ANALYTICAL BIOCHEMISTRY Analytical Biochemistry 353 (2006) 43–56 www.elsevier.com/locate/yabio
Strategies for comparing gene expression proWles from diVerent microarray platforms: Application to a case–control experiment Marco Severgnini a,b, Silvio Bicciato c, Eleonora Mangano b,d, Francesca Scarlatti b,e, Alessandra Mezzelani b, Michela Mattioli f, Riccardo Ghidoni b,e, Clelia Peano a,b, Raoul Bonnal a,b, Federica Viti a, Luciano Milanesi a,b, Gianluca De Bellis a,b, Cristina Battaglia b,d,¤ a Institute of Biomedical Technologies, National Research Council, Milan, Italy Interdisciplinary Center for Biomolecular Studies and Industrial Applications (CISI), University of Milan, Milan, Italy c Department of Chemical Process Engineering, University of Padua, Padua, Italy d Department of Sciences and Biomedical Technologies, University of Milan, Milan, Italy e Laboratory of Biochemistry and Molecular Biology, San Paolo University Hospital, University of Milan, Milan, Italy f Laboratory of Experimental Hematology and Molecular Genetics, Department of Hematology 2, Ospedale Maggiore, IRCCS, Milan, Italy b
Received 30 November 2005 Available online 3 April 2006
Abstract Meta-analysis of microarray data is increasingly important, considering both the availability of multiple platforms using disparate technologies and the accumulation in public repositories of data sets from diVerent laboratories. We addressed the issue of comparing gene expression proWles from two microarray platforms by devising a standardized investigative strategy. We tested this procedure by studying MDA-MB-231 cells, which undergo apoptosis on treatment with resveratrol. Gene expression proWles were obtained using highdensity, short-oligonucleotide, single-color microarray platforms: GeneChip (AVymetrix) and CodeLink (Amersham). Interplatform analyses were carried out on 8414 common transcripts represented on both platforms, as identiWed by LocusLink ID, representing 70.8% and 88.6% of annotated GeneChip and CodeLink features, respectively. We identiWed 105 diVerentially expressed genes (DEGs) on CodeLink and 42 DEGs on GeneChip. Among them, only 9 DEGs were commonly identiWed by both platforms. Multiple analyses (BLAST alignment of probes with target sequences, gene ontology, literature mining, and quantitative real-time PCR) permitted us to investigate the factors contributing to the generation of platform-dependent results in single-color microarray experiments. An eVective approach to cross-platform comparison involves microarrays of similar technologies, samples prepared by identical methods, and a standardized battery of bioinformatic and statistical analyses. © 2006 Elsevier Inc. All rights reserved. Keywords: Platform comparison; High-density microarray; Comparison strategy
Platform comparison of microarray measurements is one of the most interesting development Welds of arraying technology, as evidenced by the exponential increase in the number of articles on the subject [1]. Meta-analysis of microarray data from diVerent sources and platforms is
*
Corresponding author. E-mail address:
[email protected] (C. Battaglia).
0003-2697/$ - see front matter © 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.ab.2006.03.023
becoming increasingly important, especially considering the accumulation in public repositories of data sets from diVerent laboratories concerning the same biological questions. However, there is no consensus on the best methods for meta-analysis of microarray data. The agreement of data sets generated using diVerent types of microarray slides and platforms has been studied previously, yet results are confounding and conXicting. Several studies reported poor superimposition of results obtained
44
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
from two or three commercial arraying platforms [2–5]. Others found a better superimposition of commercial microarray platforms, with elevated correlation coeYcients [6–9]. Moreover, the protocols used to compare two or more microarray technologies often are case dependent. Variables that render platform comparisons diYcult include widely disparate technologies and diVerences in both the preparation of samples and the analysis of microarray raw data. In the interest of addressing critical issues in microarray cross-comparison, we designed and tested a standardized strategy to compare transcriptional analyses from two commercial single-color microarray platforms. We selected platforms that used similar technologies. AVymetrix’s GeneChip Expression Analysis Arrays [10] and Amersham’s CodeLink Bioarrays [11] both employ short-oligonucleotide DNA microarrays and a one-color detection protocol, obtained with streptavidin conjugated with phycoerythrin (PE, GeneChip)1 or cyanine 5 (CodeLink). Moreover, to have comparable RNA samples for investigation on the two platforms, we studied changes in gene expression in a metastatic breast cancer cell line (MDA-MB-231), that undergoes apoptosis when treated with the antitumoral agent resveratrol [12]. Our investigative strategy employed a battery of bioinformatic and statistical analyses, permitting us to study the extent to which gene expression proWles are platform dependent, how sequences of microarray probes inXuence outcomes, and whether gene expression signatures from microarray technology are reliable and biologically relevant. In this article, we illustrate how this meta-analysis approach is eVective in studying the factors that lead to the generation of platformspeciWc results.
RNA and complementary RNA preparation
Materials and methods
Amersham CodeLink hybridization, processing, and analysis
The human breast cancer cell line MDA-MB-231 was obtained from American Tissue Culture Collection (Rockville, MD, USA) and maintained in Dulbecco’s modiWed Eagle’s medium (DMEM) supplemented with 5% fetal bovine serum, 100 mg/ml streptomycin, and penicillin at 37 °C and 5% CO2. Resveratrol (Cayman Chemical, Ann Arbor, MI, USA) was prepared in ethanol. To evaluate the eVects of resveratrol on gene expression, cells were seeded in six-well tissue culture plates at 4 £ 105 cells/well and allowed to adhere for 24 h prior to treatment. Then resveratrol was added to a Wnal concentration of 32 M, and the cells were cultured for 48 h prior to use (treated samples). In control samples, vehicle was added.
CodeLink UniSet Human I Bioarrays were obtained from GE Healthcare (formerly Amersham Bioscience, Piscataway, NJ, USA). These DNA microarray slides contain 10,461 oligonucleotide probes (30 bases in length), of which 513 are internal controls and 9948 are “discovery” probes. Probe sequences are designed on human transcripts deposited in the National Center for Biotechnology Information’s Reference Sequence (RefSeq) and dbEST databases and on some sequences from LifeSeq Gold and Foundation databases (Incyte, Wilmington, DE, USA). The four samples of fragmented biotinylated cRNA (10 g cRNA each) were prepared for hybridization to the bioarrays using the Expression Assay Reagent Kit (Wnal volume, 260 l). The hybridization solution was heated at 90 °C for 5 min to denature the cRNA, chilled on ice, and vortexed vigorously for 5 s, and 250 l was injected into the inlet port of the hybridization chamber, previously placed in a 12-slide shaker tray. The hybridization chamber ports were sealed with 1-cm sealing strips, and the shaker trays were loaded into a shaking incubator (Innova 4080, New Brunswick ScientiWc, Edison, NJ, USA) with the hybridization chambers facing up. Slides were incubated for 18 h at 37 °C while shaking at 300 rpm.
1 Abbreviations used: PE, phycoerythrin; DMEM, Dulbecco’s modiWed Eagle’s medium; cRNA, complementary RNA; cDNA, complementary DNA; PMT, photomultiplier tube; dChip, DNA–Chip Analyzer; HGUI33A, Human Genome U133A; EDTA, ethylenediaminetetraacetic acid; BSA, bovine serum albumin; MAS, Microarray Suite; P, present; FU, Xuorescence units; RMA, robust multiarray average; PM, perfect match; A, absent; MIAME, Minimum Information about Microarray Experiment database; SAM, signiWcance analysis for microarray; DEG, diVerentially expressed gene; qPCR, quantitative real-time PCR.
RNA and complementary RNA (cRNA) were prepared using a single procedure to avoid diVerences introduced by speciWc target preparation protocols. BrieXy, total RNA was puriWed from a total of four independent wells of MDA-MB-231 cultures, treated with resveratrol (n D 2) or vehicle (n D 2), using TRI Reagent (Sigma, St. Louis, MO, USA). An additional puriWcation step was carried out using the RNAeasy kit (Qiagen, Valencia, CA, USA). Total RNA samples were quality-checked by microcapillary electrophoresis using the 2100 BioAnalyzer (Agilent, Palo Alto, CA, USA). Double-stranded complementary DNA (cDNA) was synthesized from 5 g total RNA using the Superscript Choice System (Invitrogen, Carlsbad, CA, USA). cDNA was puriWed by phenol–chloroform extraction and ethanol precipitation. In vitro transcription and biotin labeling of cRNA were carried out using the Bioarray HighYield RNA Transcript Labeling Kit (Enzo Life Sciences, Farmingdale, NY, USA). The reaction mix was incubated for 5 h at 37 °C, and cRNA was puriWed using the RNAeasy kit. cRNA was quantiWed by absorption at 260 nm and checked for quality by microcapillary electrophoresis. A typical cRNA sample had a size range of 500–3000 bases. The four samples of biotinylated cRNA (two treated and two controls) were fragmented in 40 mM Tris–acetate (TrisOAc, pH 7.9), 100 mM KOAc, and 31.5 mM MgOAc at 94 °C for 35 min. The fragmented cRNA samples were checked for quality on GeneChip Test3 array prior to the comparative study of two microarray platforms.
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
After hybridization, the 12-slide shaker tray was removed from the incubator and the hybridization chamber was removed from each slide. Each slide was brieXy rinsed in TNT buVer (0.1 M Tris–HCl [pH 7.6], 0.15 M NaCl, and 0.05% Tween 20) at room temperature and then washed in TNT buVer at 42 °C for 60 min. The signal was developed using a 1:500 dilution of streptavidin–Cy5 (GE Healthcare) in TNT buVer for 30 min at room temperature. Excess dye was removed by washing four times with TNT buVer, for 5 min each, at room temperature. Slides were rinsed in deionized water and dried under nitrogen. The processed slides were scanned on an Axon GenePix Scanner (Axon, Molecular Devices, Union City, CA, USA) with the laser set at 635 nm, the photomultiplier tube (PMT) at 600 V, and the scan resolution at 10 m. CodeLink Expression Analysis Software (version 4.1, Amersham Bioscience) was used to analyze images for each slide. Signal intensities of spots were normalized using the DNA– Chip Analyzer (dChip) invariant set algorithm [13,14]. Spot quality was evaluated using data Xagging from CodeLink Expression Analysis Software. The software allows investigators to overcome the problem of donut-shaped spots (the standard shape of CodeLink Bioarray features) with diYcult identiWcation of signiWcant pixels in each feature. Spots with intensity below that of the negative control (absence of an oligonucleotide probe) were excluded, as were those with irregular shapes or near-background intensities. In particular, we required that each included spot had intensity above the negative control threshold for at least one control sample and one treated sample and had a “good” Xag in at least two of four arrays.
45
GeneChips and replaced with 200 l hybridization solution. Hybridization took place in the Hybridization Oven 640 with rotation at 60 rpm for 16 h at 45 °C. Washing and Xuorescent labeling were performed on the Fluidics Station FS400 (AVymetrix) according to protocol EukGE-WS2 for eukaryotic samples. Staining was Wrst achieved with streptavidin-conjugated PE and was then ampliWed immunochemically with biotinylated anti-streptavidin antibody followed by streptavidin-conjugated PE Readings were carried out on the G2500A GeneArray scanner (Agilent). Data quality was controlled using Microarray Suite (MAS, version 5.0), monitoring average feature signal per chip, average background, percentage of present (P) features in the array, and 3⬘:5⬘ probe ratios. The average signal was greater than 500 Xuorescence units (FU) for control samples and greater than 700 FU for treated samples, with background intensity one order of magnitude less than the average signal. All four samples had a percentage of P features near 50% (commonly accepted threshold) and a 3⬘:5⬘ ratio for actin and GAPDH controls less than 2.8. Probe-level data were converted to expression values using the robust multiarray average (RMA) procedure [15]. Intensities of perfect match (PM) probes were backgroundadjusted and normalized using the invariant set procedure. Probe sets having an absent (A) call on all four arrays were used to estimate a threshold, Wxed at the 95th percentile of their signal distribution (190 FU), for including the remaining probe sets in the analysis. Probe sets with A calls in all four arrays and a mean intensity in both control and treated samples of less than 190 FU were Wltered out.
AVymetrix GeneChip hybridization, processing, and analysis
Microarray data analysis: intra- and interplatform comparisons
GeneChip Human Genome U133A Array (HG-U133A) from AVymetrix (Santa Clara, CA, USA) is a microarray of 22,215 probe sets, each composed of 11 pairs of 25-base probes (11 “perfect match” and 11 “mismatch” probes). The probe sets represent 68 internal controls and 18,400 human transcripts and variants deposited in GenBank, dbEST, and RefSeq databases. Prior to use, the microarrays were Wlled with 250 l prewarmed (45 °C) 1£ Mes hybridization buVer (100 mM Mes, 1 M NaCl, 20 mM ethylenediaminetetraacetic acid [EDTA], and 0.01% Tween 20) and placed for 15 min at 45 °C in the Hybridization Oven 640 (AVymetrix). Following the manufacturer’s instructions, we prepared a hybridization solution containing 150 l 2£ hybridization buVer, 5 l control oligonucleotide B2 (3 nM), 15 l 20£ eukaryotic hybridization controls (BioB, BioC, BioD, and Cre), 3 l herring sperm DNA (10 mg/ml, Promega, Madison, WI, USA), 3 l acetylated bovine serum albumin (BSA, 50 mg/ml), 15 g fragmented biotinylated cRNA, and H2O to a Wnal volume of 300 l. The hybridization solution was denatured for 5 min at 99 °C, incubated at 45 °C for 5 min, and then centrifuged at 14,000 rpm for 5 min. The Mes hybridization buVer was removed from the prewarmed
Data sets of unWltered signal intensities (four samples for each platform) were formatted according to the Minimum Information about Microarray Experiment (MIAME) database guidelines [16,17] and submitted to the ArrayExpress database [18], where they are available with accession number E-MEXP-232. Microarray data were Wrst analyzed for each platform separately. Normalized signal intensities were log-transformed for graphical analysis of distribution using Matlab (Mathworks, Natick, MA, USA). We determined intraplatform reproducibility by calculating Pearson’s and Spearman’s correlation coeYcients for treated and control duplicate samples on data sets of Wltered features. Then diVerentially expressed genes (treated vs. control samples) were identiWed using both the dChip Compare Samples procedure and signiWcance analysis for microarray (SAM). The dChip Compare Samples procedure compares mean intensities of two groups (e.g., treated vs. control) and identiWes microarray features with a more than twofold change as well as an absolute diVerence of more than 100 FU. The “use lower 90% conWdence bound” option was selected to conservatively estimate the real fold change. At SAM, usertunable parameter delta was set to 1.8682 (CodeLink) or
46
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
2.2403 (GeneChip) and a minimum twofold change was requested between control and treated groups. The null statistic for class comparison was obtained from all possible random shuZings of treated and control samples. To permit interplatform analysis, it was necessary to Wrst identify the genes represented on both microarrays (“common genes”) by comparing LocusLink IDs.2 LocusLink IDs were obtained for 21,080 (95%) of 22,215 GeneChip probe sets using the annotation Wle provided by AVymetrix’s NetAVx Analysis Center [19,20]. LocusLink IDs for CodeLink probes were obtained by searching with GenBank numbers in the Source database [21,22]. If more than one LocusLink ID was associated with a GenBank number, the correct match was determined on the basis of the corresponding CodeLink descriptions. Genes not annotated in Source were evaluated with DAVID [23,24]. Overall, we annotated 9501 CodeLink probes (95% of all discovery probes). Annotated probes (CodeLink) and probe sets (GeneChip) were characterized in terms of ontology using FatiGO [25,26] to access the Biological Process functional annotation. Interplatform comparisons, including hierarchical clustering and analysis of frequency distributions [5], were performed after normalization of signal intensities for common genes. Normalization was accomplished by z transformation; that is, the signal intensity of each microarray feature was adjusted by subtracting the mean intensity for the entire microarray and dividing by the corresponding standard deviation. To highlight similarities and diVerences in the gene expression proWles of the four data sets from each platform, z-transformed values (z scores) were analyzed with dChip’s Hierarchical Clustering function using Pearson’s correlation coeYcient and average linkage as distance and linkage methods, respectively. The z scores were also compared between platforms for the four independent samples by the calculation of correlation coeYcients. Because z transformation is expected to generate a normal distribution of intensities with mean D 0 and SD D 1, we analyzed the actual distribution of z scores for each platform using Matlab. Sequence correspondence between microarray probes and target transcripts was evaluated for diVerentially expressed genes (DEGs) identiWed on each platform. Sequences of HGU133A GeneChip probes were retrieved from the NetAVx Analysis Center (sequences were last updated 28 March 2003); sequences of CodeLink probes were not available. Using a Perl script running bl2seq of Standalone BLAST (version 2.2.9), we conducted pairwise alignments of each of the 11 probes constituting the GeneChip probe sets to the respective mRNA sequences retrieved from RefSeq [27]. To determine whether each probe was correctly aligned on its
2 EVective 1 March 2005, LocusLink has been incorporated into Entrez Gene, as explained at www.ncbi.nlm.nih.gov/entrez/query/static/help/genefaq.html#archives. Nonetheless, the ID numbers associated with the features of the two platforms studied here are the same, and the procedure for comparing the genes remains valid.
target or not, we set a threshold of 24 of 25 bases (total length of each probe) correctly matching the mRNA sequence. These probes were considered to be “exact” matches, whereas probes with less than 24 bases matching were considered to be nonspeciWc for their target. When BLAST was unable to demonstrate alignment to the target sequence, we used blastall of stand-alone BLAST to attempt alignments with sequences in the human mRNA database [28]. BLAST results were ordered according to evalue scores from lowest to highest and then queried (using RefSeq number) to Wnd a match between the AVymetrix U133A probe set annotation and the BLAST output Wle. When the query found no relevant match, the best hit was highlighted to understand whether the sequence really aligned to a diVerent gene. The distribution of DEGs into biological functional categories and the comparison between platform-speciWc results were performed using FatiGO. Step-down minP adjusted P values were calculated as in Al-Shahrour and co-workers [25]. Literature mining We used a literature mining procedure to determine whether the DEGs identiWed in resveratrol-treated cells had already been implicated to play roles in relevant biological processes. We submitted six keywords and the platformspeciWc lists of DEGs to PubMatrix [29,30] to identify similar citations in the literature indexed in Medline. The keywords corresponded to general biological functional categories (cell growth, metabolism, and physiological process) and to speciWc aspects of the cell model system (apoptosis, resveratrol, and breast cancer). Quantitative real-time PCR Quantitative real-time PCR (qPCR) evaluation of expression levels was performed on a selection of seven biologically relevant DEGs identiWed by microarray analysis and on three housekeeping genes (GAPDH, ACTB, and HPRT1). BrieXy, 1 g total RNA from control and resveratrol-treated samples and 1 g Universal Human Reference RNA (Stratagene, La Jolla, CA, USA) were reverse-transcribed using the High-Capacity cDNA Archive Kit (Applied Biosystems, Foster City, CA, USA). To determine PCR eYciencies for the 10 genes, standard qPCR curves were generated using 20, 2, 0.2, and 0.02 ng Universal Human Reference RNA. qPCR was performed in duplicate in an iCycler Thermal Cycler (Bio-Rad, Hercules, CA, USA) using TaqMan Assays-on-Demand (Applied Biosystems) in 20 l containing 1£ TaqMan Universal PCR Master Mix, No AmpErase UNG, 1£ Target Assay Mix, and cDNA template. Samples were heated for 10 min at 95 °C and then subjected to 40 cycles of denaturation at 95 °C for 15 s and annealing extension at 60 °C for 1 min. To evaluate expression levels in control and treated samples, qPCR was performed in duplicate using 20 ng cDNA.
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
Following the analysis of Vandesompele and co-workers [31], we veriWed the constitutive expression of the three housekeeping genes, evaluating the single control normalization error (E) and internal control gene stability measure (M). Constitutively expressed genes have values of E t 1 and M t 0. qPCR eYciencies from standard curves and Ct (threshold cycle) values for control and treated samples were used to calculate relative expression levels according to the algorithm of Vandesompele and co-workers [31] using the Gene Expression Macro (Bio-Rad). Then fold change was calculated as the ratio of the means of expression values for two treated samples and two control samples, with a gene with a fold change greater than 2 being considered diVerentially expressed. Results We developed an investigative strategy to address critical issues associated with microarray cross-comparison and to compare gene expression proWles obtained from diVerent single-color microarray platforms. SpeciWcally, we applied this strategy to compare properties and monitored transcriptional signals and performances of two commercial platforms, using a simple case–control biological model of resveratrol-induced apoptosis in the MDA-MB-231 breast cancer cell line. The study involved the duplicate analysis of RNA from control and treated cells, evaluated on CodeLink and GeneChip microarrays in a four-sample experiment (Fig. 1). Following hybridization of biotinylated cRNA to microarrays, Xuorescent detection of signals, invariant set normalization, and Wltering, we generated a list of 7032 probes on CodeLink microarrays (71% of 9948 discovery probes) and 8081 probe sets on GeneChip microarrays (36% of 22,215 probe sets). We Wrst evaluated the distribution of signal intensities of each array (Fig. 2). Median intensity was similar between control and treated samples and between CodeLink and GeneChip microarrays. Variance in signal was similar among the four samples within each platform but was greater on CodeLink microarrays than on GeneChip microarrays. The lack of major variations in overall signal distribution among the samples justiWed subsequent investigation for DEGs. CodeLink microarrays showed good correlation between the two replicates of resveratrol-treated samples (Pearson’s r D 0.978, Spearman’s r D 0.970), whereas correlation between duplicate control samples was somewhat less (Pearson’s r D 0.844, Spearman’s r D 0.807). The dChip Compare Sample procedure between control and treated samples identiWed 105 DEGs (107 probes); of these, 87 transcripts (89 probes) were upregulated by approximately 2- to 20-fold and 18 transcripts (18 probes) were downregulated by 2- to 110-fold (see list of CodeLink DEGs in supplementary material). Using SAM, 54 transcripts were found to be modulated by resveratrol; of these, 45 were upregulated and 9 were downregulated (data not shown). A total of 39 transcripts (37.1% of dChip DEGs)—32 upregu-
47
lated and 7 downregulated—were identiWed as diVerentially expressed by both dChip and SAM. GeneChip microarrays showed high correlation between the two replicates of both control and resveratrol-treated samples (controls: Pearson’s r D 0.994, Spearman’s r D 0.975; treated: Pearson’s r D 0.993; Spearman’s r D 0.968). The dChip Compare Samples procedure between control and treated samples revealed 42 DEGs (47 probe sets); of these, 32 transcripts (36 probe sets) were upregulated by approximately 2- to 5-fold and 10 transcripts (11 probe sets) were downregulated by 2- to 5-fold (see list of GeneChip DEGs in supplementary material). SAM identiWed 47 transcripts to be modulated by resveratrol; of these, 38 were upregulated and 9 were downregulated. A total of 26 transcripts (61.9% of dChip DEGs)—21 upregulated and 5 downregulated—were identiWed as diVerentially expressed by both dChip and SAM. To have a perspective for comparing platform-speciWc gene expression proWles, we Wrst compared the identities and biological functional categories of all genes represented on the microarrays. Although the number of human transcripts represented on GeneChip is greater than that represented on CodeLink microarrays (22,215 probe sets vs. 9948 probes), FatiGO analysis indicated that there were no signiWcant diVerences between platforms (P > 0.05) regarding the number of genes per functional category (data not shown). Furthermore, on the basis of LocusLink comparisons for annotated probes, we determined that 8414 human transcripts are represented on both CodeLink and GeneChip microarrays. On GeneChip microarrays, these 8414 “common genes” are represented by 14,934 probe sets (70.8% of annotated probe sets). On CodeLink microarrays, these transcripts correspond to 88.6% of annotated probes. The z-transformed signal intensities of common genes were used in hierarchical clustering analysis to assess the performances of the two platforms (Fig. 3). The analysis of the “heat maps” revealed similar color proWles for duplicate samples on GeneChip microarrays but more variable patterns for CodeLink microarrays, especially for the two control samples; these observations are consistent with the somewhat lower correlation coeYcients for these samples. Hierarchical clustering clearly distinguished the two types of microarrays and correctly associated resveratrol-treated samples on both platforms, whereas control samples were correctly associated only on GeneChip. The obvious diVerences in color proWles between platforms for the four independent samples were conWrmed by the low values of Pearson’s and Spearman’s correlation coeYcients: r < 0.5 and r < 0.6, respectively, for all four determinations. The two microarray platforms were also distinguished on the shape of the z score distributions for all four samples (Fig. 4), where the GeneChip distributions were smoother than those of CodeLink microarrays that showed numerous spikes. When platform-speciWc lists of DEGs were compared, only 9 human transcripts (corresponding to 21.4% and 8.5% of GeneChip and CodeLink DEGs, respectively) were identiWed on both platforms as being resveratrol sensitive.
48
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
Fig. 1. Experimental strategy for microarray platform comparison. The study compared Amersham’s CodeLink Bioarrays (left side) and AVymetrix’s GeneChip Expression Analysis Arrays (right side); the center refers to common procedures and outcomes. The horizontal line divides analyses on all genes represented on the microarrays (top) from those represented on DEGs (bottom). The main steps of the experimental strategy are numbered as follows: (1) DEGs were determined on each microarray platform; (2) genes represented on both microarrays were identiWed by LocusLink ID comparison, permitting us to deWne a population of genes that can be studied on both platforms; (3) common and platform-speciWc DEGs underwent a battery of analyses to investigate factors contributing to the generation of platform-dependent results.
These 9 transcripts were upregulated according to both platforms; for 7 genes the fold changes in expression level were comparable for GeneChip and CodeLink microarrays (CDKN1A [RefSeq: NM_000389], 4.58 vs 5.30; INSIG1
[RefSeq: NM_005542], 3.89 vs. 3.29; IRF7 [RefSeq: NM_004031], 3.14 vs. 4.06; ECM1 [RefSeq: NM_022664], 2.67 vs. 4.46; KAL1 [RefSeq: NM_000216], 2.52 vs. 4.83; JUND [RefSeq: NM_005354], 2.44 vs. 5.39; NEU1 [RefSeq:
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
49
Fig. 2. Distribution of signal intensities for four biological samples analyzed on two microarray platforms. Data were normalized by dChip invariant set algorithm and log10 transformed before plotting using Matlab. Boxes represent median (solid black horizontal line [red horizontal line in online version]) and interquartile range. Whiskers correspond to extreme values (minimum or maximum) or, when outliers were present, to 1.5 times the interquartile range. (A) CodeLink UniSet Human I Bioarrays. (B) GeneChip HG-U133A Array. (For interpretation of the references to color in this Wgure legend, the reader is referred to the web version of this paper.)
NM_000434], 2.33 vs. 4.06), and for 2 genes the fold changes were considerably higher on CodeLink microarrays (HMGCS1 [RefSeq: NM_002130], 4.60 vs. 20.61; SPHK1 [RefSeq: NM_021972], 2.29 vs. 11.22). Considering the poor overlap of DEGs identiWed by the two microarray platforms, we evaluated why the remaining genes were identiWed as diVerential only by one platform (Table 1). Of the 105 DEGs identiWed by dChip analysis on CodeLink microarrays (CodeLink DEGs), only a minority (14.3%) was not represented on HG-U133A GeneChips, whereas 45.3% of the GeneChip DEGs were not represented on CodeLink. These values are in agreement with the greater number of transcripts represented on GeneChip microarrays than on CodeLink microarrays. For the remaining DEGs that were represented on both microarrays but identiWed as diVerential on only one platform, there was a nearly equal partition of CodeLink DEGs between those Wltered out for poor quality and those with a fold change of less than 2, whereas for GeneChip DEGs, quality Wltering was less a cause of platform-speciWc calls than was lack of suYcient fold change. Therefore, on the basis of this analysis, reasons for platform-speciWc calls include lack of representation on both microarrays, inconsistent or inadequate experimental signal leading to elimination at the Wltering step, and real diVerences in measured fold change between control and treated samples. To further understand why some genes were found to be diVerentially expressed on only one microarray platform, we used BLAST sequence alignment to test the correspondence of probes with target sequences in RefSeq (Table 2). Alignments were checked for 144 probe sets, corresponding to 9 genes (9 probe sets) determined to be diVerentially expressed by both platforms, 81 transcripts (121 probe sets) found to be diVerentially expressed only on CodeLink microarrays (CodeLink DEGs), and 14 transcripts (14 probe sets) considered to be diVerentially expressed only on GeneChip microarrays (GeneChip DEGs). For the 9 genes considered to be diVerentially expressed by both platforms, exact alignment
was found with the target sequence for all 11 probes of each probe set. For 14 GeneChip DEGs, there was exact alignment of all 11 probes for 12 probe sets and partial or total misalignment of probes for 2 probe sets. For 81 CodeLink DEGs (121 probe sets), there was exact alignment of all 11 probes for 107 probe sets and partial or total misalignment of probes for 14 probe sets. Overall, at least 8 of 11 probes did not align for 13 probe sets, corresponding to 10 CodeLink DEGs excluded from evaluation on GeneChip microarrays during quality Wltering, 2 CodeLink DEGs, and 1 GeneChip DEG excluded from the other platform at dChip analysis (i.e., not diVerentially expressed). For all probes not correctly aligning with the target sequence, we nonetheless did not Wnd any compatible alignment (>50% of bases) with other genome sequences. These data indicate that misalignment of probes to target sequences is another cause of disparate calls between platforms and that sequence-veriWed probes perform more consistently. The disparate results regarding resveratrol-modulated genes obtained with the two microarray platforms were further analyzed considering the genes’ biological functional categories (Fig. 5). The functional category with the greatest number of DEGs was metabolism (15 GeneChip DEGs, 29 CodeLink DEGs). Regarding other biological functions considered to be modulated by resveratrol-induced apoptosis, we found 4 DEGs from each platform in regulation of the cellular processes category, as well as 2 DEGs in the growth category, and 4 CodeLink DEGs and 2 GeneChip DEGs were associated with the death category. The distributions of DEGs per functional category were similar between the two platforms; on the basis of adjusted P values (Fig. 5), there was no signiWcant diVerence in the number of DEGs per category identiWed by the two platforms. We used the literature mining protocol provided by PubMatrix to determine whether the platform-speciWc DEGs had been mentioned previously in Medline abstracts discussing issues relevant to the biological model studied (deWned by six keywords). Overall, at least one abstract was
50
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
Gene expression proWles obtained at microarray analysis were compared with those obtained from qPCR. Vandesompele analysis determined error (E) to be 1.04 § 0.03 (mean § SD) for the expression of three presumed housekeeping genes (GAPDH [AB: Hs99999905_m1], ACTB [AB: Hs99999903_m1], and HPRT1 [AB: Hs9999990 9_m1]) in the four biological samples. Moreover, for these genes, values of the Vandesompele stability measure (M) were close to zero (0.133, 0.115, and 0.135, respectively). These values of E and M conWrm that the three genes are constitutively expressed in the MDA-MB-231 cell line and thus could be used for normalization purposes in the following analysis. The relative expression levels and fold changes due to resveratrol treatment were then calculated for seven genes: three found to be upregulated on both platforms, two GeneChip DEGs, and two CodeLink DEGs (Table 3). qPCR eYciencies for the seven assays ranged from 74.8% to 124.8%; one of the assays having low qPCR eYciency was that for BCL2-associated X protein (BAX [AB: Hs0106548_g1]). Interestingly, standard qPCR curves for BAX had unusually long lag phases and reduced alpha angles (data not shown), suggesting a problem of assay design or the presence of splicing variants. qPCR conWrmed the upregulated expression of three common DEGs and two CodeLink DEGs, whereas it did not conWrm resveratrol sensitivity for the two GeneChip DEGs (including BAX). Pearson’s correlation coeYcient for the agreement between qPCR and GeneChip microarray results was 0.732 (0.84 excluding BAX); for the comparison with CodeLink microarray results, Pearson’s r D 0.681 (0.616 excluding BAX). Discussion
Fig. 3. Hierarchical clustering of transcripts common to both microarray platforms: dChip hierarchical clustering of signal intensities for 8414 genes commonly represented on CodeLink and GeneChip platforms for two control and two resveratrol-treated samples. Data were z-transformed before clustering and representation. The z scores were colored for each line according to their being below or above the mean for that line. Red means that the gene had a z score above the mean for that line (gene), and green means that the gene had a z score below the mean.
found for 80 (76%) of 105 CodeLink DEGs and for 37 (88%) of 42 GeneChip DEGs. Considering the three keywords for biological functional categories (metabolism, cell growth, and physiological process), at least one abstract was found for the same 80 CodeLink DEGs and 37 GeneChip DEGs, and there was good correlation (r > 0.90) between the percentages of genes with abstracts matching a keyword and the percentages of genes associated with the same biological functional category. Regarding the three speciWc keywords (apoptosis, resveratrol, and breast cancer), at least one abstract was identiWed for 41 (39%) of 105 CodeLink DEGs and for 26 (62%) of 42 GeneChip DEGs.
A diYcult aspect of interpreting microarray cross-comparison data involves the heterogeneous manner in which samples are processed and data are analyzed on diVerent platforms. Many articles have already tried to point out some of these critical aspects, with conclusions so disparate that it is hard for researchers to reach a consensus on whether microarray data are robust, reproducible, and platform independent [2–9]. To maximize the comparability of gene expression proWles from two microarray platforms, experimental designs and data mining approaches should be standardized so that all samples are handled identically and undergo parallel analyses. The approach for platform comparison that we have proposed and applied to real samples can be summarized as follows: (i) selection of platforms based on similar technologies (e.g., single color as in AVymetrix’s GeneChip Gene Expression Analysis Arrays and Amersham’s CodeLink Bioarrays); (ii) identiWcation of a simple cellular model to eliminate variability due to RNA preparation (e.g., the well-known apoptotic behavior of a cell line to the antitumor agent resveratrol [12]); (iii) elaboration of microarray data using the same statistical software to identify diVerentially expressed genes (step 1 of Fig. 1); (iv) comparison of the genes represented on each microarray to identify a population that can be studied
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
51
Fig. 4. The z score frequency distribution for transcripts common to both microarray platforms. Distribution of z-transformed signal intensities from four independent samples (two control and two resveratrol treated) for 8414 genes commonly represented on both CodeLink (black traces [blue traces in online version]) and GeneChip (gray traces [green traces in online version]) microarrays. (For interpretation of the references to color in this Wgure legend, the reader is referred to the web version of this paper.) Table 1 Correspondence between DEGs in MDA-MB-231 cells, treated with resveratrol or vehicle, according to two diVerent microarray platforms
Total DEGs (n) Common (n)
GeneChip
CodeLink
42
105
9 (21.4)
9 (8.5)
Representation of the platform-speciWc DEGs on the other platform (n)
Not represented Represented
19 (45.3) 14 (33.3)
15 (14.3) 81 (77.2)
DEGs represented on both platforms and Wltered out or not diVerentially expressed on the other platform (n)
Represented genes Excluded by quality Wltering Not diVerentially expressed
14 2 12
121a 65 56
Note. Percentages are in parentheses. a The 81 CodeLink transcripts are represented by 121 GeneChip probe sets.
with both platforms (step 2 of Fig. 1); and (v) application of multiple analyses, including sequence comparisons of probes with target genes, functional category annotation, literature mining, and qPCR validation (step 3 of Fig. 1), to the platform-speciWc DEGs as well as to “common” DEGs. This battery of analyses, applied to both platforms, permitted us to investigate the factors contributing to the generation of platform-dependent results.
Using FatiGO to analyze the genes represented on the two microarrays, we found no signiWcant diVerence in terms of distribution by biological functional category, despite the fact that GeneChip microarrays represent many more transcripts than do CodeLink microarrays. Overall, 8414 transcripts are represented on both platforms, corresponding to 70.8% of GeneChip probe sets and 88.6% of CodeLink probes with LocusLink annotations. dChip Compare
52
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
Table 2 Sequence alignment of GeneChip probes on RefSeq and comparison with results from Mecham and co-workers [39] Match with target gene DEGs according to both platforms (n D 9) 11 probes per probe set (100%)
Probe sets (n) 9
GeneChip DEGS (n D 14) 11 probes per probe set (100%) 9 probes per probe set (81.8%) 0 probes per probe set (0%)
12 1 1
CodeLink DEGs (n D 121)b 11 probes per probe set (100%) 10 probes per probe set (90.9%) 3 probes per probe set (27.2%) 1 probe per probe set (9.1%) 0 probes per probe set (0%)
107 2 1 1 10
Sequence veriWcation (Mecham et al. [39])a VeriWed VeriWed VeriWed Not veriWed 102 veriWed; 5 not veriWedc VeriWed 2 of 11 probes aligned Not veriWed Not veriWed
Note. Shown are BLAST alignment of GeneChip probes with target sequences in RefSeq database for DEGs found by both platforms (GeneChip and CodeLink microarrays), or on only one platform, and comparison with results reported by Mecham and co-workers [39]. A GeneChip probe set contains 11 pairs of probes. Within each pair, one probe is meant to have “perfect match” (PM), whereas the other is designed as a control with a 1-base mismatch. Sequence analysis regards all 11 PM probes. a VeriWed: all 11 probes were found to have 25 of 25 bases matching; not veriWed: 0 of 11 probes matched 25 of 25 bases. b 121 probe sets corresponding to 81 unique transcripts. c Genes NGLY1, VEGF, HEG, KIAA1093, and LSS.
Samples analysis of control and resveratrol-treated samples identiWed 105 DEGs on CodeLink and 42 DEGs on GeneChip microarrays; only 9 DEGs were commonly identiWed by both platforms. At FatiGO analysis, platform-speciWc DEGs were similarly distributed according to biological function, suggesting that these microarrays provide results with similar biological relevance. The functional category most aVected by resveratrol treatment was metabolism, and other categories considered to be targets of resveratrol were also aVected. Literature mining conWrmed our Wndings; resveratrol was found to aVect the expression of some genes implicated in breast cancer cells in both lists. However, at literature mining, more GeneChip DEGs than CodeLink DEGs had matching abstracts, suggesting that the GeneChip platform provides results of somewhat greater biological relevance. qPCR analysis on 3 of the 9 commonly identiWed DEGs conWrmed that the expression of these genes was modiWed by resveratrol. This Wnding is in agreement with that of Larkin and co-workers [32], who observed that when two platforms give consistent results, the outcome of qPCR analysis will also be in agreement with high correlation. The current microarray study is not the Wrst to evaluate the eVects of resveratrol treatment on gene expression. Previous studies have employed low- or medium-density microarray platforms to study changes in gene expression of breast, prostate, ovarian, and renal cancer cell lines in response to resveratrol [33–37]. These studies demonstrated that resveratrolmodulated genes fall into functional categories such as cell death, cell cycle, growth, and intracellular signaling. The current gene expression proWles, obtained using high-density microarrays, are in agreement with these earlier studies. The gene expression proWles of CodeLink and GeneChip microarrays have already been compared in various biological models [5,38]. Tan and co-workers [5] analyzed the gene expression proWles of the human pancreatic carcinoma cell
line PANC-1 at two stages of diVerentiation and found that interplatform diVerences were greater than intraplatform ones, and our results obtained by clustering and z score frequency distribution analyses conWrm this observation. Moreover, these authors found only a small overlap in the platform-speciWc gene lists (22 genes, 26% of GeneChip DEGs and 14% of CodeLink DEGs). However, Tan and co-workers reported only a partial match in the Biological Process gene ontology between platforms, whereas we found similar percentages of genes per functional category. Hollingshead and co-workers [38] compared the same platforms in the study of postmortem cortex samples from brains of normal and schizophrenic subjects. Although numerous platform-speciWc DEGs were identiWed, in this study there also was small overlap in platform-speciWc gene lists, representing only 4.2% of CodeLink DEGs and 15.6% of GeneChip DEGs. Thus, on the basis of our results and those of others, it appears that a common dichotomy of microarray cross-comparisons is a poor correspondence at the level of single genes but is a reasonable agreement when results are evaluated in terms of gene functional categories. The comparison procedure described here, therefore, leads to results that are in agreement with those obtained by the previously cited studies using the same microarray platforms. These observations, moreover, suggest that gene expression data should be compared globally between platforms on the basis of functional “enrichment” rather than at the single gene level. To infer about the expression of single genes from microarray results, it is essential to analyze the sequences of the microarray probes and to determine the extent to which they correspond to the actual target sequences and variants. Mecham and co-workers [39] analyzed the publicly available GeneChip probe sequences and found that 28% of all HGU133A probe sets had 0 of 11 probes matching sequences in
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
53
Fig. 5. Biological Process functional annotation for DEG lists: biological functional categories of DEGs identiWed on CodeLink (black bars [green bars in online version]) and GeneChip (light gray bars [red bars in online version]) microarray platforms, analyzed using FatiGO. Percentages of genes in each category were calculated with respect to the total number of DEGs identiWed with each microarray (105 and 42, respectively). DiVerences in numbers of genes per category were evaluated on the basis of step-down minP adjusted P values [25]; no signiWcant diVerences were found. (For interpretation of the references to color in this Wgure legend, the reader is referred to the web version of this paper.)
54
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
Table 3 Evaluation of gene expression of seven chosen DEGs with both microarray data and qPCR Symbol
BAX JUND SPHK1 CDKN1A IL8 PTPRG GADD45A
Gene name
BCL2-associated X protein Jun D proto-oncogene Sphingosine kinase 1 Cyclin-dependent kinase inhibitor 1A (p21, Cip1) Interleukin 8 Protein tyrosine phosphatase, receptor type, G Growth arrest and DNA damage-inducible, alpha
TaqMan assay ID
qPCR eYciency (%)
Platform indicating diVerential expression
Fold change in expression level GeneChip
CodeLink
qPCR
Hs0106548_g1 Hs00534289_s1 Hs00184211_m1 Hs00355782
74.9 124.8 82.1 75.7
GeneChip Both Both Both
2.80 2.44 2.29 4.58
NC 5.39 11.22 5.30
¡1.84 8.90 4.74 7.41
Hs01553824_g1 Hs00177193_m1
74.8 82.7
CodeLink GeneChip
NC ¡3.16
6.23 Excluded
5.92 ¡1.30
Hs00169255_m1
84.7
CodeLink
Excluded
5.00
3.50
Note. Shown are comparisons of resveratrol-induced changes in gene expression determined at microarray analysis using two commercial platforms and at qPCR for a selection of common and platform-speciWc DEGs. qPCR was performed using TaqMan Assays-on-Demand, and PCR eYciencies were determined using Universal Human Reference RNA (Stratagene). Fold change (FC) was determined with dChip for microarray data and with Vandesompele analysis for qPCR data [31]. 冷FC冷 > 2 deWned a diVerentially expressed gene. Excluded, heterogeneous signal intensities of replicate samples resulted in exclusion from analysis at dChip Compare Samples procedure; NC, not changed (e.g., 冷FC冷 < 2 at dChip analysis).
the RefSeq database. This initially surprising Wnding is partially explained by the fact that GeneChip probes are designed considering not only RefSeq, but also, GenBank and dbEST sequences. We also found some instances of probe mismatch with RefSeq sequences; BLAST analysis revealed lack of perfect correspondence between probes and target genes for 2 GeneChip DEGs and 14 CodeLink transcripts. Thus, the variability of human sequences in diVerent databases and the lack of standardization in probe design render microarray cross-comparisons diYcult. When the results of Mecham and co-workers [39] were compared with those of our own sequence analyses (Table 2), we found good agreement regarding the 9 DEGs identiWed by both platforms (i.e., all probes matched target sequences); moreover, the 13 probe sets having at least 8 mismatched probes all were scored as not veriWed by Mecham and co-workers. These observations stress the need for complete access to microarray probe sequences and emphasize that gene expression proWles obtained from microarray research must be veriWed by successive sequence analysis. Lack of access to CodeLink probe sequences renders this platform more diYcult to be investigated in its reliability and speciWcity. We observed that CodeLink microarrays gave greater fold changes in gene expression in response to resveratrol than did GeneChip microarrays. Shippy and co-workers [8] reported a similar behavior, observing a mean 3-fold gain with CodeLink. They proposed that this diVerence is due to diVerences in probe length (25 bases in GeneChip probes vs. 30 bases for CodeLink) and suggested that longer probes possess greater sensitivity in assessing diVerential gene expression. Probe length has already been recognized to inXuence platform sensitivity and speciWcity; Relogio and co-workers [40] observed a 2.0-fold gain in sensitivity and a 1.6-fold loss of speciWcity. Finally, microarray comparisons can be addressed only in case–control experiments using well-characterized and controlled biological models [32]. Because a limited number of replicate samples is a common problem in microarray research, it is important to perform quality control analyses
such as signal intensity distribution, hierarchical clustering, and the calculation of correlation coeYcients to test the reliability of the data sets. In this study, Pearson’s correlation coeYcients for biological replicates (two samples handled identically) on GeneChip microarrays were high (>0.99) and similar to those reported for technical replicates in previous studies [5,38,39]. On CodeLink microarrays, Pearson’s coeYcient was similarly high for treated samples (r D 0.978) but was lower for control samples (r D 0.844). This may be explained by sample variability or experimental error, rather than poor platform performance on the basis of the high correlation reported by Tan and co-workers for biological replicates on this platform (r D 0.9824) [5]. Larkin and co-workers [32] commented that diVerent statistical analyses may lead to diVerent outcomes from one set of microarray data and stressed that the use of multiple analyses is desirable to obtain consistent results. In the current study, we used both the dChip Compare Samples procedure and SAM [41] to identify DEGs; however, other software programs are available (e.g., TMEV, Bioconductor routines, BRB Array Tools, GeneSpring). We gave preference to dChip results in this study because this protocol gave a greater number of DEGs on which to base further analyses. We nonetheless observed a reasonable overlap of DEGs identiWed by the two algorithms (SAM identiWed 62% of GeneChip DEGs found using dChip), as did Hosack and co-workers (37%) [42] and Hansel and coworkers (67%) [43]. However, despite the application of multiple computational methods, increasing the number of replicates adds robustness to the experimental design and limits undesired eVects and bias in expression proWles, especially when comparing two or more platforms. We investigated the interesting and troublesome problem of microarray platform comparison. Our investigational strategy, proposed as a standardized protocol for evaluating the results of microarray studies, limits problems due to experimental factors (e.g., target preparation and processing, statistical analysis, choice of gene IDs for
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
identifying common transcripts). However, technical variables relating to the microarray platform (e.g., probe design, hybridization protocols, signal generation) can profoundly aVect the comparability of gene expression patterns across studies. Comparability is highest when these technical variables are standardized [44,45]. If microarray producers would standardize probe design to a common database and publicly release probe sequences, researchers would have greater possibilities of controlling variations and comparing data. Using our standardized strategy for comparing microarray data from multiple platforms, we observed a poor correspondence at the level of single genes but observed a reasonable agreement in terms of functional categories. These observations suggest that gene expression data should be compared globally between platforms on the basis of functional annotations rather than at the single gene level. Finally, gene expression proWles obtained from microarray research must be veriWed by successive sequence analysis. Acknowledgments We thank Gavin Hardy (GE Healthcare, UK) for his patience and precious help in preparing CodeLink data for submission to ArrayExpress, and we thank Valerie Matarese for editing and critically reviewing the manuscript. This work was supported by grants from the Italian Ministry of University and Research (MIUR-FIRB RBNE01TZZ8 and MIUR-FIRB RBNE01HCFK) and by funds from the Center for Biomolecular Studies and Industrial Applications (CISI, Milan, Italy). Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.ab.2006.03.023. References [1] G. Hardiman, Microarray platforms: comparisons and contrasts, Pharmacogenomics 5 (2004) 487–502. [2] R. Kothapalli, S.J. Yoder, S. Mane, T.P. Loughran Jr., Microarray results: how accurate are they? BMC Bioinform. 1 (2002) 22. [3] W.P. Kuo, T.K. Jenssen, A.J. Butte, L. Ohno-Machado, I.S. Kohane, Analysis of matched mRNA measurements from two diVerent microarray technologies, Bioinformatics 3 (2002) 405–412. [4] J. Li, M. Pankratz, J.A. Johnson, DiVerential gene expression patterns revealed by oligonucleotide versus long cDNA arrays, Toxicol. Sci. 2 (2002) 383–390. [5] P.K. Tan, T.J. Downey, E.L. Spitznagel Jr., P. Xu, D. Fu, D.S. Dimitrov, R.A. Lempicki, B.M. Raaka, M.C. Cam, Evaluation of gene expression measurements from commercial microarray platforms, Nucleic Acids Res. 19 (2003) 5676–5684. [6] A. Barczak, M.W. Rodriguez, K. Hanspers, L.L. Koth, Y.C. Tai, B.M. Bolstad, T.P. Speed, D.J. Erle, Spotted long oligonucleotide arrays for human gene expression analysis, Genome Res. 7 (2003) 1775–1785. [7] P.J. Park, Y.A. Cao, S.Y. Lee, J.W. Kim, M.S. Chang, R. Hart, S. Choi, Current issues for DNA microarrays: platform comparison, double linear ampliWcation, and universal RNA reference, J. Biotechnol. 3 (2004) 225–245.
55
[8] R. Shippy, T.J. Sendera, R. Lockner, C. Palaniappan, T. Kaysser-Kranich, G. Watts, J. Alsobrook, Performance evaluation of commercial short-oligonucleotide microarrays and the impact of noise in making cross-platform correlations, BMC Genom. 1 (2004) 61. [9] C.L. Yauk, M.L. Berndt, A. Williams, G.R. Douglas, Comprehensive comparison of six microarray technologies, Nucleic Acids Res. 15 (2004) e124. [10] D.J. Lockhart, H. Dong, M.C. Byrne, M.T. Follettie, M.V. Gallo, M.S. Chee, M. Mittmann, C. Wang, M. Kobayashi, H. Horton, E.L. Brown, Expression monitoring by hybridization to high-density oligonucleotide arrays, Nat. Biotechnol. 13 (1996) 1675–1680. [11] R. Ramakrishnan, D. Dorris, A. Lublinsky, A. Nguyen, M. Domanus, A. Prokhorova, L. Gieser, E. Touma, R. Lockner, M. Tata, X. Zhu, M. Patterson, R. Shippy, T.J. Sendera, A. Mazumder, An assessment of Motorola CodeLink microarray performance for gene expression proWling applications, Nucleic Acids Res. 7 (2002) e30. [12] F. Scarlatti, G. Sala, G. Somenzi, P. Signorelli, N. Sacchi, R. Ghidoni, Resveratrol induces growth inhibition and apoptosis in metastatic breast cancer cells via de novo ceramide signaling, FASEB J. 15 (2003) 2339–2341. [13] C. Li, W.H. Wong, Model-based analysis of oligonucleotide arrays: model validation, design issues and standard error application, Genome Biol. 8 (2001). [14] C. Li, W.H. Wong, Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection, Proc. Natl. Acad. Sci. USA. 1 (2001) 31–36. [15] R.A. Irizarry, B.M. Bolstad, F. Collin, L.M. Cope, B. Hobbs, T.P. Speed, Summaries of AVymetrix GeneChip probe level data, Nucleic Acids Res. 4 (2003) e15. [16] A. Brazma, P. Hingamp, J. Quackenbush, G. Sherlock, P. Spellman, C. Stoeckert, J. Aach, W. Ansorge, C.A. Ball, H.C. Causton, T. Gaasterland, P. Glenisson, F.C. Holstege, I.F. Kim, V. Markowitz, J.C. Matese, H. Parkinson, A. Robinson, U. Sarkans, S. Schulze-Kremer, J. Stewart, R. Taylor, J. Vilo, M. Vingron, Minimum Information about a Microarray Experiment (MIAME): toward standards for microarray data, Nat. Genet. 4 (2001) 365–371. [17] MIAME, Minimum Information about Microarray Experiment, . [18] ArrayExpress, EBI’s ArrayExpress database for microarray experiments, . [19] NetAVx, AVymetrix’s NetAVx Analysis Center, . [20] G. Liu, A.E. Loraine, R. Shigeta, M. Cline, J. Cheng, V. Valmeekam, S. Sun, D. Kulp, M.A. Siani-Rose, NetAVx: aVymetrix probesets and annotations, Nucleic Acids Res. 1 (2003) 82–86. [21] Source, Stanford’s Source functional annotation web-based tool, . [22] M. Diehn, G. Sherlock, G. Binkley, H. Jin, J.C. Matese, T. HernandezBoussard, C.A. Rees, J.M. Cherry, D. Botstein, P.O. Brown, A.A. Alizadeh, SOURCE: a uniWed genomic resource of functional annotations, ontologies, and gene expression data, Nucleic Acids Res. 1 (2003) 219–223. [23] DAVID, SAIC Frederick’s Database for Annotation, Visualization, and Integrated Discovery, . [24] G. Dennis Jr., B.T. Sherman, D.A. Hosack, J. Yang, W. Gao, H.C. Lane, R.A. Lempicki, DAVID: database for annotation, visualization, and integrated discovery, Genome Biol. 5 (2003) P3. [25] F. Al-Shahrour, R. Diaz-Uriarte, J. Dopazo, FatiGO: a web tool for Wnding signiWcant associations of gene ontology terms with groups of genes, Bioinformatics 4 (2004) 578–580. [26] FatiGO, CIPF’s FatiGO annotation tool, . [27] K.D. Pruitt, T. Tatusova, D.R. Maglott, NCBI Reference Sequence (RefSeq): a curated non-redundant sequence database of genomes, transcripts, and proteins, Nucleic Acids Res. (database issue) (2005) D501–D504. [28] Refseq, NCBI RefSeq database, .
56
Comparing gene expression profiles / M. Severgnini et al. / Anal. Biochem. 353 (2006) 43–56
[29] K.G. Becker, D.A. Hosack, G. Dennis Jr., R.A. Lempicki, T.J. Bright, C. Cheadle, J. Engel, PubMatrix: a tool for multiplex literature mining, BMC Bioinform. 1 (2003) 61. [30] PubMatrix, Literature mining tool, . [31] J. Vandesompele, K. De Preter, F. Pattyn, B. Poppe, N. Van Roy, A. De Paepe, F. Speleman, Accurate normalization of real-time quantitative RT–PCR data by geometric averaging of multiple internal control genes, Genome Biol. 7 (2002). [32] J.E. Larkin, B.C. Frank, H. Gavras, R. Sultana, J. Quackenbush, Independence and reproducibility across microarray platforms, Nat. Methods 5 (2005) 337–344. [33] A.S. Levenson, I.L. Kliakhandler, K.M. Svoboda, K.M. Pease, S.A. Kaiser, J.E. Ward III, V.C. Jordan, Molecular classiWcation of selective oestrogen receptor modulators on the basis of gene expression proWles of breast cancer cells expressing oestrogen receptor alpha, Br. J. Cancer 4 (2002) 449–456. [34] B.A. Narayanan, N.K. Narayanan, G.G. Re, D.W. Nixon, DiVerential expression of genes induced by resveratrol in LNCaP cells: P53-mediated molecular targets, Intl. J. Cancer 2 (2003) 204–212. [35] B.A. Narayanan, N.K. Narayanan, G.D. Stoner, B.P. Bullock, Interactive gene expression pattern in prostate cancer cells exposed to phenolic antioxidants, Life Sci. 15 (2002) 1821–1839. [36] T. Shi, L.S. Liou, P. Sadhukhan, Z.H. Duan, A.C. Novick, J.G. Hissong, A. Almasan, J.A. DiDonato, EVects of resveratrol on gene expression in renal cell carcinoma, Cancer Biol. Ther. 9 (2004) 882–888. [37] S.H. Yang, J.S. Kim, T.J. Oh, M.S. Kim, S.W. Lee, S.K. Woo, H.S. Cho, Y.H. Choi, Y.H. Kim, S.Y. Rha, H.C. Chung, S.W. An, Genomescale analysis of resveratrol-induced gene expression proWle in human ovarian cancer cells using a cDNA microarray, Intl. J. Oncol. 4 (2003) 741–750. [38] D. Hollingshead, D.A. Lewis, K. Mirnics, Platform inXuence on DNA microarray data in postmortem brain research, Neurobiol. Dis. 3 (2005) 649–655.
[39] B.H. Mecham, D.Z. Wetmore, Z. Szallasi, Y. Sadovsky, I. Kohane, T.J. Mariani, Increased measurement accuracy for sequence-veriWed microarray probes, Physiol. Genom. 3 (2004) 308–315. [40] A. Relogio, C. Schwager, A. Richter, W. Ansorge, J. Valcarcel, Optimization of oligonucleotide-based DNA microarrays, Nucleic Acids Res. 11 (2002) e51. [41] V.G. Tusher, R. Tibshirani, G. Chu, SigniWcance analysis of microarrays applied to the ionizing radiation response, Proc. Natl. Acad. Sci. USA. 9 (2001) 5116–5121. [42] D.A. Hosack, G. Dennis Jr., B.T. Sherman, H.C. Lane, R.A. Lempicki, Identifying biological themes within lists of genes with EASE, Genome Biol. 10 (2003) R70. [43] D.E. Hansel, A. Rahman, M. Hidalgo, P.J. Thuluvath, K.D. Lillemoe, R. Shulick, J.L. Ku, J.G. Park, K. Miyazaki, R. Ashfaq, I.I. Wistuba, R. Varma, L. Hawthorne, J. Geradts, P. Argani, A. Maitra, IdentiWcation of novel cellular targets in biliary tract cancers using global gene expression technology, Am. J. Pathol. 1 (2003) 217–229. [44] T. Bammler, R.P. Beyer, S. Bhattacharya, G.A. Boorman, A. Boyles, B.U. Bradford, R.E. Bumgarner, P.R. Bushel, K. Chaturvedi, D. Choi, M.L. Cunningham, S. Deng, H.K. Dressman, R.D. Fannin, F.M. Farin, J.H. Freedman, R.C. Fry, A. Harper, M.C. Humble, P. Hurban, T.J. Kavanagh, W.K. Kaufmann, K.F. Kerr, L. Jing, J.A. Lapidus, M.R. Lasarev, J. Li, Y.J. Li, E.K. Lobenhofer, X. Lu, R.L. Malek, S. Milton, S.R. Nagalla, P. O’Malley, V.S. Palmer, P. Pattee, R.S. Paules, C.M. Perou, K. Phillips, L.X. Qin, Y. Qiu, S.D. Quigley, M. Rodland, I. Rusyn, L.D. Samson, D.A. Schwartz, Y. Shi, J.L. Shin, S.O. Sieber, S. Slifer, M.C. Speer, P.S. Spencer, D.I. Sproles, J.A. Swenberg, W.A. Suk, R.C. Sullivan, R. Tian, R.W. Tennant, S.A. Todd, C.J. Tucker, B. Van Houten, B.K. Weis, S. Xuan, H. Zarbl, Standardizing global gene expression analysis between laboratories and across platforms, Nat. Methods 5 (2005) 351–356. [45] Y. Moreau, S. Aerts, B. De Moor, B. De Strooper, M. Dabrowski, Comparison and meta-analysis of microarray data: from the bench to the computer desk, Trends Genet. 10 (2003) 570–577.
