells that can be simultaneously interrogated and the large amount of information contained in the single-cell data. Single-cell data contain higher quality information because nonresponding cells can be removed from the analysis, and each single-cell response can be investigated separately from other cell responses. In comparison, we demonstrated that the summed data from a bulk analysis of the same cell population do not contain enough information to provide this type of classification. Bulk traces were not correctly classified by the method, and the method could not distinguish the bulk traces resulting from different stimuli from each other. The overall results when classifying single-cell traces from different cell lines are shown in Table 6. Pattern recognition analysis was performed over several fiber experiments (minimum 3), to validate the reproducibility of these agonist-specific patterns. Reproducibility when clustering data from experiments performed on different days was also investigated. The classification correctness for multiday pattern recognition was reduced and ranged from 64 to 90%. The results from the pattern recognition analysis demonstrate that, for the receptors used in this analysis, the binding of ligand (48) Witten, I. H.; Frank, E. Data mining: Practical machine learning tools and techniques, 2nd ed.; Morgan Kaufmann: San Francisco, 2005.

9052 Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

actual Mip1Beta Mip1Alpha RANTES

Mip1Beta

Mip1Alpha

RANTES

192 17 15

23 210 21

14 3 293

correctly classified instances: 695, 88% incorrectly classified instances: 93, 12% (C) FPRL2 Receptor, Single Cell, 3 Fibers classified as actual F2L peptide W-peptide

F2L peptide

W-peptide

869 84

59 749

correctly classified instances: 1618, 91.9% incorrectly classified nstances: 143, 8.1% a (A) shows results from one fiber containing CCR5 expressing cells. (B) shows results from three pooled fibers, including the fiber in (A), and (C) shows three fibers pooled containing FPRL2 expressing cells. The KNN analyses returned classifications that enabled identification of the agonist used to stimulate the different cell lines with certainty. The percent correct classification was reduced when several fibers were pooled (B, C) but was still unmistakable. KNN classification was also done using Ca2+ traces from fiber experiments performed on different days, and the percent correct classification for these experiments ranged from 64 to 90%.

to receptor and the subsequent increase of cytosolic Ca2+ follows a specific pattern over time that in a clonal population is unique for each agonist. As explained above, the ability to obtain a high number of single-cell responses using a fiber-optic sensor array results in data that are statistically significant and return distinct population classification. Previous reports have suggested that intracellular Ca2+ transients in GPCRs are agonist specific.33,34 In addition, Ca2+ monitoring experiments have demonstrated differences in Ca2+ oscillation frequencies when GPCR receptors were stimulated with different agonists.49,50 Agonist-specific GPCR responses can be monitored via intracellular Ca2+ modulation, as shown here; however, determining the underlying mechanisms (49) Breitwieser, G. E. Curr. Top. Dev. Biol. 2006, 73, 85-114. (50) Hamada, H.; Damron, D. S.; Murray, P. A. Anesthesiology 1997, 87, 900907.

Table 6. Majority Classification of Promiscuous Receptors CCR5 and FPRL2 Using a KNN Pattern Recognition Modela CCR5 and FPRL2 - Majority classification - All Fibers classified as

actual agonist receptor

Mip1Beta CCR5

Mip1Beta/CCR5 Mip1Alpha/CCR5 RANTES/CCR5 F2L peptide/FPRL2 W-peptide/FPRL2

+

Mip1Alpha CCR5

RANTES CCR5

F2L peptide FPRL2

W-peptide/ FPRL2

+ + + +

a This table shows that even if a minority of the Ca2+ traces were incorrectly identified by the model, the high number of traces used in the analysis provided the correct overall classification for the population. Agonist stimulation of CCR5 receptor and FPRL2 receptor were done on separate fibers.

usually requires collecting additional information using other techniques.51 Immunoprecipitation, FRET, FCS, and gene expression techniques are frequently used to determine the events involved in the receptor response.21,51-53 Inhibition of specific pathways has been used to determine agonist-specific activation.54,55 While the underlying mechanisms for agonist-specific responses are not identified with the method described here, the results clearly demonstrate our ability to detect and distinguish between agonist-dependent differences in dynamics of the secondary messenger Ca2+. Finally, clustering different concentrations of the same agonist was also possible using this method; however, the different concentrations would frequently be classified into separate classes. Thus, for best results, the pattern recognition analyses should be performed ideally at high concentration of agonist where a maximum stimulus is reached. This concentration should be sufficiently above the EC50 for the specific agonist. CONCLUSION A fiber-optic single-cell array containing mammalian cells expressing different cytokine receptors was demonstrated. The data extraction method employs a number of algorithms to extract information from real-time Ca2+ traces obtained simultaneously from hundreds of single cells. We demonstrated that only a subset of all live cells in a population responds to stimuli. Furthermore, we found that the natural ligands used to stimulate the receptors in these experiments did not generate an unexpectedly large percentage of outliers in responding cells, nor was there a general correlation between concentration and the number of outliers. The single-cell method can be used to demonstrate potential toxic (51) Kiselyov, K.; Shin, D. M.; Muallem, S. Cell. Signalling 2003, 15, 243-253. (52) Daaka, Y.; Luttrell, L. M.; Lefkowitz, R. J. Nature 1997, 390, 88-91. (53) Minke, B.; Cook, B. Physiol. Rev. 2002, 82, 429-472. (54) Zeng, W.; Xu, X.; Muallem, S. J. Biol. Chem. 1996, 271, 18520-18526. (55) Liu, Z.; Geng, L.; Li, R.; He, X.; Zheng, J. Q.; Xie, Z. J. Neurosci. 2003, 23, 4156-4163.

effects of drugs by exposing outliers and to uncover clonal response differences that would otherwise be masked in a bulk analysis. When applying pattern recognition algorithms, including both KNN and PCA, it was demonstrated that the Ca2+ traces from cells expressing a promiscuous receptor (CCR5 or FPRL2) were specific for each agonist used and were clearly distinguishable from patterns generated from other agonists used to activate the same receptor. The Ca2+ traces may be used to identify different activation mechanisms for different drugs binding to promiscuous receptors. Comparing the proximity of a new drug candidate’s Ca2+ traces to other known drugs could also indicate similarities or differences in activation mechanisms. The singlecell mammalian drug screening system employs a conventional fluorescence microscope adapted to hold optical fibers and should therefore be available in most laboratories. Using this fiber-opticbased sensor, several experiments can be performed serially on the same cells, and each cell can be monitored throughout the experiment. This method provides unique receptor activation data that enables monitoring of each cell’s response to different molecular challenges. Single-cell data provide the full range of complex responses from a clonal population of cells rather than the ensemble average obtained from microtiter wells or cuvettes. The method presented here may be useful in the development of new drugs and in the characterization of existing drugs. ACKNOWLEDGMENT We thank Dr. Gilbert Vassart for kindly providing us with the CHO cells expressing human GPCR receptors.

Received for review August 17, 2007. Accepted September 19, 2007. AC701744X

Analytical Chemistry, Vol. 79, No. 23, December 1, 2007

9053