Three microarray platforms: an analysis of their concordance in profiling gene expression

Background Microarrays for the analysis of gene expression are of three different types: short oligonucleotide (25–30 base), long oligonucleotide (50–80 base), and cDNA (highly variable in length). The short oligonucleotide and cDNA arrays have been the mainstay of expression analysis to date, but long oligonucleotide platforms are gaining in popularity and will probably replace cDNA arrays. As part of a validation study for the long oligonucleotide arrays, we compared and contrasted expression profiles from the three formats, testing RNA from six different cell lines against a universal reference standard. Results The three platforms had 6430 genes in common. In general, correlation of gene expression levels across the platforms was good when defined by concordance in the direction of expression difference (upregulation or downregulation), scatter plot analysis, principal component analysis, cell line correlation or quantitative RT-PCR. The overall correlations (r values) between platforms were in the range 0.7 to 0.8, as determined by analysis of scatter plots. When concordance was measured for expression ratios significant at p-values of <0.05 and at expression threshold levels of 1.5 and 2-fold, the agreement among the platforms was very high, ranging from 93% to 100%. Conclusion Our results indicate that the long oligonucleotide platform is highly suitable for expression analysis and compares favorably with the cDNA and short oligonucleotide varieties. All three platforms can give similar and reproducible results if the criterion is the direction of change in gene expression and minimal emphasis is placed on the magnitude of change.


Background
Completion of the human genome sequence has made it possible to study expression of the entire complement of 20,000-30,000 genes in a single assay. The two most common array platforms are based on collections of cDNA clones [1] or short (25 base) oligonucleotides synthesized in situ by photolithographic methods (i.e., by Affymetrix, Inc.) [2]. Partly because they are easy to use, microarrays are the most extensively used technology for studying gene expression on a global scale [3,4]. Thousands of expression studies employ one or the other microarray platform, but comparison of results between platforms has been difficult because of inherent differences in the array technologies. The situation became more complex as investigators began using long oligonucleotide arrays for expression profiling [5][6][7][8][9].
Because long oligonucleotide arrays for expression profiling are relatively new, we wished to validate them in relation to the cDNA and short oligonucleotide platforms, both of which have been used extensively in our laboratories over a number of years. The three platforms were evaluated using RNAs isolated from six cell lines and tested against a universal reference RNA. Sufficient RNA was isolated in a single harvest to supply labeling template for all experiments, so variability of RNA isolation was not an issue. However, no attempt was made to eliminate other sources of variation such as differences between lots of fluorescent dye label, microarray batch, operator, etc. We conducted these experiments under "normal" laboratory conditions so that one would not need to go to extreme lengths to reproduce the results. In almost all cases, results from the three platforms correlated reasonably well with each other. The Pearson correlation coefficients (r) ranged from 0.7 to 0.8. Because of different labeling methods and analysis algorithms, comparison of the cDNA and long oligonucleotide platforms with the short oligonucleotide system was not as straightforward, but in general all of the platforms were in reasonable agreement.

Results
This study was carried out to compare cDNA (Incyte), long oligonucleotide (Operon 70-mer), and short 25-mer (Affymetrix) array platforms, with the goal of qualifying the 70-mer arrays for general use at the National Cancer Institute. More specifically, we compared the Incyte Unigem2 set of human cDNAs (~9900 genes), the Operon human Version 2.0 set of long oligonucleotides (~21,329 genes), and Affymetrix HG-U133A arrays (~22,200 genes). RNA preparations from cell lines MCF10A, LNCaP, Jurkat, L428, SUDHL6, and OCI-Ly3 were used as probe templates, and the expression of each gene was compared directly with that of the same gene in the Human Universal Reference (HUR) RNA from Stratagene.

Genes in common across platforms
Only genes common to all three platforms were used in the comparison. Genes were matched by UniGene Cluster (UniGene Build #161), and unique cluster memberships were determined for each array type, as listed in Table 1 and enumerated in the Venn diagram in Figure 1. The intersection for the three platforms consisted of 6430 Uni-Venn diagram with number of genes present in each plat-form, genes in common between platforms, and genes in common among all three platforms Figure 1 Venn diagram with number of genes present in each platform, genes in common between platforms, and genes in common among all three platforms.  Gene clusters, and all analyses were carried out with all of these genes or a subset of them.

Comparison of expression ratios
An estimate of the concordance of the platforms was provided by the percentage of genes that showed expression ratios "in the same direction." All matched gene sets that had expression ratios (relative to the Universal Standard) greater than 1.0, irrespective of magnitude, were considered to be in the same direction, and the same for expression ratios less than 1.0. As seen in Table 2, the agreement was reasonably good, ranging from 74% to 82%. For comparison, ~50% concordance would be expected by chance. Many of the discordant values were at what might be considered "trivial" levels because a large proportion of ratios are near unity. For example, approximately 48%, 28% and 13% of the genes from Incyte, Operon and Affymetrix respectively, exhibit expression ratios within the range of 1.1 to 0.9. Thus, a ratio difference of only 0.95 versus 1.05 is designated a mismatch, which normally would be considered very concordant in the biological sense.
Statistical tests and thresholds on fold change are commonly used for the identification of altered gene expression by microarray assays. Hence, we estimated the statistical significance of differential expression of each RNA from the HUR by one sample Student-t test for a change from ratio of 1. Comparison between platforms using expression values significant at p < 0.05 shows concordance in direction of 93%-98% (Table 3 -no threshold). At the 1.5 and 2-fold threshold level, the agreement between platforms approaches 100% indicating that for reproducible data all three platforms are highly concordant in direction.  Table 5 indicate that the correlation coefficients between replicates within the same platform are closer, as   2 Comparisons between two platforms were done only when the expressions of both platforms met the p-value and threshold criteria. 3 Fraction of common expressions between two platforms to the total number of expressions found at p < 0.05 in both platforms.  expected, when compared to the correlations between platforms. In general, however, there was reasonably good concordance among the platforms.

Principal Component Analysis (PCA)
A projection on the first three principal components, which together explain 48.8 % (21%, 14%, and 13%) of the total variance, is shown in Figure 3. Close clustering of the cell samples is observed in this projection, indicating appreciable agreement among array platforms.

"Correlation of Correlations" analysis
The global concordance of the three platforms across all of the cell lines was estimated by calculating the "correlation of correlations" coefficient [10,11]. As seen in Figure 4ac, the correlations for the three platforms across all cells lines were quite good. The Pearson correlation of correlation coefficients was 0.965 between Operon and Incyte, 0.995 between Affymetrix and Incyte, and 0.956 between Operon and Affymetrix.

Clustered Image Map (CIM) visualization and analysis [12]
We used hierarchical clustering to demonstrate graphically the relationships among platforms, among cell lines, and among genes. 909 genes expressed at two times background or more in at least two of the six cell lines were included in the analysis. The resulting CIM ("heat map") is shown in Figure 5a-b. All three platforms cluster together for all six cell lines, as one would wish to find, and almost all of the gene expression values show reasonable concordance. Only a few exceptions can be seen in the cluster shown in Figure 5b. Some of the mismatches a-f. Scatter plot analysis to determine correlation coefficients between and within platforms using Jurkat RNA as an example Figure 2 a-f. Scatter plot analysis to determine correlation coefficients between and within platforms using Jurkat RNA as an example. Correlations for all cell lines are given in Table 4. (a) Operon versus Incyte (b) Affymetrix versus Incyte (c) Affymetrix versus Operon (d) GEM2 versus GEM2 replicate correlation (e) Operon versus Operon (f) HG-U133A versus HG-U133A may be due to simple errors in gene identification, rather than to the technologies of the platforms. The Incyte library is guaranteed by the manufacturer to be only 90% correct, and an unknown percentage of the Operon and Affymetrix oligonucleotides may have been designed on the basis of incorrect sequences in the public databases. Indeed, we found one oligonucleotide in the Operon set that was apparently designed from an EST sequence that has since been withdrawn from the UniGene database (see RT-PCR studies below). In any case, the concordance is quite high across all platforms with this method of analysis as well as with the others.

Quantitative real-time RT-PCR
In a pilot study with the three platforms, we compared and contrasted gene expression values for only the cell lines MCF10A and LNCaP. RT-PCR data for twelve genes are shown in Figure 6. Most of the values are in reasonable agreement except that there are differences in the magnitudes of the expression ratios. As found in other studies, the RT-PCR values are generally higher, probably because ratios are "flattened" with the microarray platforms. Affymetrix ratios are sometimes higher, but that may simply reflect the method of quantitation used in their analysis. The cDNA array ratios are generally lower than those Principal Component Analysis (PCA) of the three microarray platforms and six cell lines using expression of the 3186 genes with signals above background

L428 L428
from other platforms. Because the cDNA fragments are longer and double-stranded, they are more prone to nonspecific hybridization and can cross-hybridize more easily to related sequences. These characteristics of the probes may result in higher background signal and concomitant reduction in dynamic range of the ratios. In general, we have found that the long oligonucleotide arrays have a larger dynamic range than the do the cDNA arrays.
For further RT-PCR analysis, we chose a set of ten genes to test the accuracy of the three array platforms for all six cell lines. Those genes were chosen because there appeared to be a discrepancy among platforms in the direction of their ratios (i.e., whether they ratios were greater or less than unity). The results are shown in Figures 7a-f. Of special interest was gene ETR101, in which the Operon array was in disagreement across all cell lines. Further inquiry revealed that the sequence had been found to be incorrect and had been removed from the UniGene database. Since the oligonucleotide had been designed from the incorrect sequence, it is not surprising that it gave a different value.
Other discrepancies may be due to similar sequence errors, as even the most up-to-date databases are still being corrected and modified. In the case of AGL, the RT-PCR assay is in disagreement in several cases with two out of three of the array platforms; it appears to demonstrate an upregulation of the mRNA, whereas the arrays, with the exception of Incyte, point to a downregulation. Although RT-PCR is supposed to be the "gold standard" for measuring gene expression, this result shows that caution is indicated in interpreting results with even the PCR technology.

Discussion
The purpose of this study was to compare and contrast the three major microarray platforms, with the goal of qualifying the long oligonucleotide platform for general use in our laboratories. Our principal findings were that the magnitude of any particular expression ratio may differ from one platform to the other but that the "direction" of gene expression difference for genes with sufficient intensity of signal appears to be well correlated across platforms. The differences in apparent magnitude of expression were not unexpected. The platforms differ widely in many characteristics, including size of targets, number of targets, concentration of targets, labeling protocol, and quantitation algorithms. Nevertheless, the overall concordance of the three platforms was reasonably good, and one should be able to compare experimental data between and among the different array types. That is perhaps not surprising if one considers microarrays simply as miniaturized, multiplexed dot blots.

Only genes common to all platforms should be used in the analysis
That may seem an obvious requirement, but it is not always easy to meet in practice. For example, careful study of sequences present in various cDNA arrays indicates that 20-30% of spots may contain the wrong clone or a misidentified one [10,14,[38][39][40]. A similar situation exists for the short oligonucleotide platform (Affymetrix) because more than 19% of the sequences have been reported not to correspond to the appropriate mRNA Reference Sequence [41,42]. In these studies, measurements of cross-platform consistency were significantly improved when only sequence-matched genes were used. Similarly, discrepant results in studies using cDNA and Affymetrix platforms could be explained by errors in gene sequences [43]. The percentage of incorrect or misidentified sequences in the long oligonucleotide libraries is unknown, but we appear to have found one sequence in error during our own RT-PCR studies. Overall, then, significant disagreement in expression ratios among platforms may be due to sequence errors, not to intrinsic differences among array systems.

Only spots with signals significantly above background (and that pass stringent quality-control filters) should be used in the comparisons
With most high-density arrays, a large proportion of the genes can be expected to have signals near background, as probably fewer than half of all human transcripts are a-b. Clustered image maps showing patterns of expression relationship among genes, platforms, and cell lines  A subcluster expressed in any particular cell type or tissue [44]. Although stringent filtering decreases the number of measurements included in the analysis, the outcome will be more meaningful.
When comparing results from the three platforms, the magnitudes of ratios were often quite different, but there was generally good agreement in the direction of differences in expression (Figs. 6 &7, Tables 2 &3). Thus, it may be less productive to correlate absolute changes in expression than it is to look for agreement in direction.
Overall, the two oligonucleotide platforms were the most concordant pair. Possible reasons are (i) that the array targets are single stranded, all of the same size (25-mer or 70mer), with approximately the same melting temperature; (ii) the array elements have a much higher molarity of gene-specific sequences than do the cDNA arrays; (iii) that oligonucleotides tend to be more specific in terms of sequence and less prone to cross-hybridization than are cDNAs.

Conclusion
As the efficiency and economy of oligonucleotide arrays improves, they will probably become the platforms of choice for gene expression analysis, replacing the cDNA type entirely. Long-oligonucleotide arrays are being produced by ink-jet [6,45] and pin-spotting methods [26,27,31]. In these studies, where experiments specifically compared long-oligonucleotide arrays with the more "standard" platforms, correlations were good, and the long-oligonucleotide arrays performed as well as, or better than, the cDNA or short oligonucleotide variety. The data reported here confirm and extend those results, with the added advantage of comparing all three major platforms at one time and in the same place -something that, to our knowledge, has not been done before. As a note of caution, however, we have used materials from three commercial array platforms, and our results and conclusions Quantitative real-time RT-PCR analysis of 12 genes matched for direction of expression relative to the reference RNA for all three platforms  may not necessarily translate to other platforms and manufacturers.
Since this study began, many articles have appeared reporting the "meta-analysis" of microarray data from unrelated laboratories using several different array platforms [46][47][48][49][50][51][52][53][54]. Those studies have reported useful clinical and diagnostic findings when the data were stringently filtered before analysis. Eventually, standardization and simplification of the systems may lead to a situation in which only one or two robust methods are used in all laboratories, with a concomitant improvement in the accuracy of gene expression data.

Experimental design
Three array platforms were tested. Incyte cDNA arrays and Operon long oligonucleotide (70-mer) arrays were printed in house, and Affymetrix 25-mer HG U133A arrays were purchased from the manufacturer. The cDNA and long oligonucleotide arrays were assayed in quadruplicate, two each of red/green and green/red for dye "flips" (reverse fluor experiments) to help eliminate dye bias (except for L428 that had 3 replicates and OCI-Ly3 with 5 replicates). The Affymetrix arrays were tested in duplicate. RNA preparations from six cell lines were tested with each platform using a universal reference RNA (Stratagene) as the standard.

Preparation of cDNA and oligonucleotide arrays
Arrays with 10,000 cDNAs were prepared from ready-toprint UniGEM2 libraries obtained from Incyte, Inc.
(Wilmington, DE). Version 2.0 libraries containing 22,000 oligonucleotides of 70 bases in length were obtained from Operon, Inc. (Alameda, CA). Arrays were printed by standard protocols on Corning Ultra-GAPS II slides (Corning, NY) using a GeneMachine ® (San Carlos, CA) instrument. cDNAs were suspended at a concentration of 100 µg/ml and oligonucleotides at 25 µM in 3XSSC buffer, and the arrays were printed using SMP3 pins from Telechem International (Sunnyvale, CA). The spotted nucleic acids were fixed to the slides using protocols supplied by the manufacturers. The 25-mer oligonucleotide HG U133A Genechip ® arrays were purchased from Affymetrix, Inc (Santa Clara, CA).

Labeling and purification of targets
Labeled cDNAs from cell samples for hybridization to the long-oligonucleotide and cDNA arrays were synthesized and labeled by the indirect amino-allyl method, using reagents and protocols supplied by the manufacturer. For cDNA synthesis, we used Stratascript reagents (Stratagene, La Jolla, CA) and Cy3 and Cy5 fluorophore amino-allyl reagents from Amersham (Piscataway, NJ). Twenty micrograms of total RNA was used for each synthesis. Labeled cDNA targets were purified using Minelute purification kits (Qiagen, Valencia, CA). cRNA targets for the Affymetrix arrays were synthesized, labeled, and purified according to vendor's (Affymetrix) instructions. Briefly, 10 µg of total RNA was used to make double-stranded cDNA using reagents and protocols obtained from Invitrogen (Carlsbad, CA). Linear amplification was carried out by a modification of the Eberwine T7 method [56], and biotin was incorporated into the cRNA using the Enzo High Yield RNA Transcript labeling Kit (EnzoDiagnostics, Farmingdale, NY).

Hybridization and washing of arrays
The cDNA and long-oligonucleotide microarrays were prehybridized in 40 µl of 5XSSC with 0.1% SDS and 1% BSA at 42°C for 30 minutes. The prehybridization solution was removed, and arrays were hybridized for 16 hours at 42°C in 5XSSC buffer containing Cy3/Cy5 labeled targets, 25% formamide, 0.1% SDS, 1 µg Cot-1 DNA, and 1 µg poly A RNA. The cDNA arrays were washed at room temperature in 2XSSC, 0.1% SDS for 2 minutes, 1XSSC for 2 minutes, 0.2XSSC for 2 minutes and 0.05XSSC for 1 minute. The long-oligonucleotide arrays were treated the same except that the last wash step was omitted. The slides were dried by centrifugation at 650 rpm for 3 minutes. The Affymetrix arrays were hybridized and washed using the manufacturer's protocol. The arrays were then stained with streptavidin-phycoerythrin using the standard antibody amplification protocol (GeneChip ® Expression Manual, Affymetrix, Inc., Santa Clara, CA).

Array scanning and image processing
Long-oligonucleotide and cDNA arrays were scanned using an Axon 4000B scanner at 10-micron resolution. Images were processed, and signals from spotted arrays were quantitated using Genepix 3.0 software (Axon Instruments, Union City, CA). The Genepix result files, including signal, background, standard deviation, pixel statistics and quality parameters for both channels were deposited in the microarray database (mAdb) maintained by the NCI/CIT bioinformatics group [57]. The data were filtered on the basis of signal levels and spot quality. Local background values were subtracted from spot intensities to obtain signal values. Data were included if the signalto-background ratio was ≥ 2, the signal intensity was >100, the spot diameter was between 50 and 180 microns, at least 70% of the pixels were above their standard deviation and the spot was not flagged as "bad" visually. Arrays were normalized by median-centering the logarithmic ratios so that the median ratio of all genes that passed through the filters was equal to 1. For cDNA arrays, normalized expression ratios of 9050 genes were calculated, and the same procedure was applied to long-oligonucleotide arrays for the expression of 20,799 genes. Affymetrix HG-U133A arrays were scanned with the Affymetrix GeneArray scanner at 488 nm and 3-micron resolution. The images were analyzed using Microarray Suite 5.0 software (MAS5; Affymetrix Inc., Santa Clara, CA). Cell-line to HUR expression ratios were computed by comparative analysis of MAS5 values. The data were filtered using MAS5's signal detection and change calls generated at recommended default settings. The ratios included were those that had present calls for signal detection or an increase or decrease call associated with the ratio. The filtered data contained 17,647 genes. For all statistical calculations, logarithmic values of ratios to the base 2 were used.

Determination of genes in common among all platforms
Genes were matched by UniGene cluster methods [11,58], and expression levels were compared for only the 6,430 genes common to all platforms (Table 1 and Figure  1). UniGene clusters (Homo sapiens: UniGene Build #161) of probes of all three platforms were determined by the NCI/CIT Bioinformatics group [57] using the BLAST program from the National Center for Biotechnology (NCBI, Bethesda, MD 20894). There were multiple probes for some of the UniGene clusters. This resulted in matching of two or more probes of one platform to one or more probes of another platform. All possible combinations across three platforms were considered for each UniGene cluster. One probe from each of the platforms was selected as follows. Initially, for a given combination, the replicates were averaged to obtain expression patterns of six RNAs on three platforms. The sum correlation coefficients of these three patterns to their mean pattern was determined. The combination having the highest sum was selected for further analysis. This method relies on the assumption that the probes specific for a gene yield similar expression patterns independent of the platform. The matched expression ratios will be made available at our website http://nciarray.nci.nih.gov/publications/.

Estimation of total matched versus mismatched expression values
Expression ratios for genes in common across all cell lines and platforms were determined. If the ratio of a gene was ≥ 1 or ≤ 1 for both platforms being compared, the expression was considered matched irrespective of the magnitude. Otherwise, the ratios were considered to be mismatched (i.e., in opposite directions). These values give a rough, binary indication of the correlation between platforms (Table 2). Concordance between platforms using significant expression ratios at p-value <0.05 and 1.5 to 2-fold threshold levels are given in Table 3.

Correlation of expression among all platforms
To determine how well the data from the three platforms coincided, correlation coefficients were obtained from analysis of scatter plots of the mean expression values from the three array formats.  Tables 4 &5).

Principal Component Analysis (PCA)
The global gene expression patterns of the six cell lines in all three platforms were studied by principal component analysis (PCA) [59]. All genes (3186) with signals above background were included. Differences in signal magnitude among platforms were nulled out by normalizing the data from each to unit standard deviation. Individual platform variations were accounted for by employing unit variance. A projection on the first three principal components, which explained 48.8% (21%, 14%, and 13%) of the total variance, is shown in Fig. 3. The calculations were done using Partekpro 5.0 software (Partek Inc., St. Charles, MO)

Platform concordance by "correlation of correlations" coefficient
The global concordance of expression levels of the three platforms can be expressed in terms of the 'correlations of correlations' coefficient described previously [10,11]. To perform the computations, step 1 was calculation of the Pearson correlation coefficient across all matched genes for each of the 15 possible pairs of cell lines for each platform.
Step 2 was calculation of the Pearson correlation coefficients of those correlation coefficients for the three possible ways of pairing the three platforms. The results are shown in Figures 4a-c.

Hierarchical cluster analysis and Clustered Image Maps (CIMs)
Hierarchical clustering of individual replicates (data not shown), including 3662 genes detected in 80% of the arrays revealed a grouping of RNA samples independent of the type of array platform. A set of 909 genes expressed at >two-fold levels in all platforms in at least two cell lines was used for hierarchical cluster analysis to determine how closely the genes, cell lines and platforms corresponded in their expression values. As a distance metric, we used 1-r, where r is the Pearson correlation coefficient [60]. Cluster nodes were determined using an average linkage algorithm. In the resulting CIM (heat map) [14a], up-and down-regulation with respect to the reference RNA are color-coded as red and green, respectively (Figure 5a-b).

Quantitative real-time RT-PCR
RNA preparations from each of the six cell lines and Stratagene Human Universal Reference RNA were converted into single-stranded cDNA using the Applied Biosystems High-Capacity cDNA Archive kit (ABI, Foster City, CA). Primer and probe sets were obtained as ABI Assays-on-Demand™ Gene Expression Products (TaqMAN ® MGB probes, FAM™ dye-labeled) for a set of genes to be studied, as well as GAPDH and BACT, which were used as comparative controls. All quantitative PCR reactions were performed in quadruplicate, with two carried out in an ABI Prism ® 7000 sequence detection system and two in a Corbett Research R-300 instrument (Corbett Research, Sidney, Australia). The results were analyzed using the "Relative Quantitation of Gene Expression" method described in ABI Prism 7700 Sequence Detection System User Bulletin #2, Rev B. An initial study comparing only two cell lines, LnCaP and MCF10A, was carried out for the three platforms. Twelve genes that matched in direction of change were chosen for Q-PCR analysis, and the results are shown in Figure 6. Subsequently, ten genes were chosen for analysis using all six cell lines and three platforms. Genes in that set were deliberately chosen for mismatched ratios to determine if any platform was in error more often than the others. The results are shown in Figures 7a-f.