Prognosis-Determinant Candidate Genes Identified by Whole Genome Scanning in Eyes with Pterygia

Abstract

purpose. To identify the genes that can differentiate primary from recurrent pterygia.

methods. The transcriptional differences of primary and recurrent pterygia were first determined by microarray analyses. Computational analyses were used to extract the biological significance of the genes accurately, and a significant functional classification of the genes was made by unsupervised methodologies. After confirming the functional classification for primary and recurrent pterygia by a clustering algorithm, a support vector machine (SVM) algorithm was applied. Based on a machine learning technique, the minimum number of genes that can accurately classify primary and recurrent pterygia was determined.

results. Clustering analyses classified primary and recurrent pterygia transcriptomes and identified 10 clusters associated with distinct biological processes. When the SVM algorithm was applied to the microarray-analyzed products from three primary and three recurrent cases, periostin, TIMP-2, and l-3-phosphoserine phosphatase homolog (PSPHL) were identified as the minimum set of predictors with 100% accuracy. A differential expression of these genes in primary and recurrent pterygia was confirmed by immunohistochemistry. When the 24 patients with primary disease and the 8 patients with recurrent disease were analyzed with this gene set, an accuracy of classification of 84.38% was achieved.

conclusions. Periostin, TIMP-2, and PSPHL can be used as predictor genes for the recurrence of pterygia. Their biological activities may explain the events leading to recurrences of pterygia and thus may be genes to target for pharmaceutical interventions.

Apterygium is a common ocular surface disease that is characterized by an invasion of conjunctival tissue onto the clear cornea with elastosis and breakdown of Bowman’s layer. Recent findings suggest that during the development of pterygia, there is a remodeling of the extracellular matrix, alterations of inflammatory cytokines or growth factors, and involvement of antiapoptotic mechanisms. ¹ Unfortunately, these findings are not sufficient to develop efficacious pharmacologic treatments of pterygia. 

To determine the cause of recurrences more precisely, it is necessary to undertake a detailed analyses of the etiology of pterygia at the molecular level. Our initial analysis of the pterygia transcriptome ² and the observations of others ^{3 4} confirmed that the mechanisms associated with the development of pterygia are multifaceted, and thousands of transcripts are differentially expressed in pterygia and normal conjunctiva. 

Histopathologically, the tissues of recurrent and primary pterygia are very similar although quite different from normal conjunctival tissues. Eyes with recurrent pterygia are known to have a higher rate of recurrence after surgery with increased cellular proliferative responses in the subepithelial fibrovascular layers. ⁵ This indicates that the mechanisms determining the genesis and the recurrence of a pterygium may not be the same. Because this also suggests transcriptional alterations in the tissues of recurrent pterygia, it seemed reasonable to assume that the tissues of recurrent pterygia possess molecular signatures relating to their recurrence. 

To identify the recurrence-related genes, we used a differential whole genome scanning approach with a computational analysis of gene expression to isolate a functionally significant classification of genes. A clustering algorithm identified functional gene clusters and accurately classified tissues obtained from primary and recurrent cases. With a support vector machine (SVM) algorithm, ⁶ we extracted the minimum gene set that would accurately classify primary and recurrent pterygia. We showed that only three genes are necessary to differentiate primary and recurrent pterygia: periostin, tissue inhibitor of metalloproteinase-2 (TIMP-2), and l-3-phosphoserine phosphatase homolog (PSPHL). 

Materials and Methods

Patients’ Background and Specimen Collection

Pterygia heads were collected from 32 eyes of 32 patients undergoing pterygium excision with conjunctival autografting at the Tottori University Hospital (Yonago, Japan). Of these, 24 eyes of 24 patients had primary pterygia and 8 eyes of 8 had recurrent pterygia. There were 11 men and 21 women with a mean age of 65.4 ± 2.0 years (range, 31–85), and all were Asians who were residing in Tottori or Shimane Prefecture. Patients with primary cicatricial ocular surface disease and inflammatory ocular surface disease of unknown etiology were excluded. 

Reference RNA was obtained from pooled RNA extracted from 12 normal nasal conjunctival specimens of volunteers undergoing cataract surgery. The study protocol conformed to the tenets of the Declaration of Helsinki, and the procedures used were approved by the Tottori University Ethics Committee. A signed informed consent was obtained from all patients. 

Microarray Procedures

To minimize the influence of confounding factors (e.g., age, gender, hormonal effects, race, UV exposure time) and the effects of geographical location, we initially studied three primary and three recurrent pterygia from postmenopausal women. Their mean age of the patients with primary pterygia was 66.3 ± 3.4 years and of those with recurrent pterygia was 74.0 ± 2.1 years. 

Pterygia head tissues (∼3 × 3 mm) were collected during surgery by two authors (DM and YI) who used, as exactly as possible, the same surgical procedures. The tissues were immediately transferred into storage medium (RNAlater; Ambion, Austin, TX). Total RNA was isolated from the tissue samples (RNA STAT-60; Tel-Test, Inc., Friendswood, TX) and purified (RNeasy Mini Kit; Qiagen, Hilden, Germany), according to the manufacturers’ protocols. Total RNA was reverse-transcribed and amplified (Amino Allyl MessageAmp aRNA kit; Ambion) for Cy5 and Cy3 dye labeling. The Cy5 dye was used to generate the experimental cRNA probe from the pterygia tissues, and the Cy3 dye was used to generate the reference cRNA probe from normal conjunctiva. The labeled cRNA probes were hybridized on oligo microarrays (AceGene; Hitachi, Tokyo, Japan) corresponding to 30,336 genes, and scanned (FLA-8000 scanner; Fuji Film, Tokyo, Japan). The intensities of the fluorescent signals were quantified with a computer program (DNASIS Array program; Hitachi). After background subtraction, the signal from the gene spots was adjusted to compensate for excitation differences between the two dyes. Then, the fluorescent signal was corrected for image intensity, background–spatial artifacts, and chip-to-chip comparisons, using a custom database constructed by the gene array program (DNASIS Array). 

Cluster Analysis

Before clustering and display, the logarithm of the ratio of the measured fluorescence for each gene was centered by subtracting the arithmetic mean of all ratios measured for that gene (DNASIS STAT program; Hitachi). The centering made all subsequent analyses independent of the amount of each gene’s mRNA in the reference pool. We extracted the genes by normalized fluorescence ratios for the 30,336 transcriptions on the arrays by applying different selection criteria for subset selection. 

We applied a hierarchical clustering algorithm on the genes using the matching coefficient of Weinstein et al. ⁷ and Eisen et al. ⁸ as a measure closely linked to the average clustering. For visual display of the rows and columns in the initial data, tables were reordered to conform to the structures of the dendrogram. The data in the table were represented graphically by coloring each gene by the measured fluorescence ratio. Each gene in the cluster-ordered data table was replaced by a graded color, pure red through black to pure green, representing the mean-adjusted ratio for each gene. 

Construction and Evaluation of Support Vector Machine-Based Classifier

We attempted to extract the essential genes by transforming this task into one that differentiated the two groups using a minimum number of genes (i.e., a minimum number to construct a binary classifier). For this purpose, a machine learning algorithm is a well-known approach used for its efficiency, similar to that of a neural network. Of these, we applied the SVM algorithm developed by Vapnik. ⁹  

The SVM algorithm design facilitates the generation of efficient classifiers especially for untrained data points. Another advantage is that the SVM relies only on machine-selected data points as support vectors, thus minimizing the dimensional complexity of the classifier. Using the DNASIS STAT program, classifiers were generated with defined sets of genes. To evaluate their performance, we used the leave-one-out cross-validation method. In this procedure, classifiers for n data points were generated by supervised training with n− 1 data points. Next, data not used for training, were tested for correctness of the answer by the generated classifier. This procedure was repeated n times to test all untrained data points. The mean percentage of questions answered correctly was defined as the classification accuracy. For comparison with other classification algorithms, including the multilayer perception (MLP) method and the k-nearest-neighbor (KNN) method, the leave-one-out cross-validation experiment was performed to evaluate the effectiveness of the classification. 

Extraction of Classifier Genes for Recurrence by SVM

To generate the initial classifiers, genes were selected by application of the Mann-Whitney test. When the classifier was 100% accurate in the classification, gene sets that were used to generate the classifier served as the starting template. We then randomly discarded a gene from the classifier to remove any redundant genes from the template, and then the accuracy of classification was tested by the leave-one-out cross-validation method. The discarding procedure was repeated as long as the remaining gene sets maintained 100% accuracy of classification. If the discarding of a gene resulted in a loss of accuracy, the procedure was canceled and returned to the previous step of random discarding. The resultant minimum gene set was defined as classifier genes of pterygia recurrence. 

Real-Time RT-PCR

One microgram of 32 pterygia head RNA samples and 11 normal nasal conjunctiva tissues were reversed transcribed using random hexamers (Superscript III; Invitrogen, Carlsbad, CA). The c-DNAs were amplified and quantified by thermocycler (LightCycler; Roche, Mannheim, Germany, with a QuantiTect SYBR Green PCR kit; Roche). The sequences of the used real-time PCR primer pairs are listed in Table 1 . Primers were designed using Primer 3 (http://fokker.wi.mit.edu/primer3/ developed by Steve Rozen and Helen Skaletsky, Whitehead Institute Center for Genome Research, Cambridge, MA) setting the importance on primer dimers, self-priming formation, and mispriming. All primer pairs that have an equal optimal annealing temperature of 58°C and similar GC content were selected and generated by Sigma-Aldrich Corp. (St. Louis, MO). The specificity of the primers was verified by agarose gel electrophoresis and melting curve analysis. To ensure equal loading and amplification, we normalized all products to GAPDH transcript as an internal control. Actual copy numbers of the products were calculated based on cloned templates of respective genes using the second derivative maximum method applied to threshold cycles of fluorescence detection. 

Immunohistochemistry

A universal immunoenzyme polymer method was used for immunostaining of periostin, TIMP-2, and PSPHL. To maintain antigenicity and integrity of soft conjunctival tissues, fresh-frozen sections (5 μm) were prepared according to reported procedures. ^{10 11} The sections were treated with 3% H₂O₂ for 15 minutes to quench endogenous peroxidase activity, and nonspecific binding was blocked by normal serum for 30 minutes. After they were washed, the sections were incubated with primary antibodies overnight in a moist chamber at 4°C. The primary antibodies were; rabbit anti-human periostin (1:1000; Biovendor, Brno, Czech Republic), mouse anti-human TIMP-2 (1:200; clone 50-3D2; Fuji Chemical Industries, Takaoka, Japan), and goat anti-human PSPHL (1:200; Imgenex, San Diego, CA). Washed slides were then incubated with secondary antibodies conjugated with peroxidase (N-Histofine Simple Stain MAX PO; Nichirei Co., Tokyo, Japan) for 30 minutes at room temperature. Signals were made visible with peroxidase substrate 3-amino-9-ethylcarbazole (AEC), and counterstained with hematoxylin. A negative control slide was included in each immunostaining. 

Statistical Analyses

Data are presented as the mean ± SEM. Statistical analysis was performed by unpaired Student’s t-test (two tailed) or the Mann-Whitney test, as appropriate. P < 0.05 was considered significant. 

Results

Microarray Analyses

When a genome-wide scanning by microarray was applied to age- and gender-matched primary and recurrent tissues, 26,657 genes were differentially detected in the two groups. To extract candidate genes sets, we first applied the Mann-Whitney test and detected 1724 genes that were significantly different in the two groups (P < 0.05); 1621 genes were upregulated and 103 were downregulated in the recurrent cases. 

Cluster Analyses

When these genes were used for cluster analyses, the resultant dendrogram successfully generated two clusters for tissues from primary and recurrent pterygia (Fig. 1) . This finding indicated that the transcriptional information in the extracted genes set was accurate and sufficient to provide a basis for the classification. 

To reduce the complexity of the model and the number of required genes sets for pattern recognition, we set a threshold of a twofold change. This returned 184 genes, which again accurately differentiated the two groups of tissues. These 184 genes in the primary and recurrent tissues were summarized into 10 categories that play distinctive roles in the pathophysiology of recurrent pterygia (Table 2) . The 10 clusters were genes associated with tissue development, cellular function maintenance, cell death, cell signaling, cell-to-cell signaling and interaction, cell cycle, molecular transport, tissue development, hematologic system development, and cell cycle/cancer. 

To reduce the complexity of the model further, we set more rigorous thresholds for gene selection. However, this strategy did not result in a more simplified model without reducing the accuracy of classification. 

Supervised Classification

We initially hypothesized that complex biological processes associated with recurrences might well be simplified to representative genes that might trigger or regulate recurrences. These genes may still be traceable after surgical intervention. If this interpretation is valid, an accurate classification may be achieved when a suitable algorithm is applied, even though a minimum number of genes is used as input. 

To apply a classification algorithm, we initially used the 184 genes of a twofold change. The performance of the classifier was tested by using leave-one-out cross-validation tests. The SVM generated classifiers with 100% accuracy. In contrast, MLP generated classifiers with 83.33% accuracy while KNN generated classifiers with 66.67% accuracy. Because the SVM classifiers were totally accurate, it was used to extract the recurrence-associated classification genes. 

We extracted three classifier genes—periostin, TIMP-2, and PSPHL—that showed 100% accuracy of classification. Of interest, no other minimum sets of genes were found to have 100% accuracy. Therefore, we reasoned that these three genes are the minimally required parameters for detecting recurrences in the initial small sample groups. When these genes were tested to generate classifiers using MLP and KNN methods, accuracy of the resultant classifiers was 100% and 83.33%, respectively (Table 3) . 

Verification of Gene-Expression Data

Next, we evaluated the expression levels of these three genes in both primary and recurrent cases. The expression levels were normalized by cloned templates of a defined copy number, and the actual copy numbers were calculated. First, we examined the copy numbers of periostin, TIMP-2, and PSPHL in the 14 primary and four recurrent cases in postmenopausal women. This population included the original microarray samples and equivalent cases. Consistent with the microarray data, all periostin, TIMP-2, and PSPHL transcripts showed significant changes between primary and recurrence cases. Periostin and TIMP-2 had significantly higher copy numbers in the recurrent group, whereas the PSPHL copy numbers were significantly lower in the recurrent group (Fig. 2a) . Of the three genes, nine times more transcripts of periostin were detected in recurrent than in primary cases (primary: 2.2 × 10⁹ ± 3.7 × 10⁸ copies/μg RNA; recurrent: 2.3 × 10¹⁰ ± 1.5 × 10¹⁰ copies/μg RNA; P < 0.05). Six times more transcripts of TIMP-2 were detected in recurrent than in primary cases (primary: 3.2 × 10⁸ ± 4.6 × 10⁷ copies/μg RNA; recurrent: 2.1 × 10⁹ ± 1.5 × 10⁹ copies/μg RNA; P < 0.05). For PSPHL, the recurrent cases had significantly fewer transcripts of PSPHL than did primary cases (primary: 6.5 × 10¹⁶ ± 1.5 × 10¹⁶ copies/μg RNA; recurrent: 7.0 × 10¹⁴ ± 6.3 × 10¹⁴ copies/μg RNA; P < 0.05). 

When the copy numbers determined by real-time PCR were used to generate classifiers of recurrence by SVM, the resultant classifier again showed 100% accuracy by the leave-one-out cross-validation method. 

To determine whether all the 32 pterygia RNAs were correctly classified by these genes, they were analyzed by real-time PCR and compared with control conjunctiva (Fig. 2b) . Of the three genes, five times more transcripts of periostin were detected in recurrent than in primary cases (primary: 3.5 × 10⁹ ± 8.4 × 10⁸ copies/μg RNA; recurrent: 2.1 × 10¹⁰ ± 1.0 × 10¹⁰ copies/μg RNA, P < 0.05; control: 5.2 × 10⁵ ± 3.3 × 10⁵ copies/μg RNA). Recurrent cases also had more TIMP-2 transcripts, but the difference was not statistically significant (primary: 1.1 × 10⁹ ± 5.6 × 10⁸ copies/μg RNA; recurrent: 1.5 × 10⁹ ± 9.2 × 10⁸ copies/μg RNA, P = 0.746; control: 1.9 × 10⁷ ± 7.4 × 10 ⁶ copies/μg RNA). For PSPHL, the recurrent cases had less transcripts, but the difference was not statistically significant (primary: 6.4 × 10¹⁶ ± 1.0 × 10¹⁶ copies/μg RNA; recurrent: 4.2 × 10¹⁶ ± 2.0 × 10¹⁶ copies/μg RNA; P = 0.321, control: 1.5 × 10¹⁰ ± 7.8 × 10⁹ copies/μg RNA). 

To validate that these transcriptional alterations are actually translated, immunohistochemical analysis was performed. Immunoreactivity to the three proteins was detected in all pterygia tissue. Consistent with the transcriptional analysis, periostin and TIMP-2 were more prominently expressed in the recurrent tissue, whereas the expression of PSPHL was reduced (Fig. 3) . Their distinctive localization suggested an interesting perspective on their involvement in the disease process. First, periostin was localized to the basement membrane of pterygia epithelial cells and diffusedly distributed in the superficial stroma (Figs. 3a 3b) . In recurrent pterygia tissues, periostin staining was more intense and diffuse but centered on the basement membrane. Together with the most remarkable change in periostin expression, this strongly suggests that recurrence involves alterations of pathologic processes, especially in the basement membrane of the diseased epithelia. 

In contrast, TIMP-2 was specifically detected in extracellular areas of the diseased epithelia (Figs. 3c 3d) . Recurrence tissues showed diffuse staining in all epithelial cells. The expression of TIMP-2 in primary tissues was confined to the basal epithelial cells near to the epithelial basement membrane, some of which were also present in the connective tissue matrix near the basement membrane. PSPHL was predominantly expressed in the most superficial layers of the epithelia and was reduced in the recurrence tissue (Figs. 3e 3f) . 

Evaluation of SVM-Based Classifier

To test whether the three genes provide sufficient dimensionality, the copy number–based SVM classifiers generated by the three primary and three recurrent cases were evaluated by the accuracy for all the pterygia samples (n = 32; Table 3 ). The performance of the classifier was tested by the leave-one-out cross-validation procedure. These classifiers predicted with 84.38% accuracy in differentiating the two groups. When one of the three genes was omitted in training the SVM-based classifiers, the classification accuracy decreased to 50%, 83.33%, and 83.33% without periostin, TIMP-2, and PSPHL, respectively. The high accuracy of the three genes further indicates that they are candidate genes in recurrent pterygia. 

To evaluate the classification efficacy of SVM-based classifier, the MLP and KNN methods were also used to generate classifier using the three genes copy numbers for all the samples. The performance of the classifier was again tested by the leave-one-out cross-validated procedure. Under these conditions, the MLP-based classifier predicted a recurrence with 83.33% accuracy. The KNN-based classifier predicted with 16.67% accuracy, again showing that the SVM based-classifier performed with the highest accuracy. 

Discussion

By application of the SVM algorithm, pattern recognition of transcriptomes in primary and recurrent pterygia tissue was achieved surprisingly by only three genes: (periostin, TIMP-2, and PSPHL). 

A pterygium is known to have altered growth characteristics that may indicate an involvement of neoplastic pathways in the pathophysiology of genesis and recurrence. For example, pterygia fibroblasts display characteristics of transformed cells including the loss of heterozygosity and microsatellite instability. ^{12 13} In addition, the pterygial epithelium displays reduced apoptosis and in recurrent pterygia, an increased proliferation is noted in the fibrovascular layer. ⁵  

Previous observations suggest that the development and recurrence of pterygia have certain properties of carcinogenesis. How do the identified prognostic factors, periostin, TIMP-2, PSPHL, influence its recurrence after surgical removal? We identified periostin as a statistically significant determinant of recurrence. It is a 90-kDa heparin-binding N-glycosylated protein, originally isolated as an osteoblast-specific factor, that functions as a cell-adhesion molecule for preosteoblasts. ^{14 15 16} Periostin contains fasciclin domains, interacts with αvβ3 and αvβ5 integrins, and increases fibronectin-dependent motility of epithelial cells. ¹⁷ Its interaction with αvβ5 at sites of focal adhesion also suggests its contribution to cell adhesion and motility. 

Of interest, periostin has a high amino acid homology with the TGF-β-induced protein β-igh3, which promotes the adhesion and spread of fibroblasts and is a gene mutated in granular and lattice dystrophy. Recently, periostin has also been shown to be overexpressed in colon cancer, and it also promotes metastasis and angiogenesis. At the signaling level, periostin interacts with αvβ3 to activate the Akt kinase/protein kinase B (Akt/PKB) pathway, leading to increased cell survival. ¹⁸ Thus, the presumed neoplastic nature of pterygia recurrence may well be explained by periostin. 

Collagen types I, II, III, and IV constitute a large part of the extracellular matrix of pterygial tissue where extensive collagen remodeling is observed. ^{1 19} In the degradation and remodeling of extracellular matrix, the MMPs, a family of zinc-dependent endopeptidases, are known to be critical components. The MMPs also participate in tumor cell invasion, metastasis, and angiogenesis. ²⁰ MMPs have also gained significant interest for their involvement in pterygia. Earlier, an abundant expression of MMPs was found in pterygia tissue, whereas little or no expression was observed in normal conjunctiva. ¹ Of the MMPs, MMP-1, -2, -3, -7, and -9 are known to be involved. Di Girolamo et al. ²¹ suggested that MMP-1 may be one of the principle MMPs in the UVB-related pathogenesis of pterygia, because MMP-1 and not MMP-2, MMP-9, or the TIMPs, was strongly and dose dependently induced in cultured pterygia epithelial cells by UVB. 

The activities of MMPs are physiologically regulated by alterations of gene expression, activation of latent zymogens, tissue localization, and inhibition by tissue inhibitors of metalloproteinase (TIMPs). Many studies have shown that the TIMPs have a central role in pterygia. ^{22 23 24 25} Based on the physiological roles of TIMPs, synthetic inhibitors of MMPs have been developed for potential therapeutic use. Our results identified TIMP-2 as one of prognostic determinant of pterygia recurrence. TIMP-2 complexes with MMPs, including MMP-1, -2, -3, -7, -8, -9, -10, -13, -14, and -15, contribute to the control of MMP activity. ²⁶ In addition to MMP inhibition, TIMP-2 also acts as an activator of MMP-2 via the ternary MT1-MMP/MMP-2/TIMP-2 complex. ²⁷ Independent of the well-known MMP-dependent activities, TIMP-2 has been shown to abrogate angiogenic factor-induced endothelial cell proliferation in vitro and angiogenesis by its α3β1 integrin-mediated binding. ²⁸ This binding has also been shown to result in a negative regulation of growth factor activation that required the activity of the protein tyrosine phosphatase (PTP) SHP-1. These observations suggest a suppressive role of TIMP-2 in pterygia recurrence. Our observations of relatively higher copy numbers of TIMP-2 in recurrent cases might be explained by its presumable counterbalancing role in recurrences. 

TIMP-2 is also known to be pluripotential. It stimulates quiescent cells to proliferate and functions as a metanephric mesenchymal growth factor. ^{29 30} These properties may further complicate its role as a recurrence-suppressing factor. 

PSPH (PSPHL) is the other classification–determinant factor. It is the rate limiting enzyme for the synthesis of serine by hydrolysis of phosphoserine, ³¹ and is a known marker of neoplasticity in the lung and colon cells. ³² Considering the possible neoplastic aspects of pterygia, the reduction of PSPHL in recurrent pterygia is not easily explained. Our detection of PSPHL may reflect the nonregulated nature of pterygia tissue outgrowth, possibly originating from limbal stem cells of the primary pterygia. However, the present data do not answer the question of whether PSPHL is actually recurrence promoting or suppressing. 

Our immunohistochemical data showed another perspective of PSPHL. The location of periostin and TIMP-2 expression strongly suggests that the recurrence process would significantly affect the basement membrane of the epithelial cells (Fig. 3) . In the context of basement membrane involvement, differentiation of adenocarcinoma cells is promoted by contact with the basement membrane. Of interest, PSPH is involved in this process, ³³ which may suggest that the impaired induction of PSPHL in recurrent pterygia is secondary to drastic alterations of the basement membrane structure. 

Our finding of differential expression patterns in primary and recurrent pterygium tissues by immunohistochemistry further confirmed the protein level and supports the importance of genes filtered by transcriptional differences. Our findings of periostin by immunochemistry in recurrent pterygia may well explain the neoplastic nature of pterygia’s recurrence and may result in increased proliferation. Based on the location of the classifier genes, we propose that interaction of pterygia epithelial cells and basement membrane is critical in the process of pterygia recurrence. 

In conclusion, of the large number of genes that are differentially expressed in primary and recurrent pterygia tissues, we extracted periostin, TIMP-2, and PSPHL as classification determinants. Understanding the mechanisms of action of these factors may help develop methods leading to less recurrence of pterygia. 

 Submitted for publication September 25, 2006; revised December 26, 2006 and February 8, 2007; accepted May 21, 2007.

 Disclosure: C.-H. Kuo, None; D. Miyazaki, None; N. Nawata, None; T. Tominaga, None; A. Yamasaki, None; Y. Sasaki, None; Y. Inoue, None

 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.

 

Oligonucleotide Primer Sequences

	PCR Primers	PCR Fragment (bp)
Periostin		199
Forward	5′-TTGAGACGCTGGAAGGAAAT-3′
Reverse	5′-AGATCCGTGAAGGTGGTTTG-3′
Tissue inhibitor of metalloproteinases-2 (TIMP-2)		183
Forward	5′-AAAGCGGTCAGTGAGAAGGA-3′
Reverse	5′-CTTCTTTCCTCCAACGTCCA-3′
l-3-Phosphoserine phosphatase homologue (PSPHL)		196
Forward	5′-TCCAAGGATGATCTCCCACT-3′
Reverse	5′-AGGCAGGAGAATTGCTTGAA-3′
Glyceraldehyde-3-phosphate dehydrogenase (GAPDH)		66
Forward	5′-AGCCACATCGCTCAGACAC-3′
Reverse	5′-GCCCAATACGACCAAATCC-3′

Figure 1.

 

Unsupervised classification of primary and recurrent pterygia transcripts by clustering analysis. Differentially expressed genes were analyzed for clustering by the Mann-Whitney test (P < 0.05, n= 184). Chip direction analysis showed that the data sets accurately classified them into primary and recurrent groups. Gene direction analysis detected 10 clusters of genes. The gene expression levels are represented by a red-green color scheme, with red to higher and green to lower than median expression levels.

Table 2.

 

Functional Categories of the 184 Differentially Expressed Genes Obtained by Cluster Analysis

	Molecular Function	Genbank ID	Gene Definition
Cluster 1	Tissue development	NM_014058	DESC1 protein
		XM_029802	Testis-development related nyd-sp27
		ENSG00000108081	AGhsC071408
		ENSG00000119285	Protein BAP28
		NM_019081	Limkain b1
		NM_020327	Activin A receptor, type 1B AGhsC220204
		NM_003201	Transcription factor A, mitochondrial AGhsC171107
		XM_037759	KIAA0376 protein
		NM_004857	A-kinase anchor protein 5
		AL136125	Chromosome 6 open reading frame 33
		NM_004953	Eukaryotic translation initiation factor 4 gamma, 1
		NM_000561	Glutathione S-transferase M1
		U49349	Phosphoinositide-3-kinase, regulatory subunit, polypeptide 1(PIK3R1)
		NM_013326	Ensembl genscan prediction AceGene oligo matches these RefSeq numbers
		ENSG00000109967	AGhsB080604
		AF247168	npd014 AGhsC141219
		ENSG00000099581	AGhsC151606
Cluster 2	Cellular function and maintenance/cell death	Y12041	Immunoglobulin heavy chain; igz30vh3
			ENSG00000115700	Septin 10
			AF349679	B-factor, properdin
		AF116637	Calcium/calmodulin-dependent protein kinase II
		NM_007076	Huntingtin interacting protein E AGhsC240602 AGhsC231312
		NM_001283	Adaptor-related protein complex 1, sigma 1 subunit
		NM_005766	FERM, RhoGEF (ARHGEF) and pleckstrin domain protein 1 (chondrocyte-derived)
		NM_021133	Ribonuclease L (2″,5″-oligoisoadenylate synthetase-dependent)
		NM_005319	Histone 1, H1c
		NM_004571	PBX/knotted 1 homeobox 1
		NM_004491	gLucocorticoid receptor DNA binding factor 1 AGhsC180506
		AF144054	Apoptosis related protein apr-4 AGhsC181317
		BC007817	Uroplakin 1A
		XM_042029	Hypothetical protein xp_042029; flj20552 AGhsC211404
		AB020713	Nucleoporin 210
		NM_004286	Ensembl genscan prediction AceGene oligo matches these RefSeq numbers
			AGhsC050114
		XM_044987	Hypothetical protein xp_044987; loc92415
		AF161393	SEC31-like 1 (S. cerevisiae)
Cluster 3	Cell signaling	ENSG00000126549	Statherin
		NM_014918	Carbohydrate (chondroitin) synthase 1
		NM_006393	Nebulette
		AB023163	Huntingtin interacting protein 14 AGhsC050920
		NM_002698	POU domain, class 2, transcription factor 2
		XM_044972	Hypothetical protein xp_044972; acf7
		AB023191	KIAA0974 mRNA
		AF072098	Hdcmb21p
		XM_049198	Hypothetical protein xp_049198; kiaa0640
		XM_028367	Core-binding factor, runt domain, alpha subunit 2; translocated to, 3
		AF305826	Pro2972
		XM_029313	Hypothetical protein xp_029313; cpsf2 AGhsC201422
		AB021643	Zinc finger protein 325
		NM_001507	G protein-coupled receptor 38
		NM_004580	RAB27A, member RAS oncogene family
		NM_012162	F-box and leucine-rich repeat protein 6
		NM_017618	Hypothetical protein FLJ20006
		XM_007419	Similar to unknown (protein for image:3138445) (h. sapiens); loc115738
		NM_012265	Chromosome 22 open reading frame 3
		BC011264	DEAD/H (Asp-Glu-Ala-Asp/His) box polypeptide 11 (CHL1-like helicase homolog, S. cerevisiae) AGhsC250905
		BC001744	Special AT-rich sequence binding protein 1
		NM_000067	Carbonic anhydrase II
		AK022797	Hypothetical protein FLJ12735
		M10612	Apolipoprotein C-II AGhsC060605
		XM_048806	Hypothetical protein xp_048806: cdle
		M99439	Transducin-like enhancer of split 4 (E(spl) homologue, Drosophila)
		M14910	Simian sarcoma associated virus ssav-related pol region dna;
		AB049915	Single chain fv fragment AGhsC231615
Cluster 4	Cell-to-cell signaling and interaction	NM_000777	Cytochrome P450, family 3, subfamily A, polypeptide 5
			XM_015435	Btg family, member 2; btg2
		NM_000024	Beta-2 adrenergic receptor
		NM_002704	Pro-platelet basic protein (chemokine (C-X-C motif) ligand 7)
		NM_006890	Carcinoembryonic antigen-related cell adhesion molecule 7
		XM_028728	1–3-phosphoserine phosphatase homolog; PSPHL
Cluster 5	Cell cycle	S68860	TIMP-2
		NM_002519	Nuclear protein, ataxia-telangiectasia locus
		AL031705	c305c8.3.2 (kiaa0683); c305c8.3
		XM_036607	Hypothetical protein xp_036607; loc91169
		BC000967	Chronic myelogenous leukemia tumor antigen 66
		NM_021737	Chloride channel 6
		NM_003899	Rho guanine nucleotide exchange factor (GEF) 7 AGhsC050313
		AB007871	SLIT-ROBO Rho GTPase activating protein 2
		BC008920	Hypothetical protein from EUROIMAGE 2021883
		AK027484	DEAD (Asp-Glu-Ala-Asp) box polypeptide 31
		AF269144	Gamma-aminobutyric acid (GABA) A receptor, gamma 3
		XM_036813	Hypothetical protein xp_036813; snap29
		S82667	Cyclin d1-igh sγ2
		XM_008617	Hypothetical protein xp_008617; loc114958 AGhsC211302
		BC010084	Hypothetical protein SB153
		XM_034232	Hypothetical protein xp_034232; loc90806
Cluster 6	Molecular transport	XM_050864	Hypothetical protein xp_050864; loc93367
		XM_045453	Hypothetical protein xp_045453; loc92500
		AF346816	Envelope protein
		XM_049348	Hypothetical protein xp_049348; loc93119 AGhsB170620
		ENSG00000129987	AGhsC070908
		NM_016204	Growth differentiation factor 2
		XM_051247	Hypothetical protein xp_051247; loc93411
		NM_022557	Growth hormone 2
		XM_040690	Hypothetical protein xp_040690; nyd-sp20
		BC015202	Hypothetical protein FLJ13111
		XM_051857	Hypothetical protein xp_051857; kiaa1199
		XM_038468	Hypothetical protein flj14640; flj14640
Cluster 7	Tissue development	XM_046163	Hypothetical protein xp_046163; tc11a
		ENSG00000104677	AGhsC070905
		NM_001611	Tartrate resistant acid phosphatase 5
			AGhsC260514
		NM_020389	Transient receptor potential cation channel, subfamily C, member 7
		NM_006984	Claudin 10
		NM_024820	K 1AA1608
		AF090934	Maternally expressed 3
		NM_015900	Phospholipase A1 member A
		NM_016619	Placenta-specific 8 AGhsC050124 AGhsB140124
		AB018739	Neurochondrin
		NM_001877	Complement component (3d/Epstein Barr virus) receptor 2
Cluster 8	Hematological system development/cell cycle	L12088	iMmunoglobulin kappa-chain; igk
			BC009851	Immunoglobulin heavy constant mu
			ENSG00000073711	AGhsC071306
		NM_021950	Membrane-spanning 4-domains, subfamily A, member 1
		AB032952	KIAA1126 protein
		NM_000061	Bruton agammaglobulinemia tyrosine kinase
		AC004382	Unknown gene product; a-152e5.9
		AL390736	ba209j19.1.1 (gw112 protein)
			AGhsC190501
		AL133055	Hypothetical protein DKFZp434J1015
			AGhsC240112
		XM_010574	Hypothetical protein xp_010574; ptger3
			AGhsC030113
			AGhsC140506
		X98707	Mma encoding chimaeric transcript of 1 collagen type alpha and platelet derived growth factor beta 314 bp
		NM_000840	Glutamate receptor, metabotropic 3
Cluster 9	Cell cycle/cancer	D13665	Periostin, osteoblast specific factor
		AK002166	A kinase (PRKA) anchor protein 11
		Z27202	Tcr delta chain (vj)
		AF202144	Transcription factor cyclin D-binding
			Myb-like protein 1
		NM_006152	Lymphoid-restricted membrane protein
		NM_000460	Thrombopoietin
		NM_000626	CD79B antigen (immunoglobulin-associated beta)
		ENSG00000109721	AGhsB040323
		AB029020	Ubiquitin specific protease 33
		BC004921	Hypothetical protein BC004921
		XM_030808	Hypothetical protein xp_030808; dnase113
		AF202635.	pp1200
		XM_039244	Hypothetical protein xp_039244; loc91566
		L23270	Ig heavy chain
		NM_000738	Cholinergic receptor, muscarinic 1
			AGhsC100812
		BC001199	Thioredoxin domain containing 5
		X98214	Antibody heavy chain vdj region AGhsC190215
		AL139421	dj717i23.1 (novel protein similar to xenopus laevis sojo protein)
		NM_006795	EH-domain containing protein 1
			AGhsC181305
		NM_078481	CD97 antigen
		NM_020230	Peter pan homolog (Drosophila)
		NM_002423	Matrix metalloproteinase 7
		ENSG00000131865	AGhsB130523
		NM_002091	Gastrin-releasing peptide
		AJ133439	Glutamate receptor interacting protein 1
		AF190825	Purinergic receptor P2X, ligand-gated ion channel, 2
			AGhsC130424
		XM_053866	Serum amyloid a4 constitutive saa4
		AK027689	Tularik gene 1
		NM_005308	G protein-coupled receptor kinase 5
Cluster 10	Cell cycle/cancer	D13666	Periostin, osteoblast specific factor
		M13509	Matrix metalloproteinase 1
		AL138828	Ba472e5.2.1 (novel protein, isoform 1)
		NM_022787	Ensembl genscan prediction AceGene oligo matches these RefSeq numbers
		NM_001677	ATPase, Na+/K+ transporting, beta 1 polypeptide
		XM_050275	Idn3 protein
		XM_053121	Similar to erythroblast macrophage protein musculus loc94776

Table 3.

 

Accuracy of the Different Classification Methods

Classification Method	Sample Size (n)	SVM-Based	MLP	KNN
Overall	6	100%	83.33%	66.67%
3-Gene training set	6	100%	100%	83.33%
3-Gene training set	32	84.38%	83.33%	16.67%
2-Gene training set
TIMP-2, PSPHL	32	50%
Periostin, PSPHL	32	83.33%
Periostin, TIMP-2	32	83.33%

 Two popular classifiers, k-nearest-neighbor (KNN) and multilayer perception (MLP) were applied to various test datasets, to compare the classification accuracy of the SVM-based classifier with that of the other methods. “Overall” denotes the training set that includes all genes. The performance of all the classifiers was tested by the leave-one-out cross-validation procedure.

Figure 2.

 

Differential mRNA expressions of periostin, TIMP-2, and PSPHL in primary and recurrent pterygia tissues. (a) Periostin, TIMP-2, and PSPHL showed significantly different transcripts in primary and recurrent pterygia tissues (* P < 0.05, unpaired Student’s t-test). n= 14 eyes for primary tissue, n = 4 eyes for recurrent tissue from menopausal women. (b) Recurrent tissues showed significantly more periostin transcript (*P < 0.05, unpaired Student’s t-test). n = 24 eyes with primary pterygia, n= 8 eyes with recurrent pterygia.

Figure 3.

 

Localization of periostin, TIMP-2, and PSPHL in primary and recurrent pterygia by immunohistochemistry. Periostin was expressed more prominently in the basement membrane area of recurrent pterygia (b) compared with primary pterygia (a). Periostin expression is more intense in the superficial stroma. The expression of TIMP-2 is more prominent in the epithelial layer of recurrent pterygia (d) compared with primary pterygia (c). (e, f) Expression of PSPHL in superficial epithelia of pterygia. Recurrent pterygia (f) showed reduced expression compared with primary pterygia (e). (g) Pterygia tissues stained by control IgG showed no signals. Original magnification: (a, b, e, f) ×400; (c, d, g) ×1000.