Classification analysis of the transcriptosome of nonlesional cultured dermal fibroblasts from systemic sclerosis patients with early disease

Abstract

Objective

To compare the transcriptosome of early-passage nonlesional dermal fibroblasts from systemic sclerosis (SSc) patients with diffuse disease and matched normal controls in order to gain further understanding of the gene activation patterns that occur in early disease.

Methods

Total RNA was isolated from early-passage fibroblasts obtained from nonlesional skin biopsy specimens from 21 patients with diffuse SSc (disease duration <5 years in all but 1) and 18 healthy controls who were matched to the cases by age (±5 years), sex, and race. Array experiments were performed on a 16,659-oligonucleotide microarray utilizing a reference experimental design. Supervised methods were used to select differentially expressed genes. Quantitative polymerase chain reaction (PCR) was used to independently validate the array results.

Results

Of the 8,324 genes that passed filtering criteria, classification analysis revealed that <5% were differentially expressed between SSc and normal fibroblasts. Individually, differentially expressed genes included COL7A1, COL18A1 (endostatin), DAF, COMP, and VEGFB. Using the panel of genes discovered through classification analysis, a set of model predictors that achieved reasonably high predictive accuracy was developed. Analysis of 1,297 gene ontology (GO) classes revealed 35 classes that were significantly dysregulated in SSc fibroblasts. These GO classes included anchoring collagen (30934), extracellular matrix structural constituent (5201), and complement activation (6958, 6956). Validation by quantitative PCR demonstrated that 7 of 7 genes selected were concordant with the array results.

Conclusion

Fibroblasts cultured from nonlesional skin of patients with SSc already have detectable abnormalities in a variety of genes and cellular processes, including those involved in extracellular matrix formation, fibrillogenesis, complement activation, and angiogenesis.

It is generally accepted that fibroblasts play an important role in the development of cutaneous and visceral fibrosis that is the sine qua non of scleroderma (systemic sclerosis [SSc]). Abnormalities in a variety of cellular processes in SSc fibroblasts have been described; most prominent among them is the exaggerated accumulation of collagen and other extracellular matrix (ECM) components (1). Because of these properties, SSc fibroblasts have been described as being “activated.” Much of what has been learned about the pathophysiology of fibroblast activation has been made possible by detailed investigations of lesional tissue or studies of explanted fibroblasts from such tissues in cell culture systems. Because of the systemic nature of SSc, however, it is likely that even in nonlesional skin the events triggering fibroblast activation may have already taken place at the molecular level. In support of this concept, Claman et al found that fibroblasts and endothelial cells from nonlesional skin specimens from patients with SSc already display patterns of activation characteristic of SSc (2). Thus, one of our primary aims in the present study was to gain a fundamental understanding of the fibroblast activation in SSc by ascertaining which genes are dysregulated early in the course of disease development, as opposed to genes that may be activated in the later stages of fibrosis.

A second aim was to gain further insight into the gene expression profile of SSc fibroblasts in culture. Cultured dermal fibroblasts are an important model system in the investigation of the cellular and molecular pathogenesis of SSc. Studies with cell culture models have yielded important insights into the disease, including the key roles of cytokines such as transforming growth factor β (TGFβ) and connective tissue growth factor, and possible molecular defects in the Smad intracellular signaling pathways (3). However, normal human dermal fibroblasts undergo significant changes in global gene expression patterns in response to manipulations in culture (4). In addition, the SSc fibroblast activation is not a stable phenotype. Instead, these fibroblasts gradually lose their activated phenotype, such that by the tenth passage, collagen synthesis declines to normal values (5). These background changes induced by the cell culture environment add a layer of complexity to the interpretation of results from in vitro studies, sometimes making it difficult to unambiguously attribute a given observation to intrinsic defects in the SSc fibroblast itself, or to a differential response of the SSc fibroblast to the manipulations in culture. Thus, knowledge of the global gene expression of resting SSc fibroblasts can provide a context whereby results from in vitro studies can be better interpreted.

To achieve these aims, we examined the transcriptosome of nonlesional fibroblasts explanted from skin biopsy samples from SSc patients with disease duration of <5 years and compared them with fibroblasts explanted from biopsy specimens from matched, healthy controls. We used oligonucleotide microarrays because this technology can be used to ascertain the expression level of thousands of genes in parallel. Because array experiments are inherently “noisy,” we used supervised methods and rigorous statistical criteria to discover differentially expressed genes. The results show that of several thousand genes expressed by fibroblasts, <5% are differentially expressed. Some genes that were most discriminating for SSc fibroblasts were the basement membrane nonfibrillar collagen genes collagen type VII α1 (COL7A1) and XVIIIα1 (COL18A1) or endostatin, and DAF, (which protects cells from complement-mediated injury). Moreover, the data suggest that even in nonlesional fibroblasts, dysregulation of multiple genes and cellular processes, including those affecting ECM formation, fibrillogenesis, angiogenesis, and complement activation, are already detectable.

PATIENTS AND METHODS

Patients and controls.

Biopsies of nonlesional skin from 21 SSc patients (6 African American, 7 Hispanic American, and 8 white) were performed. Skin biopsy specimens were obtained from the dorsal upper arm or scapular area in all subjects, avoiding obviously sun-exposed skin. All patients fulfilled the American College of Rheumatology (formerly, the American Rheumatism Association) criteria for SSc (6) and also met criteria for diffuse SSc (7). Disease duration was <5 years in all but 1 patient. The mean ± SD modified Rodnan skin score (MRSS) (8) among the SSc patients was 27 ± 12. Pulmonary fibrosis was present in 11 patients (52%), and 8 (38%) were anti–topoisomerase I autoantibody positive by passive immunodiffusion. The 18 normal controls were screened for absence of autoimmune disease by questionnaire, medical history, and physical examination. The controls were matched by age (±5 years), sex, and race to the SSc cases. In 3 instances, a control was matched to 2 different SSc cases.

Fibroblasts and RNA.

Dermal fibroblasts were explanted from 3-mm punch biopsy specimens and cultured as previously described (9). Total RNA was extracted from the cells before passage 4, using a combination of TRIzol (Invitrogen, San Diego, CA) and RNeasy (Qiagen, Chatsworth, CA).

Probe labeling and microarrays.

For each SSc RNA sample labeled and hybridized, a matched normal control RNA sample was processed in parallel. One microgram of fibroblast total RNA was subjected to reverse transcription and a single cycle of linear amplification with T7 polymerase in the presence of aminoallyl nucleotide triphosphates to generate aminoallyl amplified RNA (aRNA; Ambion, Austin, TX). RNA amplification accurately reflects the transcript population in the parent sample with improved sensitivity for detecting low-copy-number genes, resulting in better data quality (10). To generate labeled targets, 5 μg of aminoallyl aRNA was labeled with monoreactive Alexa Fluor 555 (Molecular Probes, Eugene, OR) and column purified. A common reference RNA was similarly labeled with Alexa Fluor 647. This reference RNA was purchased commercially (Universal Human Reference RNA; Stratagene, La Jolla, CA) and consists of a mixture of total RNA from 10 human cell lines (mammary gland adenocarcinoma, hepatoblastoma, cervical adenocarcinoma, testicular embryonal carcinoma, glioblastoma, melanoma, liposarcoma, histiocytic lymphoma, lymphoblastic leukemia, and myeloma). One hundred picomoles each of Alexa Fluor 555 – and Alexa Fluor 647–labeled probe were combined and hybridized with a custom microarray as previously described (11). The microarray was printed with 16,659 human genes represented as 70-mer oligonucleotides from the Qiagen Array-Ready Oligo Set, version 1.1. The complete gene list can be downloaded from the Qiagen Web site (http://omad.qiagen.com/download/index.php). Array images were obtained on a Perkin Elmer (Emeryville, CA) Scanarray Lite scanner and features extracted using fixed-circle segmentation (QuantArray, version 3.1; Perkin Elmer). Since class comparison and class prediction were the primary aims, we favored studying larger numbers of biologic samples over fewer samples with multiple technical replicates (12). However, a few technical duplicates were performed on a subset of the samples (4 SSc and 4 corresponding matched controls) to ensure that protocols were working properly.

Data preprocessing and normalization.

Features with poor signal-to-noise ratio (<3) or that were obvious artifacts were flagged for exclusion. To filter out nonexpressed genes, we used the t-test to compare the signal from each gene with the signals from 30 negative control features on the array across all experiments for each channel. Any gene whose signal was not significantly different from negative control features (P > 0.05) was excluded. In addition, any gene in which the expression data had been filtered out in 25% of the experiments was also excluded from further analysis. The signal intensities from both channels were transformed into log₂ ratios and local weighted least squares regression (or “lowess”) normalized to adjust for differences in labeling intensities of the 2 fluorescent dyes (13).

Data analysis.

Classification.

Supervised classification methods were used to detect differentially expressed genes between the 2 classes. The program BRB ArrayTools, version 3.2 (designed by Richard Simon and Amy Peng Lam, Biometric Research Branch, National Cancer Institute) was used for these analyses. This program implements a variety of analytic methods including classification, class prediction, quantitative trait analysis (14), significance analysis of microarrays (SAM) (15), and comparative gene ontology (GO) analysis.

For classification, the random-variance model t-test was used to discover differentially expressed genes (14). In addition, a global test was performed to determine whether the overall expression profiles differed significantly between the classes. This was done by randomly permuting the labels of which arrays corresponded to which classes. Then, for each of the 10,000 permutations, the P values were recomputed and the number of genes differentially expressed at a given level of significance. The proportion of permutations that gave at least as many significant genes as the actual data is the significance level of the global test.

In order to minimize false-positives, we specified that for a given significance level, the number of false discoveries must not exceed 10 and the proportion must not exceed 0.1, with 95% confidence. To estimate the actual number of false discoveries at a given significance level, we utilized spacings locally weighted regression smoother histogram (SPLOSH) to calculate the conditional false discovery rate (cFDR) (16).

Class prediction.

Models that utilized gene expression profiles to predict the class of a given fibroblast strain (SSc or normal) were developed. The models used were the compound covariate predictor (17), diagonal linear discriminant analysis (18), nearest neighbor classification (18), and support vector machines (19). The prediction error of each model was estimated using leave-one-out cross-validation (LOOCV) (20). For each LOOCV training set, the entire model building process was repeated, including the gene selection process. We also evaluated whether the cross-validated error rate estimate for a model was significantly less than would be expected from random prediction. The class labels were randomly permuted 10,000 times, and the entire LOOCV process was repeated. The significance level is the proportion of the random permutations that gave a cross-validated error rate no greater than the cross-validated error rate obtained with the real data.

Quantitative trait analysis.

Genes whose expression might be significantly related to the patient's MRSS were identified by computing the statistical significance level for each gene, testing the hypothesis that the Spearman's correlation between gene expression and MRSS was 0. These P values were then used in a multivariate permutation test in which the MRSS were randomly permuted among arrays (21).

GO analysis.

To discover cellular processes or pathways whose expression was differentially regulated between SSc and normal fibroblasts, genes were analyzed by GO groups rather than individually. This reduces the number of tests conducted and enables findings among biologically related genes to reinforce one another. In this analysis, the statistical significance values are based on testing the null hypothesis that the list of genes in a given GO category is a random selection from the genes on the array, against the alternative hypothesis that the GO category contains more genes differentially expressed between the 2 classes being compared. For a given GO category, 2 statistics that summarize the P values for genes in the group are computed: the Fisher least significance (LS) statistic and the Kolmogorov-Smirnov (KS) statistic (as defined in ref. 22). Genes are randomly selected from the array and the summary statistic computed for those random samples. This process is repeated 100,000 times to obtain a distribution of these statistics. The number of genes randomly selected for permutation is equivalent to the actual number of genes in a GO category being considered. The significance level for a GO category is the proportion of the random samples giving as large a value of the summary statistic as in the actual genes contained in that GO category. A GO category was considered significantly differentially regulated if either the LS or the KS significance level was ≤0.01.

Quantitative real-time polymerase chain reaction (PCR).

DNase-treated fibroblast total RNA was reverse transcribed (High-Capacity cDNA Archive kit; Applied Biosystems, Foster City, CA), and quantitative PCR was performed using commercially developed semiquantitative TaqMan assays (Assays-on-Demand; Applied Biosystems) on an ABI 7900 Sequence Detection System with Sequence Detection software according to the instructions of the manufacturer (Applied Biosystems). Cyclophilin A (PPIA) was used as the endogenous control.

RESULTS

Classification analysis and class prediction.

We found that 8,324 genes on the array passed data preprocessing and filtering criteria when all experiments were considered. The most strongly expressed genes in fibroblast strains of both types were reversion-inducing cysteine-protein with kazal motifs (RECK), stromelysin 1 or matrix metalloproteinase 3 (MMP3), extracellular sulfatase 1 (SUFL1), COL3A1, COL1A2, COL6A3, thrombospondin 1 (THBS1), a disintegrin-like and metalloprotease (reprolysin type) with thrombospondin type 1 motif 5 or aggrecanase 2 (ADAMTS5), ADAMTS2, platelet-derived growth factor receptor β polypeptide (PDGFRB), and osteonectin or secreted protein, acidic and rich in cysteine (SPARC).

We then used the random-variance t-test to assess the 8,324 genes expressed by the fibroblasts for differential expression. Since selection of a threshold level of significance in which a gene is declared differentially expressed represents a compromise between the proportion of tolerable false-positives and the proportion of tolerable false-negatives, we used SPLOSH to establish a threshold significance of α = 0.043 as the level that minimizes the estimated total of false discoveries and false-negatives (16). This resulted in 832 genes (10%) that could potentially be declared as differentially expressed. At this level of significance (α = 0.043), the global probability of finding this number of genes to be different by chance if there was in fact no real difference between the 2 fibroblast types was P = 0.0039. However, the cFDR of 0.384 at this level of significance gives rise to ∼320 false potential discoveries. We also estimated, using SPLOSH, that ∼1,040 false-negatives, i.e., differentially expressed genes that do not achieve this level of significance, may be present. A supplementary table listing the first 100 genes that were significantly differentially expressed with the random-variance t-test is available online at http://www.uth.tmc.edu/scleroderma, or upon request from the corresponding author. The maximum corresponding cFDR for these 100 genes was 0.241. The first 71 genes could be declared differentially expressed between SSc and normal fibroblasts with a reasonably conservative number of false-positives (∼10). These genes are listed in Table 1.

Table 1. First 71 differentially expressed genes ordered by significance, and the associated conditional false discovery rate (cFDR)*

P†	cFDR‡	% cross-validation correct§	Geometric mean of ratios (systemic sclerosis: normal)	Name	GenBank	UniGene	Symbol
* All genes listed except 2 (EH domain–containing 3 and 3-α hydroxysteroid dehydrogenase type IIb) were also differentially expressed by significance analysis of microarrays (SAM). There were a total of 166 genes that were significant by SAM, with a median false discovery rate of 0.0854. † By random variance t-test. ‡ Estimated by SPLOSH (spacings locally weighted regression smoother histogram) (see Patients and Methods). § Samples correctly classified by leave-one-out cross-validation (see Patients and Methods). The first 26 genes were found to comprise the nominal set for class prediction. With this gene panel, the 1- and 3-nearest neighbors (both P = 0.001), nearest centroids (P = 0.001), and support vector machines (P = 0.001) achieved 100% estimated predictive accuracy, and the compound covariate (P = 0.003) and diagonal line linear discriminant analysis predictor (P = 0.002) achieved estimated 96% accuracy.
0.0000001	0.000	100	1.553	Collagen, type VII, α1	L02870	1640	COL7A1
0.0000002	0.000	100	0.621	Decay—accelerating factor for complement (CD55)	M30142	1369	DAF
0.0000006	0.000	100	1.453	6-phosphofructo-2-kinase/fructose-2,6-biphosphatase 3	AF109735	195471	PFKFB3
0.0000015	0.000	100	0.548	Serum and glucocorticoid–induced kinase	AJ000512	296323	SGK
0.0000015	0.000	100	0.548	UDP-N-acetyl-α-D-galactosamine:polypeptide N-acetylgalactosaminyltransferase 12	NM_024642	47099	GALNT12
0.0000052	0.001	100	0.584	Neuronal cell adhesion molecule	AB002341	7912	NRCAM
0.0000314	0.007	100	0.811	cDNA DKFZp434B1620	AL137548	43112
0.0000314	0.007	100	0.728	Heme-binding protein	NM_015987	108675	HEBP
0.0000318	0.008	100	0.465	Aldo-keto reductase family 1, member C3 (3-α hydroxysteroid dehydrogenase, type II)	D17793	78183	AKR1C3
0.0000376	0.009	100	0.708	Inositol polyphosphate-1-phosphatase	L08488	32309	INPP1
0.0000382	0.009	100	0.62	Hypothetical protein FLJ20546	AK000953	279896	FLJ20546
0.0000388	0.009	100	0.768	Alcohol dehydrogenase 5 (class III), chi polypeptide	M81118	78989	ADH5
0.0000389	0.009	100	0.517	Aldo-keto reductase family 1, member C2 (dihydrodiol dehydrogenase 2; bile–acid binding protein; 3-α)	U05598	201967	AKRIC2
0.0000462	0.011	76	0.751	Tetraspan 3	AK001326	100090	TSPAN-3
0.0000477	0.011	60	1.422	Metallothionein 1X	X65607	278462	MTIX
0.000051	0.012	60	1.251	Heterogeneous nuclear RNP C (C1/C2)	M16342	182447	HNRPC
0.0000587	0.014	40	1.352	Desmoplakin (DPI, DPII)	P15924	349499	DSP
0.000066	0.016	36	1.249	Aldehyde dehydrogenase 2 family (mitochondrial)	X05409	195432	ALDH2
0.0000768	0.018	36	1.314	Ras-related C3 botulinum toxin substrate 2 (rho-family, small GTP binding protein Rac2)	Z82188	173466	RAC2
0.0000796	0.019	24	0.771	Hypothetical protein FLJ12436	NM_024661	69485	FLJ12436
0.000081	0.019	40	1.757	Collagen, type XVIII, α1	AF018081	78409	COL18A1
0.0000814	0.019	36	1.689	Platelet-derived growth factor C	NM_016205	43080	PDGFC
0.0000837	0.020	36	0.721	Cellular repressor of E1A-stimulated genes	AF084523	5710	CREG
0.0000841	0.020	32	0.724	Receptor tyrosine kinase–like orphan receptor 1	M97675	274243	ROR1
0.0000957	0.023	32	0.689	Glycophorin C (Gerbich blood group)	NM_002101	81994	GYPC
0.0000978	0.023	36	0.742	KIAA0469 gene product	AB007938	7764	KIAA0469
0.000102	0.024	–	0.629	cDNA DKFZp434B1272	AL162032	158258
0.0001051	0.025	–	1.436	Metallothionein 1A (functional)	K01383	203967	MT1A
0.0001231	0.029	–	0.664	Retinal short-chain dehydrogenase/reductase retSDR3	NM_016246	18788	LOC51171
0.0001365	0.032	–	0.815	V-maf musculoaponeurotic fibrosarcoma (avian) oncogene family, protein G	AF059195	252229	MAFG
0.0001408	0.034	–	0.706	A kinase (PRKA) anchor protein (gravin) 12	U81607	788	AKAP12
0.0001438	0.034	–	1.23	KRAB-associated protein 1	X97548	228059	TIFIB
0.0001535	0.037	–	1.333	Fibroblast growth factor receptor–like 1	AK000530	193326	FGFRL1
0.0001574	0.037	–	0.702	KIAA0537 gene product	AB011109	200598	KIAA0537
0.0001615	0.038	–	1.292	Phosphatidylglycerophosphate synthase	NM_024419	278682	PGS1
0.0001766	0.042	–	1.483	EH domain–containing 3	NM_014600	87125	EHD3
0.0001804	0.043	–	0.773	Predicted osteoblast protein	D87120	29882	GS3786
0.0001883	0.045	–	0.631	Histamine N-methyltransferase	U44111	81182	HNMT
0.000199	0.047	–	1.66	Homo sapiens clone N97 immunoglobulin heavy chain variable region mRNA, partial CDs	AF103295	283882
0.0002046	0.049	–	1.386	Hypothetical protein FLJ20364	AK000371	32471	FLJ20364
0.0002073	0.049	–	0.792	Limbic system–associated membrane protein	U41901	26479	LSAMP
0.000236	0.056	–	0.735	Nidogen 2	AB009799	82733	NID2
0.0002425	0.058	–	1.375	Homo sapiens CD44 isoform RC	AF098641	306278	CD44
0.0002495	0.059	–	0.704	Glutathione S-transferase M3 (brain)	AF043105	2006	GSTM3
0.0002629	0.063	–	0.685	Solute carrier family 11 (proton-coupled divalent metal ion transporters), member 3	NM_014585	5944	SLC11A3
0.000273	0.065	–	1.292	Hypothetical protein FLJ10292	AK001154	104650	FLJ10292
0.0002776	0.066	–	0.68	Acid sphingomyelinase–like phosphodiesterase	AK000184	42945	ASM3A
0.000283	0.067	–	0.803	Niemann-Pick disease, type C1	AF002020	76918	NPC1
0.0003011	0.072	–	1.255	KIAA1118 protein	AB029041	209646	KIAA1118
0.0003091	0.074	–	1.211	Hsp75	NM_016292	182366	TRAP1
0.000312	0.074	–	1.374	Metallothionein 1B (functional)	M13485	211924	MTIB
0.0003147	0.075	–	0.742	Lipin 2	D87436	166318	LPIN2
0.0003323	0.079	–	0.592	3-α hydroxysteroid dehydrogenase type IIb	NM_016253	279801	LOC51708
0.0003559	0.085	–	0.786	Progesterone membrane-binding protein	AJ002030	9071	PMBP
0.0003599	0.086	–	0.775	Neural precursor cell expressed, developmentally down-regulated 4	D42055	1565	NEDD4
0.0003609	0.086	–	0.805	Osteopetrosis-associated transmembrane protein 1	NM_014028	163724	HSPC019
0.0003735	0.089	–	0.774	Paraoxonase 2	AF001601	169857	PON2
0.0003803	0.091	–	0.538	Platelet-derived growth factor receptor–like	D37965	170040	PDGFRL
0.000396	0.094	–	1.243	Neuropathy target esterase	AJ004832	5038	NTE
0.0004317	0.103	–	0.785	GABA-A receptor–associated protein–like 1	AF087847	282654	GABARAPL1
0.0004521	0.108	–	1.427	KIAA0008 gene product	D13633	77695	KIAA0008
0.0004652	0.111	–	1.279	Kinesin family member 4A	AF071592	279766	KIF4A
0.0004899	0.117	–	1.451	Cartilage oligomeric matrix protein	AC003107	1584	COMP
0.0004923	0.117	–	0.623	1,2-α-mannosidase IC	NM_020379	8910	HMIC
0.0004929	0.117	–	0.768	Destrin (actin depolymerizing factor)	S65738	82306	ADF
0.0005009	0.119	–	0.756	Vascular endothelial growth factor B	U43368	78781	VEGFB
0.00052	0.124	–	1.189	Homo sapiens HspC337 mRNA, partial CDs	AF161451	284295	HSPC333
0.0005649	0.134	–	0.786	Very-long-chain acyl–coenzyme A synthetase homolog 2	NM_012254	111401	VLCS-H2
0.00058	0.138	–	0.742	Neutral sphingomyelinase (N-SMase) activation–associated factor	X96586	78687	NSMAF
0.0005998	0.143	–	1.367	KIAA1329 protein	AB037750	21061	KIAA1329
0.0006111	0.145	–	0.804	Inositol(myo)-1(or 4)-monophosphatase 2	NM_014214	5753	IMPA2

As an alternative approach to the random-variance t-test, we used the SAM method to discover genes that were differentially regulated (15). This method revealed 166 genes that were significantly differentially expressed between SSc and normal fibroblasts, with a median FDR of 0.0854 (http://www.uth.tmc.edu/scleroderma). Of the first 100 genes significantly differentially expressed by the random-variance t-test, 97 were found to be differentially expressed by the SAM method as well. Genes in which expression in SSc fibroblasts was significantly increased by both methods included COL7A1, COL18A1, COMP, CD44 (CDw44, lymphocyte homing antigen), and 5 metallothionein genes (MT1X, MT1A, MT1B, MT2A, and MT1F). Genes that were most significantly decreased included SGK, VEGFB, DAF, and PTX3. Among the 71 genes that were differentially regulated in SSc fibroblasts, there were 11 uniquely expressed sequence tags or as-yet-uncharacterized genes (Table 1).

Having established that there are genes that are differentially expressed between SSc and normal fibroblasts, we sought to determine if these genes might be useful for class prediction. Six different multivariate classification models were tested on the actual data to determine in which class a given sample belonged. Using multiple univariate parametric significance thresholds, we empirically determined the nominal number of genes to be included in the predictor. During this iterative process, the average estimated predictive accuracy increased from 91% to 99% as the significance thresholds became more stringent (α = 10⁻²–10−⁴), and the number of genes included in the classifier fell from 223 to 26. The optimum panel consisted of these 26 genes (Table 1) and demonstrated an average estimated predictive accuracy of 99% with all 6 models. Restricting significance thresholds further to α = 10⁻⁵ (15-gene classifier), however, reduced estimated classification accuracy to 89%.

Quantitative trait analysis.

All but 2 of the 21 SSc patients were members of a prospective outcomes cohort; thus, information on the MRSS at the time of enrollment was available on 19 patients. Since SSc patients with an MRSS >20 tend to have worse prognosis (23, 24), we grouped the SSc fibroblast strains based on the MRSS of the donor patient (>20 versus <20). Fifty-seven percent of the SSc patients had MRSS >20, while 43% had MRSS <20. Classification analysis did not reveal any genes to be significantly differentially expressed between these 2 groups, and the global test of whether the overall expression profiles between fibroblasts from these patients were indeed different did not reveal statistically significant differences. Quantitative trait analysis did not reveal any genes that were significantly correlated with MRSS. Similarly, no genes were significantly differentially expressed when the SSc cases were subgrouped based on anti–topoisomerase I status or pulmonary fibrosis.

GO analysis.

To investigate possible mechanistic defects in cellular processes in SSc fibroblasts, we performed GO analysis of genes. Of the 1,297 GO classes considered, 35 displayed significant differential regulation between SSc and normal fibroblasts (Table 2). Some of the GO categories contained highly overlapping gene lists. These included 323 (lytic vacuole), 5764 (lysosome), and 5773 (vacuole); 5581 (collagen), 30934 (anchoring collagen), and 5201 (extracellular matrix structural constituent); 6956 (complement activation) and 6958 (complement activation, classical pathway); and 6692 (prostanoid metabolism), and 6693 (prostaglandin metabolism).

Table 2. GO categories that discriminate between cultured systemic sclerosis and control fibroblasts*

GO category	GO description	No. of genes	P, LS permutation	P, KS permutation	Geometric mean of ratios†
* Detailed annotations of gene ontology (GO) categories are available at http://www.geneontology.org. LS = least significance statistic; KS = Kolmogorov-Smirnov statistic. † Average of all genes on the array that passed filtering criteria in that GO class.
8047	Enzyme activator activity	81	0.1143	0.0060	1.002
15078	Hydrogen ion transporter activity	73	0.2950	0.0081	1.004
30246	Carbohydrate binding	71	0.0429	0.0093	0.972
323	Lytic vacuole	54	0.0039	0.0256	0.939
5764	Lysosome	54	0.0039	0.0256	0.939
5773	Vacuole	54	0.0039	0.0256	0.939
5201	Extracellular matrix structural constituent	36	0.0050	0.0984	1.006
1871	Pattern binding	40	0.0231	0.0058	0.939
16477	Cell migration	26	0.0033	0.0066	0.960
4867	Serine-type endopeptidase inhibitor activity	23	0.0031	0.0510	0.972
5581	Collagen	20	0.0092	0.4945	1.066
6665	Sphingolipid metabolism	17	0.0088	0.0047	0.903
6956	Complement activation	16	0.0051	0.2052	0.896
6958	Complement activation, classical pathway	13	0.0009	0.0838	0.874
6690	Eicosanoid metabolism	11	0.0057	0.0273	0.946
4028	Aldehyde dehydrogenase activity	9	0.0095	0.0622	0.993
46519	Sphingoid metabolism	9	0.0055	0.0184	0.888
7411	Axon guidance	9	0.0082	0.2469	0.883
6000	Fructose metabolism	8	0.0073	0.2051	1.023
8154	Actin polymerization and/or depolymerization	8	0.0187	0.0095	0.980
16628	Oxidoreductase activity, acting on the CH-CH group of donors, NAD or NADP as acceptor	8	0.0001	0.0002	0.790
8645	Hexose transport	7	0.0213	0.0004	0.952
15749	Monosaccharide transport	7	0.0213	0.0004	0.952
15758	Glucose transport	7	0.0213	0.0004	0.952
9986	Cell surface	7	0.0066	0.3479	0.932
6672	Ceramide metabolism	7	0.0016	0.0008	0.856
30934	Anchoring collagen	6	0.0036	0.5163	1.101
5087	Ran guanyl-nucleotide exchange factor activity	5	0.0221	0.0016	1.044
4030	Aldehyde dehydrogenase (NAD[P]+) activity	5	0.0021	0.0055	0.980
30166	Proteoglycan biosynthesis	5	0.0849	0.0029	0.962
45664	Regulation of neuron differentiation	5	0.0083	0.3449	0.942
51094	Positive regulation of development	5	0.0087	0.3311	0.930
45597	Positive regulation of cell differentiation	5	0.0087	0.3311	0.930
6692	Prostanoid metabolism	5	0.0021	0.0122	0.866
6693	Prostaglandin metabolism	5	0.0021	0.0122	0.866

Of interest, in GO 5201, in addition to the interstitial collagens (types I and III), there were many genes important in microfibril formation. These include types V, VI, and IX collagen, fibrillin 1, fibrillin 2, the microfibril-associated proteins (MFAPs) MFAP2 and MFAP3, microfibrillar-associated glycoprotein 2 (MAGP), fibulin 2 (FBNL2), elastin microfibril interface–located protein (EMILIN), and epidermal growth factor–containing fibulin-like extracellular matrix protein 2 (EFEMP2). Of these, fibrillin1, MAGP2, and MFAP2 were significantly decreased in SSc fibroblasts (P = 0.012, P = 0.025, and P = 0.011, respectively). Likewise, in GO 6956, there were several genes whose products inhibit complement activation, including membrane cofactor protein (MCP or CD46) and decay-accelerating factor for complement (DAF), both of which also were significantly decreased in SSc fibroblasts (P = 0.008 and P = 2 × 10⁻⁷, respectively). In GO 1871, thrombospondins 1 (THBS1) and 3 (THBS3) and vascular endothelial growth factor B (VEGFB) expression were decreased, while CD44 was significantly overexpressed, in SSc fibroblasts.

Real-time PCR validation of microarray results.

We performed quantitative PCR using the same fibroblast total RNA as was used for the microarray experiments (Figure 1). Overall, the quantitative PCR results were concordant with the array results in 7 of 7 genes tested, in terms of significant differences in expression between SSc and normal fibroblasts. These included COL1A2, which was not differentially expressed between the 2 cell types by classification analysis. The quantitative PCR box plots showed that the median level of COL1A2 transcript was indeed higher in the SSc fibroblasts compared with controls (Figure 1), but because of variation in COL1A2 levels among the SSc fibroblast cell strains, the overall difference was not significant.

Figure 1. Box plots of genes selected for validation by quantitative polymerase chain reaction (PCR). Total RNA from the systemic sclerosis (SSc) and normal fibroblast lines used in the array studies was subjected to quantitative PCR using commercially available semiquantitative TaqMan assays (Assays-on-Demand; Applied Biosystems, Foster City, CA). Cyclophilin A was used as the endogenous loading normalizer. The plots show the normalized transcript levels, where the boxes represent the 25th to 75th percentiles, the horizontal lines within the boxes represent the medians, and the lines outside the boxes represent the 10th to 90th percentiles. Circles indicate outliers. TRAP1 is a synonym for Hsp75. TRAP1, COL7A1, COL18A1, and RUVBL2 transcripts were increased by 1.7-, 2.4-, 5.3-, and 1.6-fold, respectively, and SGK and VEGFB transcripts were decreased by 1.9- and 1.6-fold, respectively, in SSc fibroblasts compared with normal fibroblasts. P values were determined by Mann-Whitney U test. On the arrays, RUVBL2 was increased by 1.2-fold (P = 0.0032).

DISCUSSION

This is the largest series to date comparing the transcriptosome of early-passaged cultured dermal fibroblasts explanted from nonlesional skin of SSc patients and normal controls. A reference experimental design was used since it was the best suited for classification analysis (12). The number of genes passing preprocessing criteria, i.e., 8,324, compares favorably with estimates of 10⁴ genes expressed at ≥1 copy per cell in a recent study of gene expression in 2 human cell lines (HB4a and HCT-116) by massively parallel signature sequencing (25).

Our results demonstrate that the majority of genes are expressed at similar levels in SSc and normal fibroblasts in culture and that ∼18% of genes represented on the arrays are potentially significantly differentially regulated, but our study had sufficient statistical power to detect only a fraction of these. In contrast, in a recent array study that included 10 SSc patients and 4 controls, the investigators failed to detect any differentially expressed genes in cultured SSc fibroblasts but were able to find many genes differentially expressed in biopsy specimens from intact SSc skin (26). One consideration may be the insufficient statistical power to detect differences given the small sample size. Other explanations might be the loss of host factors (e.g., autoantibodies or inflammatory or endothelial cells) in the cell culture environment which help maintain the scleroderma fibroblast phenotype in vivo (5) or the heterogeneity of SSc fibroblasts in culture systems (27, 28).

Although our observations are limited to the transcript level, the data offer a glimpse into the gene activation patterns in dermal fibroblasts that occur early in the course of SSc before the skin involvement becomes clinically apparent. Of particular interest were the results of the analysis of differential gene expression by GO classes (Table 2). GO classes are tree-structured, controlled vocabularies (ontologies) that describe gene products in terms of their associated biologic processes, cellular components, and molecular functions in a species-independent manner. Thus, analysis of GO classes provides biologic insight into possible cellular functions that may be dysregulated in SSc fibroblasts. It can be appreciated from the data that while only a small number of genes are differentially expressed by SSc fibroblasts, these genes impact a wide variety of biologic processes and cellular components. It is not surprising, given the nature of SSc, that 3 GO classes related to ECM are significantly differentially expressed. Genes in these classes include interstitial and basement membrane collagens, as well as genes important in microfibril formation (29). Twelve genes in GO 5201 are directly involved in the formation of ECM microfibrils. There was subtle (∼1.3-fold), but statistically significant, decrease in FBN1, MFAP2, and MAGP2 expression in SSc fibroblasts. This is noteworthy given the results of previous genetic and functional studies of fibrillin 1 in human SSc (30) and the critical role of fibrillin 1 in the sequestration of TGFβ and the regulation of its signaling in the ECM (31).

Unexpectedly, GO classes representing complement were also significantly differentially regulated, and among these were 4 genes, DAF, I factor (IF), clusterin (CLU), and MCP that encode proteins that inhibit complement activation. This is the first report of significantly decreased DAF and MCP expression in SSc fibroblasts, although decreased expression of DAF and MCP in endothelial cells from biopsy samples of SSc skin has been reported previously (32). It is also interesting to note that the expression of CD44, a lymphocyte homing receptor, was significantly increased in SSc fibroblasts. CD44 is a multistructural cell-surface glycoprotein that has multiple isoforms and is involved in mediating inflammatory cell function, cell migration, as well as tumor growth and metastasis (33). The CD44RC isoform that is overexpressed in SSc fibroblasts has been shown to have particularly enhanced binding affinity for hyaluronan (34). Earlier histologic studies have demonstrated striking overexpression of CD44 in stratum granulosum, stratum spinosum, lymphocytes, and macrophages from SSc patients (35). On the other hand, we observed ∼2-fold decreased expression of serum and glucocorticoid–induced kinase (SGK), a stress-response protein kinase that mediates cell survival signals by phosphorylating and negatively regulating the proapoptotic gene FOXO3a (36, 37). Abnormal regulation of these genes might increase susceptibility to immune-mediated injury, cell death, and perhaps subsequent fibrosis.

Increased COL7A1 was one of the most discriminating features of fibroblasts from patients with early SSc. COL7A1 transcripts were, on average, increased in SSc fibroblasts by 2.4-fold by quantitative PCR (Figure 1). Type VII collagen is present in the basement membrane zone of the dermal–epidermal junction and is an integral component of the anchoring fibrils providing adhesion of the lamina densa to its underlying stroma (38, 39). Consistent with our observations, Rudnicka et al found that COL7A1 transcripts were increased in cultured SSc fibroblasts (40). Moreover, the protein is aberrantly expressed in the dermis of SSc patients, accompanied by elevated expression of immunodetectable TGFβ (40). Another highly discriminating feature was the increased expression of COL18A1, accompanied by decreased expression of VEGFB. On average, COL18A1 was increased by 5.3-fold and VEGFB was decreased by 1.6-fold in SSc fibroblasts by quantitative PCR (Figure 1). Like COL7A1, COL18A1 (endostatin) is also found in the skin along the dermal–epidermal junction and around small vessels (41). Endostatin is derived from a 20-kd C-terminal fragment of type XVIII collagen and is a potent inhibitor of angiogenesis and endothelial proliferation that has been reported to also induce endothelial cell apoptosis (42, 43). It is synthesized in a variety of tissues, but is particularly abundant in the lung, liver, blood vessels, and kidney (41, 44, 45).

Our finding that COL7A1 and COL18A1 were differentially expressed in nonlesional fibroblasts suggests that they might potentially be useful biomarkers for disease. Indeed, Hebbar and colleagues have reported that SSc patients have increased levels of circulating endostatin, which correlate positively with the extent of skin disease (46). Although the link between endostatin and fibrosis has yet to be completely investigated, results of some studies suggest that it may be involved in some animal models of hepatic fibrosis (47). VEFGB, on the other hand, promotes angiogenesis and endothelial cell growth, but its effects are inhibited by endostatin (48). Dysregulation of these genes in SSc fibroblasts might inhibit angiogenesis, leading to microvascular abnormalities in the dermis. It is interesting to also note the induction of 5 metallothionein (MT) genes in the SSc fibroblasts. MTs are a family of stress-induced proteins with diverse physiologic functions whose expression is induced by a variety of conditions including heavy metals, oxidative stress, and hypoxia (49, 50). It has been shown that MT positively regulates the cellular level and activity of NF-κB, which, in turn, is an important regulator of genes that are involved in inflammation, immune response, and apoptosis (51, 52).

Taken together, these array data suggest that SSc fibroblasts from nonlesional skin already have subtle, but detectable, abnormalities in a variety of cellular processes, especially those affecting ECM formation, fibrillogenesis, angiogenesis, and complement activation. The fact that some of the most discriminating genes (COL7A1 and COL18A1) are normally expressed at the dermal–epidermal junction suggests that this site, where fibroblasts are in close proximity to the microvasculature, may be an important location in which early pathologic processes that lead to SSc take place.

Using these expression data, we developed a set of model predictors that achieved high predictive accuracy with simple binary classes. These models could be further developed for clinical use in the future, for example, by incorporation of expression data from fibroblasts explanted from patients with limited SSc or with SSc in different stages, or from patients with other cutaneous fibrosing diseases. Ultimately, to be clinically relevant, these models will need to take into account outcome parameters such as disease survival or surrogate markers of survival in SSc, such as MRSS and pulmonary fibrosis (23, 24). With the existing data set, we are unable to demonstrate significant correlations of gene expression with simple binary outcomes (death, pulmonary fibrosis, MRSS >20, etc.). This is likely due to loss of power with division of the SSc patients into subgroups. Using much larger samples, many investigators have been able to correlate gene expression with disease outcomes in various types of cancer (53–55).

Finally, it should be pointed out that a variety of methods are available to extract useful data from microarray studies, and new analytical approaches continue to be developed while the field matures. While some methods are more widely used than others, there is no standardized approach. The criteria used in this study for feature extraction, preprocessing, data filtering, normalization, etc. might result in the misclassification of some genes that may, in fact, be truly differentially expressed but are just below detectable thresholds. By using other analytical approaches, additional genes that are aberrantly regulated in SSc fibroblasts might be discovered. In addition, this report can only cover a small fraction of the large amount of data that is generated by microarray studies. Supplemental data (such as a comprehensive list of differentially expressed genes and GO listings, as well as other expression data) are available online at http://www.uth.tmc.edu/scleroderma and may be used to help direct future studies.

Acknowledgements

The authors would like to acknowledge Gina Tercero Isada and Pravitt Gourh for their excellent technical assistance, the scleroderma patients and healthy volunteers who participated in this study.