The imaging system we have described is the third generation in a developmental series (4
). We desired to minimize the need to apply computational corrections to the images in order to reduce the risk of introducing systematic errors into the biological results. Thus we chose to image large areas in a single frame with sufficient resolution so that tens of thousands of elements could be distinguished for ‘whole genome’ experiments, and to provide a chromatically stable illumination distribution. Arrays that do not fit into a single image field are analyzed field by field and the data subsequently merged. Composite images of an entire array are produced only for display purposes. While the emphasis here has been on measurements of nucleic acid microarrays, our system has been used also for analysis of proteins (5
) and low-resolution imaging of tissue specimens.
The design choices for our system are based on the view that the information from the array is carried in the ratios of the total integrated intensities produced by hybridization to the array elements. Experience has supported this expectation, and we find that many of the array parameters of concern to others, such as uniformity of array element morphology or size, do not substantially affect the results obtained with the arrays that we produce (3
). Cleanly resolving neighboring array elements and attributing the proper total emission to each is critical, but high-resolution examination of the signal distribution within an array element is not. Several tens of pixels per array element are sufficient to apply general quality control measures to the hybridization, for example the correlation of the different fluorochrome intensities over the element (2
). Thus we chose to employ ~100 pixels to image an array element and its surrounding background. A high-dynamic range genome-scale measurement involving a tiling path of genomic BAC clones or a complete set of genes requires ~30–40 000 array elements, would then require an image with ~2K × 2K pixels with large electron capacity. We chose a back-thinned CCD in order to obtain the highest possible quantum efficiency. Array printing technology that is based on using standard microtiter plates for holding the printing solutions and on microscope slides as the printing substrate, suggests that a convenient dimension for an array is 18 mm. These criteria require fluorescence detection optics with ~10 µm resolution or better across the entire field. Moreover, we desired to compare more than two nucleic samples simultaneously (6
), to use the blue-emitting stain DAPI to identify array elements independently of the hybridization signals, and to simplify operations by avoiding refocusing as wavelengths were changed. Our experience with high-quality commercially available photographic (4
), television, dissecting microscope lenses and the like in the first generation versions of this imaging system (4
) indicated that custom lens designs would be required to meet our performance criteria. – illustrate the performance of the resulting system, and demonstrate in particular that arrays with many more than 30
000 elements per field can be well imaged.
The maximum number of array elements that can be imaged with this system in its current configuration depends on the desired measurement accuracy. Most array elements in our hybridizations show pixel to pixel intensity correlations of the test and reference signals that are very close to 1.0. This indicates that the ratio on any single pixel within the element would provide a good measurement. The rest are ‘redundant’, except in so far as one might want to apply quality control criteria. As stated above, the optical resolution of the current optics keeps essentially all of the light from a sub-resolution point source within a 3 × 3 block of pixels. The minimum useful spacing between such point sources for a real measurement depends on the intensity variations among the sources, with the most challenging task involving accurate measurement of a dim array element that is next to a bright one. Deconvolution image sharpening, as illustrated in , can assist with this task. Given these considerations it seems very likely that high-quality data could be obtained from an array with elements on 50 µm centers, ~130 000 elements per image, and perhaps even closer. The measurement precision for such an array can be estimated using the relationship derived in the Supplementary Data and illustrated . Assuming an area of 1 pixel for an array element and a typical background level of 200 digital units, the contribution to the standard deviation of the Log2Ratio from the imaging system should be below 0.1 for signal intensities above ~300. Thus the useful dynamic range of the system, assuming this accuracy criterion, would be ~200, and could be extended substantially by summing multiple images.
The imaging system contributes two sources of ‘background’ that must be removed during analysis in order to obtain the most accurate results—the inherent offset of the camera output, and light originating from the array substrate and the coverslip. illustrates typical magnitudes of the offset and background in our system for standard measurement configurations. Our image analysis employs the common technique of subtracting the local background from the intensity of an array element prior to calculating fluorescence ratios. The appropriateness of local background subtraction for these images is evident in and , where the ‘M
’ plots remain straight and essentially identical for a wide range of exposure times without employing computational corrections that are frequently instituted in array analysis (8
). The uniform backgrounds produced by some array substrates and coverslips, while esthetically displeasing, have limited impact on data quality. The contribution of a background to the measurement noise level arises through the uncertainty in establishing the level to subtract, which is fundamentally determined by photoelectron shot noise from the background. The predicted effect of a wide range of background levels is illustrated in .
Subtraction of local background has a problematic aspect since it also subtracts background due to non-specific binding of the labeled nucleic acids to the array substrate. This correction may not be accurate because the non-specific affinities of the substrate and the array elements are likely to be different, as evidenced the occurrence of hybridizations in which the substrate affinity is so high that it is brighter than the array elements. The amount of this background signal can be estimated using the data in to determine the contribution from the camera offset and bare substrates, and subtracting that from the measured background levels. If the signal from the non-specific binding to the substrate is a substantial fraction of the apparent signal on the array elements, data quality may be compromised.
The maximum useful sensitivity of an array-imaging instrument is determined by the distribution of fluorescent contaminants in the measurement environment, including autofluorescence from the array elements and the like. Once these can be detected, increases in system sensitivity will not result in improved data since the measurement problem is not signal detection, but distinguishing the portion of the signal that represents the desired hybridization from the rest of the emission sources. Better data require improved experimental technique related to array production and hybridization, not increased instrument sensitivity. demonstrated that our instrument maintains a linear response down to fluorochrome densities well below 1 molecule/µm2, corresponding to tens of molecules per pixel. This limit was set by difficulty in producing a cleaner measurement environment. Extended integration times would allow us in principle to detect single fluorochromes, but the large pixel size results in a substantial number of other weakly emitting molecules in the pixel that overwhelm the desired signal.
The performance of the imaging system described here can serve to estimate what could be achieved with different designs that meet different optimization criteria. Alternative choices for excitation source and design of the excitation optics could improve performance and operational convenience, and potentially reduce cost. With the current system, adding anti-reflection coatings on the surfaces of the four elements of Lens 1 and on Lens 2, increasing the proportion of the excitation beam used for illumination by reducing the focal length of Lens 3, and using the reflector system to return excitation light to the array for a second pass, could increase the effective excitation intensity by more than a factor of 5, but at the potential cost of needing to institute computational corrections for chromatic variation in the illumination pattern. These changes would reduce the routine exposure times sufficiently so that the camera readout time, currently 4 s, would be a significant factor in determining the data acquisition rate. Given these simple changes, it should be possible to acquire DAPI, Cy3 and Cy5 images at >3 fields per minute. Alternative excitation designs using non-imaging optics, and potentially light emitting diode sources, might provide even higher useful illumination levels and/or operational simplicity compared to the arc lamp and imaging optics of the current excitation system.
The current detection optics are capable of collecting more than enough light to reach the practical sensitivity limits of arrays we are now using. These lenses were designed in two phases, with the more recently designed 105 mm focal length lens having reduced chromatic aberrations compared with the other two. Updating the design of the other lenses would slightly increase the performance of the current system, most notably by extending the useful wavelength range. Changing the optical design to increase the aperture would potentially allow compensating reductions in performance of other components of the system by collecting more light, but keeping the lens aberrations to acceptable levels would be difficult and expensive. Alternatively, reducing the aperture somewhat would reduce the lens costs and not compromise resolution since the current resolution is set by aberrations and not diffraction.
The camera used in this system provides quantum efficiencies >90% across the useful spectral region, has a large electron capacity for high dynamic range imaging, and is cooled to allow long integration times. We chose this camera so that we could obtain measurement accuracy of a few percent over a dynamic range of >1000 in signal intensity. The dynamic range can be extended if necessary by reducing the readout noise by slowing the readout time or using electron multiplying readout methods (Andor Technology EMCCD cameras), and/or summing multiple images. However, many applications have sufficiently intense signals and dynamic ranges limited to ~2 orders of magnitude, so that a lower performance, less expensive, camera would be adequate. If the excitation intensity, camera readout rate, and the speed of the stage and filter wheels were increased, and microarrrays with element densities that exploited the full-resolution capability of the optics were employed, highly quantitative multi-fluorochrome data could be obtained from tens of thousands of array elements per second with a CCD array imaging system.