Identification of Novel Genes Preferentially Expressed in the Retina Using a Custom Human Retina cDNA Microarray

Abstract

purpose. To construct a custom cDNA microarray for comprehensive human retinal gene expression profiling and apply it to the identification of genes that are preferentially expressed in the retina.

methods. A cDNA microarray was constructed based on the predicted human retina gene expression profile according to expressed sequence tag (EST) databases. Gene expression profiles were obtained from five human retinas, two livers, and the cerebral cortical regions of two brains. Each sample was studied in duplicate, using a reference sample experimental design. Retina-enriched genes were identified by using the significance analysis for microarray (SAM) algorithm. Quantitative real time PCR was used to confirm microarray results. Bioinformatic analysis was performed to compare the array results with expression data available from public databases.

results. The cDNA microarray contains 10,034 sequences: 67% represent known genes and 33% represent ESTs. Differential hybridization with the array identified, in addition to known retinal genes, 186 retina-enriched genes that do not have known retinal function. Of these, 96 represent novel genes. Quantitative real-time PCR of 11 of the identified genes and ESTs confirmed their retina-enriched expression pattern. Bioinformatic analysis of EST databases suggests that of the 186 genes, approximately 40% are predominantly expressed in the retina, whereas the remainder show significant expression in other tissues. Comparison of this study’s microarray-based retina-enriched gene set with three published similar sets identified using complementary high-throughput approaches demonstrated only limited overlap of the identified genes.

conclusions. Because previous studies have demonstrated that many retina-enriched genes are crucial for maintaining normal retinal function, the genes identified here are likely to include ones that have important roles in the retina and ones that when mutated can cause or modulate retinal disease. In addition, the retina custom array should provide a useful resource for comparing expression profiles between normal and diseased human retinas.

The retina’s unique histologic, metabolic, and physiological characteristics are, in large part, due to its distinctive gene expression profile. Studies of genes that are preferentially expressed in the retina (retina-enriched genes) have provided important insights into processes such as phototransduction and the visual cycle, and into photoreceptor structure. The crucial role of such genes is underscored by the fact that approximately half of all retinal degeneration–associated genes identified to date are preferentially expressed in the retina. ¹ In an effort to learn more about retinal function and disease, identification of additional retina-enriched genes has become a major focus in recent years. ^{1 2 3 4 5 6 7 8} These efforts are being aided by advances in the human genome project combined with the development of high-throughput gene expression profiling approaches that facilitate the simultaneous identification of multiple tissue-specific genes. ⁹  

Sequencing and data mining of cDNA libraries is one of the high-throughput expression profiling approaches that has been applied to identify retina-enriched genes. ^{4 5 6 7 8} However, potential bias and limited cost effectiveness has hampered its application for gene expression level comparison across multiple samples. ¹⁰ Two additional recent techniques, serial analyses of gene expression (SAGE) and microarrays, have gained popularity for high-throughput gene expression profiling. ^{11 12 13} Both methods have been efficient in identifying novel genes and in providing insights into complex pathologic processes, such as cancer and disease of the central nervous system. ^{14 15} Unlike microarrays, SAGE is not dependent on preselection of sequences for analysis. It theoretically can identify all mRNA molecules in a sample. In practice, however, SAGE is often limited by the presence of orphan and ambiguous tags and by the considerable sequencing costs involved in its application for the study of multiple samples. By contrast, microarrays can be more readily applied to analyses of multiple samples, and follow-up studies on particular sequences are simplified by the immediate availability of the relevant cDNA clones. In contrast to SAGE, a disadvantage of microarrays is that they are limited to analyses of the sequences represented on the array. This limitation is particularly problematic when studying highly specialized tissues such as the retina, because the differentially expressed genes that underlie the unique structure and function of the retina are underrepresented on commercial microarrays. 

One approach to circumvent this problem of limited representation of genes of importance to the retina is to construct custom cDNA microarrays enriched for genes that are likely to be relevant for retinal research. Recently, the feasibility of constructing a mouse retina custom cDNA microarray was reported. ^{16 17} Herein, we report construction of a custom human cDNA microarray that allows for comprehensive expression profiling of genes that are likely to be of interest in studies of the normal and diseased human retina. As one potential application of this microarray, we describe its use to identify human genes with a retina-enriched expression pattern. 

Methods

Generation of PCR Products

We were able to obtain clones representing 10,034 of the 12,000 genes and ESTs included in our gene list (see the Results section for details about the generation of the gene list). Approximately 6,000 of the clones were obtained from the sequence verified 40,000-clone set from Research Genetics (Invitrogen Life Technologies, Carlsbad, CA). Clones were also amplified from two arrayed retina cDNA library sets: Soares (University of Iowa Health Care, Iowa City, IA) N2b4HR retina library (Unigene library number 178; http://www.ncbi.nlm.nih.gov/UniGene; provided in the public domain by the National Center for Biotechnology Information [NCBI], Bethesda, MD), roughly 1500 clones, and a library generated and arrayed by Jeremy Nathans and Amir Rattner (NCBI Unigene library number 226 and 228), approximately 2000 clones. Five hundred additional clones, most of which were sequence verified, that were not found in these sources were purchased individually from Research Genetics. Bacterial clones were rearrayed and cultured for 12 hours at 37°C in 96-well plates containing Luria-Bertani [LB] medium/10% glycerol. Plates were then stored at −80°C. Products for spotting were generated by PCR amplification. A disposable 96-pin replicator was used to add small aliquots from the bacterial growth plates directly to duplicate 96-well PCR plates, with each well containing 100 μL of reaction mix composed of primers M13AEK forward and reverse for the Research Genetics and Soares retina library clones (ctgcaaggcgattaagttgggtaac and gtgagcggataacaatttcacacaggaaacagc), or GT −10 forward and reverse (ggttaagtccaagctgaattc and gggtaaaaagcaaaagaattc) for Nathans’s and Rattner’s retina library, and Taq polymerase (Invitrogen). The resultant duplicate PCR products were combined, purified (MultiScreen; Millipore, Bedford, MA), and resuspended in 50 μL of Tris-EDTA. PCR product concentration after resuspension ranged from 100 to 1000 ng/μL. A sample from each PCR product was evaluated by gel electrophoresis. The overall PCR success rate was 95%. The purified PCR products were then dried, resuspended in 50% dimethyl sulfoxide (DMSO), and arrayed in 384-well plates. 

Probe Labeling

Probes were prepared by minor modification of the method described by Hedge et al. ¹⁸ RNA was extracted using isolation reagent (TRIzol; Invitrogen) according to the manufacturer’s protocol, and 20-μg aliquots of total RNA were treated with DNase (DNAfree; Ambion, Austin, TX), followed by RNA purification (RNeasy kit; Qiagen, Valencia, CA). Purified RNA samples were then incubated for 10 minutes at 70°C with 6 μg random hexamers (Invitrogen), cooled on ice, and used as a template for first-strand cDNA synthesis (5 μL first-strand buffer, 3 μL dithiothreitol [DTT], 2 μL reverse transcriptase (SuperScript II; Invitrogen) and 0.6 μL of a mixture of 10 mM aminoallyl-UTP (Sigma-Aldrich Corp., St. Louis, MO), 15 mM dTTP and 25 mM of dATP, dCTP, and dGTP (Invitrogen)). The 30-μL reverse transcription reaction was incubated overnight at 42°C. The 2:3 aminoallyl-UTP to dTTP ratio used in the RT reaction was selected after testing various ratios. For human RNA, the 2:3 ratio yielded the best results in terms of dye incorporation and hybridization (data not shown). After the RT reaction, 10 μL 1 M NaOH and 10 μL 0.5 M EDTA (pH 8) were added. After incubation for 15 minutes at 65°C, 25 μL of Tris–HCl (pH 7.4) was added followed by a purification step (PCR purification kit; Qiagen) in which the PE wash buffer in the kit was replaced with a phosphate wash buffer (1 M KPO₄ [pH 8] in 80% ethanol). 

The purified cDNA was dried, resuspended in 4.5 μL of 0.1 M carbonate buffer (pH 9), and incubated for 2 hours in room temperature with 12.3 μg of either Cy3 or Cy5 monoreactive dye (Amersham Biosciences, Inc., Piscataway, NJ) suspended in 4.5 μL of DMSO. The labeled probe was then purified (PCR purification kit; Qiagen) and dried in a heated speed vac. The total amount of dye incorporation (measured in picomoles of dye per probe) and the ratio of unlabeled to fluorescent labeled nucleotide in the probe were assessed by measuring probe absorbance at 260, 550, and 650 nm to assess DNA, Cy3, and Cy5 concentration, respectively. ¹⁸ In typical reactions, dye incorporation is more than 300 picomoles and the unlabeled/labeled nucleotide ratio is in the range of 10 to 20. 

Slide Printing, Postprocessing, and Hybridization

Amine-coated slides (SuperAmine; TeleChem, Sunnyvale, CA) were printed using an arrayer with 100-μm-tip pins (MicroGrid II; Biorobotics, Cambridge, UK) in a 60% humidity and 22°C environment. Slides were stored in a desiccator for up to 6 months and treated with 60 mJ of UV light (UV Stratalinker 2400; Stratagene, La Jolla, CA) before use. 

Prehybridization and hybridization were preformed as described by Hedge et al. ¹⁸ Briefly, slides were incubated for 40 minutes in 1% bovine serum albumin (BSA), 5× SSC, and 0.1% SDS at 42°C, washed in water and isopropanol, and dried using compressed air. Cy3 and Cy5 labeled probes were combined in 40 μL hybridization buffer (50% formamide, 5× SSC, 0.1% SDS buffer, 10 μg poly dA, and 10 μg Cot-1 DNA) and heated at 95°C for 3 minutes. Hybridization was performed under a coverslip (Fisher Scientific, Suwanee, GA) in a humidified hybridization chamber (GeneMachine, San Carlos, CA) for 18 hours in a water bath at 42°C. After hybridization, slides were washed in 1× SSC/0.2% SDS at 42°C, followed by washes in 0.1× SSC/0.2% SDS, and 0.1× SSC both at room temperature. Each wash lasted for 4 minutes, and then the slides were dried with compressed air. 

Data Acquisition and Analysis

Slides were scanned using the a confocal laser scanner (ScanArray 5000; Perkin Elmer, Boston, MA). Scans using the maximum laser power setting that did not produce saturated spots were used for analyses. Spot finding and quantification of fluorescence intensity were performed on computer (Imagene software; Biodiscovery, Inc., Marina del Rey, CA). Local background subtraction and normalization were also performed on computer (Genesight software; Biodiscovery, Inc.) by dividing each spot intensity by the mean intensity value of that particular fluorescent channel (either Cy3 or Cy5). 

Normalized study sample–reference sample ratios were analyzed using the significant analyses for microarray (SAM) algorithm. ¹⁹ Two-class unpaired analysis was applied to compare the five retina samples to the liver and cortex samples separately. SAM calculates a false discovery rate (FDR), which is the median percentage of genes from a list that are likely to be mistakenly identified as differentially expressed. According to the SAM algorithm, genes are identified as differentially expressed based on the difference in expression among the sample groups and the consistency of this expression difference. We used similar, but not identical, FDRs to identify common genes in the retina–cortex and retina–liver expression pattern comparisons, because the SAM algorithm does not produce continuous FDR values but rather produces increments that depend on the imputed data. 

In this study novel genes were defined as EST or group of ESTs that are not part of a cloned gene, regardless of whether the EST was assigned to a Unigene cluster. 

Hierarchical clustering was performed and viewed using the Cluster and Treeview (both softwares are available at http://rna.lbl.gov.Eisensoftware.htm; softwares were developed by M. Eisen and provided in the public domain by the Lawrence Berkeley National lab and the University of California at Berkeley, Berkeley, CA) programs, respectively. ²⁰ Principal component analyses (PCAs) were performed on computer (Partek; Partek Inc., St. Charles, MO). Assessment of the tissue distribution of the expressed sequences was performed by BLAST search of each sequence against the human EST database (http://www.ncbi.nlm.nih.gov/dbest/ NCBI, Bethesda, MD). Hits with an E-value less than 1^e−10 were accepted and mapped to the library name and further to the tissue of origin (http://www.ncbi.nlm.nih.gov/UniLib). For each clone the number of unique tissues in which matches were found were counted. 

Donor Tissue

Five eyes were obtained through the National Disease Research Interchange (NDRI, Philadelphia, PA). Whole retinas were dissected and RNA extracted as described under probe labeling. RNA was also extracted from two liver and cortex tissues that were obtained from three donors (Table 1) . 

Quantitative Real-Time PCR

The expression levels of 11 genes and ESTs identified as retina-enriched according to the array experiments were studied by real-time PCR to validate the microarray results. RNA was DNase treated (DNAfree; Ambion), and cDNA was synthesized from 1 μg of retinal, liver, and cortex RNA (retina 5 and liver and cortex 1 in Table 1 ), using commercial reverse transcriptase (Superscript II; Invitrogen). Control reactions without addition of RT enzyme were performed with each cDNA synthesis, and aliquots from these tubes were tested by PCR to ascertain the absence of DNA contamination. Primers used are listed in Web Table 2. Real time PCR reactions were performed using a PCR cycler (Light-Cycler; Roche, Nutley, NJ). Six serial twofold dilutions of the retina sample were defined as standard in all the reactions. PCR products were quantified using the second derivate maximum values calculated by the system’s analysis software (Roche). Expression levels were normalized to the acidic ribosomal phosphoprotein P0 (ARP) mRNA levels. ²¹ Each EST was tested in duplicate PCR reactions, and the mean of the two reactions was used for calculating the expression levels. 

Results

Microarray Design, Optimization, and Validation

The first criterion we used for selection of genes to be included in the human retinal cDNA microarray was expression pattern—specifically, representation in a human retina and/or retinal pigment epithelium (RPE) cDNA library. The initial gene list was generated using EST databases from NCBI (http://www.ncb.nlm.nih.gov), The Institute of Genomic Research (TIGR; http://www.tigr.org/ Rockville, MD), and sequences from a human RPE cDNA library that was constructed in our laboratory (Chang et al., unpublished results, 1996). A second and complementary criterion was gene function. Included were groups of genes involved in phototransduction, visual cycle, photoreceptor structure, and retinal and neuronal development. Genes from other functional classes such as receptors, signal transduction molecules, cell adhesion molecules, and transcription factors were also included, even if their expression in the retina and/or RPE had not been previously demonstrated. Last, genes known or suspected to be involved in retinal diseases (retinal and macular degeneration, glaucoma, retinal neovascularization), and genes related to general pathologic processes (inflammation, apoptosis, ischemia) were also included. The databases that were used to identify these genes included the NCBI, TIGR, SOURCE (http://genome-www5.stanford.edu/cgi-bin/SMD/source/ hosted in the public domain by Stanford University, Stanford, CA) and Ret Net (http://www.sph.uth.tmc.edu/Retnet/ hosted by University of Texas Houston Health Science Center, Houston, TX). The resultant gene list was made nonredundant based on each gene’s Unigene cluster. ESTs that were not assigned to any Unigene cluster were blasted against the nr database (http://www.ncbi.nlm.nih.gov/Blast; NCBI) and were included on the array if no other identical sequence from a different EST was on the array. The final list included approximately 12,000 genes and ESTs. 

We were able to obtain plasmids representing 10,034 of the sequences on our master list. Of these, 67% are from known genes and 33% are from ESTs. About half of the known genes on the array have been characterized. Their encoded proteins are involved in variety of biological (Fig. 1A) and molecular (Fig. 1B) processes. 

After generating and purifying PCR products representing the 10,034 cDNAs, we tested various conditions for spotting and hybridization, and obtained stronger and more consistent signals using 50% DMSO compared with aqueous spotting buffers (data not shown). cDNA labeling methods were also optimized and compared. We found that the indirect method (with aminoallyl-UTP), compared with direct dye incorporation, was superior in terms of amount of dye incorporated per probe as well as the ratio of labeled/unlabeled nucleotide in the probe, and yielded higher signal-to-noise ratios on hybridization (data not shown). In addition, indirect labeling yielded similar incorporation of Cy3 and Cy5, whereas with the direct method Cy5 incorporation was significantly less than that of Cy3. 

To assess overall performance and reliability of our microarray analyses, we performed self–self hybridizations. Forty micrograms of reference sample total RNA was divided into two aliquots: Half was labeled with Cy3, and the other half was labeled with Cy5, and then the two were mixed and hybridized together (Fig. 2) . A high degree of correlation was achieved between the two channels (correlation coefficient, R ² = 0.9432), demonstrating the reproducibility of the labeling, hybridization, and image analysis processes. However, it should be noted that a small fraction of the genes artifactually appeared to be differentially expressed: 21 of the 10,034 sequences showed twofold expression difference in the hybridization presented in Figure 2 . This is a common finding in microarray studies that underscores the importance of performing replicate arrays with dye swapping for each sample. The signal-to-background ratios varied across the spots on each slide and across arrays, with average signal-to-background ratios of the Cy3 and Cy5 channels on different slides varying from 2 ± 1.9 to 9.8 ± 6.7, and the percentage of spots with signal-to-background ratios higher than 1.4-fold varied from 36% to 85%. 

Identifying Retina-Enriched Genes

The custom microarray was then used to identify retina-enriched genes by comparing the expression profile of retina to that of brain (cortex) and liver. Analysis of each tissue was performed with at least two donors to decrease bias that may be introduced by donor-specific gene expression patterns. A reference experimental design was used; retina (n = 5), liver (n = 2) and cortex (n = 2) RNA samples were studied in duplicate with two microarray slides for each sample (total of 18 microarrays). In the first hybridization, the study sample was labeled with Cy3 and a human reference sample composed of 80% cortex RNA, 15% retina RNA and 5% retinal pigment epithelium RNA was labeled with Cy5. In the second hybridization, dyes were swapped between the study and the reference samples. This design enables comparison across all samples (by using a common reference sample) and corrects for possible differential effects of Cy3 and Cy5 dye on labeling and hybridization. 

As would be expected based on their shared neuronal identity, the expression pattern of the retina was more similar to that of cortex than that of liver. This was reflected by the finding that, at similar significance levels, more than twice as many genes were preferentially expressed in the retina compared with liver than in the retina compared with cortex based on the SAM algorithm (Web Tables 3 and 4). As outlined in the Methods section, we have assessed the statistical significance of the differential gene expression patterns by using the FDR as determined with the SAM algorithm.¹⁹ The FDR is the median number of genes likely to be falsely identified as differentially expressed between samples. For example, at a moderately stringent FDR of 26% for the retina–cortex and 27% for the retina–liver comparisons (the design of the SAM program makes it difficult to achieve identical FDRs), there were 452 and 1097 genes that were preferentially expressed in the retina in the retina–cortex and retina–liver comparison, respectively (Web Tables 3 and 4). Fig. 1 shows the functional annotation of genes from Web Tables 3 and 4. Known photoreceptor genes, which constitute the best-defined group of retina-specific genes, composed a larger portion of the identified retina-enriched genes from the retina–cortex versus retina–liver comparison (46% vs. 26%, respectively; Fig. 1 ). In contrast, a larger absolute number of known photoreceptor genes were correctly identified by the retina–liver versus the retina–cortex comparison (31 vs. 25 genes, respectively). These findings presumably stem from the higher retina–cortex than retina–liver similarity and reflect better specificity for the retina–cortex comparison, and better sensitivity for the retina–liver comparison for identifying retina-specific genes. Of note, both inflammation and immune response genes were rare or absent among the retina-enriched genes compared with their significant representation among the sequences on the array (Fig. 1) , perhaps reflecting the so-called immune-privileged status of the retina. 

To focus more on individual differentially expressed genes, so as to select genes of potential interest for further evaluation, we reanalyzed the data with a more stringent FDR. With FDRs of 10.3% and 9.5% for the retina–cortex and retina–liver comparisons, 345 and 720 retina-enriched genes were identified, respectively. Of these genes, 211 were common to both lists (Table 5 lists the top 40 of these genes; the full list is available in Web Table 5). Of these, 46% (98 genes) were ESTs representing novel retina-enriched genes, and 36% (77 genes) and 17% (36 genes) were cloned genes with known or unknown function, respectively. Of those with known function, 32% (25 genes) represent already characterized photoreceptor-enriched or photoreceptor-specific genes (Fig. 3) , compared with 10% percent of known photoreceptor-enriched genes of the genes with known function on the array (Fig. 1) . This was a reassuring finding and suggests that a significant fraction of the remaining 186 genes in the group (Table 5) are likely to represent retina-enriched genes. 

The gene expression patterns identified by the microarray studies were also analyzed by both hierarchical clustering and principle component analysis (PCA) to ascertain the consistency of the observed tissue-specific expression patterns identified by the array experiment (Fig. 4) . When all sequences on the array were used for clustering, the five retina samples clustered together and were distinct from both the cortex and the liver. When only the retina-enriched genes set identified by the SAM analysis (Table 5) was used for clustering, the overall pattern was similar, but the retinas were more separated from the cortex and liver samples. This is apparent by the longer branches that separate the retina samples from the cortex and liver samples in the hierarchical clustering (Figs. 4A 4B) . The liver and cortex samples cluster together, probably reflecting the bias of our microarray toward retina-enriched genes, which masks the differences between cortex and liver gene expression. Furthermore, the overwhelmingly different gene expression pattern of the retina versus cortex and liver observed our custom retina array also probably results in blunting of the closer similarity of the retina–cortex expression pattern compared with the retina–liver expression pattern that was obvious with the SAM analyses. PCA results were consistent with the hierarchical clustering (Figs. 4C 4D) . Again, retina samples segregated more closely together and further from the liver and cortex samples when only retina-enriched genes as opposed to all genes on the array were used for clustering: The mean retina–cortex and retina–liver distance by PCA were 0.84 and 0.72, respectively, when all genes were used for clustering. When only retina-enriched genes were used, the mean retina–cortex and retina–liver distances were 2.1 and 2.5, respectively. 

As a complementary approach to assessing the expression pattern of the genes identified as retina-enriched by the array analysis, we determined the distribution pattern of each identified gene in tissue EST libraries available at NCBI. As a summary value, we counted how many different tissue libraries contained a representative of each microarray-identified gene. We compared the distribution pattern of the sequences identified as retina-enriched to that of 1000 randomly selected sequences from the array and to that of 1000 randomly selected sequences from the human genome. The analysis showed that approximately 38% of the retina-enriched genes were identified only in one tissue (Web Table 6, shows the library distribution of the full set of retina-enriched genes). Thirteen of the 35 sequences that were present in two tissues were found only in the central nervous tissue in addition to the eye. By comparison, approximately 7% of the sequences in the genome were cloned from a single tissue (not necessarily retina). Most of the retina-enriched genes were expressed in several different tissues. Approximately 25% of such genes (compared with 64% of the randomly selected genes from the human genome) are expressed in more than 10 tissues (Fig. 5) . The total set of genes represented on the retina custom array showed corresponding values between the retina-enriched gene and the randomly selected gene sets, presumably reflecting the selection bias used in generating the retina array (Fig. 5) . 

To validate our microarray results using an independent experimental method, the retina-enriched expression pattern of 11 of the retina-enriched genes and ESTs was studied by quantitative real-time RT-PCR (QPCR) on one of the retina samples (retina 5). The differential expression pattern of all 11 genes was confirmed, but as noted previously, ^{22 23} the expression ratio obtained by array analysis often underestimated that obtained by QPCR. For the 11 genes and ESTs studied, the array analysis indicated retina-enriched expression ratios of 2.3- to 24.8-fold, whereas the QPCR results indicated ratios of 1.4- to 7023-fold (Table 7) . 

Comparison of High-Throughput Retinal Gene Expression Studies

To assess the similarities and differences between the various approaches that are being used to profile retinal gene expression patterns, we compared our results with those of two recent SAGE studies, one of mouse ¹ and one of human retina, ³ and another study based on EST data mining. ⁴ The comparison served to validate the identification of novel retina-enriched genes by each of the studies, as a number of genes were found in common, but also revealed significant differences between the approaches, probably reflecting both different experimental design and technical issues. 

We included in the analysis the genes from the retina–cortex comparison at an FDR of 26% (Web Table 3 in the present study), genes in Table S11 from Blackshaw et al., ¹ genes in Table 11 from Sharon et al., ³ and genes in Table 2 from Stohr et al. ⁴ In all cases the genes from these tables were identified by the authors as retina-enriched, but it should be noted that not only different methods but also different species (human and mouse) and different statistical criteria were used to identify retina-enriched genes in each of the studies. Known photoreceptor genes were excluded from our analysis, both because they are generally the easiest to detect and because they were not included in two of the studies. ^{1 4} In addition, because we used Unigene clusters to compare findings across the different studies, only known genes and ESTs that were assigned to such clusters were compared. 

Table 8A shows the number of genes that were identified in two or more studies. The degree of overlap ranged from 23.9% (percent of genes from Sharon et al. ³ that were identified by us as well) to 0% (no overlap between Blackshaw et al. ¹ and Stohr et al. ⁴ ). Overall, the present study showed more overlap with the other studies than any of the other studies showed with each other. We identified 44 retina-enriched genes that were also identified in one of the other studies, and three genes that were identified in two of them (Table 8B) . No genes were identified that were shared by two other studies but not by ours, and, perhaps surprisingly, there were not any genes that were identified as differentially expressed in all four studies. 

Discussion

Several recent high-throughput studies have focused on identifying retina-enriched genes. ^{1 2 3 4 5 6 7 8} Blackshaw et al. ¹ used SAGE analyses to identify 264 novel mouse rod-enriched or -specific genes. Sharon et al. ³ used SAGE to obtain gene expression profiles of two human retinas. By comparing their results to 100 nonocular SAGE libraries, they identified 89 retina-specific or -enriched SAGE tags. ³ In another recent study, Stohr et al. ⁴ used EST data mining to identify several hundred retina-enriched or -specific Unigene clusters. The expression patterns of ESTs from 180 of these Unigene clusters were then assessed by RT-PCR, and 39 retina-specific ESTs were identified. ⁴ However, because the RT-PCR used was not quantitative, it is possible that additional ESTs from the group of 180 may exhibit a retina-enriched expression pattern. 

We have constructed a cDNA microarray designed for analysis of human retinal gene expression under normal and pathologic conditions. As an initial use for the microarray, we have applied it to characterize gene expression pattern differences between retina, another central nervous system–derived tissue (cerebral cortex) and liver. This comparison enabled the identification of more than 100 novel retina-enriched genes, as well as a large group of known genes that were not previously reported to have a retina-enriched expression pattern. Several of these known genes may play important roles in the physiology of the retina. For example, thrombomodulin, an endothelial cell surface glycoprotein that forms a complex with thrombin and converts it into a potent anticoagulant, was previously detected in chorioretinal blood vessels and in the anterior eye segment. ^{24 25} The finding of its preferential expression in the retina suggests a possibly significant role in maintaining retinal and choroidal blood flow. 

Both bioinformatic and experimental data support the validity of the microarray-derived expression results. First, the microarrays correctly identified multiple known retina-enriched genes. Second, the bioinformatic analysis we used, based on the SAM algorithm, is tested for internal consistency in the microarray data by looking for expression differences that were consistent across the five retina samples compared with the brain and liver samples. ¹⁹ Third, the microarray analysis identified 47 genes that have been reported recently as retina-enriched by other investigators. ^{1 3 4} Fourth, we confirmed the microarray findings for 10 of the newly identified differentially expressed genes and ESTs with an independent experimental method, QPCR. Consistent with previous reports, the QPCR data in fact suggest that the microarray results tended to underestimate the identified differential expression patterns. ^{22 23 26} Fifth, analysis of available databases revealed that ESTs corresponding to the microarray identified retina-enriched genes were identified in fewer tissue sources than the overall clones represented on the array, which in turn were present in fewer tissues than sequences randomly selected from the genome. Last, clustering of the microarray data by two different algorithms successfully identified the five retina samples as being distinct from the liver and cortex samples. 

Several caveats about our study should be acknowledged. First, although the retina custom microarray contains a substantial representation of the genes (both novel and known) that have been identified in retina cDNA libraries, additional retina-enriched genes not represented on the microarray are likely to exist. Second, we have studied whole retinas, but because the retina contains approximately 50 different cell types, ²⁷ conclusions regarding the gene expression patterns in particular retinal cell types cannot be directly drawn from our study. This problem may be addressed by expression profiling of homogeneous cell populations isolated by techniques such as laser capture microscopy ^{28 29} or single-cell isolation methods ³⁰ (Wahlin et al., manuscript in preparation). A challenge in using these methods for microarray studies, however, is that they require amplification, a process that potentially introduces significant bias to the analysis. 

Although there was considerable overlap between our microarray results and the previous SAGE and bioinformatic studies, ^{1 3 4} it seems significant that many of the novel retina-enriched genes were identified by only one of the studies. Several factors may account for this surprising finding. Species differences in gene expression pattern may contribute to the relatively small overlap between the mouse SAGE study ¹ and the human SAGE, EST data mining, and microarray studies. ^{3 4} As an example of potential species differences, we have identified several genes that are highly expressed in bovine and human RPE, but only minimally if at all in murine RPE, as well as a human and bovine retinal transcription factor that appears not to even exist in the mouse genome (Wang et al., manuscript submitted). However, given the major structural and functional similarities between murine and human retina, species differences probably account for only a small fraction of the observed expression profile differences. A more likely factor is the probable existence of a large number of unknown retina-enriched genes, together with the likely possibility that each method and approach is sampling only a limited and partially overlapping subset of this larger set. Additional microarray, SAGE, and bioinformatic studies, together with proteomic and other approaches, should eventually more definitively define the unique expression profiles of the murine and human retina. 

 Supported in part by grants from the National Eye Institute, Macular Vision Foundation, Foundation Fighting Blindness, Fight For Sight, and Santen Pharmaceutical, and by generous gifts from Marshall and Stevie Wishnack and from Robert and Clarice Smith. PAC is the George S. and Dolores Dore Eccles Professor of Ophthalmology and Neuroscience, and DJZ is the Guerrieri Professor of Genetic Engineering and Molecular Ophthalmology.

 Submitted for publication October 23, 2002; revised April 8, 2003; accepted April 18, 2003.

 Disclosure: I. Chowers (P); T.L. Gunatilaka, None; R.H. Farkas, None; J. Qian, None; A.S. Hackam, None; E. Duh, None; M. Kageyama, Santen Pharmaceutical (E, P); C. Wang, None; A. Vora, None; P.A. Campochiaro, None; D.J. Zack, Santen Pharmaceutical (F, P)

 The publication costs of this article were defrayed in part by page charge payment. This article must therefore be marked “advertisement” in accordance with 18 U.S.C. §1734 solely to indicate this fact.

 

Table 1.

 

Tissue Donors

Tissue	Donor Age (y)	Gender	Race	Cause of Death
Retina 1	46	Female	White	Myocardial infarction
Retina 2	90	Female	White	Myocardial infarction
Retina 3	58	Female	White	Myocardial infarction
Retina 4	68	Female	White	Cerebral vascular accident
Retina 5	44	Male	White	Smoke inhalation
Cortex 1	80	Female	White	Congestive heart failure
Liver 1
Cortex 2	58	Female	White	?
Liver 2	51	Female	Black	Sepsis

Figure 1.

 

Classification of the gene set represented on the array (□), and retina-enriched genes compared with cortex ( ; FDR = 26%) and liver (▪; FDR = 27%), based on biological (A) and molecular (B) function. Gene function was classified based on the gene ontology database (http://www.geneontology.org/). Only those genes that were present in the gene ontology database were classified. In (A), the number of genes classified from the array, retina–cortex, and retina–liver comparisons was 1106, 49, and 86, respectively. In (B), the number of genes classified from the array, retina–cortex, and retina–liver comparisons was 2098, 71, and 177, respectively. As indicated, the known photoreceptor gene set is enriched for genes that are differentially expressed in the retina, whereas the inflammation and immune response categories are deficient in such differentially expressed genes.

Figure 2.

 

Scattergram showing self–self hybridization to human retinal microarray. A 40-μg reference RNA sample was divided into equal aliquots, one of which was labeled with Cy3 (x-axis), whereas the other was labeled with Cy5 (y-axis). Each axis shows the normalized expression level expressed in log base 2. The correlation coefficient is shown.

Table 5.

 

Top 40 Retinal Enriched Genes and ESTs Compared with Cortex and Liver Sorted by SAM Score of the Retina–Cortex Comparison

Accession Number	Unigene Cluster	Gene Name	Gene Symbol	Cytoband	Expression Ratio
										Retina/Cortex	Retina/Liver
AA057232	Hs.32721	S-antigen: retina and pineal gland	SAG	2q37.1	40.0	5.1
NM_002602	Hs.1857	phosphodie sterase 6G, cGMP-specific, rod, gamma	PDF6G	17q25	17.6	4.1
H38198	Hs.391386				12.5	6.6
H87212					6.3	2.6
W26299	Hs.139041	ESTs			18.3	7.4
H87729					6.9	4.8
AA057185					10.4	8.1
AA971179	Hs.128096	Melanin-concentrating hormone receptor 1 interacting zinc-finger protein	MIZIP	9q34.3	11.8	4.3
AA015930	Hs.63085	Membrane protein, palmitoylate d4	MPP4	2q33.2	7.6	7.0
AA233809	Hs.169300	Transforming growth factor, beta 2	TGFB2	1q41	6.3	5.0
H87222		EST			5.1	5.8
W96259	Hs.182648	Homo sapiens cDNA FLJ14444 fis, clone HEMBB100 1153	4.78526	2.98927
AI656410	Hs.89606	Neural retina leucine zipper	NRL	14q11.1-q11.2	4.1	1.6
W26499					5.6	2.9
H87232	Hs.135058	ESTs			4.7	4.2
H69531	Hs.396489	Transferrin	TF	3q2.1	6.3	7.3
AI393043	Hs.36973	Guanine nucleotide binding protein, alpha transducing activity polypeptide 2	GNAT2	1p13.1	8.7	6.1
H59861	Hs.2030	Thrombomodulin	THBD	20p12-cen	7.8	2.9
R85753	Hs.144156	EST			10.9	8.2
AA284693	Hs.3005	Transcription factor AP-4	TFAP4	16p13	5.4	4.2
R40057	Hs.112360	Prominin-like 1 (mouse)	PROML1	4p15.33	5.6	2.8
AA058563	Hs.195851	Actin, alpha 2, smooth muscle, aorta	ACTA2	10q23.3	5.2	3.2
AA039370	Hs.42458	TEA domain family member 1	TEAD1	11p15.4	5.5	3.2
AA291577	Hs.1706	Interferon-stimulated transcription factor 3, gamma 48kDa	ISGF3G	14q11.2	5.7	2.9
H38839	Hs.129882	Interphotoreceptor matrix proteoglycan 1	IMPG1	6q14.2-q15	10.6	14.4
AA015841	Hs.73112	Guanine nucleotide binding protein gamma transducing activity polypeptide 1	GNGT1	7q21.3	3.6	2.7
W96535	Hs.137556	MT-protocadherin	KIAA1775	10q22.1-q22.3	4.3	5.4
AA018809	Hs.194673	Phosphoprotein enriched in astrocytes 15	PEA15	1q21.1	3.7	2.2
AA021348	Hs.66727	Potassium inwardly rectifying channel, subfamily J, member 10	KCNJ10	1q22-q23	5.8	5.9
W29004	Hs.81728	Unc-119 homologue	UNC119	17q11.2	6.3	6.1
AF024711	Hs.249186	Cone-rod homeobox	CRX	19q13.3	3.8	5.4
W23098	Hs.373518	ESTs			5.3	3.3
H50186	Hs.184793	Hypothetical protein DKFZp434 G2311	DKFZp434 G2311	9q34.3	4.1	3.9
AA954230	Hs.281564	Retinal outer segment membrane protein 1	ROM1	11q13	3.6	2.5
H95537	Hs.146067	ESTs			7.3	10.1
W26244					6.4	3.0
W23262	Hs.20930	Poly(rC) binding protein 4	PCBP4	3p21	7.3	3.3
AA069770	Hs.84244	Potassium voltage-gated channel, Shab-related subfamily, member 1	KCNB1	20q13.2	5.3	8.2
AA017021	Hs.232072	Usher syndrome 2A	USH2A	1q41	4.3	5.1
W27799	Hs.71642	Guanine nucleotide binding protein G	GNB3	12p13	3.7	3.4

Figure 3.

 

Classification of retina-enriched genes common to retina–cortex and retina–liver comparison at false discovery rates of 10.3% and 9.5%, respectively (total of 224 sequences which are listed in Table 5 and Web Table 5).

Figure 4.

 

Hierarchical clustering (A, B) and PCA (C, D) of the individual retina, cortex, and liver samples. Cluster analysis was performed with the program Cluster, using average linkage clustering of all genes with more than 80% of values present and with at least one of the expression ratios having a value greater than 1.4. Analyses were performed with both the full set of genes on the array (A, C) and with the subset of retina-enriched genes (B, D). Difference in expression profile between samples is represented by branch length in (A) and (B), and by distance in space between samples in (C) and (D). In (C) and (D): retina (▴), cortex (•), liver (▪). With both hierarchical clustering and PCA, the separation of the retina samples from the other samples is clearer when the analysis is restricted to the retina-enriched gene subset.

Figure 5.

 

Expression pattern of retina-enriched genes, microarray gene set, and random ESTs, as assessed by a bioinformatic analysis of representation in ESTs tissue libraries. Analysis was performed as described in the text. The x-axis is the number of tissues from which a sequence identical with the one that was submitted for a BLAST search was cloned. The y-axis represents the fraction of sequences in each of the three groups (random ESTs □; retina custom array ; and retina-enriched genes ▪) that were found in each number of tissues. The microarray-defined retina-enriched genes, as would be expected, are overrepresented in the subsets of genes that are found in only one or two tissue libraries, and are underrepresented in the subsets that are found in more than 10 tissue libraries.

Table 7.

 

Comparison of Microarray and Quantitative PCR Results

Accession No.	Gene Name	Retina/Cortex Change (×)			Retina/Liver Change (×)
				Microarray (Retina 5)	Quantitative PCR	Microarray (Retina 5)	Quantitative PCR
H38838	Aryl hydrocarbon receptor interacting protein-like 1; AIPL1	3.1	49.1	6.0	17.2
AA058563	Actin, alpha; ACTA2	12.5	5.7	3.9	77.3
H59861	Thrombomodulin	7.4	10.2	2.3	8.3
AA971179	Melanin-concentrating hormone receptor 1 interacting zinc-finger protein; MIZIP	10	89	4.6	306.9
AA057232	Arrestin; SAG	24.8	7023.0	25.2	4565.0
H85173	PRO2405 hypothetical protein	5.8	8.8	3.5	1.4
W26299	KIAA1337 protein	43.1	1383.3	Undetectable in liver	1493.5
H38089	EST	2.3	20.8	5.5	18.6
H28997	EST	2.5	16.0	Undetectable in liver	1566.0
AA012852	EST	3.6	2.3	Undetectable in liver	935.7
AA017422	EST	6.2	2.9	Undetectable in liver	3.1

Table 8A.

 

Comparison of Four High-Throughput Studies Focused on Identifying Retina-Enriched Genes

	Chowers et al.	Blackshaw et al.	Sharon et al.	Stohr et al.
Chowers et al. (284)	—	8	11	28
Blackshaw et al. 1 (203)	8	—	1	0
Sharon et al. 3 (46)	11	1	—	2
Stohr et al. 4 (180)	28	0	2	—

Table 8B.

 

Retina-Enriched Genes Identified in the Present and in Three Previous High-Throughput Retina Gene Expression-Profiling Studies

Accession No.	Unigene Cluster	Gene Name	Study
W96259	Hs.35493	ESTs	1, 3 (R)
AA521470	Hs.13768	Homo sapiens mRNA; cDNA DKFZp434I1216	1
W27668	Hs.154131	Voltage-gated potassium channel Kv11.1	1
AA015799	Hs.33792	ESTs	1, 3
W96535	Hs.137556	MT-protocadherin	1, 2
AA058694	Hs.63085	Membrane protein, palmitoylated 4 (MAGUK p55 subfamily member 4)	1
AA664180	Hs.336920	Glutathione peroxidase 3 (plasma)	1
AA450189	Hs.146580	Enolase 2, (gamma, neuronal)	1
R65782	Hs.153684	Frizzled-related protein	1
AA487505	Hs.70337	Immunoglobulin superfamily; member 4	1
H69531	Hs.284176	Transferrin	1
AA011139	Hs.134503	PR domain containing 8	2
AA291577	Hs.1706	Interferon-stimulated transcription factor 3, gamma (48 kDa)	2
AA013300	Hs.198726	Cold shock domain protein A	2
AA057520	Hs.30824	Leucine zipper transcription factor-like 1	2
W92507	Hs.332633	Bardet-Biedl syndrome 2	2
W26014	Hs.4082	Lectin, galactoside-binding, soluble, 8 (galectin 8)	2
AA021101	Hs.7857	Erythrocyte membrane protein band 4.1-like 2	2
AA600189	Hs.7957	Adenosine deaminase, RNA-specific	2
R56082	Hs.8071	Synaptic vesicle protein 2B homolog	2
AA069770	Hs.84244	Potassium voltage-gated channel, Shab-related subfamily, member 1	2
AA017482	Hs.138944	ESTs	3 (B/R)
R85753	Hs.144156	ESTs	3 (R)
AA016013	Hs.60673	Homo sapiens ALS2CR5 mRNA, splicing variant	3 (R)
AA053945	Hs.40814	ESTs	3
AA019773	Hs.24901	ESTs	3
AA019201	Hs.269210	ESTs	3 (B/R)
AA056082	Hs.40918	ESTs	3
AA897784	Hs.64616	Chromosome 12 open reading frame 3	3 (R)
H38836	Hs.32840	ESTs	3 (R)
AA010974	Hs.60473	RFamide-related peptide precursor	3 (R)
AA054222	Hs.40400	ESTs	3 (R)
AA021504	Hs.269223	ESTs	3
AA054260	Hs.28411	ESTs	3 (R)
H84529	Hs.40594	ESTs	3 (R)
H40626	Hs.32795	ESTs	3
AA046909	Hs.60677	ESTs	3
R17747	Hs.20935	Hypothetical protein DKFZp761D221	3
H92356	Hs.271692	ESTs	3 (R)
AA001426	Hs.40863	ESTs	3
AA020741	Hs.171611	ESTs	3
AA017607	Hs.60802	ESTs	3
AA018469	Hs.40486	ESTs	3 (B/R)
H85885	Hs.40838	ESTs	3
AA053974	Hs.269238	ESTs	3 (R)
R85821	Hs.269224	ESTs	3
R85422	Hs.268813	ESTs	3 (B/R)

Supplementary Materials

Table 2 (PDF) - 10KB 

Web Table 2 (Excel) - 16KB 

Web Table 3 (Excel) 

Web Table 4 (Excel) - Description 

Web Table 5 (Excel) 

Web Table 6 (Excel) 

The authors thank Jeremy Nathans and Amir Rattner for generously providing one of the retina cDNA libraries that was used for the construction of the microarray.