Online Fluorescence Suppression in Modulated Raman Spectroscopy

Dec 17, 2009 - SUPA, School of Physics and Astronomy, University of St Andrews, North .... of the Raman signal for online suppression of the fluoresce...
0 downloads 0 Views 3MB Size
Download PDF
Anal. Chem. 2010, 82, 738–745

Online Fluorescence Suppression in Modulated Raman Spectroscopy Anna Chiara De Luca,† Michael Mazilu,† Andrew Riches,‡ C. Simon Herrington,‡ and Kishan Dholakia*,† SUPA, School of Physics and Astronomy, University of St Andrews, North Haugh, St. Andrews, Fife, KY16 9SS, United Kingdom, and Bute Medical School, University of St Andrews, St. Andrews, Fife, KY16 9TS, United Kingdom Label-free chemical characterization of single cells is an important aim for biomedical research. Standard Raman spectroscopy provides intrinsic biochemical markers for noninvasive analysis of biological samples but is often hindered by the presence of fluorescence background. In this paper, we present an innovative modulated Raman spectroscopy technique to filter out the Raman spectra from the fluorescence background. The method is based on the principle that the fluorescence background does not change whereas the Raman scattering is shifted by the periodical modulation of the laser wavelength. Exploiting this physical property and importantly the multichannel lock-in detection of the Raman signal, the modulation technique fulfills the requirements of an effective fluorescence subtraction method. Indeed, once the synchronization and calibration procedure is performed, minimal user intervention is required, making the method online and less time-consuming than the other fluorescent suppression methods. We analyze the modulated Raman signal and shifted excitation Raman difference spectroscopy (SERDS) signal of 2 µm-sized polystyrene beads suspended in a solution of fluorescent dye as a function of modulation rate. We show that the signal-to-noise ratio of the modulated Raman spectra at the highest modulation rate is 3 times higher than the SERDS one. To finally evaluate the real benefits of the modulated Raman spectroscopy, we apply our technique to Chinese hamster ovary cells (CHO). Specifically, by analyzing separate spectra from the membrane, cytoplasm, and nucleus of CHO cells, we demonstrate the ability of this method to obtain localized sensitive chemical information from cells, away from the interfering fluorescence background. In particular, statistical analysis of the Raman data and classification using PCA (principal component analysis) indicate that our method allows us to distinguish between different cell locations with higher sensitivity and specificity, avoiding potential misinterpretation of the data obtained using standard background procedures. When a laser beam is incident upon a sample, a small proportion of the incident photons are inelastically scattered, * To whom correspondence should be st-andrews.ac.uk. † SUPA, School of Physics and Astronomy. ‡ Bute Medical School.

738

addressed.

E-mail:

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

kd1@

resulting in an optical frequency shift. Raman spectroscopy is based on the detection of these inelastically scattered photons. The frequency difference between the incident and scattered photons gives information on the internal vibration modes of the illuminated molecules. Therefore, the Raman spectrum constitutes an intrinsic molecular fingerprint of the investigated sample, revealing detailed information on its chemical composition and structural conformation. Recent years have seen an increasing interest in the use of Raman spectroscopy for biomedical applications such as discrimination between various cellular components,1 detection of biochemical changes related to the cell cycle,2,3 identification of normal and diseased cells,4,5 and diagnosis of cancer.6 This technique shows multiple and complementary advantages when compared to other optical techniques such as fluorescence or confocal microscopy normally used to study biochemical properties of living cells. Indeed, Raman spectroscopy is noninvasive and nondestructive such that the samples can be maintained under appropriate physiological conditions, not requiring the addition of labels or chemical agents for the sample identification. Additionally, it gives insight into the chemical processes in single cells. Although Raman spectroscopy is a useful technique to identify and quantify species in a given matrix, it has been severely limited in its applicability by fluorescence. Spectrally, this fluorescence occurs at the same wavelength as the Raman signal and is often several orders of magnitude more intense than the signal from the weak chemical vibrations. Often, this fluorescence background and its natural variability makes biochemical analysis using Raman spectroscopy impractical. Numerous experimental methods have been introduced in order to extract the Raman information from the raw acquired spectrum and to completely or partially reject the fluorescence background.7-10 (1) Snook, R. D.; Harvey, T. J.; Faria, E. C.; Gardner, P. Integr. Biol. 2009, 1, 43–52. (2) Singh, G. P.; Volpe, G.; Creely, C. M.; Gro¨tsch, H.; Geli, I. M.; Petrov, D. J. Raman Spectrosc. 2006, 37, 858–864. (3) Singh, G.; Creely, C.; Volpe, G.; Gro ¨tsch, H.; Petrov, D. Anal. Chem. 2005, 77, 2564–2568. (4) De Luca, A. C.; Rusciano, G.; Ciancia, R.; Martinelli, V.; Pesce, G.; Rotoli, B.; Selvaggi, L.; Sasso, A. Opt. Express 2008, 16, 7943–7957. (5) De Jong, B. W. D.; Schut, T. C. B.; Maquelin, K.; Van Der Kwast, T.; Bangma, C. H.; Kok, D. J.; Puppels, G. J. Anal. Chem. 2006, 78, 7761– 7769. (6) Jess, P. R. T.; Mazilu, M.; Dholakia, K.; Riches, A. C.; Herrington, C. S. Int. J. Cancer 2009, 124, 376–380. (7) Genack, A. Z. Anal. Chem. 1984, 56, 2957–2960. (8) Stellman, C. M.; Bucholtz, F. Spectrochim. Acta, Part A 1998, 54, 1041– 1047. 10.1021/ac9026737  2010 American Chemical Society Published on Web 12/17/2009

Time resolved Raman spectroscopy is one of the most straightforward solutions to this problem.9,11 This method takes advantages of different lifetimes between Raman scattering and fluorescence: Raman scattering is almost instantaneous while fluorescence has a finite lifetime (typically ps-ns). Therefore, a dominant part of the fluorescence signal can be rejected from the raw signal by the use of a pulsed laser and limiting the signal collection time. However, the involved instruments, pulsed laser and detection system, are sophisticated and expensive. In addition, when the fluorescence lifetime is comparable to the laser pulse duration, the method is not effective, while for fast enough intervals the methods reveal photocollection problems. The polarization modulation technique is another conventional method for fluorescence rejection.12 The technique is based on the different polarization properties of the Raman and the fluorescence signals. In fact, the Raman lines are highly polarized, and the fluorescence is totally depolarized. In this experiment, two different spectra are acquired with the use of light polarized parallel and perpendicular to the plane of incidence. Therefore, it is only necessary to subtract the two spectra to obtain a Raman spectrum free from any fluorescence contribution. However, not all of the Raman features can be detected by the use of this method.12 An other innovative approach to fluorescence rejection in Raman spectroscopy is based on the use of the annular excitation beam profile.13 Using this novel beam shape, it is possible to demonstrate that the fluorescence from the annular excitation beam has a reduced generation and collection efficiency compared to the standard Gaussian beam. This helps both reduce the need for prolonged acquisition as well as reduce the overall noise of the Raman signal. However, this method allows only a partial removal of the fluorescence background. Recently, Rusciano et al. have demonstrated the ability to remove the background from Raman spectra by the use of a phasesensitive detection scheme.14,15 The method is based on the periodic modulation of the position of the analyzed object and single channel lock-in detection of the signal using a photomultiplier. To acquire the full spectra using a photomultiplier, it is necessary to scan the monochromator grating with a consequent increase of the acquisition time, making this method slow. Mathematical methods such as polynomial fit of the background are standard techniques for fluorescence rejection.16 However, such background-subtraction procedures can cause artifacts in the processed data that could affect the validity of the collected biochemical information. Finally, shifted excitation Raman difference spectroscopy (SERDS) corresponds to a very useful approach to reject the (9) Mandal, D.; Mizuno, M.; Tahara, T. J. Mol. Struct. 2005, 735, 189–195. (10) Hasegawa, T.; Nishijo, J.; Umemura, J. Chem. Phys. Lett. 2000, 317, 642– 646. (11) Campani, E.; Gorini, G.; Masetti, G. J. Phys. D: Appl. Phys. 1981, 14, 2189– 2197. (12) Angel, S. M.; Dearmond, M. K.; Hanck, K. W.; Wertz, D. W. Anal. Chem. 1984, 56, 3000–3001. (13) Cormack, I. G.; Mazilu, M.; Dholakia, K.; Herrington, C. S. Appl. Phys. Lett. 2007, 91, 0239031-0239033. (14) Rusciano, G.; De Luca, A. C.; Pesce, G.; Sasso, A. Appl. Phys. Lett. 2006, 89, 261116–261118. (15) Rusciano, G.; De Luca, A. C.; Pesce, G.; Sasso, A. Anal. Chem. 2007, 79, 3708–3715. (16) Lieber, C. A.; Mahadevan-Jansen, A. Appl. Spectrosc. 2003, 57, 1363–1367.

fluorescence in Raman spectroscopy.16-19 This technique takes advantage of the property that the use of two slightly different excitation frequencies induces a frequency shift of the Raman spectrum leaving unchanged the fluorescence background. Therefore, a difference Raman spectrum can be obtained by subtracting the two acquired spectra, while the fluorescence background is removed. The SERDS is the closest technique to our new presented method. However, the use of only two or a few excitation frequencies provide a poor performance in the retrieved Raman signal. A major advance would be a simple, widely applicable method to remove the background improving on the above-mentioned methods. In this paper, we present the theory and the implementation of a novel modulated Raman spectroscopy technique to filter out the Raman spectra from the fluorescence background by the modulation of the excitation wavelength. The wavelength position of the Raman peaks is directly linked to the excitation wavelength. Thus, changing the excitation wavelength shifts the Raman peaks while the fluorescence background remains essentially constant. Exploiting this physical property allows us to clearly distinguish between the Raman signal and the fluorescence background. Our method is related to classic shifted excitation Raman difference spectroscopy (SERDS) but incorporates two important novel elements: (i) the use of a continuously modulated excitation wavelength (∆ν = 60 GHz) and (ii) multichannel lock-in detection of the Raman signal for online suppression of the fluorescence background. Indeed, by synchronizing the detection with the modulated excitation wavelength (multichannel lock-in detection), it is possible to distinguish the Raman peaks from the background fluorescence signal. This gives online access to the differential spectra with minimal user intervention, making the method more practical and less time-consuming with respect to the standard methods. To illustrate the power of our technique, we used a sample of 2 µm-sized polystyrene beads suspended in a solution of dye and water, to form a fluorescent sample with a known Raman spectrum. In this case, the raw acquired spectrum is dominated by the fluorescence contribution, but the pure Raman spectrum of polystyrene can be easily observed by the use of our modulated Raman spectroscopy which completely removes the background fluorescence. We systematically studied the Raman signal and its associated noise showing their dependence on the modulation rate of the excitation wavelength. Our results demonstrate that the modulated Raman method significantly enhances the acquisition of the Raman spectrum with respect to the classic SERDS method for fluorescence suppression. Importantly, by analyzing separate spectra from the membrane, cytoplasm, and nucleus of Chinese hamster ovary cells (CHO), we are able to identify characteristic Raman features associated with DNA, protein, and lipid molecular vibrations for discriminating between different locations inside the cell, away from interfering fluorescence background. Finally, using the principal component analysis (PCA) for the interpretation and classification of the (17) Shreve, A. P.; Cherepy, N. J.; Mathies, R. A. Appl. Spectrosc. 1992, 46, 707–711. (18) Mosier-Boss, P. A.; Lieberman, S. H.; Newbery, R. Appl. Spectrosc. 1995, 49, 630–638. (19) McCain, S. T.; Willett, R. M.; Brady, D. J. Opt. Express 2008, 16, 10975– 10991.

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

739

Raman data, we demonstrate that our modulated Raman spectroscopy facilitates spectral assignment and increases detection sensitivity, opening the way for biomedical imaging. THEORY In the following, we consider multiple Raman spectra Sj(λi) measured at the discrete wavelength λi where the subscript j refers to the spectra number. Each individual Raman spectra Sj(λi) ) SF(λi) + SR(λi + ∆λj) can be seen as the superposition of two parts: the fluorescent part SF(λi) and the Raman peaks SR(λi + ∆λj) where ∆λj represents the Raman excitation wavelength shift of the corresponding spectrum number j. Here, the fluorescent part is assumed to be constant regardless of the excitation wavelength while the Raman peaks shift as a whole following the excitation wavelength. Mathematically, the individual Raman spectra can be represented as Sj(λi) ) SF(λi) + Mik(∆λj)SR(λk)

(1)

where we implied summation over repeating indices. The matrix Mik(∆λj) corresponds to a fractional shift matrix that depends on the wavelength shift ∆λj and is defined in the following way: Mik(∆λj) ) (∆λj - ∆λj)δi+∆λj,k + (∆λj - ∆λj)δi+∆λj,k

(2)

where ∆λ_j and ∆λj correspond, respectively, to the next smallest and next largest integer with respect to ∆λj. Here, we have generalized the standard integer shift matrix to account for fractional pixel wavelength shift. Indeed, the standard shift matrix corresponding to a shift of one “pixel” upward is given by a matrix with ones on the superdiagonal and zeros elsewhere (δi+1, k). We introduce the fractional shift matrix through the weighted superposition of the two integer shift operators. For example, a shift of 1.2 wavelength pixels gives the following shift matrix: Mik(1.2) ) 0.2δi+2,k + 0.8δi+1,k

(3)

that is 20% of a two pixel shift and 80% of a one pixel shift. Knowledge of the excitation wavelength shift allows us to count 2n unknowns (SF(λi) and SR(λk)) where n is the number of pixels or wavelength acquired on the CCD camera. Acquiring more than two spectra, each giving n equations, defines an over determined linear system that we solve using the least-squares procedure and retrieve at the same time the Raman and the fluorescence spectra. Conceptually, the least-squares fit corresponds to a lock-in procedure where the modulation reference is given by fractional wavelength shift operator Mik. From this retrieved Raman spectra, we deduce the modulated Raman spectra through Dk ) SR(λk) - SR(λk+1). EXPERIMENTAL SECTION Raman Microscope. In Figure 1 is shown a schematic of the experimental setup developed in the present work. Except for the laser, the setup used for the modulation technique is identical to 740

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

Figure 1. Schematic of our Raman microscope system. A tunable diode laser at 785 nm is used to excite Raman scattering. The laser beam is introduced into an inverted microscope through a high numerical aperture objective (100×). The scattering light from the sample is collected by the same objective and coupled into a spectrometer equipped with a cooled CCD camera. Abbreviation: M, mirror; L, lens; DM, dichroic mirror; NF, notch filter.

the standard Raman setup.20 The laser source was a tunable diode laser (Sacher Lasertechnik, Model no. TEC-520-0780-100, External Cavity Diode Laser in a Littman configuration, total tuning range 25 nm) centered at 785 nm and with a maximum power of 100mW. The 5 mm-diameter laser beam is passed through a bandpass filter (Semrock optic, Max line 785), to remove any spontaneous emission from the laser source, and reflected from a holographic notch filter (Tydex) into a homemade inverted microscope. A 100× objective lens (Olympus, oil immersion, NA ) 1.2) was used to focus the laser light on the sample and to collect the back scattered photons. The scattered signal from the sample was then filtered by the same notch filter and imaged into the monochromator (Triax 550, Jobin-Yvon). The spectrometer employed a 300 line/mm grating, blazed at 600 nm, and is equipped with a cooled CCD camera (Symphony, Jobin-Yvon, 1024 × 256 pixels) for detection of the Raman spectrum. The Raman signal was focused on the entrance slit (set at an aperture of 50 µm) of the monochromator by the use of a lens ( f ) 50 mm). The entrance slit aperture in combination with the objective defines a cylinder of examination in the focal plane with diameter ∼0.7 µm and depth of ∼1 µm. Spectral resolution of the system was around ∼2 cm-1. To allow observation of the sample, the light from a LED, focused on the sample by a 20× objective, was used to generate bright field images of the sample on a CCD camera. For cell-based studies, Chinese hamster ovary cells (CHO-K1) were used and were prepared as follows: CHO cells were cultured in T25 flasks at 37 °C and 5% CO2 in modified Eagles medium (MEM) with 10% fetal calf serum (Invitrogen) and 20 µg/mL streptomycin (Sigma, UK). The cells were routinely split three times a week. A day before the experiment, cells were plated on 35 mm glass-bottomed culture grade dishes (World Precision Instruments) to achieve 40-50% confluency. The cells were incubated at 37 °C in 5% CO2 for 24 h to allow cell attachment at the bottom of the glass dishes. (20) Jess, P. R. T.; Smith, D. W.; Mazilu, M.; Dholakia, K.; Riches, A. C.; Herrington, C. S. Int. J. Cancer 2007, 121, 2723–2728.

Figure 2. Voltage signals from the TTL electronic output of (a) the laser and of (b) the CCD camera. The duration between rising edges of the two triggers allows us to determine the precise delay between the laser modulation and the acquisition rate.

The experiment consists of acquiring many short duration spectra (acquisition time 100 ms) each one corresponding to a different known excitation wavelength while the tunable laser periodically modulates its wavelength (the modulation frequency is 0.4 Hz, and ∆ν = 60 GHz). There are two technological challenges in the experiment. One is the delay necessary to reinitialize the CCD camera between successive acquisitions. The second challenge is to precisely determine the excitation laser wavelength for each spectrum. This second problem is actually linked to the first one inasmuch as the initialization and transfer of the spectral data makes the two processes potentially lose synchronicity. In the following, we show how we resolved this synchronization issue, by the use of multichannel lock-in detection, to obtain the retrieved differential spectra online. Multichannel Lock-in Detection and Data Analysis. To determine the laser wavelength for each acquired spectrum, we use a “stroboscopic” technique, i.e., both the laser modulation and the acquisition run at their respective frequencies which in general are different. Typically the laser is modulated at 0.4 Hz. The CCD camera acquires spectra with an integration time of 100 ms and a delay between two consecutive acquisitions of ∼80 ms (acquisition rate is about 5 Hz). The difference between the two repetition rates enables the two systems to have a variable delay between each other. We determine this delay by measuring the duration between the CCD acquisition trigger and the laser modulation trigger. Figure 2 shows the voltage on these two trigger lines. After synchronizing the modulation to the acquisition, the next step is to calibrate the modulation with respect to the output wavelength of the Raman excitation laser. This ensures that we do not need to detect the excitation spectra when acquiring the Raman spectra. This is advantageous as the Rayleigh scattering from the sample can easily saturate the detector. Figure 3 shows the calibration curve linking the delay between the two triggers and the output wavelength. The red curve corresponds to a Fourier series interpolation that reduces the calibration data to just 20 Fourier coefficients. Additionally, the interpolation renders the calibration data more robust to noise.

Figure 3. Wavelength calibration data and its interpolation. The x axis corresponds to the relative delay between the two triggers normalized to the laser modulation period, and the y-axis corresponds to wavelength variation in CCD pixels.

Using this calibration data, it is now possible to uniquely and exactly associate an excitation wavelength to each Raman spectra. Once this is known, there are multiple approaches to retrieve the differential spectrum from the raw data. We have thoroughly studied, compared, and contrasted multiple approaches: Fourier filtering to retain the varying components, principal component analysis (PCA) to extract the largest variation of data, multipoint numerical differentiation, and least-squares fitting of overdetermined spectra. Comparing these methods on our present system, we find the latter is the most efficient and precise. Once the raw spectra are acquired, the retrieved Raman spectrum was obtained online by applying the least-squares fitting algorithm. For this purpose, a Labview program was written on the basis of a Mathematica algorithm used to calibrate and analyze the spectra. RESULTS AND DISCUSSION Demonstration of the Modulated Raman Method with Polystyrene Beads. In the present study, to characterize the performance of the modulated Raman (MR) method, we use 2 µm-sized polystyrene beads suspended in a solution of dye (Fluorescent Red near-infrared (NIR), Sigma, 1 × 10 -8 M) to form a fluorescent sample with a known Raman spectrum. Figure 4 shows the acquired Raman spectrum of a polystyrene bead (a) and the reference spectrum recorded without the polystyrene bead (b) in the spectral region between 300 and 2500 cm-1. The spectra are acquired with an acquisition time of 20 s, and the Raman laser power on the sample is ∼10 mW. Several Raman features due to the polystyrene are observed on the top of a broad fluorescence background. Therefore, the spectrum in Figure 4 is an ideal candidate to assess the effectiveness of the MR spectroscopy technique in rejecting fluorescence. The same spectral region is, therefore, investigated by modulating sinusoidally the laser wavelength (modulation frequency 0.4 Hz) and detecting the signal by the use of the multichannel lock-in system. We acquire 200 spectra with an acquisition time of 100 ms. These parameters are chosen to maximize the signalto-noise ratio, as it will be clearly shown later. In Figure 5, we show the retrieved polystyrene Raman (a) and the reference Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

741

Figure 4. (a) Raman signal of a 2 µm polystyrene bead, as acquired by the standard Raman spectroscopy. (b) Reference spectrum. The features are due to a solvent scattering contribution. Figure 6. (a) Modulated Raman spectrum of a 2 µm polystyrene bead, obtained by acquiring 200 spectra with an integration time of 100 ms and a modulation rate of 0.4 Hz. (b) The same modulated Raman spectrum acquired by the use of a modulation rate of 0.1 Hz and keeping the total acquisition time constant. (c) Shifted excitation Raman difference spectrum obtained by acquiring only two Raman spectra with an integration time of 100 s each at different laser wavelength. All the spectra are obtained by the use of a laser power on the sample of ∼1 mW. The horizontal red lines correspond to the noise standard deviation used to measure the signal-to-noise ratio.

Figure 5. Retrieved polystyrene Raman (a) and reference spectrum (b) obtained by the use of the modulation method. For the sake of clarity, this last signal was vertically shifted.

spectrum (b) acquired by the use of the modulation method. The spectra are acquired in the same experimental conditions (acquisition time and Raman probe power) as the spectra of Figure 4. Clearly in the retrieved spectra, the fluorescence is now completely removed, leaving only the derivative-like peaks due to the Raman contribution, testifying to the validity of the assumptions made earlier. As matter of fact, Figure 5a is clearer than the raw spectra of Figure 4a, and it is free of the fluorescence background affecting the raw spectra. The described measurements demonstrated that the MR spectroscopy can provide a sensitivity high enough to measure Raman spectra and that it can be applicable to the fluorescence rejection in Raman spectroscopy. Characterization of the Modulated Raman Technique and Comparison with the SERDS Method. To quantify the advantage of the MR spectroscopy, we acquired the modulated Raman spectra of a 2 µm sized polystyrene bead in a dye for different modulation rates of the laser wavelength. More specifically, we demonstrate the progressive improvement of the signal-to-noise ratio as the modulation rate increases from 0.03 to 0.4 Hz. The upper frequency value (0.4 Hz) is limited by the CCD acquisition rate (100-80 ms) while the lower frequency value (0.03 Hz) is limited by the number of modulation periods in the acquisition. 742

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

Figure 7. Signal-to-noise ratio of the polystyrene Raman peak at 1001 cm-1 as a function of the modulation rate of the laser wavelength, f. The green dot represents the limit case f ) 0.005 Hz that corresponds to the classic SERDS method.

In order to do a fair comparison between the different modulation rates, we adjusted the number of spectra acquired such that the total acquisition duration remains constantly 200s. Figure 6a,b show two modulated Raman spectra acquired with two different modulation rates. We estimated the signal-to-noise ratio by measuring the polystyrene Raman peak at 1001 cm-1 with respect to the standard deviation of the surrounding background. Direct comparison of these modulated Raman spectra demonstrates a strong increase of the signal-to-noise ratio for the higher modulation rate. Indeed, Figure 7 shows a linear relationship between modulation rate and signal-to-noise ratio consistent with 1/f noise behavior.

Figure 8. CHO cell image. The Raman probe location during the spectra acquisition is also shown.

Finally, we directly compared the modulation Raman with the SERDS method by acquiring two spectra with an integration time of 100 s, each having two different excitation wavelengths (∆ν = 60 GHz). By subtracting the two spectra, a derivative-like signal is obtained, see Figure 6c. As observed in Figure 6a,c, the SERDS spectrum presents a 3 times lower signal-to-noise ratio. This decreased signal-to-noise ratio is not only due to the 1/f noise and the advantage of the “multi-channel lock-in” technique in this context but also due to the continuous wavelength shift employed in our technique as opposed to the two excitation wavelengths in standard SERDS. As matter of fact, the modulation Raman method is more robust than the SERDS method for extracting relevant Raman features. Characterization of Single Living CHO Cells. To demonstrate the robustness of the MR method, we apply our technique to a real biological sample. Figure 8 shows an image of a CHO cell in aqueous medium, grown upon a glass substrate. The CHO cells assume a characteristic shape and size, typically between 15 and 30 µm in the x-y dimension and ∼3 µm thick, and the examination cylinder in the sample resulted in a diameter of ∼0.7 µm and a depth of ∼1 µm. It is easy to recognize the nucleus, surrounded by cytoplasm, and the membrane position. By the use of the standard Raman spectroscopy and our modulation method, we recorded separate spectra from the nucleus, cytoplasm, and membrane of CHO cell (12 spectra were taken for each point). The background spectra after the cell had been moved out of the Raman laser beam waist were also acquired. The resultant mean spectra, obtained by the use of standard Raman spectroscopy, are shown in Figure 9. Each spectrum was recorded in the range from 300 to 2500 cm-1 with an acquisition time of 200 s. The Raman spectra in Figure 9 show an intense fluorescence background that is due to intrinsic interference effects from the glass surface where the cells are attached. Clearly, using a standard Raman spectroscopy method, it was not possible to assign precisely the various chemical constituents in the cell. The corresponding retrieved mean spectra, from the nucleus, cytoplasm, and membrane of the cell, and the reference mean spectrum acquired by the use of our modulation method are reported in Figure 10. All the spectra are acquired in the same experimental conditions as the spectra of Figure 9 (we acquire 1000 spectra with an acquisition time of 200 ms). With respect to the standard Raman spectra, in the retrieved spectra, it is possible to observe a flat background and also the Raman peaks are more evident. In particular, observing the spectra in Figure 10, it is immediately noticeable that the Raman features from the three

Figure 9. Raman spectra obtained from three different positions within a CHO cell: nucleus, cytoplasm, and membrane. The spectra are characterized by the presence of a strong background that does not allow us to discern the Raman features in more detail. The background spectrum acquired after the cell had been moved out of the Raman laser beam waist is also shown.

Figure 10. Raman spectra obtained from three different positions within a CHO cell (nucleus, cytoplasm, and membrane) by the use of the modulation method. The background Raman spectrum is also shown.

different locations inside the cell are visually quite different. Typical bands in these spectra are indicative of nucleotide conformation and electronic structure (785, 1095, and 1376 cm-1), C-C and C-H modes due to proteins and lipids (940, 1130, 1450 cm-1), amide vibrations (1620 cm-1), and finally backbone geometry and phosphate ion interactions (1130 cm-1). Specific assignments of individual peaks, based on previous data,21,22 can be found in Table 1. Here, we focus on and discuss some of the principal distinct features that highlight the most significant differences between the three different locations inside a single cell. First of all, peaks in the nucleus spectrum that are almost exclusively due to ring breathing modes in DNA bases, such as 785 and 1376 cm-1, are reduced in the cytoplasm spectrum. (21) Zhang, X.; Yin, H.; Cooper, J. M.; Haswell, S. J. Anal. Bioanal. Chem. 2008, 390, 833–840. (22) Jess, P. R. T.; Garce´s-Cha´vez, V.; Smith, D.; Mazilu, M.; Paterson, L.; Riches, A.; Herrington, C. S.; Sibbett, W.; Dholakia, K. Opt. Express 2006, 14, 5779– 5791.

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

743

Table 1. Band Assignment for Raman Spectra of CHO Cellsa bands (cm-1)

assignment DNA/RNA

785 940 1004 1075

T/C BkB:(O-P-O)

1095 1130 1376 1450 1620

DNA (O-P-O) sym

proteins

lipids

ν(C-C) phenilalanine ν(C-C) chain, phospholipids ν(C-N) BkB ν(C-C) T/G/A δ(CH) amide I

δ(CH)

expected location NC-CY NC-CY NC-CY M NC NC-M NC-CY M-CY-NC CY-NC

a

Abbreviations: T, thymine; C, cytosine; G, guanine; A, adenine; NC, nucleus; CY, cytoplasm; and M, membrane.

These reduced heights of the peaks indicate that the DNA concentration is significantly lower in the cytoplasm than in the nucleus. This is also confirmed by the similarly reduced intensity of the 1095 cm-1 mode of the symmetric PO2 stretching vibration of the DNA backbone. If we look at the membrane spectrum, we can again see that, in comparison to the other two spectra, the DNA Raman peaks are vanishingly small but not completely absent. This is probably due to interference from surrounding structures. It is also interesting to note that some peaks that are due to protein vibrations and amino acids are stronger in intensity in the cytoplasm spectrum. This is the case for the 1130 cm-1 C-N stretching vibration, the 1450 cm-1 CH2 deformation mode, and the 1620 cm-1 amide I, which indicates a higher protein concentration in the cytoplasm than in the nucleus. Finally, as expected, the spectrum taken from the membrane area shows significant features corresponding to lipids (1075, 1130, and 1450 cm-1). Principal Component Analysis. The spectra acquired through the modulation technique present additional advantages when used in multivariate analysis such as the principal component analysis (PCA). PCA is a statistical technique that takes the large amount of spectral data and reduces it to a few major differences between the spectra, known as principal components.23 These components form a model that can be used for a graphical representation of data variations. The individual samples are analyzed and separated into clusters for which the tightness of the clusters indicates the ability to discriminate the analytes. Most importantly, PCA relies on a normal distribution of the noise. In the case of standard Raman spectroscopy, the noise cannot have a perfect normal distribution because the spectral intensity measured is always positive, thus skewing the noise spread. In contrast, the Raman modulation spectra naturally oscillate between positive and negative values (see Figure 5). The noise of the modulated Raman spectra can, thus, have a normal distribution. To check this, we apply PCA to the standard Raman and modulated Raman spectra at different locations within a single stained CHO cell. Twelve Raman spectra were taken at each cell location (nucleus, cytoplasm, and membrane) and were then subjected to PCA. The results from the standard and modulated Raman spectra can be seen in Figures 11 and 12, respectively, where we show the first three main principal components against each other. The (23) Chan, J.; Taylor, D.; Zwerdling, T.; Lane, S.; Ihara, K.; Huser, T. Biophys. J. 2006, 90, 648–656.

744

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

Figure 11. Principal component analysis of all individual spectra of nucleus, cytoplasm, and membrane of CHO cells, acquired by the use of the standard Raman spectroscopy.

Figure 12. Principal component analysis of all individual spectra of nucleus, cytoplasm, and membrane of CHO cells, acquired by the use of the modulated Raman spectroscopy. Table 2. Confusion Matrix Giving the Classification for Each Cell Location for Standard and Modulated Raman Spectroscopya standard Raman spectroscopy

modulated Raman spectroscopy

(# sample)

NC

CY

M

NC

CY

M

NC (12) CY (12) M (12)

11 0 1

0 9 5

1 3 6

12 0 0

0 12 1

0 0 11

a Rows give the number of each cell location classified by principal component analysis (PCA) as the cell locations given in columns. Abbreviations: NC, nucleus; CY, cytoplasm, and M, membrane.

PCA method can be checked in a predictive way through a leave one out cross validation technique. In this case, we calculate the principal components of the whole data set without one spectra. This forms a training set that is used to predict the location of the left out spectra. This cross validation approach is used for each spectra in the set, and we construct a confusion matrix which summarizes the correct and incorrect spectra classification. Each row of the confusion matrix gives the predicted classification for a specific cell location. The diagonal terms of the confusion matrix give the number of correct predictive classification for the three different cell locations, and by taking into account the average of these values, it is possible to obtain information on the efficiency of the method. As seen in Table 2, for the standard Raman spectroscopy data, the nucleus (NU) is misidentified only one time, the cytoplasm (CY) three times, and the membrane (M) five times on all 12

acquisitions. By analyzing the diagonal values of the confusion matrix, we can obtain an overall predictive efficiency for the standard Raman method of about 72%. In contrast to standard Raman spectroscopy, the MR method was able to identify different locations inside a CHO cell with an efficiency of 99%. More specifically, a subdivision of a cell in three locations, representing nucleus, cytoplasm, and membrane, shows that modulated Raman spectroscopy identifies cell locations in these three categories correctly 12, 12, and 11 times, respectively, out of 12. The improved efficiency of the modulation technique might originate in the difference in the amount of variance taken into account by each of the first principal components (PC). Indeed, the first three PC of the standard Raman spectra account, respectively, for 87%, 9%, and 3% of the total variance while PC of the modulate spectra account for 23%, 8%, and 6%. The difference that can be observed in the first PC can be explained through the reduced background variability of the modulated spectra with respect to the standard spectra. Eliminating this variability enhances the relative importance of the pure Raman peaks in the PCA. This explains the improved PCA prediction efficiency of the modulated spectra versus the standard ones.

provement of the signal-to-noise ratio with respect to the SERDS procedure. In addition, once the synchronization and calibration procedure is performed, minimal user intervention is required making the method more practical and less time-consuming than other techniques. Finally, we evaluated the benefits of our method by applying the MR approach in the presence of an intense fluorescence background normally observed in real biological experiments. A major pragmatic implication of the present work is, in fact, the application of our method to the biochemical characterization of single cells. We showed that our modulation method can provide, with high efficiency, Raman spectra of different locations (nucleus, cytoplasm, and membrane) within a single CHO cell, suggesting that this minimally invasive optical technology has potential for biomedical diagnosis and imaging. It should be emphasized, in addition, that our method avoids potential misinterpretation of the data due to the standard background subtraction procedure. These results are also particularly encouraging, because they provide evidence of the potential that our modulated Raman spectroscopy method has for biomedical applications and imaging, including disease diagnosis.

CONCLUSION A new modulation method of separating Raman scattering from fluorescence has been developed that uses the principle of multichannel lock-in detection. The method allows the suppression of the fluorescent background and improves spectral quality of the data. It can remove online fixed background and render visible weak Raman features that are masked by the fluorescence background in the standard spectrum. In a test experiment, the Raman spectrum of a polystyrene bead of 2 µm in fluorescent dye was measured. We demonstrated that the use of a continuously modulated Raman excitation wavelength provides an im-

ACKNOWLEDGMENT We thank M. L. Torres and X. Tsampoula of the School of Physics and Astronomy, University of St. Andrews, for help with cell culture. We acknowledge the assistance of Dr. P. Jess, University of Berkeley, in early stages of the work. The UK EPSRC and Cancer Research UK are acknowledged for funding this research. K.D. is a Royal Society-Wolfson Merit Award Holder. Received for review November 23, 2009. Accepted November 23, 2009. AC9026737

Analytical Chemistry, Vol. 82, No. 2, January 15, 2010

745