Quantitative analysis of DNA array autoradiographs

Abstract

DNA arrays and chips are powerful new tools for gene expression profiling. Current arrays contain hundreds or thousands of probes and large scale sequencing and screening projects will likely lead to the creation of global genomic arrays. DNA arrays and chips will be key in understanding how genes respond to specific changes of environment and will also greatly assist in drug discovery and molecular diagnostics. To facilitate widespread realization of the quantitative potential of this approach, we have designed procedures and software which facilitate analysis of autoradiography films with accuracy comparable to phosphorimaging devices. Algorithms designed for analysis of DNA array autoradiographs incorporate 3-D peak fitting of features on films and estimation of local backgrounds. This software has a flexible grid geometry and can be applied to different types of DNA arrays, including custom arrays.

INTRODUCTION

Two general classes of DNA matrices for hybridization-based expression profiling are in widespread use: filter arrays where a matrix of probes is loaded on a flexible membrane, such as nylon, and higher density ‘chips’ on a rigid substrate, such as silicon or glass. Data from flexible arrays are obtained by autoradiography, using film or phosphorimaging. In each case the result of the readout procedure is an image which must then be analyzed either visually or using appropriate computer software. Even for relatively small arrays, visual analysis involves many hundreds of image features and quickly becomes impractical. The potential for quantitation based on visual analysis is also limited. Software analysis of autoradiographs is thus essential for all but the most rudimentary applications.

The lowest cost, and therefore most widespread, experimental technique for imaging of cDNA arrays is film-based autoradiography. Film autoradiography is not generally considered a quantitative method of analysis, since the response of film to radiation is not only non-linear, but also dependent on a number of experimental conditions that are difficult or inconvenient to control. Most film-based analysis is limited to binary decisions: if a signal in a given location is much darker (less dark) on one film than another, the corresponding gene is significantly up (down)-regulated. DNA array film autoradiography is thus largely a screening technique to identify differences in expression, which are then quantified using other experimental methods. Adryan et al. (1) have recently described image processing techniques which encode information from two autoradiographs as different colors. When the images are registered and combined, the resulting colors encode the relative intensity of corresponding positions in the two images and thus the relative expression levels of the corresponding genes. The observer is relieved of the need to recognize whether one spot is darker than another, instead relying on colors in the combined images. Although these authors have accounted for a number of experimental problems associated with film autoradiography, their technique is still described as ‘semi-quantitative’.

One potentially powerful use of cDNA arrays is to identify and analyze patterns in the expression of multiple genes (2,3). In order for patterns to be compared from one cell line, time point or other experimental condition to another, it is necessary for the underlying expression values to have a reasonably high degree of numerical accuracy and consistency. If radioactive filter arrays are to be analyzed in these terms, development of methods which allow reliable quantitative analysis of autoradiographs is needed. There are several conditions which data acquisition procedures and analysis software for autoradiography ideally should fulfill. First, the approach should be sensitive enough to detect changes in low abundance mRNA, which produce faint signals on the films. Second, a wide range of linearity should allow simultaneous comparison of changes in both high and low abundance signals. Finally, comparison and normalization of data obtained at different times with different samples of RNA should be possible. Each of these comparisons needs to be made with as much accuracy as possible and with reasonable estimates of uncertainty in the results.

Quantitative radiation dosimetry using film is a standard technique in radiotherapy physics. The calibration procedures and experimental methods required to achieve consistent accuracy of measured radiation dose of the order of a few percent are well established (4,5). These methods are particularly well suited for electron dosimetry, where difficulties due to the variation in response of the film at very low energy do not pose a problem, as they do in X-ray dosimetry. For β-emitters such as ³²P, the radioisotope used in most cDNA hybridization experiments, the highly developed technology of quantitative electron film dosimetry used in radiotherapy is directly applicable to the autoradiography analysis problem. In this paper we describe an approach for reading and analysis of DNA array film autoradiographs, differing from others which have been described in that it incorporates experimental procedures adapted from quantitative electron film dosimetry as utilized in radiotherapy. The essential features of this technique are as follows. First, the film scanner is calibrated in terms of physical units, namely optical density (OD), which can be directly related to dose. Second, optical density is further calibrated in terms of radioactivity, by exposure to specially prepared arrays with known amounts of activity. Finally, quantitative analysis is based on fitting of observed signals with a peak shape hypothesis and a non-uniform background ‘fog’ level, after application of these calibration steps. The advantages this approach brings to analysis of DNA arrays are improved quantitative accuracy and the ability to work with a wide range of film exposures.

These techniques are directly applicable to radiolabeled filter arrays such as those produced by Clontech and Research Genetics. Due to the inherent flexibility of the analysis software, it can readily be applied to any format of custom filter-based DNA array.

MATERIALS AND METHODS

Human glioma cell lines were maintained at 37°C in 5% CO₂ in RPMI 1640 medium supplemented with 10% fetal calf serum, 2 mM glutamine, 50 U/ml penicillin, 50 mg/ml streptomycin and non-essential MEM amino acids (Gibco BRL, Gaithersburg, MD).

Preparation of RNA was performed as described elsewhere (6), with a few modifications. Total RNA was extracted with TRIzol reagent (Gibco BRL) according to the manufacturer’s instructions. DNase I (Worthington) cleavage was performed at concentrations of 5 U/µl, in 5 mM MgCl₂-containing buffer. RNA samples were normalized by concentration and further checked by electrophoresis in neutral 1.5% agarose in TAE buffer, with preliminary denaturation in formaldehyde/formamide sample buffer at 65°C for 5 min. In subsequent experiments for synthesis of cDNA we used samples with intact 18S and 28S RNA. cDNA synthesis was performed with MMLV reverse transcriptase (Gibco BRL), in the presence of gene-specific primers (Clontech) and [³²P]dATP, following the manufacturer’s protocols. Each sample was normalized by amount of RNA and contained 5 µg RNA, prepared as described above. Aliquots of 20 × 10⁶ c.p.m. of cDNA were used in hybridizations for all results described here.

Autoradiography was performed using Kodak BioMax MS film and Kodak HE intensifying screens at –80°C. At least three films were exposed for each membrane, for periods of 16, 48 and 72 h. An automatic film developer with standard reagents (Kodak) was used.

Films were scanned using a Vidar VXR-8 digitizer (Vidar Systems, Herndon, VA). Each film was scanned together with an OD standard (Kodak Step Tablet 609 ST150) with 21 calibrated steps from diffuse density 0.04 to 3.04. Data were transferred to a Silicon Graphics Indigo-2 or O2 workstation for analysis using software developed with the AVS5 visualization toolkit (AVS, Waltham, MA).

Activity standards were prepared using Millipore positively charged nylon membranes. Grids were laser printed directly onto the membranes. [³²P]dATP was diluted, counted in a liquid β-spectrometer (Beckman LS 3801) and loaded on the membranes with microsyringes or micropipettes. Quantities of from 0.6 to 10³ d.p.m. were used in calibration experiments. Each standard was exposed to film for 16, 72 and 120 h. Corrections for decay were made for membranes prepared or exposed at different times. With 72 or 120 h of exposure at –80°C as little as 0.6 d.p.m. of radioactivity could be detected.

The software described in this paper is available upon request through the corresponding author.

RESULTS AND DISCUSSION

An example film autoradiograph for human glioma cell line U87MG is shown in Figure 1, where the upper panel corresponds to untreated (control) cells and the lower to cells γ-irradiated with 10 Gy. In a typical experiment ∼150 signals corresponding to expressed messengers were detected. Differences between signals within a single array varied over a wide range (compare positions A3a and F4d in the upper panel). On the other hand, differences in the same signal before and after treatment may be much less pronounced (compare position A6f in the upper and lower panels).

Figure 1.

Autoradiograph of Clontech Atlas Human Expression Arrays for human glioma cell line U87MG. (Top) Untreated cells. (Bottom) Cells irradiated with 10 Gy ⁶⁰Co γ-rays.

The OD standard was scanned together with each autoradiograph. The characteristic scanner curve of pixel value versus actual OD was thus obtained and used to convert the scanned autoradiograph to OD units with an accuracy of the order of 1 or 2%. Example characteristic scanner curves from three independent measurements at different times are illustrated in Figure 2. Note that scanner pixel values are related to brightness, so high OD (dark) regions have low pixel values. The usable dynamic range of the film digitizer was ∼2.7 OD.

Figure 2.

Film scanner calibration curves from five scans between January 1999 and March 2000.

There are two issues we wish to address through calibration: accuracy and linearity. If the analyzed signal is proportional, for a given time of exposure, to the amount of radioactivity, then the ratio for the same location on two arrays should be equal to the ratio of activity of hybridized cDNAs, a direct measure of relative change in gene expression. If in addition we can establish a calibration of the film OD produced by a known amount of ³²P activity, then we can produce a measure from our autoradiographs that is not only proportional, but can also be converted to units of absolute quantity of labeled cDNA in each cell. To address these needs, we used standard amounts of radioactivity loaded on nylon membranes (see Materials and Methods). Figure 3 shows the correlation between analyzed integrals of specific peaks and their known activity. Linearity is observed over the entire range of activities utilized, spanning three orders of magnitude between 0.6 and 780 d.p.m. In Figure 3B the raw data (gray) and fitted function (red) are plotted as 3-D surfaces for a portion of the activity standard array. In this mode of presentation one can readily appreciate the quality with which the fitted functions match the film data.

Figure 3.

Activity calibration data and results. (Top) Example image of activity standard with grid overlay. (Center) Surface plot of image optical density (grey scale) and fitted peaks (color). (Bottom) Analysis output plotted versus known activity.

A key to achieving good results is the use of image processing techniques to obtain a reliable estimate of the integrated activity in each array cell from OD, taking into account non-uniform background density (‘fog’), peak spreading and film saturation. The film used in autoradiography is extremely sensitive and thus prone to background darkening, which is difficult to control and varies across the film. For faint signals it is important to make an accurate correction for this local background. Individual signals are also non-uniform; darker in the center, fading to the background level far from the center. As the exposure becomes longer and/or the activity greater, the central density becomes higher and the low density fringe larger in area. To obtain a signal linear with radiation exposure, it is necessary to integrate over the entire spot, since all density above background is presumably due to the radiation. For cells containing low and moderate density spots, an easily calculated measure is the sum of the OD at each pixel within the cell minus the estimated background at each pixel.

As the exposure increases, the peak OD saturates at a value which depends at least on the film type and processing technique. For saturated peaks, simply summing the OD no longer estimates radiation exposure. Instead, we fit a mathematical model of the expected peak shape, ignoring those central pixels which are saturated. The analytical integral of the fitted function provides a measure of integrated peak density, the peak of the fitted function filling in the missing (i.e. saturated) central pixels of the observed peak. Fitting a function which has a peaked component, added to a locally uniform component, the latter component also estimates the local background. Thus the image intensity on the film is modeled as follows:

I(x, y) = I_bg + ∑^Nc_{j = 1}I_j(x, y)

where I(x, y) is the total OD at pixel position (x, y), I_bg is a global constant background value, N_c is the total number of cells and I_j(x, y) is the contribution of cell j to the intensity at point (x, y). The contribution of each cell is modeled as a sum of Gaussian peaks plus a local background:

I_j(x, y) = I_bj(x, y) + ∑^Np_{i = 1}a_ij exp{–[(x – x_ij)² + (y – y_ij)²]/σ_ij²}

where N_p is the number of peaks per cell, I_bj(x, y) is the local background, constant within cell j and 0 outside, a_ij is the amplitude of the ith peak within cell j, (x_ij, y_ij) is the center position of the ith peak in cell j and σ_ij is the r.m.s. width of the ith peak in cell j.

The image intensity model thus has three distinct components: (i) a global constant background value, I_bg; (ii) a mosaic of local backgrounds, constant within each cell and varying from cell to cell; (iii) one or more individual Gaussian peaks per cell. Figure 4 illustrates three examples of peak behavior which frequently occur. Figure 4A is the ‘normal’, well-behaved situation where symmetrical peaks are well separated, with surrounding low density pixels which approximate the local background. In Figure 4B the features are not symmetrically positioned within the cell and correspond to background noise or spurious signals. If the analysis does not locate N_p similar peaks along the cell center line, the cell is not counted. Very intense multiple signals can merge into a single spot, which may extend beyond the cell border, as in Figure 4C. Here, the algorithm models the large merged spot as a superposition of multiple peaks. The center of the image spot in this situation is also typically saturated. Pixels with OD above a saturation value are assigned very low weight in the fitting procedure, since their OD inaccurately represents their radiation exposure.

Figure 4.

Typical appearances of spots in DNA array cells. (Left) Nominal appearance, two well-separated symmetrical spots of nearly equal intensity. (Center) Artifact, unmatched spot. (Right) Overexposed, two spots merged into a single feature.

As shown above, using this model areas of fitted peaks for our activity standards yield a measure of exposure which is linear over at least three orders of magnitude in activity, for a given exposure time. Errors from peak saturation at high activity are effectively overcome. Furthermore, linearity persists down to very low activity, suggesting that weak signals can be reliably measured. Therefore, changes in expression which are large in relative terms, but to signals which are small in absolute terms, can be quantified. Both small and large changes in signals which are very strong can also be quantified, which would normally be impossible due to saturation (data not shown). Finally, overall linearity of the computed signal with exposure means that signals from films exposed to the same array but for different times can be reliably compared or combined, using one or more peaks for normalization. Short exposure films with unsaturated peaks from strong signals can be combined with long exposure films, which reveal peaks due to weaker signals. This would be impossible without use of the procedures described above.

To verify the linearity of our analysis with actual array data, we compared our film results with phosphorimage analysis, which does not suffer from saturation and fog effects. The same arrays were analyzed using a Storm phosphorimager (Molecular Dynamics) and film. Figure 5 compares the signals detected and quantified by phosphorimaging with the same signals detected and quantified by film. After 16 h exposure the gross correlation of film with phosphor signals is very strong (r² = 0.9773) over the entire range of detected signals. A logarithmic plot (Fig. 5, bottom) shows that this good fit persists down to very low signal intensities. However, after 48 h exposure the overall correlation is significantly decreased. Examination of the logarithmic plot reveals two regions of lower quality fit after 48 h exposure; the most intense signals and some of the peaks in the low intensity region. At the high intensity end, even after 48 h exposure, when the film was significantly overexposed, only six of 588 cells were sufficiently saturated that they were not accurately analyzed. Of course, these features can be easily quantified after shorter times of exposure and excluded from analysis of the overexposed films. Correlation of 48 h data with phosphor screen data after exclusion of the six most overexposed signals produced a significantly better result (r² = 0.8847; plotted in Fig. 5, center) than when the six strongest signals were not excluded. Weak signals with values close to background levels may present a more serious problem than overexposed signals. First, they often produce false positive values of changes in comparisons of control and treated samples. Second, Figure 5 shows that such weak signals, which do not correlate well with the phosphor data, are more numerous than oversaturated features. Our data show that there are two major sources of such dispersion in weak signals; location of a weak peak adjacent to a strong one and significant fluctuations in local background (non-uniform washing, mechanical damage to membranes or films, etc.). In Figure 5 we did not exclude any low intensities, however, our experience as well as other published studies show that the weakest signals should be removed by thresholding to minimize false positive values (7). The dashed line in Figure 5C has a slope of 1.0 on the log plot, indicating that outside the low intensity region the 48 h film analysis results and the phosphorimager data are linearly related. With preliminary filtering of the data and application of peak fitting functions other than Gaussian, measurements of overexposed films will be further improved. However, the presented data show that with appropriate ranges of exposure (16–24 h) our current approach allows reliable quantitation of DNA array data detected by autoradiography.

Figure 5.

Comparison of phosphorimager output with film peak areas using the present method. The same array was imaged with a Storm phosphorimager and with film at two exposures, 16 and 48 h (see explanation in text). (Top) 16 (left) and 48 h (right) films. Note severe oversaturation of the 48 h film. (Center) Plot of peak volumes from film analysis versus corresponding phosphorimager peak volumes. (Bottom) The same data plotted on a logarithmic scale.

These data show that with application of appropriate procedures and software, the precision of analysis based on film autoradiographs may be comparable with that available using electronic data acquisition, but with significantly decreased expense. Another advantage of this approach is the ability to accomodate any array format, including custom arrays. Rapidly growing demand for DNA array technology may establish approaches such as we have described here as important research tools.