Design, construction, characterization, and application of ... - CiteSeerX

Design, construction, characterization, and application of a hyperspectral microarray scanner Michael B. Sinclair, Jerilyn A. Timlin, David M. Haaland, and Margaret Werner-Washburne

We describe the design, construction, and operation of a hyperspectral microarray scanner for functional genomic research. The hyperspectral instrument operates with spatial resolutions ranging from 3 to 30 ␮m and records the emission spectrum between 490 and 900 nm with a spectral resolution of 3 nm for each pixel of the microarray. This spectral information, when coupled with multivariate data analysis techniques, allows for identification and elimination of unwanted artifacts and greatly improves the accuracy of microarray experiments. Microarray results presented in this study clearly demonstrate the separation of fluorescent label emission from the spectrally overlapping emission due to the underlying glass substrate. We also demonstrate separation of the emission due to green fluorescent protein expressed by yeast cells from the spectrally overlapping autofluorescence of the yeast cells and the growth media. © 2004 Optical Society of America OCIS codes: 170.3880, 170.2520.

1. Introduction

Changes in gene expression, resulting in the production of protein, are significant components of an organism’s response to factors such as environmental perturbations or disease. Measurement of a gene expression profile can therefore provide a detailed snapshot of the biological state of an organism. However, obtaining complete gene profiles for an entire organism requires measurement of the expression levels of thousands of genes, and until the development of high-throughput, parallel assays, determination of such profiles was not possible. With the advent of DNA microarray technology,1–5 gene expression measurements can be performed in a highly parallel fashion, leading to greatly increased experimental throughput. In an increasing number of cases, entire genomes can be profiled.6 Although microarrays are already widely employed in genomic research, this technology is still relatively young; many efforts are currently devoted to the improve-

M. B. Sinclair 共[email protected]兲, J. A. Timlin, and D. M. Haaland are with the Sandia National Laboratories, Albuquerque, New Mexico 87185-1405. M. Werner-Washburne is with the Department of Biology, University of New Mexico, Albuquerque, New Mexico 87131. Received 27 August 2003; revised manuscript received 5 January 2004; accepted 15 January 2004. 0003-6935兾04兾102079-10$15.00兾0 © 2004 Optical Society of America

ment of all aspects of this technology to further increase sensitivity, throughput, and accuracy. In a typical microarray experiment,1– 4 glass slides are spotted with gene fragments from the organism under study with a micropen printing apparatus. The spot diameters are of the order of ⬃100 ␮m with ⬃200-␮m center-to-center spacings. Each spot contains a distinct gene fragment, and up to tens of thousands of spots can be printed per slide. Messenger RNA resulting from gene expression within the organism is harvested from two cell groups: control cells 共e.g., normal tissue兲 and experimental cells 共e.g., cancerous兲. The messenger RNA is then transcribed to complementary DNA 共cDNA兲 by a process that incorporates a different fluorescent tag for each of the cell groups. For example, the control cDNA might be tagged with a green fluorophore and the experimental cDNA with a red fluorophore. The two cDNA samples are subsequently mixed together and incubated on the slide so that the cDNA can hybridize with the bound DNA on the slide. During the hybridization process, the labeled cDNA in solution attaches to the complementary immobilized gene fragment located within one of the spots on the array. To measure the relative gene expression, the hybridized arrays are scanned with an instrument similar to a fluorescence microscope.1,2 For each fluorophore, the excitation source is chosen to coincide with its absorption band, and the emission filter passes the fluorescence in a band centered on the fluorophore’s emission peak. The gene expression ratios for the 1 April 2004 兾 Vol. 43, No. 10 兾 APPLIED OPTICS

2079

two cell groups are then determined by calculation of the fluorescence ratios for the two tags for each spot on the array. Although the current microarray technology is of great utility in functional genomics, there are still several aspects of this technology that can be significantly improved. Of interest in the current study is the operation of the microarray scanners. As described above, the current generation of scanners simply measure the band-integrated emission corresponding to the bandpass filter utilized for each scanned fluorophore.1,2 This measurement approach has several significant drawbacks. First, to minimize cross talk, the two fluorescent tags must be chosen such that there is no spectral overlap of their emission spectra and minimal overlap of their absorption spectra. In principle, microarray measurements could be performed with three or more cell groups or experimental conditions 共hence requiring three or more tags兲, leading to even higher throughput for gene expression profiling. However, in practice it is extremely difficult to identify sets of three or more tags that possess a sufficient degree of spectral separation, and a separate laser wavelength is required for each tag. A second drawback of the current scanner measurements is associated with the univariate nature of the measured data: There is no reliable mechanism to determine if the recorded data have been corrupted by extraneous emission sources. The presence of such sources can greatly compromise the determination of expression ratios for weakly expressed genes. Extraneous sources can include emission from the glass substrate, emission from residual impurities from the array printing process, and cross talk associated with other fluorescent tags.7 For background removal, current scanners typically image the fluorescence with high spatial resolution 共⬃2–10 ␮m兲. Computer algorithms are then employed to accurately determine the perimeter of each spot on the array, and the signal in regions immediately adjacent to each spot is used for background subtraction.8,9 This approach, although acceptable for uniform background emission, fails with spatially varying background emission and if the signal around a spot is not representative of the background within the spot area.7 In this paper we describe the design, construction, and operation of a hyperspectral microarray scanner and we address the shortcomings mentioned above. In contrast to conventional scanners, the new hyperspectral scanner records the entire emission spectrum for each pixel of the image. Measurement of full spectral information provides several key advantages, particularly when coupled with multivariate data exploitation. Because multivariate techniques can accurately determine the contribution of highly overlapping emission sources, the requirement for widely separated absorption and emission of the tags is lifted. This leads directly to the ability to probe more than two tags, with a single laser, within a single scan. In addition, with multivariate analysis we can identify and remove all extraneous emission from the array data, even in cases in which the ex2080

APPLIED OPTICS 兾 Vol. 43, No. 10 兾 1 April 2004

traneous emission exhibits large spatial variations 共for example, emissions from spot-localized impurities兲. It should be emphasized that, under ideal conditions of widely separated tag emission and no impurity emission, filter-based and hyperspectral instruments should exhibit comparable signal-to-noise ratio 共SNR兲 performance. The primary advantages of the hyperspectral instrument are the ability to separate the contributions of multiple, highly overlapping fluorescent tags and the ability to detect and remove impurity emission. The application of hyperspectral imaging to biological systems has been of increasing interest over the past several years, and a number of hyperspectral instruments have been described in the literature.10 –13 These include a full-field imager with a Sagnac interferometer for spectral measurement11; confocal instruments12; and a lineillumination, pushbroom system.10 The results obtained with these instruments demonstrate the many advantages of hyperspectral technology. However, for the most part, these instruments were optimized for high-spatial-resolution imaging of biological samples and not for high-throughput, highsensitivity, wide-area imaging. The hyperspectral scanner described in this paper has been specifically optimized for microarray scanning. This paper is organized as follows. First, important performance requirements and design trade-offs are discussed. In particular, the need for high sensitivity and high acquisition speed are emphasized. Next, detailed descriptions of the design, construction, and operating principles of the new scanner are presented. Results obtained with the hyperspectral scanner are then presented, including measurements of image quality and sensitivity, as well as actual microarray measurements. Finally, the operation of the new scanner as a general-purpose hyperspectral imaging instrument is briefly mentioned. 2. Design Considerations

Commercially available microarray scanners are fast and sensitive.1 These instruments are capable of scanning a 20 mm by 60 mm region at a resolution of ⬃10 ␮m for a total of ⬃107 data points per fluorescent tag in approximately 5 min, implying a single-pixel acquisition time of ⬃30 ␮s. A single-pixel datum is recorded with 16-bit resolution, and minimumdetectable concentrations are specified to be of the order of 0.1 fluorophore 共fluor兲兾␮m2. Although a hyperspectral scanner is, by definition, required to acquire a significantly larger data set—the entire emission spectrum at each pixel—it cannot do so at the expense of speed and sensitivity. The primary trade-off that was chosen for the current design was to decrease the spatial resolution 共i.e., increase the pixel size兲 to keep data rates reasonable. The instrument is capable of acquiring images at ⬃3-␮m resolution; however, for microarray applications, 30-␮m resolution is usually acceptable and leads to shorter scan times. One of the primary drivers for the 10-␮m 共or better兲 resolution of the commercial

scanners is the need to accurately define the spot perimeter so that the background can be sampled in the immediate vicinity of the spot.8,9 The full spectral data provided by the hyperspectral scanner, when coupled with multivariate data analysis techniques, allows the background to be modeled and removed for all pixels within the scanned image; thus the need to record the image at a high spatial resolution is relaxed. An extremely important issue to consider in the design of a scanner is the photochemical stability of the organic chromophores used to tag the cDNA. Degradation, or bleaching, of the chromophores is commonly observed when slides are scanned multiple times or with high-intensity excitation. Numerous studies have shown that, at fixed excitation intensity, the probability of irreversible photobleaching is proportional to the number of absorption or emission cycles a fluorescent molecule undergoes.14 Typical values14 for the number of cycles before degradation are of the order of 105–106. Thus any sensitivity increase obtained by an increase in the excitation flux will be offset by a reduced lifetime of the chromophore. A notable exception to this behavior is the class of newly developed semiconductor quantum dots that are highly resistant to photochemical bleaching and are beginning to see application in biological labeling.15 The philosophy adopted in the current study is to consider the total number of photons that can be emitted by each tag molecule before degradation as a strictly limited quantity. This leads directly to the strategy of maximizing both the photon collection efficiency and the detection efficiency. The incident excitation flux can then be adjusted to ensure the detection of a sufficient number of photons within one sampling period. The photon collection efficiency increases as the square of the numerical aperture 共NA兲 of the primary objective. Therefore it is advantageous to utilize the highest NA objectives available for the desired magnification. In general, NA increases with magnification; however, the maximum magnification that can be employed will be limited by other system requirements 共for example, scan times will become prohibitively long for extremely high magnification兲. In addition to collecting the fluorescence at high NA, it is imperative to utilize a high-throughput optical system to deliver the collected photons to the detector and to use a high-sensitivity, low-noise detector. Although many scan architectures have been demonstrated for hyperspectral imaging, the specific requirements for microarrays quickly narrow the possible choices. For example, the requirement of high detection efficiency prohibits use of wavelength scanning systems for which all photons whose wavelengths do not match the current scan wavelength are wasted. Point-focus raster-scanning instruments, such as confocallike systems, can be employed for hyperspectral imaging, but for microarray applications would require rapid motion of the microarray or laser scanning system. In addition, such systems will require spectral detection with an extremely high readout rate, high-sensitivity detector array. As

described in Section 3, a line-excitation, pushbroom scan architecture was adopted for the hyperspectral microarray scanner. This arrangement enables rapid array scanning, maintains reasonable scan velocities 共⬃1 mm兾s兲, and allows use of highperformance charge-coupled device 共CCD兲 detector arrays. 3. Hyperspectral Scanner Design

In a pushbroom hyperspectral fluorescence imaging system, the excitation source is focused to a line at the plane of the sample. The optical system then images the illuminated region onto the entrance slit of an imaging spectrograph, which in turn projects a spectrally dispersed image of the line onto a twodimensional 共2-D兲 detector array. Each data frame obtained in this manner contains the emission spectra corresponding to all points along the lineilluminated region of the sample 共the y direction, see Fig. 1兲. To obtain a full 2-D image, multiple data frames are recorded with the sample’s position along the x direction 共i.e., perpendicular to the line of illumination兲 incremented for each frame. Once the full region of interest is scanned, the data can be assembled into a data cube containing the full emission spectrum for each 共x–y兲 location on the sample. Figure 2 shows the layout of the hyperspectral microarray scanner. The optical system was constructed on an optical breadboard 共Newport Corporation兲 and primarily employed a commercial lens and mirror mounts 共Newport Corporation兲. Because of its modular construction, a variety of lasers can be placed on the breadboard to serve as excitation sources. The microarray results presented in this paper were obtained with 532-nm excitation provided by a 25-mW frequency-doubled solid-state Nd:YVO4 laser 共Crystal Laser兲. In addition, an air-cooled argon-ion laser 共Ion Laser Technology兲 was utilized to produce 25 mW of 488-nm excitation for green fluorescent protein 共GFP兲 imaging. The laser fluence is adjusted to fit experimental conditions by use of a filter wheel populated with a set of neutral-density filters. The desired line illumination is obtained when the laser is directed through a Powell lens16,17 共StockerYale兲. This optic is similar in appearance to a prism; however, the apex is slightly rounded with the contour containing a conic constant.17 The apex contour causes the TEM00 laser mode to be redistributed into a diverging fan of light that will produce a uniform intensity line at the focus of a converging lens. Achieving high-quality line illumination is simpler with a Powell lens than with scanning mirror systems, which would be required to operate at extremely fast scan rates and which produce only uniform illumination near the center of the scan line. The Powell lens approach is also more efficient than approaches based on cylindrical lenses that require beam expansion and truncation for the production of uniform lines. Generation of a highly uniform line illumination with a Powell lens requires that the input laser mode size match the mode size for which the 1 April 2004 兾 Vol. 43, No. 10 兾 APPLIED OPTICS

2081

Fig. 2. Schematic diagram of the hyperspectral scanner layout.

Fig. 1. 共Top兲 Diagram of the illumination and scan geometries for the hyperspectral scanner. The microarray is scanned in the x direction, whereas the y direction is parallel to the focused line of laser light as well as the spectrograph entrance slit. 共Bottom兲 Illustration of a data cube comprising a complete emission spectrum from each spatial pixel of the 2-D image.

lens was designed. This was achieved by means of adjusting the optical path length between the laser output and the Powell lens. The 3.6° fan angle produced by the Powell lens is slightly modified with a simple telescope to adjust the line length to 1.6 mm at the focus of a 10⫻ objective 共 f ⫽ 20 mm兲. A dichroic beam splitter 共Chroma Inc.兲, characterized by high reflectivity at the laser wavelength and high transparency at longer wavelengths, is used to direct the diverging fan of light through the primary objective. We obtained the results presented in this paper using an infinity-corrected 10⫻, NA 0.45, apochromatic objective 共CFI Plan Apochromat 10⫻, Nikon Inc.兲 as the primary objective. For higherresolution fluorescence microscopy applications, a 20⫻, NA 0.75, apochromatic objective is also available 共CFI Plan Apochromat 20⫻, Nikon Inc.兲. These objectives were chosen for three reasons. First, they possess high NAs for their magnifications, thereby providing higher collection efficiencies. Second, by virtue of their apochromatic color correction they 2082


minimize chromatic aberration effects in the hyperspectral data. Third, these objectives exhibit fairly long working distances that reduce the potential for inadvertent contact with the microarray slides. The 10⫻ primary objective focuses the incident laser light to a line 1.6 mm long and 8 –10 ␮m wide at the surface of the microarray. The microarray slide is mounted on a computer-controlled X–Y scan stage that is capable of 100 mm of travel in both the x and y directions and has a resolution of 0.5 ␮m. The flatness of the stage motion 共less than ⫾2 ␮m兾100 mm of travel兲 combined with the relatively large depth of focus of the ⬃10-␮m-wide laser line and the slight oversizing of the slit 共100 ␮m兲 render this arrangement relatively immune to Z-axis variations during scanning. Fluorescent photons emitted from the line-illuminated area are collected by the primary objective and directed through the dichroic filter toward the imaging system. A holographic notch filter 共Kaiser Inc.兲 matched to the laser wavelength is utilized to eliminate unwanted scattered and reflected laser light. The fluorescent light then passes through an f ⫽ 200-mm tube lens 共Nikon Inc.兲, which produces an intermediate line image of the fluorescing sample. For direct video imaging of the sample, the intermediate image can be directed onto a video CCD camera 共XC-ST50, Sony Inc.兲 by means of a mirror attached to a pop-up mount. Otherwise the line image is relayed to the entrance slit of an f兾3 imaging spectrograph 共CP200, ISA Inc.兲 by means of a 0.6⫻ reducing lens 共HRP060-CMT, Diagnostic Instruments Inc.兲. The optical system illuminates the spectrograph at f兾6, somewhat underfilling the grating. The spectrograph is equipped with a 133line兾mm concave holographic grating that projects a spectrally dispersed image of the entrance slit onto

the surface of a cooled, scientific-grade CCD array. Thus the overall system magnification is 6⫻ with the 10⫻ objective and 12⫻ with the 20⫻ objective. A CCD detector located at the focal plane of the spectrograph records the spectral and spatial data. For sensitive detection of low-level fluorescence at high readout speeds, a CCD detector enhanced with on-chip electron multiplication 共DV465, Andor Technologies兲 is used. This camera differs from traditional CCD systems through the inclusion of a serial gain register between the usual serial register and the charge detection node.18,19 Electrons are clocked through the gain register with larger voltages than normal clock voltages, which results in the generation of additional electrons through impact ionization. Adjusting the magnitude of the high-voltage clock phase controls the total electron gain obtained during traversal of the entire length of the gain register. In this manner, the total output charge corresponding to a single photoelectron can be adjusted to a level greater than the detector read noise. The excess noise associated with the on-chip gain increases the signal variance by a factor of 2 relative to the shot-noise-limited variance.20 This is to be compared with the excess noise factors of 1.2–1.4 characteristic of the photomultiplier tubes used in commercial confocal scanners.21 The utilization of this type of on-chip gain is particularly advantageous for applications such as microarray scanning that require high readout rates because the increased read noise associated with high-speed operation can be offset by the electron gain. In this manner, lowlight-level images can be obtained even at high readout speeds. The CCD detector in the hyperspectral imager was operated at a readout rate of 1 MHz and with an electron gain of approximately 20. The DV465 detector is a frame transfer device with 576 pixels along the spectral direction and 288 pixels along the y direction. Individual detector elements are 20 ␮m long by 30 ␮m high. The 11.5-mm length of the detector in conjunction with the dispersion of the 133-lines兾mm grating allow for a spectral coverage of 490 –900 nm. The effective spectral resolution of the spectrograph– detector system with a 100-␮m entrance slit was measured to be approximately 3 nm 共full width at half-maximum兲 by use several spectral lines from low-pressure krypton and neon lamps. The 8.6-mm height of the detector along the y direction 共the slit direction兲 corresponds to a length of 1.4 mm on the microarray sample. With the 10⫻ objective, 30-␮m spatial resolution corresponds to six fold on-chip pixel binning along the vertical direction 共two fold binning for 10-␮m spatial resolution兲. The entrance slit is slightly oversized relative to the ⬃60-␮m-wide image of the illuminated region of the sample. Such oversizing could potentially cause spectral artifacts if the illuminating line is not straight or moves with time. However, use of the Powell lens results in a uniform, straight, and stable line illumination and thereby prevents these spectral artifacts. The hyperspectral microarray scanner is operated

under computer control by a C⫹⫹ control application. The application reads the user input such as the size and location of the scan region, the scan rate, and the binning parameters for the detector. A sequence of programming instructions is then downloaded to the digital signal processor-based controller board for the X–Y stage system. The stages are scanned along the x direction at constant velocity, and the stage controller issues tightly synchronized trigger pulses at regular intervals corresponding to the desired spatial resolution. The trigger pulses from the stage controller are used to initiate frame transfer and readout of the CCD detector. At the conclusion of each x scan, the stage position is incremented in the y direction and a new x scan is initiated. The y increment is typically 1.0 mm with the 10⫻ objective. After the entire region of interest has been scanned, the scan data are written to the computer’s hard disk. A full microarray scan of 20 mm ⫻ 60 mm at 30-␮m resolution generates approximately 3 Gbytes of data. Another C⫹⫹ based application is then used to visualize the data cube. 4. Hyperspectral Image Analysis

Prior to multivariate analysis of the hyperspectral data, calibration data frames are acquired with the laser source replaced by low-pressure discharge lamps 共krypton and neon兲. The spectral lamp data are used to calibrate the wavelength scale and to remove the effects of image curvature caused by the spectrograph. Multivariate curve resolution 共MCR兲 techniques are then used to analyze calibrated, curvature-compensated data.22,23 A thorough discussion of the MCR analysis is beyond the scope of this paper and is presented elsewhere.22,23 Briefly, MCR is a powerful multivariate method that allows the emission spectra of all components 共including species for which no a priori information is available兲 to be separately estimated and their relative concentration maps to be determined, even when the emitting species are highly overlapped spectrally or spatially. MCR analysis can fail in the case of strongly covarying components 共i.e., components that appear everywhere with the same relative concentrations兲; however, this is highly improbable with microarrays. Independent concentration maps obtained from the MCR analysis are 2-D images representing the spatial location and the relative concentration of each of the emitting sources. Thus MCR is ideal for analysis of hyperspectral microarray data. The particular implementation of MCR used in the present study is a constrained alternating least-squares analysis that iteratively solves for the pure emission spectra and the relative concentrations of each of the emitting components.22 5. System Characterization

Figure 3 presents examples of the image quality obtained with the hyperspectral microarray scanner. Figure 3共a兲 is an image of a U.S. Air Force standard resolution target obtained by broadband illumination, with the scanner set for 10-␮m spatial resolution. As can be seen from the image, element 6 of 1 April 2004 兾 Vol. 43, No. 10 兾 APPLIED OPTICS

2083

Fig. 3. 共a兲 and 共b兲 Images of a U.S. Air Force resolution target and 共c兲 and 共d兲 a microarray taken with 10-␮m resolution and 30-␮m resolution. The microarray spots are still clearly resolved at the coarser resolution setting.

group 4 of the resolution target, corresponding to ⬃30 line pairs兾mm, is resolved in both the horizontal and the vertical direction. The image presented in Fig. 3共b兲 was obtained under identical conditions as that in Fig. 3共a兲, with the exception that the vertical binning of the CCD detector and the trigger interval along the x direction were adjusted for 30-␮m resolution. As expected, the decreased resolution has blurred all the high-frequency spatial detail, with even the coarsest lines 共element 1 of group 4, 16 line pairs兾mm兲 only partially resolved. However, as mentioned in Section 1, the size and spacing of the printed DNA spots on the microarray are of the order of 100 ␮m and should be easily resolved with 30-␮m resolution. The spectrally integrated microarray images shown in Figs. 3共c兲 and 3共d兲, obtained with the scanner set for 10- and 30-␮m resolution, respectively, confirm this assertion. Although the image obtained at 30-␮m resolution does appear more pixilated than the 10-␮m image, sufficient detail remains to clearly identify the spot boundaries. The coffee-ring structure of the microarray spots 共i.e., highest intensity near the circumference of the spots兲 is a common feature of microarrays produced by micropen printing. Note that both the acquisition time and the data file size for the 10-␮m resolution image are nine times larger than that of the 30-␮m resolution image. Under normal conditions, 30-␮m resolution hyperspectral images can be obtained with a linear x-scan velocity of 0.4 mm兾s, corresponding to an ⬃75-ms exposure time. Note that use of a frame transfer device is advantageous in this application because the readout of a given frame and exposure of the subsequent frame occur simultaneously. At the 0.4mm兾s scan velocity, data-acquisition rates are approximately 1.5 Mbytes兾s. The maximum x-scan velocity for a desired spatial resolution is currently limited by the detector readout time and can be improved to 4 mm兾s through use of faster detector sys2084


Fig. 4. Emission spectra recorded from the Cy3 dilution standard. The emission recorded within the Cy3 spots contains emission from both the Cy3 and the substrate glass. The emission recorded in regions between the spots contains only glass emission. Subtraction of the glass emission from the spot emission yields the true Cy3 emission spectrum.

tems 共10-MHz systems are now available兲. With a 10-MHz detector, a 60 mm by 20 mm microarray can be scanned in 5 min by the hyperspectral scanner. Thus, at these acquisition speeds, the hyperspectral scanner will be comparable in speed to commercial scanners. In addition, the ability to measure more than two fluorophores by a single excitation wavelength within a single scan further enhances the throughput of the hyperspectral scanner. The sensitivity of the hyperspectral scanner was characterized by commercially prepared standards24 containing a dilution sequence of Cy3 spots. Cy3 is a cyanine dye with an absorption maximum at 550 nm and an emission maximum at 570 nm that are commonly used in microarrays.1 Each spot in the sequence contains a different areal concentration of Cy3 fluorophores, with concentrations ranging from 0.01 to 6200 fluors兾␮m2. Figure 4 shows the average raw emission spectrum, the average background 共glass兲 emission, and the average difference spectrum obtained from a spot with a Cy3 concentration of 48 fluors兾␮m2. The background emission spectrum arising from the glass substrate was measured for a region immediately adjacent to the Cy3 spot. To obtain an estimate of the sensitivity of the scanner, the difference spectrum from each of the spots was band integrated from 558 to 595 nm. The results, shown in Fig. 5共a兲 show that the integrated signal counts decrease with the fluorophore concentration in a linear fashion and that, even for a concentration as low as 1.5 fluor兾␮m2, the scanner recorded several thousand signal counts. Thus the hyperspectral scanner is capable of detecting spots with concentrations well below 1 fluor兾␮m2. A more conservative measure of performance is the SNR of the Cy3 emission obtained as a function of the

Fig. 5. 共a兲 Recorded signal counts as a function of fluorophore concentration for the Cy3 dilution standard obtained with the hyperspectral scanner. 共b兲 The SNR calculated with Eq. 共1兲 as a function of Cy3 concentration for both the hyperspectral scanner and a commercial microarray scanner.

fluorophore concentration. In the microarray literature, the SNR is defined as1 SNR ⫽

ns ⫺ nb , ␴b

(1)

where ns is the average number of counts obtained from the spot of interest, nb is the average number of background counts obtained in regions not containing any fluorophores, and ␴b is the pixelwise standard deviation of the background regions. Typically a SNR of 3 is considered as the detection limit for a fluorophore. The dependence of the SNR on the fluorophore concentration is shown in Fig. 5共b兲. For each Cy3 spot, a region of the background in the immediate vicinity of that spot was used to determine the average and standard deviation of the background emission. In this way, effects due to large length-scale variations in the illumination or collection efficiency are minimized. Also shown in Fig. 5共b兲 is the SNR that we obtained using the same reference slide with a commercial confocal scanner. It is difficult to directly compare the two data sets because of the differences in the intensity of the laser excitation, spatial resolution, and differing scan

speeds. However, it is clear from the data of Fig. 5共b兲 that the hyperspectral scanner is capable of comparable 共if not better兲 SNR performance as that of the commercial instrument. It must be emphasized that the analysis of the hyperspectral data 共i.e., simple band integration兲 was chosen to allow the closest comparison with filter-based scanners. Full, multivariate analysis of the data will further increase the SNR performance and accuracy of the hyperspectral scanner, particularly in cases in which single or multiple impurity emission sources are present and overlap with the emission of the fluorophore labels. Although commercial scanners are specified to have sensitivities of the order of 0.1 fluor兾␮m2, in practice sufficient SNR for detection is almost never achieved at this concentration. This is also true for the hyperspectral scanner if multivariate analysis is not employed. The primary factor that limits the measured SNR 关as defined by Eq. 共1兲兴 is not the instrument sensitivity 共ns兲, but the background emission due to the glass substrate 共nb and ␴b兲. Sensitivities of the order of 0.1 fluor兾␮m2 are readily achievable for both types of scanner; however, doing so requires multivariate analysis in the case of hyperspectral data or emission-free substrates in the case of univariate data. Figure 5共a兲 shows that the hyperspectral scanner measured ns ⬇ 3000 signal counts 共after background subtraction兲 for a concentration of 1.5 fluor兾␮m2; however, because of the strong glass emission, the SNR is only 7.5. Significant improvements in the SNRs, as defined by Eq. 共1兲, could be obtained immediately for both the hyperspectral scanner and the commercial scanners by use of emission-free substrates such as fused silica. However, to date, most commercial microarray substrates are glass slides with substantial emission. 6. Selected Results

Figure 6 shows an image of a yeast genome microarray slide 共GAPSII, Corning, Inc.兲 hybridized with Cy3- and Cy5-tagged cDNA. Note that, because this image is generated from a hyperspectral data cube, each pixel of the image contains an entire emission spectrum. The graphs inset in Fig. 6 show the raw emission spectra corresponding to several locations of the image. Within the printed DNA spots, the spectra are representative of differing relative concentrations of the Cy3 and Cy5 fluorophores, whereas in the region between the spots emission from the glass slide is clearly resolved. Note that emission from the cyanine dye Cy5 at ⬃670 nm is observed even in this case where the laser wavelength is not optimized for excitation of Cy5 关the absorption maximum of Cy5 occurs at 649 nm 共Ref. 1兲兴. In an ideal hyperspectral microarray experiment, multiple fluorophores with strongly overlapping absorption spectra would all be excited by a single laser wavelength and quantitated in a single scan. In addition to the observation of emission from the substrate glass of the commercially prepared microarray, other initial studies7 have revealed emission from a spot-localized impurity whose emission spectrum strongly overlapped with that of Cy3. Although this 1 April 2004 兾 Vol. 43, No. 10 兾 APPLIED OPTICS

2085

Fig. 6. 共Left兲 Portion of a Cy3–Cy5 microarray image recorded with the hyperspectral scanner. Although the excitation wavelength 共532 nm兲 used in this scan is optimized only for Cy3, residual Cy5 absorption allowed observation of Cy5 signatures from several DNA spots. 共Right兲 The raw emission spectra obtained at various locations 共labeled a– c兲 within the image. Spectrum a is obtained in a region between the DNA spots and contains only glass emission. Spectrum b is obtained from a DNA spot and contains both Cy3 and Cy5 emission. Spectrum c is obtained from a DNA spot containing only Cy3.

does not present a difficulty for the hyperspectral scanner coupled with MCR analysis, conventional scanners are incapable of separating such an extraneous emission source from the true Cy3 emission. The increased throughput possible with the hyperspectral scanner has been tested with two closely overlapping dyes 共Cy3 and Alexa 532兲 printed on a glass slide at various concentrations of pure and mixed dyes. The MCR analysis of the spectra obtained from this slide demonstrates the ability of the scanner to separate and quantify these two dyes even though their emission maxima are separated by only 12 nm. Figure 7 shows the separated concentration maps of the two dyes and the pure-emission spectra extracted for the glass, Cy3, and Alexa 532. These results clearly demonstrate the potential for higherthroughput capabilities with gene expression microarrays with multiple overlapping dyes. Although there are many dyes available that could readily be incorporated in microarray processes, they have not traditionally been used in microarrays. Direct demonstration of the ability of the hyperspectral scanner to dramatically increase microarray throughput currently awaits the generation of the hybridized microarray slides incorporating more than two overlapping dyes. Although the hyperspectral scanner was specifically designed for microarray applications, it has shown great utility as a general-purpose hyperspectral imager. As a demonstration, images of live yeast cells expressing GFP while immobilized within a microfluidic channel were obtained with a 488-nm 2086


Fig. 7. Results of a hyperspectral scan and multivariate analysis of a microarray containing two spectrally overlapping fluorophores. In this microarray, the top four spots contained only Cy3, the middle four spots contained only Alexa 532, and the bottom four spots contained a mixture of Cy3 and Alexa 532. 共Top兲 The emission spectra of Cy3 共dashed curve兲, Alexa 532 共solid curve兲, and the glass substrate 共dotted curve兲 obtained from the MCR analysis. 共Bottom兲 The relative concentration maps obtained from the MCR analysis demonstrate the ability to separate the two overlapping fluorophores.

excitation wavelength and 10-␮m spatial resolution. Galactose-based growth media were conducted through the channel to stimulate the GFP production and to maintain cell viability. Although the spatial resolution used for these measurements was not sufficient to resolve the features of individual yeast cells, it was sufficient to observe the clumps of cells abutting the packing material within the channel. Although the yeast cells are expressing GFP in the steady state, the GFP emission was almost impossi-

Fig. 8. Emission spectra obtained from a hyperspectral scan and MCR analysis of yeast cells expressing GFP. An average raw emission spectrum containing contributions from the GFP and autofluorescence from the growth medium is shown, along with the primary pure-component spectra of GFP emission and the autofluorescence obtained from the MCR.

ble to observe directly because of significant autofluorescence from the growth media. Figure 8 shows a typical emission spectrum obtained from the sample containing an admixture of GFP emission and autofluorescence. Note that the strength of the GFP emission is, at best, comparable to that of the autofluorescence in this spectrum. Application of the MCR algorithms to the raw image data yielded the pure-component spectra for the GFP and autofluorescence shown in Fig. 8, as well as the concentration maps shown in Fig. 9. Significant concentrations of GFP are observed only in the vicinity of the yeast cells whereas the emission from the growth media is observed everywhere throughout the channel. Observation of GFP emission in the presence of autofluorescence or other impurity emissions is an outstanding problem in bioimaging.25–27 This example clearly highlights the advantage of combining hyperspectral imaging with multivariate data analysis tools for bioimaging applications. 7. Conclusion

We have described the design, construction, characterization, and application of a new hyperspectral microarray scanner that records the entire emission spectrum for each pixel of the imaged area. The new scanner uses a pushbroom scan architecture with the line focusing of the excitation laser achieved by a Powell lens system and also uses a high-sensitivity electron multiplying CCD array to record the spectral and spatial data. In normal operation, the hyperspectral scanner records data at a 30-␮m spatial resolution. Although this is lower than the spatial resolution of commercial scanners 共⬃5–10 ␮m兲, this resolution is sufficient to resolve the printed DNA spots of the microarray, and use of multivariate back-

Fig. 9. Relative concentration maps of the GFP and autofluorescence obtained from the MCR. The GFP emission is observed only in the region of the microfluidic channel occupied by the yeast cells whereas the autofluorescence from the growth medium is observed everywhere in the channel.

ground compensation circumvents the need for higher spatial resolution. Characterization of the hyperspectral scanner with commercially prepared dilution standards24 shows that its sensitivity and SNR performance are comparable to, or better than, those of a commercial scanner. With use of new high readout rate CCD detectors, the scan speed of the hyperspectral scanner can be comparable to the speed of commercial instruments. We have also demonstrated that analysis of hyperspectral data with multivariate techniques allows for quantitative separation of the contributions of multiple, spectrally overlapped fluorophores. This leads to the ability to measure more than two spectrally overlapping dyes in a single scan 共by use of a single excitation laser wavelength兲 and to quantitatively analyze the concentrations of all the emitting species. We are currently developing microarray protocols that will employ more than two fluorescent labels that can be excited with a single laser wavelength to demonstrate the improved throughput of the hyperspectral scanner. In addition, we have shown that hyperspectral data, when combined with multivariate analysis techniques, allow for the identification and removal of extraneous emission sources such as substrate and impurity emission, even when these sources are spatially varying. Spatially varying impurity emissions are difficult to detect with commer1 April 2004 兾 Vol. 43, No. 10 兾 APPLIED OPTICS

2087

cial 共univariate兲 scanners and can significantly degrade the results of microarray measurements obtained with these instruments. Finally, we have demonstrated that the hyperspectral scanner can be used as a general-purpose bioimaging system and have shown that hyperspectral techniques are effective for separating GFP emission from large autofluorescence backgrounds. The authors gratefully acknowledge Juanita Martinez, Anthony Aragon, Gavin Pickett, and Marilee Morgan for preparing the microarrays shown in this paper and for collecting the commercial scanner data; Jeb Flemming and Monica Manginell for preparing the GFP and yeast specimens; and Gary Jones for his aid in constructing the hyperspectral scanner. This research was funded in part by the W. M. Keck Foundation and the Laboratory Directed Research and Development program from Sandia National Laboratories. A portion of this research was also funded by the U.S. Department of Energy’s Genomes to Life program 共www.doegenomestolife.org兲 under project Carbon Sequestration in Synechococcus sp.: From Molecular Machines to Hierarchical Modeling 共www.genomes-to-life.org兲. Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the U.S. Department of Energy under contract DE-ACO4-94AL85000. M. Werner-Washburne was supported by the National Science Foundation 共MCB-0092374兲 and the National Institutes of Health 共HG02262 and GM67593兲. References 1. M. Schena, Microarray Biochip Technology 共Eaton, Natick, Mass., 2000兲. 2. V. C. Cheung, M. Morley, F. Aquilar, A. Massimi, R. Kucherlapati, and G. Childs, “Making and reading microarrays,” Nat. Genet. 21共Suppl.兲, 15–19 共1999兲. 3. N. L. W. van Hal, O. Vorst, A. M. M. L. van Houwelingen, E. J. Kok, A. Peijnenburg, A. Aharoni, A. J. van Tunen, and J. Keijer, “The application of DNA microarrays in gene expression analysis,” J. Biotechnol. 78, 271–280 共2000兲. 4. D. J. Duggan, M. Bittner, Y. Chen, P. Meltzer, and J. M. Trent, “Expression profiling using cDNA microarrays,” Nat. Genet. 21共Suppl.兲, 10 –14 共1999兲. 5. P. O. Brown and D. Botstein, “Exploring the new world of the genome with DNA microarrays,” Nat. Genet. 21共Suppl.兲, 33–37 共1999兲. 6. J. L. DeRisi, V. R. Iyer, and P. O. Brown, “Exploring the metabolic and genetic control of gene expression on a genomic scale,” Science 278, 680 – 686 共1997兲. 7. M. J. Martinez, A. D. Aragon, A. L. Rodriguez, J. M. Weber, J. A. Timlin, M. B. Sinclair, D. M. Haaland, and M. WernerWashburne, “Identification and removal of contaminating fluorescence from commercial and in-house printed DNA microarrays,” Nucl. Acids Res. 31, e18 共2003兲. 8. P. H. Tran, D. A. Peiffer, Y. Shin, L. M. Meek, J. P. Brody, and K. W. Y. Cho, “Microarray optimizations: increasing spot accuracy and automated identification of true microarray signals,” Nucl. Acids Res. 30, e54 共2002兲. 9. C. S. Brown, P. C. Goodwin, and P. K. Sorger, “Image metrics in the statistical analysis of DNA microarray data,” Proc. Natl. Acad. Sci. USA 98, 8944 – 8949 共2001兲.

2088


10. R. A. Schultz, T. Nielsen, J. R. Zavaleta, R. Ruch, R. Wyatt, and H. R. Garner, “Hyperspectral imaging: a novel approach for microscopic analysis,” Cytometry 43, 239 –247 共2001兲. 11. H. Tsurui, H. Nishimura, S. Hattori, S. Hirose, K. Okumura, and T. Shirai, “Seven-color fluorescence imaging of tissue samples based on Fourier spectroscopy and singular value decomposition,” J. Histochem. Cytochem. 48, 653– 662 共2000兲. 12. M. J. Stimson, N. Haralampus-Grynaviski, and J. D. Simon, “A unique optical arrangement for obtaining spectrally resolved confocal images,” Rev. Sci. Instrum. 70, 3351–3354 共1999兲. 13. Q. S. Hanley, P. J. Verveer, and T. M. Jovin, “Spectral imaging in a programmable array microscope by Hadamard transform fluorescence spectroscopy,” Appl. Spectrosc. 53, 1–10 共1999兲. 14. S. A. Soper, H. L. Nutter, R. A. Keller, L. M. Davei, and E. B. Shera, “The photophysical constants of several fluorescent dyes pertaining to ultrasensitive fluorescence spectroscopy,” Photochem. Photobiol. 57, 972–977 共1993兲. 15. A. J. Sutherland, “Quantum dots as luminescent probes in biological systems,” Curr. Opin. Solid State Mater. Sci. 6, 365– 370 共2002兲. 16. K. A. Christensen and M. D. Morris, “Hyperspectral Raman microscopic imaging using Powell lens line illumination,” Appl. Spectrosc. 52, 1145–1147 共1998兲. 17. A. Bewsher, I. Powell, and W. Boland, “Design of singleelement laser-beam shape projectors,” Appl. Opt. 35, 1654 – 1658 共1996兲. 18. C. D. Mackay, R. N. Tubbs, R. Bell, D. Burt, and I. Moody, “Subelectron read noise at MHz pixel rates,” in Sensors and Camera Systems for Scientific, Industrial, and Digital Photography Applications II, M. M. Blouke, J. Canosa, and N. Sampat, eds., Proc. SPIE 4306, 289 –298 共2001兲. 19. P. Jerram, P. Pool, R. Bell, D. Burt, S. Bowring, S. Spencer, M. Hazelwood, I. Moody, N. Catlett, and P. Heyes, “The LLCCD: low-light imaging without the need for an intensifier,” in Sensors and Camera Systems for Scientific, Industrial, and Digital Photography Applications II, M. M. Blouke, J. Canosa, and N. Sampat, eds., Proc. SPIE 4306, 178 –186 共2001兲. 20. M. S. Robbins and B. J. Hadwen, “The noise performance of electron multiplying charge coupled devices,” IEEE Trans. Electron Devices 5, 1227–1232 共2003兲. 21. Photomultiplier Tubes: Basics and Applications, 2nd ed. 共Hamamatsu Photonics KK, Japan, 1999兲. 22. D. M. Haaland, J. A. Timlin, M. B. Sinclair, M. H. Van Benthem, M. J. Martinez, A. D. Aragon, and M. WernerWashburne, “Multivariate curve resolution for hyperspectral image analysis: applications to microarray technology,” in Spectral Imaging: Instrumentation, Applications, and Analysis, R. M. Levenson, G. H. Bearman, and A. MahadevanJansen, eds., Proc. SPIE 4959, 共2003兲. 23. R. Bro and S. DeJong, “A fast non-negativity-constrained least squares algorithm,” J. Chemometr. 11, 393– 401 共1997兲. 24. Cy3 dilution slide DS3001, Panomics Inc., Redwood City, Calif., http:兾兾www.panomics.com. 25. N. Billinton and A. W. Knight, “Seeing the wood through the trees: a review of techniques for distinguishing green fluorescent protein from endogenous autofluorescence,” Anal. Biochem. 291, 175–197 共2001兲. 26. D. Bumann, “Examination of Salmonella gene expression in an infected mammalian host using the green fluorescent protein and two-colour flow cytometry,” Mol. Microbiol. 43, 1269–1283 共2002兲. 27. K. P. Doyle, R. P. Simon, A. Snyder, and M. P. Stenzel-Poore, “Working with GFP in the brain,” Biotechniques 34, 492– 494 共2003兲.