Global Gene Expression Analysis of the Developing Postnatal Mouse Retina

Abstract

purpose. Postnatal mouse retinal development involves glial and neuronal differentiation, vascularization, and the onset of vision. In the current study, the gene expression profiles of thousands of genes in the developing postnatal mouse retina were analyzed and compared in a large-scale, unbiased microarray gene expression analysis.

methods. For each of eight different time points during postnatal mouse retinal development, two separate sets of 30 retinas were pooled for RNA isolation, and gene expression was analyzed by hybridization to gene chips in triplicate (Mu74Av2; Affymetrix, Santa Clara, CA). Genes were sorted into clusters based on their expression profiles and intensities. Validation was accomplished by comparing the microarray expression profiles with real-time RT-PCR analysis of selected genes and by comparing selected expression profiles with predicted profiles based on previous studies.

results. The Mu74Av2 chip contains more than 6000 known genes and 6500 estimated sequence tags (ESTs) from the mouse Unigene database. Of these, 2635 known gene sequences and 2794 ESTs were expressed at least threefold above background levels during retinal development. Expressed genes were clustered based on expression profiles allowing potential functions for specific genes during retinal development to be inferred by comparison to developmental events occurring at each time point. Specific data and potential functions for genes with various profiles are discussed. All data can be viewed online at http://www.scripps.edu/cb/friedlander/gene_expression/.

conclusions. Expression analysis of thousands of different genes during normal postnatal mouse retinal development as reported in this study demonstrates that such an approach can be used to correlate gene expression with known functional differentiation, presenting the opportunity to infer functional correlates between gene expression and specific postnatal developmental events.

Retinal development is a complex process involving glial and neuronal differentiation, the establishment of specific neuronal pathways, functional synapse formation, the formation of vascular plexuses, and, finally, the onset of vision. In the mouse, much of this development takes place during the first 3 weeks after birth in a process very similar to third-trimester human retinal development. ¹ Mouse retinal development is initiated at approximately embryonic day (E)12 and involves proliferation, migration, and differentiation of neuronal precursor cells, a process that continues through final neuronal differentiation and the maturation of neuronal synapses at approximately postnatal day (P)8. ² Various factors including sonic hedgehog, ^{3 4} fibroblast growth factors, ^{5 6} math5, ⁷ reelin, ⁸ and cadherins ⁹ are known to be essential for neuronal differentiation and guidance during retinal development. Coincident with the final stages of neuronal differentiation and retinal vascular development, the mice open their eyes, and vision is initiated at approximately P14. Thus, it is reasonable to believe that the genes controlling the expression of many factors critical to establishing the visual machinery and pathways are turned on during the first two postnatal weeks. 

Another key physiological process that occurs during postnatal mouse retinal development is vascularization. ^{1 10} The mouse retina is avascular at birth. A normal adult-like vascular pattern is then formed through a triphasic process as vessels migrate from the central retinal artery and form a primary or superficial vascular network from P0 to P10. Vessels from the superficial layer branch at approximately P8 and migrate to form a secondary, deep vascular plexus during the second postnatal week. Finally, a tertiary, intermediate vascular plexus forms between the primary and secondary layers during the third postnatal week, producing the final mature vascular pattern (see Fig. 1A ). ¹⁰ The mouse retinal development model can thus be used to study factors involved in developmental vascularization. Recent studies in our and other laboratories have begun to suggest that many of the factors involved in neuronal guidance and differentiation may also be involved in directing retinal angiogenesis. ^{10 11 12}  

Microarray gene expression analysis of a single time point during initial mouse retina neuronal formation from E14.5 has been reported ¹³ and serves to illustrate the utility of this approach in analyzing retinal development. In the present study, the retinal gene expression profiles of thousands of genes were analyzed during postnatal mouse retinal development. Mu74Av2 chips (Affymetrix, Santa Clara, CA) were used to screen most of the characterized mouse genes and estimated sequence tags (ESTs) available in the mouse Unigene database (build 74). This in vivo analysis not only allows for the identification of genes that are expressed coincident with the onset of various important functions during retinal development, but also may serve to identify gene products that are common to several different retinal processes, such as neuronal guidance and differentiation, vascularization, and subsequent functioning of the visual cascade. 

Materials and Methods

RNA Preparation

All animal experiments were performed according to the ARVO Statement for the Use of Animals in Ophthalmic and Vision Research. Eyes were enucleated from BALB/c mice killed on P0, P4, P8, P10, P12, P14, or P21, or from adult mice (6 weeks of age). Retinas were carefully dissected in RNase-free medium, transferred immediately into extraction reagent (TRIzol; Invitrogen-Gibco, Rockville, MD) on ice, and stored at −80°C for less than 1 month. In two separate experiments, 30 retinas from each time point were pooled, and total RNA was isolated according to the manufacturer’s protocol. Briefly, retinas were completely lysed in the extraction reagent, chloroform was added, and RNA was isolated from the aqueous phase by phenol-chloroform separation. RNA was then precipitated with isopropanol, washed with 70% ethanol, and redissolved in RNase-free buffer. Isolated RNA was further purified with a kit (RNeasy; Qiagen, Valencia, CA). Integrity of the RNA was verified by visualization of the 28S and 18S ribosomal RNA bands. 

Hybridization to Microarray Chips

All RNA samples for microarray analysis were prepared and hybridized to Mu74Av2 gene chips (Affymetrix, Santa Clara, CA) according to the manufacturer’s instructions. Briefly, total RNA was used for reverse transcription synthesis of double-stranded cDNA. The first strand was primed using a T7-(dT)₂₄ oligonucleotide primer containing a T7 promoter sequence on the 5′ end. After second-strand synthesis, the cDNA was purified by a phenol-chloroform extraction and ethanol precipitation. All cDNA reactions were performed with a commercial system (Superscript Choice System; Invitrogen-Gibco) according to the manufacturer’s protocol. The purified, double-stranded cDNA was then used for in vitro transcription, incorporating biotin-labeled NTPs with an RNA transcript labeling kit (BioArray High Yield; Enzo Biochem, New York, NY). The cRNA was then purified, fragmented, mixed with manufacturer-supplied control polynucleotides (a biotin-labeled oligonucleotide and four control cRNA sequences) and hybridized to the chip. After hybridization, the chip was washed, stained with streptavidin-conjugated phycoerythrin dye (Molecular Probes, Eugene, OR) and scanned on a chip reader (Affymetrix). 

Statistical Analysis

Coefficient of variance (COV) levels were calculated for the expressed genes by dividing the standard deviation by the mean expression level of each cRNA replicate. In addition, the correlation between expression levels and noise variance was calculated by correlating θ with ς where θ = mean [LN(expression)] and ς = SD [LN(expression)]. In each case, our data fell well within acceptable limits. ¹⁴ These graphs can be found in the online supplemental data, in the Statistical Analysis section (http://www.scripps.edu/cb/friedlander/gene_expression/stat.html). 

Data Analysis

Computer software that accompanied the microarray (Suite 5 analysis software; Affymetrix) was used to process the data from the gene chips, and the data were scaled to a mean intensity of 250. Chip-to-chip normalization was initially achieved by comparing the hybridization intensities of control probe sets on each of the microarray chips. During analysis, a final normalization step was performed by setting the median expression signal for each chip to a similar arbitrary value. Because less than 50% of the genes (∼43%) were expressed above background, the median value for each separate experiment fell near upper noise regions. By performing this small, final normalization step, we ensured that all background levels were comparable, minimizing the likelihood of generating false positives without losing important global expression changes of expressed genes. This allowed the raw expression values generated by each experiment to be directly comparable. 

For profile clustering, each gene was normalized to itself; the median expression value for each gene across the time series was set to 1, so that the relative changes in gene expression could be easily observed. A computer was used (GeneSpring software; Silicon Genetics, Redwood City, CA) for bioinformatics analysis and profile clustering. The gene profile clusters were generated using both k-means and self-organizing map (SOM) algorithms based on average linkage using the Pearson correlation. Gene trees were also generated based on the Pearson correlation, average linkage. 

Real-Time RT-PCR

Primer-probe pairs specific for each gene of interest were designed on computer (PrimerExpress software; PerkinElmer, Boston, MA). Polynucleotide sequences of 100 to 120 nucleotides (amplicons) corresponding to the region covered by the primers and probe were made synthetically (Operon, Alameda, CA), quantified, and used for the creation of a standard curve for gene quantification of each unknown gene tested. Expression levels of GAPDH were also tested along with each test sample, for normalization during relative quantification. Leftover RNA from the microarray analysis (described earlier) was used so that the validity of the microarray data could be directly assessed. Three micrograms of total RNA was used for first-strand synthesis using oligo dT primers according to the manufacturer’s protocol (Amersham Pharmacia, Piscataway, NJ). The equivalent of 20 ng original total RNA was used from each time point to test gene expression, using real-time RT-PCR systems (TaqMan technique, model 5700 sequencer [PerkinElmer]; and iCycler [Bio-Rad, Hercules, CA]). Each gene was tested in two sets of triplicates, using amplicons to create a standard curve. No-template controls and total RNA samples (before first-strand reverse transcription cDNA synthesis) were used as negative controls. 

Gelatin Zymography

Gelatin zymography was performed as previously described. ¹⁵ Briefly, after dissection, retinas were lysed in lysis buffer (300 mM NaCl, 50 mM Tris-HCl [pH 7.4], and 1% Triton X-100). Equivalent 10-μg amounts of retina lysate from each time point were loaded and run on a 10% SDS-polyacrylamide gel with 0.2% gelatin. After electrophoresis, the SDS was removed from the gel by washing with 2.5% Triton X-100 followed by several H₂O washes to remove detergent. The gels were then incubated for 36 hours in buffer (50 mM Tris-HCl [pH 7.4], 200 mM NaCl, and 50 mM CaCl₂) to allow enzymatic digestion of the gelatin, followed by Coomassie blue staining. Cleared bands are formed as a result of MMP-9 enzymatic activity. 

Results

Mouse Retinal Development Model

To generate comprehensive gene expression profiles during postnatal mouse retinal development, eight time points were chosen: P0, P4, P8, P10, P12, P14, and P21, and adult (P42). The early time points cover the final stages of glial and neuronal proliferation and migration, axonal migration, and synaptogenesis. The time points selected also allowed us to analyze genetic expression during all phases of retinal vascularization, because the key events during retinal vascular development occur during this timeframe. At P0 and P4, the superficial vascular network is forming. In P8 mice, the primary vascular plexus begins to mature. Branches from this plexus begin the formation of a deep, secondary vascular plexus, which develops through the second week after birth. In P14 retinas, a final, intermediate vascular plexus begins to form. Finally, analysis of gene expression in P21 retinas provided information during the period of vascular remodeling and final maturation, whereas young adult retinas were representative of a quiescent, mature vasculature. Coincident with final neuronal differentiation and synaptogenesis, as well as retinal vascularization, is the onset of vision, which generally occurs at approximately P14 in BALB/c mice. Thus, this developmental molecular profiling study also facilitated the analysis of genes involved in preparing the retina for vision, as well as genes involved in retinal differentiation itself (Fig. 1) . 

Microarray Data Results

In two separate experiments, 30 retinas from each time point were pooled, and RNA was isolated. Gene expression variability due to individual differences, rather than the involvement of a specific gene in a particular physiological process, was minimized by analyzing pooled samples. Gene expression in these two separate experiments was then analyzed by hybridization to Mu74Av2 gene chips (Affymetrix) during which the integrity of the RNA was analyzed at multiple steps. From this it was determined that data from the two separate experimental pools of RNA were comparable (probability greater than 90% for each time point), and any possible errors during harvesting and preparation of the retinas and retinal RNA had been effectively eliminated by pooling multiple retinas. RNA from the two separate pools was then combined and hybridized to a third chip so that an effective measure of error during hybridization and staining was generated by analysis of each time point in triplicate. The samples were normalized chip to chip, to facilitate comparisons between the different samples. Each gene was then normalized to itself by setting its median expression value to 1 (per gene normalization), allowing differences in expression to be visualized more easily. Thus, a value of 2 indicates a twofold increase in expression over the gene’s median value, whereas a value of 10 indicates a 10-fold increase in expression over the median value. For time points with lower expressions, a value of 0.5 indicates a twofold decrease in expression from the median value, and a 0.1 value indicates a 10-fold decrease in expression from the median value. Thus, a gene with a low normalized expression value of 0.5 and a high normalized expression value of 2 would have a fourfold overall change in expression. 

For stringent expression analysis, a minimum threshold raw expression value was set at 300 (before per gene normalization), a figure representative of expression levels two- to threefold above estimated background levels. Of all sequences, 5429, including 2635 known gene sequences and 2794 ESTs had expression levels above this arbitrary threshold in at least one of the eight time points tested. Thus, despite our stringent criteria required for a gene to be classified as present, approximately 43% of the 12,500 gene sequences tested were considered to be expressed at sometime during the mouse retinal development. All subsequent statistical and clustering analysis was performed on this subset of expressed genes. Approximately 49% of the expressed genes (n = 2649) had consistent expression throughout the developmental time course (Table 1A) . These genes exhibited minimal expression changes that varied less than threefold and included many common housekeeping genes with expression that remained constant even during extreme physiological changes such as those occurring during development. Although these genes are less interesting from the standpoint of gene expression analysis, they are a key validation component of the data. 

Statistical Significance

Several analyses were performed to determine the statistical significance of observed gene expression differences across the data time series. An experimental tree generated by using the Pearson correlation algorithm demonstrated that replicates within a single time point were significantly more similar than experiments from different time points (Fig. 1B) . Coefficient of variance (COV) and correlations between expression values and SD among replicates were also calculated. In each case, the statistical evaluations fell within established limits of acceptable variation. ¹⁴ The median COV for each time point fell within 5% to 8% (Table 1B) . Assuming a full 10% error between each time point, the maximum total error across the entire times series would be approximately 80% (less than a twofold difference). Because at least 75% of the genes within each time point have a COV less than 25%, at least 75% of the genes that exhibit expression variances of more than threefold are significant. It is important to note that this is an extremely modest estimation of significance, because it is highly unlikely for maximum variance to occur in a single direction across all time points. In addition, not all the “expressed genes” are expressed at each time point. Higher COVs for overall “expressed” genes that have low or no expression at certain specific time points are misleading, as demonstrated by correlating variance with gene expression levels. Thus, it is likely that threefold differences are significant for a much greater percentage of genes. Finally, confidence levels quickly become exponentially higher as differences in expression increase beyond threefold. Full graphic data of the statistical analyses can be found in the online supplemental data with the URL provided in Statistical Analysis in the Methods section. 

Validation of the Gene Expression Data

To further validate the microarray data, the expression profiles of a few select genes were validated by real-time RT-PCR and protein expression analysis (Fig. 2) . It is important to note that we have chosen genes representing different expression profiles, as well as genes with expression levels that are at the low end of our positive expression criteria (integrin αv and TIMP-2). These genes generally have lower confidence levels than genes expressed at higher levels. For each gene tested, the expression profiles are very similar between the microarray and RT-PCR data (Fig. 2A) . Where slight differences are noted, changes in expression generally appear less dramatic in the microarray data, demonstrating that false positives are unlikely. The microarray gene expression data for MMP-9 also correlates well with protein expression as analyzed by gelatinase zymography (Fig. 2B) . Because the gene expression profiles obtained by microarray for genes near the lower threshold limits are strongly validated by real-time RT-PCR, we are confident that the profiles generated for most expressed genes, especially those with higher expression and therefore greater innate confidence levels, reflect actual gene expression profiles. 

Validation of the data was also achieved by analyzing the expression profiles of genes with retinal function and expression that have already been well characterized. In nearly all cases, the expression profiles of these genes fit the predicted expression profiles based on previous experimental evidence. Many of these are discussed further later in the article. In the few cases in which gene expression profiles did not match the profiles predicted from previous studies, expression levels generally either fell below our threshold for expression, or remained fairly constant throughout the developmental time course, even though the particular genes’ activity levels are expected to change. The expression of these products may well be regulated at the posttranscriptional level. 

Gene Clustering

Approximately 51% of the expressed genes and 22% of the overall tested gene sequences (n = 2780) had at least a threefold change between the lowest and highest expression levels (Table 1) . To gain insight into the potential significance of the observed changes in gene expression, gene-cluster analysis was performed leading to classification of the expressed genes into 10 different clusters. These final clusters were generated based on expression levels and temporal expression profiles (Fig. 3) . Because of the complex nature of this analysis, the expressed genes were initially incorporated into 25 separate clusters using both k-means and SOM clustering algorithms. Similar clusters were combined and then subclustered into new groups, again using k-means and SOMs. By repeatedly clustering, combining similar clusters, comparing the resultant subclusters, and arbitrarily recombining the most appropriate clusters, six large clusters were eventually generated. However, during the clustering analysis, it became apparent that single clusters were not adequate to represent genes with expression levels that consistently increased or decreased throughout postnatal retinal development. Thus, these two clusters were subclustered into three separate clusters each, generating the final 10 expression profile clusters (Fig. 3) . Subclustering the genes with constant increasing or decreasing expression was important so that significant differences in the intensity of expression change could be easily observed. This also prevented profiles of the relatively few genes with expression that increased or decreased dramatically from becoming statistically buried within clusters where most genes would have similar expression trends, but less dramatic overall changes in expression. 

Cluster 1 contained genes with constant expression levels throughout development. Clusters 2, 3, and 6 contained genes with consistently increasing expression through the time course of study, whereas clusters 4, 7, and 8 contained genes with consistently decreasing expression throughout. Clusters 5 and 9 contained genes with expression levels that were lowest and highest at the mid time points, respectively (Fig. 3) . Most of the genes that were classified within cluster 10 had multiple expression peaks and valleys throughout the developmental time course and, thus, expression profiles were too varied for simple classification. Although the genes within cluster 10 do not follow a single general trend, using a tree algorithm, these genes were organized based on expression profile similarities. This allowed individual expression profiles to be observed, and genes with similar profiles to be found (Fig. 4) . By correlating expression profile clusters and events occurring at various points during postnatal retinal development, we can begin to associate specific genes with different retinal developmental processes. 

Genetic Events That Occur during Later Postnatal Development

Genes that fell within clusters 2, 3, and 6 generally had low expression levels at birth, with expression consistently increasing throughout postnatal development (Fig. 2) . Their expression was generally highest in adult mouse retinas when retinal development is complete. Many genes with known functions involving preparation of the retina for vision, and the subsequent transfer of visual information, were found within these three clusters. 

Many studies have investigated the mechanism of phototransduction (see reviews in Refs. ^{16 17 18} ). Many of the proteins involved in the phototransduction process, as well as the deactivation process, exhibited expression profiles in clusters 2, 3, or 6 (Table 2) . Each was expressed at very low levels at birth, with expression dramatically increasing throughout postnatal development. The identification of novel genes, with similar expression profiles to the genes with known visual functions, may help fill some of the gaps remaining in our understanding of the phototransduction mechanism. For example, HRG4 (UNC-119, GenBank accession number AF030169 [all accession numbers are GenBank; http://www.ncbi.nlm.nih.gov/Genbank; provided in the public domain by the National Center for Biotechnology Information, Bethesda, MD]) exhibits expression profiles and levels comparable to that of opsin and other phototransduction related molecules (Table 2) . Consistent with its expression, localization of HRG4 to photoreceptor synapses has suggested that it may have a function in photoreceptor neurotransmission. ¹⁹ The voltage-gated sodium channel 1B (SCN1B; accession no. L48687) plays an important role regulating neuronal excitability in the central nervous system, and mutations have been shown to cause epilepsy in certain individuals. ^{20 21} Although there is no reported function for this molecule in the retina, based on similarities in expression to genes involved in phototransduction, it seems reasonable to speculate that SCN1B may also play a role in regulating neuronal excitability and ion levels in the retina. 

Although many of the genes that fell into class 3 (clusters 2, 3, and 6) are directly related to the visual process, there are genes with other functions as well. For example, neuroleukin (accession no. M14220) expression increased throughout postnatal retinal development with the highest levels of mRNA expression in the adult retina (Table 2) . Neuroleukin is known to be a neurotrophic factor in other neuronal systems. ^{22 23} With no known function in the visual process, it is likely that neuroleukin may have a neurotrophic effect of helping to protect retinal neurons from degeneration. Other molecules such as peripherin-2 and rod outer segment membrane protein 1 (ROM1) are critical to maintaining photoreceptor outer segment integrity. They associate at the rims of photoreceptor outer segments and are involved in maintaining the structural integrity of photoreceptor outer segments, as well as mediating outer segment disc morphogenesis and regeneration. ^{24 25 26} Although these molecules are not directly involved in phototransduction per se, formation, renewal, and maintenance of photoreceptor outer disc segments is critical for the phototransduction machinery to function correctly. 

Genetic Events That Occur during Early Postnatal Development

Genes categorized into clusters 4, 7, or 8 generally had high expression levels in the mouse retina at P0 and P8 that then decreased consistently throughout the first 3 weeks after birth (Fig. 2) . Many of the genes that fell into these clusters and that have known functions are mediators of cell proliferation or are involved in neuronal differentiation and migration. 

The Ki67 antigen, a marker for proliferating cells, ²⁷ exhibited high expression in the retina through the first week after birth before declining below detection threshold limits at P14. Many cell cycle genes also clustered to expression profiles 7 and 8 (Table 2) . This included cell cycle protein p34 (CDC2; accession no. M38724) and many other cell division cycle (CDC) and minichromosome maintenance (MCM) gene family members required for the correct progression through cell cycle checkpoints. ^{28 29} The mitotic checkpoint kinases Bub1 (accession no. AF002823) and Bub3, ³⁰ as well as c-myb (M12848), a proto-oncogene known to be critical for proliferation and differentiation of hematopoietic progenitor cells, ³¹ were also clustered within profile cluster 8 (Table 2) . Inhibitor of DNA-binding 1 (Id1; accession no. M31885), a helix-loop-helix protein that lacks a normal DNA binding domain but can inhibit other transcription factors through the formation of heterodimers, was also highly expressed during early postnatal retinal development with subsequent expression levels declining dramatically. This protein is required for G1 progression in cultured fibroblasts ³² and is likely to suppress the activity of many other transcriptional regulators involved in preventing cell cycle progression. 

Many genes within clusters 4, 7, or 8 (Fig. 3) have demonstrated properties involved in various aspects of neuronal development in the central nervous system. Based on their classification within our database, many of these genes are likely to be involved in development of the neural retina as well (Table 2) . Six9 (accession no. AJ011787), a homeobox gene involved in neuronal differentiation ³³ was highly expressed at P0 and P4 before its expression dropped sharply at P8. Little or no expression was observed in retinas from mice older than 1 week. Neuronatin (accession no. X83569), a gene proposed to participate in the development and maturation of the brain and central nervous system, ³⁴ and neurogenin (math4a), a gene broadly expressed in neuronal precursor cells of the central and peripheral nervous systems ³⁵ were also expressed during the final stages of retina neuronal development. Math5, essential for ganglion cell development and differentiation, ^{7 36} was found in cluster 8. Math3 and Pax6, basic helix-loop-helix transcription factors that regulate amacrine cell fate specification in the retina and neuronal differentiation in other systems, ³⁷ both exhibited cluster-4 expression profiles (Table 2) . Thus, consistent with the occurrence of the final neuronal differentiation and migratory stages during the first week after birth, many of the genes within clusters 4, 7, or 8 (Fig. 3) that have identified functions participate in neuronal proliferation, migration, or cell fate determination. This classification may be useful for the identification of novel genes that participate in retinal neurogenesis. 

Although most gene expression profiles correlated with known functions, there were some surprises as well. For example, cysteine and glycine rich protein 2 (Crp2; accession no. D88792), also known as double Lim protein-1, was highly expressed early, with expression decreasing fairly consistently throughout development (Table 2) . This gene is expressed during embryonic development of mouse, ³⁸ rat, ³⁹ and chick aortas, ⁴⁰ with expression correlated with vascular smooth muscle cell development and differentiation. However, in the neonatal mouse retina, Crp2 expression was highest before arteries in the superficial vascular layer have begun to develop smooth muscle cell association. Our expression data, however, agree with recent studies suggesting that Crp may participate in the growth and development of other systems as well. ^{41 42} Members of the Crp family, including Crp2, may even be involved in signaling pathways that couple extracellular signals with changes in cytoskeletal architecture. In fact, it has recently been shown that Crp2 may affect cytoskeleton dynamics by phosphorylating cofilin, ⁴³ another gene with expression patterns similar to that of Crp2 in our database. This demonstrates an example of the potential utility of microarray analysis in generating helpful insights regarding potential functions for new genes and in suggesting additional functions for previously described genes. 

Intermediate Genetic Events

Unlike the other gene clusters, the expression levels of genes within clusters 5, 9, and 10 did not follow a single consistent trend throughout postnatal retinal development (Fig. 3) . For genes in cluster 5, expression was lowest during the mid development time points, with higher expression levels during early and late retinal development. These genes are likely to function during neuronal development and then continue to have important functions in the mature retina. The distal-less homeobox 2 protein (DLX2) exhibited decreasing expression levels during early postnatal development until P10, when expression levels began to increase once again (Table 2) . The high expression level of DLX2 at early postnatal time points is not surprising, because DLX2 is known to participate in the formation of the forebrain, as well as craniofacial patterning and morphogenesis. ^{44 45} Thus, it may also play a role in patterning and morphogenesis of the neural retina. However, increasing expression levels after P10, when retina neuronal patterning is set, suggests that DLX2 may also play a role in the maintenance of retinal morphology. 

In cluster 9, genes had the highest expression levels at the middevelopmental time points when retinal vascularization is the main ongoing developmental process (Fig. 3) . Genes within cluster 10 have multiple changes in expression at various time points throughout postnatal development (Figs. 3 4) . Although the known functions for genes within these clusters were more varied, many of the genes within clusters 9 and 10 that have known functions are involved in retinal vascularization. Neuronal differentiation in the retina is in its final stages at birth and therefore the expression of genes solely involved in neuronal differentiation generally decreased consistently during early postnatal development, causing these genes to cluster within gene profile clusters 4, 7, or 8. The process of vision begins after birth and, thus, genes involved in that process generally exhibited expression levels that increased consistently throughout postnatal retinal development. Most of these genes therefore fell within clusters 2, 3, or 6. Of these three major processes, vascularization is the only one that occurs in the mouse retina throughout the first 3 weeks after birth. In addition, the process of vascularization goes through different stages during this time, as the superficial vascular layer develops during the first week and subsequent deep vascular layers develop during the second and third weeks after birth. Therefore, genes with expression levels that variably increase and decrease at specific time points during postnatal development are likely to be involved in retinal vascularization, or at least have multiple overlapping processes. 

Platelet-derived growth factor α-receptor (PDGFαR), angiopoietin-2, and claudin 5 were all classified within cluster 9. PDGF is expressed by retinal astrocytes, ^{11 46} which are known to migrate ahead of the endothelial cells and are important for vascular guidance during developmental retinal vascularization. ¹⁰ PDGFαR is expressed on the endothelial cells and mediates vascular development of the retina. Of note, PDGFα was classified in cluster 1 (Table 2) , with fairly consistent expression throughout retinal development, indicating that PDGFα activity may be regulated by specific expression of the receptor, rather than variation of ligand expression levels. Angiopoietin plays a critical role in stabilizing new vessels during developing and pathologic retinal vascularization. ⁴⁷ Claudin 5 is a member of the claudin family of tight junction membrane receptors. Whereas other members of the family are found mainly in tight junctions of epithelial cells, claudin 5 expression has been found solely in endothelial tight junctions and is thought to be important for formation of the blood–brain barrier. ⁴⁸ Based on our observations, it is possible that claudin 5 is important for the formation of endothelial tight junctions during formation of the blood–retinal barrier as well. 

Vascular endothelial growth factor (VEGF), a major growth factor involved in endothelial cell proliferation and migration during neovascularization, was also found within cluster 10 (Table 2) . Like PDGF, VEGF (accession no. M95200) is also expressed by astrocytes ahead of the vessels during formation of the superficial vascular network and may be involved in recruiting the vessels of the superficial layer to branch and form the subsequent deep vascular layers as well. ¹² Insulin-like growth factor (IGF)-1 is another growth factor that is critical for the normal development of the retinal vasculature in both mouse ⁴⁹ and human retinas. ⁵⁰ EphB4, a molecule critical for the differentiation of vessels into arterial or venous endothelial cell lineages, ⁵¹ exhibited an expression profile within cluster 10. Platelet endothelial cell adhesion molecule (PECAM /CD31; accession no. L06039) and vascular cell adhesion molecule (VCAM-1, M84487), common endothelial cell markers, were also found within cluster 10 or 9, respectively. The expression profile of R-cadherin, a cell adhesion molecule, also fit into cluster 10 (Table 2) . This observation led to further studies regarding the potential role of R-cadherin during retinal neovascularization. ¹⁰ In addition to endothelium-specific genes, numerous hematopoietic cell markers, including EphB6, interleukin-4 receptor (IL-4r), and leukemia inhibitory factor (LIF) receptor, were also expressed within clusters 9 or 10. ^{52 53 54}  

Although genes within cluster 10 do not follow a single consistent trend, the genes within this cluster were organized by profile similarities (Fig. 4) . This allowed us to determine which genes have similar expression profiles and thus may be a useful determinant for genes involved in similar processes, including retinal angiogenesis during development. The identification and localization of each of these genes within the gene tree can be found at http://www.scripps.edu/cb/friedlander/gene_expression/tree.gif. 

Discussion

Historically, studies of retinal development have focused on known, specific factors and their putative roles in various aspects of retinal development. With the development of microarray technologies, there has been a growing interest in large, global-scale analysis of factors that may be involved in various aspects of retinal functional development. In several recent studies, investigators have created ocular-specific gene array chips ^{55 56} and analyzed gene expression in different regions of the eye. ⁵⁷  

The use of global microarray gene expression analysis to study the molecular events occurring during developmental or disease-related processes has not been restricted to study of the eye. Similar microarray chips (Affymetrix) were used to study gene expression during development of the brain hippocampus. ⁵⁸ Microarray analysis was also used to profile the molecular signatures of various carcinomas ⁵⁹ and was critical in finding potential factors involved in hemangioendothelial tumor progression. ⁶⁰ Recently cDNA arrays have also been used to study ocular diseases. Genes involved in retinal inflammation, leukocyte adhesion to the retinal endothelium, metabolic and detoxifying enzymes, and various proliferation and apoptosis genes were identified in the retina during early stages of diabetes. ⁶¹ Changes in gene expression in the aging retina have also been found by using gene arrays. ⁶²  

Functional information can be inferred from gene expression data. However, protein expression and localization, as well as functional data, must be obtained before any conclusions regarding a gene’s function can be made. Many proteins are regulated by translational regulation or posttranslational modifications such as phosphorylation or proteolytic cleavage. The activities of many of these important factors are not regulated at the transcriptional level and thus, microarray analyses may not capture such critical alterations in functional gene expression. As proteomic analysis progresses, high-throughput global analysis of protein expression levels in a complex system such as retinal development may become possible and be a useful correlate to microarray analysis. Meanwhile, microarray global gene expression analysis provides an unbiased (albeit a non–hypothesis-driven) approach to the identification of factors that may have novel functions during specific physiological and pathologic processes, potentially leading to novel mechanistic insights and identification of potential drug targets. 

This is the first report to describe the global gene expression of thousands of genes in the developing retina. These data can serve as a database and reference point to improve the cataloging of potential gene functions, through analysis of expression profile changes. Most studies using microarray analysis of gene expression use a two-point system in which changes in expression between two different states of the same tissue are compared—for example, diseased tissue is compared with healthy tissue, or angiogenic is compared with quiescent vasculature. Although these studies are valuable for finding genes that are potentially involved in a specific process, they provide relatively little evidence or insight for the genes’ specific roles. By investigating multiple time points throughout postnatal retinal development, rather than analyzing the gene expression from a single developmental time point and comparing it to mature retinas, we can begin to associate specific genes with specific developmental events and, it is hoped, establish a functional correlation between the event and the gene expression. In addition, our analysis suggests that there are certain genes that have multiple, overlapping functions during several different processes during retinal development. This is interesting from a developmental standpoint, because many of the observed processes have historically been regarded as distinct. Recent studies, however, suggest that these distinct developmental systems are not isolated, but may share similar factors and mechanisms. In addition, this knowledge may also be relevant to clinical practice in the design of specific drugs that not only affect the targeted condition, but others as well. 

Many ocular disease processes that alter retinal physiology are likely to do so as a result of abnormal gene expression. For example, many of the diseases that can lead to a catastrophic loss of vision involve the abnormal growth of new blood vessels. ⁶³ Pathologic angiogenesis is likely to use basic angiogenic mechanisms similar to those used during normal developmental angiogenesis. Thus, novel factors found to be involved in developmental angiogenesis may have an important role in pathologic angiogenesis as well. By analyzing a broad range molecular signature of different phases during postnatal mouse retinal development, we have identified genes of potential interest during neuronal differentiation, neuronal maintenance, retinal vascularization, and visual cycling. These include genes that have previously been well characterized with regard to retinal function, genes whose functions are characterized in other tissues but have yet to be described in the retina, and genes with functions that have yet to be identified. As genes are identified with potentially important functions for each of these processes, new insight will be gained regarding potential areas to study genes whose altered expression and function can lead to neurodegenerative and/or neovascular diseases associated with loss of vision. 

A few of the genes with interesting expression profiles during retinal development have been discussed briefly, along with a general interpretation of the clustering profile results. However, because of the high number of genes with interesting expression profiles in the retina during this complex process, it is impossible to discuss the relevance and potential function of each one. Thus, we have developed a Web site available at http://www.scripps.edu/cb/friedlander/gene_expression/ where the data for each gene, as well as complete list and data for every gene within each individual cluster (Fig. 3) , can be found. Based on our analysis, this information should give researchers initial clues to the potential functions of certain genes during retinal development, according to their classification. In addition, one can search this database for the expression profile of any specific gene present on the Mu74Av2 gene chips (Affymetrix). Users can then find the 10 genes with the most closely related expression profiles during postnatal mouse retinal development in an effort to identify genes with potentially similar roles during retinal development. Finally, the localization and individual expression profiles of cluster-10 genes within the gene tree (Fig. 4) can be found online in a larger format. It is our hope that other researchers with expertise regarding the various retinal developmental processes or certain ocular diseases will be able to use these data to search for and identify genes that warrant further study. 

 Supported by Grants to MF from the National Eye Institute (R01 EY11254), an administrative supplement (R01EY11254) to support DNA microarray facilities, the Robert Mealey Program for the Study of Macular Degenerations, and an unrestricted grant from Merck. MID was supported by a fellowship from the Achievement Recognition for Collegiate Scientists Foundation (ARCS), San Diego chapter.

 Submitted for publication July 29, 2003; revised October 21, 2003; accepted December 3, 2003.

 Disclosure: M.I. Dorrell, None; E. Aguilar, None; C. Weber, None; M. Friedlander, Merck (F)

 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.

 

Figure 1.

 

Major developmental processes in the postnatal mouse retina. (A) During the first 3 weeks after birth, final neuronal differentiation, developmental vascularization, and vision development occurs. Neuronal differentiation and patterning occurs before birth and continues until approximately P8. Retinal vascularization begins just after birth and continues until the complete development of three vascular layers by P21. Finally, the mouse opens its eyes, and the visual process begins around P14. (B) An experimental tree algorithm generated using a Pearson correlation algorithm demonstrates that replicates within each time point are more closely related than experiments measuring gene expression from different time points.

Table 1.

 

Microarray Analysis Results

A	Cluster	Description of Cluster Profile	No. of Genes	% of Total “Expressed”	% of Total Genes
	1	Flat	2649	49.8	21.2
	2	Increasing (shallow)	288	5.3	2.3
	3	Increasing (medium)	341	6.3	2.7
	4	Decreasing (shallow)	733	13.5	5.9
	5	Lowest in middle	83	1.5	0.7
	6	Increasing (steep)	133	2.5	1.1
	7	Decreasing (medium)	290	5.4	2.3
	8	Decreasing (steep)	133	2.5	1.1
	9	Highest in middle	301	5.5	2.4
	10	Varying expression	418	7.7	3.3

B	Time Point	Median COV (%)
	P0	6.2
	P4	7.9
	P8	6.0
	P10	5.2
	P12	5.9
	P14	5.7
	P21	5.8
	Adult	6.1

 (A) The number of genes within each cluster, along with the percentage of total genes tested and the total “expressed” genes are given. (B) The median COV for each time point is shown. COVs were calculated by dividing the standard deviation by the mean expression value for the triplicate measurements of each gene tested.

Figure 2.

 

Validation of the microarray gene expression profiles. (A) The expression profiles of integrin αv, R-cadherin, opsin, TIMP-2, MMP-9, and ki-67 were validated by comparing the microarray data with gene expression data from real-time RT-PCR. Various profiles were represented and in each case the data from the microarray analysis closely resembles the data generated by real-time RT-PCR. (B) Protein expression data was generated for MMP-9 using gelatin zymography. MMP-9 protein expression also correlates with mRNA expression profiles.

Figure 3.

 

Gene expression profile clustering. The 5429 “expressed” genes were organized into 10 general clusters. Cluster 1 contains the genes with expression that remains constant (less than a threefold difference) throughout postnatal retina development. Clusters 2, 3, and 6 contain genes with expression that consistently increases. Many of these genes were found to be directly vision-related as part of the phototransduction machinery or recycled components of the photoreceptors. Clusters 4, 7, and 8 contain genes with consistently decreasing expression. Many of these genes were found to be involved in neuronal patterning, migration, or differentiation. Genes in cluster 5 have the highest expression in the intermediate time points, and genes in cluster 9 have a decrease in expression during intermediate time points. Cluster 10 contains all the genes with expression that variably decreases and increases at different time points during postnatal retina development. Many of the genes in clusters 9 and 10 were determined to be associated with retinal vascularization. For clusters 1 to 9, solid lines represent the cluster’s mean normalized expression value, whereas the two dashed lines represent the general ranges in expression levels for individual cluster members. In cluster 10, a single trend does not exist, and thus the profile of each individual gene is shown. The percentage of genes represented by each cluster is indicated at the bottom of each profile.

Figure 4.

 

Expression profile tree for cluster 10. Genes within cluster 10 were analyzed according to differences in expression profiles and expression levels. Genes with similar expression profiles are proxi- mate, whereas distant genes’ expressions are more varied. A color bar is included to indicate the different relative levels of expression. Genes with similar expressions may be involved in similar postnatal development processes. A higher-magnification image, gene names, and individual data for each gene within the tree can be found online at the URL provided in Statistic Analysis in the Methods section.

Table 2.

 

Microarray Expression Data of Discussed Genes

Common Name	Cluster	P0	P4	P8	P10	P12	P14	P21	adult	GenBank
Ki-67	8	57.10	23.40	5.20	1.50	0.30	0.03	0.10	0.06	X82786
Cdc2a (p34)	8	49.60	30.60	5.80	1.50	0.17	0.16	0.08	0.09	M38724
Cdc25c	8	22.30	13.30	4.60	0.20	0.18	1.90	0.12	0.16	U15562
Cdc20 (EST)	8	15.40	6.50	2.20	1.10	0.83	0.81	0.68	0.10	AW061324
Cdc6	7	14.60	3.70	1.70	1.00	1.20	1.40	0.95	1.60	AJ223087
Cdc45-like (EST)	8	7.60	3.00	2.80	0.85	0.80	0.18	0.23	0.20	AF098068
Mcm4 (cdc21)	7	9.70	2.50	1.80	1.70	1.10	1.20	1.70	0.20	D26089
Mcm5 (cdc46)	8	6.80	4.70	1.30	0.70	0.84	0.68	0.65	0.11	D26090
Mcm7	7	6.70	2.80	2.00	1.40	0.67	0.62	0.54	0.45	D26091
Bub1	8	45.20	25.30	6.00	1.10	0.47	0.14	0.07	0.09	AF002823
Bub3	4	3.90	2.60	2.00	2.20	1.80	1.90	1.60	1.30	U67327
c-Myb	8	17.40	4.20	2.60	1.70	0.44	0.35	0.11	0.04	M12848
Id1	8	10.40	5.00	2.90	3.00	2.50	2.30	0.85	0.15	M31885
Six9	8	8.70	7.60	1.10	0.22	0.23	0.14	0.20	0.23	AJ011787
Neuronatin	7	3.60	2.60	3.00	2.20	0.95	0.67	0.21	0.08	X83569
Neurogenin (math4a)	7	4.10	3.00	2.40	1.10	0.63	0.55	0.40	0.43	Y07621
Math5	8	105.00	17.90	0.21	0.19	0.21	0.22	0.25	0.36	AF071223
Math3	4	1.70	3.30	2.70	2.10	1.10	1.30	1.10	0.87	AF036257
Pax-6	4	4.20	3.10	1.80	1.90	1.50	1.60	1.60	1.70	X63963
Crp2, SmLim	8	6.60	4.90	2.10	0.80	0.40	0.44	0.35	0.05	D88792
Cofilin	4	2.50	2.20	2.10	2.10	1.50	2.10	1.40	0.66	L29468
Opsin	6	0.01	0.06	0.43	0.94	2.85	3.01	3.40	3.02	M36699
Gnat-1, transducin	6	0.01	0.01	0.06	0.14	1.33	3.41	4.47	3.33	M25513
Gnat-2, transducin	2	0.45	0.59	0.89	0.89	1.26	2.10	3.24	2.21	L10666
Pde-6α (cGMP-specific)	6	0.08	0.07	0.25	0.44	1.33	2.21	4.31	3.95	X60664
Pde-6β (cGMP-specific)	6	0.01	0.29	1.16	1.46	2.11	2.42	2.71	2.83	X55968
Pde6b-like (EST)	3	0.07	0.38	0.65	0.73	1.33	1.24	1.41	1.35	AV361022
Pde-6γ (cGMP-specific)	6	0.04	0.21	0.99	1.30	2.07	2.39	2.54	2.66	Y00746
Phosducin	3	0.25	0.92	2.03	2.09	2.23	2.49	2.26	1.86	L08075
CNCG	6	0.01	0.03	0.67	1.27	2.64	4.14	4.82	2.43	M84742
Arrestin	6	0.06	0.49	1.35	1.60	2.84	2.52	2.46	2.70	M24086
Recoverin	6	0.02	0.11	0.63	1.10	2.72	2.66	2.98	2.58	X66196
HRG4 (Unc-119)	3	0.23	0.66	1.26	1.51	2.20	2.21	2.41	2.44	AF030169
SCN1B	6	0.08	0.10	0.66	0.93	2.12	1.98	2.49	3.45	L48687
Neuroleukin	3	0.97	1.79	1.59	1.78	2.38	2.29	3.43	5.27	M14220
Peripherin 2	6	0.04	0.84	1.88	1.97	2.82	2.94	4.02	4.55	X14770
Rom 1	6	0.07	0.46	1.29	1.64	2.21	2.37	2.60	2.60	M96760
DLX2	5	6.02	1.69	1.12	1.07	0.91	0.42	1.76	1.81	M80540
Angiopoietin 2	9	1.43	1.80	2.61	2.92	2.31	1.55	1.07	1.76	AF004326
PDGF-alpha-R	9	0.16	2.38	2.62	2.54	2.07	1.11	0.65	0.13	M57683
PDGF-alpha	1	1.37	1.36	1.62	1.81	2.00	1.74	1.77	2.24	M29464
Claudin 5	9	1.69	1.10	1.40	1.67	3.30	2.73	1.27	0.90	U82758
VEGF	10	1.87	1.29	2.54	1.79	1.32	2.24	3.10	1.54	M95200
IGF-1	10	2.37	2.37	1.13	1.30	2.27	0.13	0.19	1.34	X04480
EphB4	10	2.21	1.42	2.08	1.56	1.20	1.57	1.58	0.89	U06834
Pecam	10	1.64	1.13	1.84	0.84	1.55	2.69	1.81	0.20	L06039
VCAM-1	9	0.07	0.11	1.33	4.05	6.18	5.42	2.34	2.65	M84487
R-Cadherin	10	2.11	2.95	0.22	2.02	2.33	1.55	0.17	1.32	X69966
EphB6	10	2.20	2.45	1.93	1.98	2.15	0.20	1.33	2.02	L77867
IL-4 receptor	10	1.52	2.38	0.29	2.78	2.48	0.17	0.12	0.17	M27960
LIF receptor	10	0.87	1.65	2.53	2.12	1.68	1.91	2.14	1.06	D17444

 The normalized gene expression data is given for discussed genes. The data are presented after chip-to-chip and per-gene normalization in which the median expression value for each gene was set to 1, and all other values are presented relative to the median.

The authors thank Marilyn Leonard for assistance in dissection of the mouse retinas and Matthew Ritter, Steve Head, and Tom Whisenant for helpful bioinformatics discussions.