Flow cytometry is capable of multiparameter measurements of cells and other particles, and is a versatile platform for the study of cell and biomolecular function. Over the past decades, there have been steady incremental increases in the numbers of simultaneous measurements that can be performed with a flow cytometer. These increases have been achieved by increasing the number of light sources and detectors, and of fluorescent probes that can be attached to ligands such as antibodies. Today, instruments capable of simultaneously measuring as many as nine fluorescence signals are commercially available from a number of manufacturers, and specially configured instruments have been used to measure nearly twice that number (1). However, such instruments are complex to operate and further increases in the numbers of parameters are limited by the range of accessible spectral wavelengths with common light sources and detectors, and the spectral widths of the emissions of available fluorescence probes.
One optical signal that is potentially compatible with flow-based measurement but which has not yet been widely exploited is Raman scattering. Raman scattering originates from the interaction of light with chemical bonds in a sample to produce a characteristic spectrum of molecular vibrations. The spectral features of the Raman scattering are much narrower than fluorescence spectra, and contain a wealth of information about the chemical composition of the sample, which makes it a widely used technique in analytical chemistry. Surface-enhanced Raman scattering (SERS), is a special case where the interactions of a Raman-active compound with a metal surface results in orders-of-magnitude enhancements in scattering intensities (2, 3), giving this approach the potential for applications requiring sensitive detection.
Biological applications of SERS-based detection are growing, including those that employ functionalized SERS nanoparticles (4–11). We would like to exploit the features of SERS to expand the multiplexing capability of both cell- and particle-based applications of flow cytometry. Historically, optical signals in flow cytometry have been spectrally separated using interference filters and dichroic mirrors. This is suitable for the separation of a few broad emission spectra, but not for the many narrow features of Raman spectra. To increase spectral resolution, several groups have incorporated prisms or spectrographs to disperse the collected light in fluorescence flow cytometry applications (12–16). Here, we describe a Raman Spectral Flow Cytometer and its application to the detection and discrimination of several SERS-tags as a first step in implementing Raman-based detection in flow cytometry.
The Raman Spectral Flow Cytometer (schematic presented in Fig. 1) used here employs a 250 μm (id) square bore flow cell from an Epics Elite flow cytometer (Beckman Coulter). Sheath delivery is provided by a commercial syringe pump (PhD2000, Harvard Biosciences) and sample delivery is provided by a custom-built stepper motor driven high resolution syringe pump controlled via LabView (National Instruments). Sheath and sample volumetric flow rates were between 25 and 100 μL/min and 25 and 100 nL/min, respectively resulting in linear particle flow rates of 7–27 mm/s and particle transit times of 250–1000 ms through the probe volume. Excitation is provided by a high powered HeNe laser beam (∼18 mW, Melles Griot) passing through a adjustable neutral density filter, laser linepass filter (633 nm), and iris, and focused through a long focal length cylindrical lens (f = 80 mm) and short focal length cylindrical lens (f = 15 mm) resulting in an elliptical beam ∼16 μm high by 85 μm wide. Forward angle light scatter (FALS) is collected through an aspheric condenser lens (F = 18 mm, Linos) onto a photodiode (PDA100, ThorLabs). A lens mounted to one face of the flow cell and a mirrored lens opposite, both orthogonal to the excitation beam increase the light collected from the probe volume. A microscope objective (Olympus LMPlanFl 50×, NA = 0.50) images collected light onto the face of a multimode optical fiber (600 μm id, ThorLabs). The output of the fiber is either focused through a bandpass filter (680/50 nm, Chroma) onto a photomultiplier tube (PMT, R3896 Hamamatsu) or coupled to an imaging spectrograph (HoloSpec f/1.8, Kaiser), which disperses the collected light and images it onto an EMCCD camera (Newton, Andor). The PMT output signal passes through a transimpedance amplifier and is sent to the flow cytometer's data acquisition system (MICA, National Flow Cytometry Resource, Los Alamos National Laboratory), while the output from the packaged photodiode is sent directly to the data system. The FALS signal is used to trigger data acquisition by the camera (Andor Solis software), which has 16 μm square pixels arranged in a rectangular aspect (1,600 pixels wide by 200 pixels tall). A horizontal region of interest (ROI) was defined corresponding to the spectral image from the optical fiber (50 pixels tall for the 600 μm ID fiber), and in most cases pixels were binned in the horizontal dimension while in all cases pixels were fully vertical binned. The exposure time was adjusted to match the transit time of particles through the probe volume. Typically, 1,000 events were analyzed per sample.
Aqueous silver nanoparticles were prepared using a sodium citrate reduction of silver nitrate as described (17). The core particle size distribution was determined to be 67 ± 17 nm by TEM analysis (JEOL 3000F). Exact conditions for coating the nanoparticles vary between dyes and particle batches, and will be published elsewhere (Brown et al., in preparation). In a typical preparation of dye-tagged nanoparticles, 3-mercaptopropyltrimethoxysilane (MPTMS, 750 μL, 25 μM) was added to silver nanoparticles (15 mL, 1.1 × 10–10 M) with vigorous stirring. After 45 s, the selected dye (110 μL, 2.8 μM) was introduced. Aggregation was then initiated with the addition of sodium chloride (90 μL, 1%). The solution was stirred for 6 min, at which time sodium silicate (1,800 μL, 0.54%) was added to begin the glass coating process. After 6 days, the majority of excess sodium silicate was removed by centrifuging the solution once (14.5g, 25 min), and replacing the supernatant with water. To biotinylate the nanoparticles, 3-aminopropyltrimethoxysilane APTMS (300 μL, 25 μM) was added to the coated nanoparticle solution, and the pH adjusted with sodium bicarbonate (one drop, 10%). A solution of PFP-biotin (540 μL, 12 μM) was introduced after 12 h. Centrifugation (6 × 25 min at 14.5g) was used to purify the samples after an additional 2 h of reaction time. The SERS spectra of the biotinylated nanoparticles were characterized with a DeltaNu Advantage 200A Raman spectrometer using 633 nm excitation at 1.72 mW.
Streptavidin-coated polystyrene microspheres (3.3 μm) were purchased from Spherotech. Avidin-coated microspheres were prepared by washing carboxylated polystyrene microspheres (5.3 μm diameter, Bang's Laboratories or 6.7 μm diameter, Spherotech) twice in PBS, incubated with 100 μg of avidin, followed by the addition of 5 mg EDAC (Pierce), and incubated on ice for 60 min. The beads were washed twice before labeling of aliquots with fluorescent biotin or biotin-coated SERS NPs.
Spectral data, typically spectra from 1,000 events per sample, was converted to ASCI text format using the Andor Solis software. Spectra from individual particles was parsed, and saved in both ASCI and Excel2007 formats. Flow cytometry data (10 parameters including peak height and area for four channels, peak width for one, and time) were exported as text using FCS Express (Denovo Software). Flow cytometer (10 parameters) and spectral data (200 parameters) were combined in a text file and a 210 parameter FCS3.0 file was written using a custom utility (DeNovo Software). It should be noted that the trigger thresholds for the light scatter detection and the spectral detection were independent, so that while the data were acquired simultaneously, they are not correlated event-by-event. The parameter math functions of FCS Express were used to create new parameters, five pixels in width, which corresponded to informative regions of the Raman spectra. To account for the broad background that underlies the Raman features, the intensity of adjacent but featureless regions of the spectra were subtracted to give new parameters that corresponded to the peak intensity distinguishing Raman spectral features. These new parameters are as follows: 450 cm−1 (−563 cm−1), 563 cm−1 (−450 cm−1), 1,150 cm−1 (−1,000 cm−1), 1,630 cm−1 (−1,685 cm−1). The frequency histograms for these parameters were used to set markers and define gates that identified each of the SERS tags used in this study according to the logic outlined in Table 1. For multivariate analysis of the Raman spectra, spectral data were pre-processed by smoothing (Savitsky–Golay three point), normalization (standard normal variate, SNV), and calculation of the Norris-Gap first derivative (one-point gap size) using the Unscrambler program (CAMO Software). Principal component analysis (PCA) was performed using four principle components, also using Unscrambler.
Table 1. Spectral bands used for SERS tag identification
Instrument Design and Operation
The Raman Spectral Flow Cytometer used in these studies was built around a square quartz flow cell (4 mm × 4 mm outer dimensions, 250 μm × 250 μm inner dimensions, Biosense, Beckman Coulter). Cylindrical optics were used to form an elliptical focused beam within the flow cell, and forward angle scattered light was collected after an obscuration bar, in a configuration common for flow cytometry. Raman scattered light was collected at a right angle to the excitation beam through a microscope objective and focused into a multimode optical fiber. This fiber could be coupled either to a PMT with a bandpass filter to make integrated intensity measurements in the conventional manner, or to a spectrograph which disperses the light over the photo-sensing chip of an EMCCD camera. FALS was used to trigger data collection by the PMT or EMCCD of signals from individual particles. The detector spectral response, calibrated with an HgNe calibration lamp, is presented in Figure 2. The spectral resolution of the system is ∼80 cm−1, defined by the diameter of the input fiber, as we are using the spectrograph without a slit to maximize light throughput. The pixels on the CCD detector are binned prior to readout to maximize signal to noise while accurately sampling the spectra. Unless otherwise noted, the data presented here were acquired with 8× pixel binning in the horizontal direction resulting in 200 data points per spectrum.
Measurement of a Fluorescent Alignment Particle
We used 3.1 μm diameter Nile Blue-stained microspheres to assess the alignment of the system. Detection was triggered by light scatter and bead fluorescence was measured with either a PMT or the EMCCD. Presented in Figure 3A are the fluorescence spectra from 1,000 beads arranged in order of increasing intensity for clarity. A frequency histogram of integrated emission intensity is presented in Figure 3B. The major singlet peak is clearly distinguishable from false trigger events and doublets and higher aggregates. This singlet peak had a CV of 5.82% compared to a CV of 5.84% for the same bead measured on the PMT, and a 5.86% measured on the FL4 channel of a Becton Dickenson FACSCalibur. The average spectrum of particles in the singlet peak is presented in Figure 3C.
Detection of Raman Spectra from SERS Nanoparticle Tags
To demonstrate our ability to detect individual Raman vibrations we prepared microspheres bearing SERS nanoparticle tags. Colloidal silver nanoparticles were incubated with a Raman-active compound, aggregated, coated with silica, and functionalized with biotin. The bulk spectra of the nanoparticles used in this study are presented in Figure 4. The biotinylated SERS tags, which had an average diameter of 67 nm, were then incubated with streptavidin microspheres (3.3 μm diameter) at a ratio of ∼7,800 SERS tags per microsphere, and the Raman spectra measured from these as described earlier for fluorescent microspheres. Presented in Figure 5A are spectra from 1,000 Ag-Nile Blue-labeled particles, arranged in order of increasing intensity. The x-axis is presented in units of the wave number shift relative to the exciting 632.8 nm laser. In the foreground of this collection of spectra are several background signals due to false triggers. At the back of the collection a few carryover fluorescent alignment beads are seen to saturate the detector. In Figure 5B is a frequency histogram of the intensity of the 563 cm−1 peak. The CV for this population was 25%. The variance in the population distribution may come from four sources: (1) the size/surface area of the bead; (2) variation in the amount of streptavidin bound to the bead; and (3) variation in the number/brightness of the SERS particles bound to the bead and (4) photon counting statistics. The size of the beads is uniform as demonstrated by information provided by the manufacturer and the low CV (3.9%) for the forward light scatter parameter. The fluorescence CV of these beads labeled at saturation with biotinylated phycoerythrin is 6.8%, providing an estimate of the variance due to differing streptavidin content. The remainder of the variance might be attributed to variation in the binding of the SERS particles, which was sub-saturating, or to photon counting statistics, but a more thorough investigation of the properties and behavior of SERS nanoparticles as reporter is required before the quantitative resolution of this approach can be defined. Presented in Figure 5C is the average spectrum of the singlet SERS-labeled particle peak, with the background spectra from blank microspheres is also shown. The fine structure of the SERS spectra rides on top of background from several sources. First, the SERS particles themselves have an underlying broad fluorescence background (Fig. 4). Second, the polystyrene beads have a broad autofluorescence background (Fig. 5C, dashed line). The cuvette itself exhibits a low level of fluorescence. Finally, the CCD detector has a baseline offset and dark signal.
Dependence of SERS Signal Intensity on Integration Time, Pixel Binning, and Laser Power
We examined our ability to exchange analysis speed for increased signal by changing the particle transit time. The sheath volumetric flow rate was varied from 250 μL/min to 50μL/min, which resulted in an increase in particle transit times through the probe volume from ∼200 μs to 1 ms. The exposure time of the EMCCD was increased to match the increased transit times, and spectra were acquired as above. Presented in Figure 6A are the average spectra of ∼1,000 particles obtained at three different transit times. The measured SERS signal increased in proportion to the increase in integration time, confirming our ability to improve measurement sensitivity by slowing the transit of particles through the probe volume.
We also examined the dependence of on-chip pixel binning on the measured signal intensity. Spectra were obtained as above at a single integration time (420 μs) with 4×, 8×, or 16× horizontal pixel binning. Presented in Figure 6B are the average spectra of ∼1,000 microspheres taken at three different levels of pixel binning. The intensity of the Ag-Nile Blue SERS peak increased with increasing binning, demonstrating our ability to exchange spectral resolution for signal without compromising our ability to discern important spectral features. Finally, we evaluated the dependence of SERS signal intensity on laser power. Spectra from particles bearing Ag-Nile Blue SERS tags were acquired at three different laser powers and average spectra computed as above. As presented in Figure 6C, the intensity of the SERS signal increases with increasing laser power, indicating the potential to maximize sensitivity of Raman-based applications by using more powerful light sources. The minor deviations from linearity are likely due to uncertainties in the absolute power incident on the sample during measurement, and the differential laser power dependencies of the SERS and various autofluorescence background signals.
Discrimination of Particles Bearing Different SERS Tags
The implementation of Raman Flow Cytometry for multiplexed detection applications requires the ability to distinguish particles bearing different tags. To test our ability to do this, we labeled streptavidin-coated microspheres with four different biotinylated SERS tags (bulk spectra shown in Fig. 4) and analyzed these using the Raman Spectral Flow Cytometer as described above. Presented in Figure 7 are the average spectra from ∼1,000 particles labeled with each of the four SERS tags. In many cases it is possible to identify a distinguishing spectral feature or features that allow a particular SERS tag to be identified. For example, Ag-Thionin has a strong peak at ∼450 cm−1 and a relatively weak peak at ∼1,620 cm−1, whereas Ag-Rhodamine800 has a relatively weak peak at 450cm−1 and a stronger peak at 1,620 cm−1. Ag-Ox170 and Ag-NB are more similar, showing strong vibrations at ∼650 cm−1 and 1,620 cm−1 but with Nile Blue having a moderate intensity peak at 1,150 cm−1 that Ag-Ox170 lacks The presence and absence of these spectral features can be considered to comprise a spectral “barcode” for identifying specific SERS tags, a concept we sought to exploit using standard flow cytometry data analysis methods.
To test our ability to use these spectral features to identify SERS-tags using conventional flow cytometry data analysis software, we combined the 200 channels of spectral data with 10 parameters from the flow cytometry data acquisition to create new FCS3.0 format files with 210 total parameters. Several new parameters corresponding to Raman spectral features of interest were then created by the application of virtual spectra bandpass filters produced by summing the relevant spectral channels (Table 1 and highlighted in Fig. 7). Because a broad (presumably fluorescent) signal underlies the spectral features, the signal from an adjacent, but featureless region of the spectrum was also calculated (e.g., 1,000 cm−1 and 1,685 cm−1) and subtracted as a background from each Raman feature. The resulting intensities were displayed as one parameter histograms (Fig. 8), and gates set to differentiate between positive and negative signals. Combination gates were defined according to the logic outlined in Table 1 to correspond to the spectral barcode of each SERS tag. This approach allowed the identification of each of the four of SERS tags employed here.
Multivariate Statistical Approaches to Resolving Raman Spectral Data
Although the earlier demonstration shows that it is possible to use virtual bandpass filters applied to spectral data to read spectral barcodes and identify individual SERS tags, this approach is fairly cumbersome and the complexity will increase rapidly as more SERS tags are introduced and the number and overlap of spectral features increases. To enhance our ability to resolve SERS spectra more efficiently, we evaluated Principal Component Analysis (PCA), an approach often used to reduce the dimensionality of spectroscopic data (18–20), to resolve our four SERS tags. Spectra acquired from the Raman Spectral Flow Cytometer were smoothed, normalized and the first derivative calculated and subjected to PCA. As presented in Figure 9, the derived parameters, or principal components identified by PCA enable the differentiation of all four SERS tags, demonstrating the potential to efficiently discriminate between SERS tags using more sophisticated data analysis techniques, and indicates the potential for performing highly multiparameter analyses in flow cytometry using SERS tags and a single detector and light source.
We have developed a flow cytometer capable of detecting discrete Raman vibrations, and used it to identify particles bearing different SERS tags. To our knowledge this is the first report demonstrating the measurement of Raman spectra from individual particles in a flowing sample stream and the first report of the use of SERS as a reporter modality in multiparameter flow cytometry.
A key difference between Raman signals and fluorescence signals is the feature density in the optical spectra. Although fluorophores tend to have broad emission spectra that are rather featureless, Raman spectra are characterized by sharp spectral features that correspond to specific classes of molecular vibrations. In many fluorescence measurement applications, including flow cytometry, the breadth of the emission spectra limit the number of different reporter tags that can be used for a particular spectral range. This necessitates the use of multiple light sources, each of which excites a few different tags that can be resolved within a spectral range of a couple of hundred nanometers of the optical spectrum. Recently, semi-conductor quantum dots have gained attention for flow applications (21) because of their broad absorption spectra and emission spectra that are slightly narrower than the emission spectra of small organic fluorophores. However, the spectra are only slightly narrower, and the increase in the number of labels that can be discriminated within a given spectral region is limited. The more information rich-Raman spectra have the potential to be exploited to enable a higher degree of multiplexing for a given spectral range. This is illustrated by comparison of the fluorescence spectra of the Nile Blue-labeled alignment particles shown in Figure 3, with the SERS spectra of the microsphere-bound nanoparticles in Figure 5, which span the same wavelength range.
Another difference between fluorescence emission and Raman scattering is the intensities. Intrinsic Raman scattering from molecules is orders of magnitude less intense than the fluorescence from fluorophores commonly used in biological applications. This necessitates the use of high laser powers and/or long signal integration times. Surface enhanced Raman scattering, or SERS, provides a partial solution to this problem, by offering orders of magnitude enhancement of Raman scattering to many Raman-active compounds in the presence of a metal surface such as gold or silver. The wavelength at which the SERS effect is maximal varies with the composition and dimension of the metal surface. An additive resonance effect arising from coupling of the laser-excited plasmon resonance of the metal with the absorbance maximum of the scattering compound can provide a further enhancement, in an effect termed surface-enhanced resonance Raman (or SERRS). Very frequently, these Raman-active compounds are chromophores, and single molecule detection has been reported under certain conditions.
The potentially very bright signals from surfaces exhibiting SERS or SERRS, including nanoparticles have been employed in a number of different biological detection applications, including immunoassays, nucleic acid detection, and cellular measurements. Their potential for use in flow cytometry, a platform noted for its highly multiparameter and multiplexed applications was one of the motivations for building a flow cytometer with Raman spectral measurement capabilities.
The key challenge to Raman flow cytometry is to resolve and measure the very narrow features in the Raman spectra. While it might be theoretically possible to achieve this using specially designed dichroic mirrors, bandpass filters, and discrete detectors such as PMTs, in a manner analogous to the spectral separation of and detection of multiple fluorescence signals in conventional flow cytometry, cost and light throughput issues make this impractical. A more practical approach is to use dispersive optics, such as a prism or a grating in combination with an array type detector, as has been reported for measuring fluorescence spectra in flow cytometry. We chose a spectrograph with holographic gratings and filters that provide very high light throughput and dispersion and efficient rejection of Rayleigh scattered light at the illumination wavelength. For detection, we chose a thermoelectrically cooled CCD with a readout speed compatible with flow cytometry data acquisition rates. The potential resolution of this combination of spectrograph and detector is greater than required for our use. We traded some of the excess spectrograph resolution for improved signal by operating without a slit, which allowed the maximum amount of light entering the spectrograph through the fiber to reach the detector and the diameter of the fiber to define the output resolution. The output of the spectrograph is dispersed over 1,600 columns of pixels of the detector, resulting in oversampling of the spectrograph output, so we took advantage of the detector's on-chip binning capabilities to bin columns to appropriately sample the coarse resolution of the imaged spectra and maximize signal to noise of the detector output. In most cases, pixels were binned 8× in the horizontal direction, resulting in 200 points for each spectra.
Another challenge of Raman spectral measurement by flow cytometry is acquiring enough signal from a particle while it flows through the probe volume. While microscopy-based approaches have the luxury of integrating signal for seconds or longer as needed to obtain the required signal to noise, flow-based measurements typically involve sample integration times of less than a millisecond, and sometimes much less. Here, by implementing an efficient light collection setup, a high throughput spectrograph, and a sensitive detector, we demonstrate the collection of resolvable Raman spectra with integration times down to 100 μs. Further, by using increased laser power and higher pixel binning, trading resolution for increased signal, we expect to be able to operate with integration times on the order of 10 μs, comparable with conventional commercial flow cytometers.
The ultimate goal of this work is to extend the advantages of Raman scattering discussed above to the analysis of single particles by flow cytometry. While there are a variety of commercial software tools available to the spectroscopist to analyze and interpret spectral data, these are not well suited for the multiparameter analysis of many samples of thousands of spectra each. On the other hand, flow cytometry data analysis software does not have dedicated spectral analysis capabilities. The existing FCS data standard will accommodate large numbers of parameters and several commercial flow cytometry data analysis programs do provide the ability to work with these parameters. To demonstrate the identification of particles bearing distinct SERS tags, we merged the output of the flow cytometry data acquisition (10 parameters) with that of the CCD detector (200 parameters) to make a new FCS data file containing 210 parameters. We used the parameter math capabilities of FCSExpress to calculate new parameters corresponding to the discrete spectral bands highlighted in Figure 7 and to subtract an appropriate background. Histograms displaying these derived parameters were used to set markers and define gates that defined each SERS tag in terms of being positive or negative for each of the spectral features examined as outlined in Table 1. These gates were used to determine the number and intensity for particles bearing each SERS tag. This demonstrates the ability of Raman flow cytometry to distinguish particles bearing one of a set of SERS tags, and for commercial flow cytometry software to analyze this data and report results.
Like fluorescence tags, SERS tags can be envisioned to be used in a number of different contexts in flow cytometry. The simplest use is the enumeration of particles that are positive for a given tag, and has been demonstrated here. A slightly more demanding application is determining how many of a given tag are on each particle. We estimate that there is an average of ∼7,800 SERS tags each on the particles shown in Figure 5, and that we could detect ∼100 on a particle. Converting these reasonable, but still rough, estimates into numbers that would support quantitative analysis in the same way fluorescence flow cytometry measurements do will require further work. It will be necessary to reproducibly make SERS tags of uniform intensity and to accurately measure their concentrations. Additionally, robust methods for the conjugation of the SERS tags to ligands such as antibodies, and for the characterization of those conjugates, must be developed. These are not trivial tasks, but given the wide interest in employing SERS tags on a variety of measurement platforms, should be achievable. Beyond the detection and quantification of a single SERS tag on a particle, many flow cytometry applications require particles to be scored positive or negative for multiple tags, and for the intensities of each of those tags to be determined. The simple “virtual bandpass” approach employed above will not likely meet these requirements, but there are a number of multivariate statistical approaches that should. An example presented in Figure 9 shows the ability of principle component analysis (PCA) to resolve the different tags used here. Spectra from 600 individual particles labeled with one each of the four SERS tags were acquired using the Raman Spectral Flow Cytometer, smoothed, normalized, and the first derivative spectra calculated. The processed spectra were then subjected to PCA, and the scores for the first three principle components are presented in Figure 9, illustrating the ability of this approach to distinguish between the four SERS tags. Coupled with classification and least squares fitting routines, these approaches can enable the identification of components in mixed signals and extraction of the relative intensity contributions to the measured spectra. These capabilities are not part of any commercial flow cytometry software that we know of, but there is no reason why they could not be.
In summary, we have developed a Raman Spectral Flow Cytometer that can measure Raman spectra from individual particles. We measured the spectra of particles labeled with each of several SERS nanoparticle tags, and have shown that these tags could be distinguished on the basis of their spectral features using commercially available flow cytometry data analysis software. The feature-rich spectra of the Raman signals offers the possibility of highly multiparameter measurements using a relatively simple optical detection approach.
The authors thank Gary Durack (iCyt) and Vince Shankey (Beckman Coulter) for providing the Coulter Biosense flow cells used.