Novel approaches to gene expression analysis of active polyarticular juvenile rheumatoid arthritis

Abstract

Juvenile rheumatoid arthritis (JRA) has a complex, poorly characterized pathophysiology. Modeling of transcriptosome behavior in pathologic specimens using microarrays allows molecular dissection of complex autoimmune diseases. However, conventional analyses rely on identifying statistically significant differences in gene expression distributions between patients and controls. Since the principal aspects of disease pathophysiology vary significantly among patients, these analyses are biased. Genes with highly variable expression, those most likely to regulate and affect pathologic processes, are excluded from selection, as their distribution among healthy and affected individuals may overlap significantly. Here we describe a novel method for analyzing microarray data that assesses statistically significant changes in gene behavior at the population level. This method was applied to expression profiles of peripheral blood leukocytes from a group of children with polyarticular JRA and healthy control subjects. Results from this method are compared with those from a conventional analysis of differential gene expression and shown to identify discrete subsets of functionally related genes relevant to disease pathophysiology. These results reveal the complex action of the innate and adaptive immune responses in patients and specifically underscore the role of IFN-γ in disease pathophysiology. Discriminant function analysis of data from a cohort of patients treated with conventional therapy identified additional subsets of functionally related genes; the results may predict treatment outcomes. While data from only 9 patients and 12 healthy controls was used, this preliminary investigation of the inflammatory genomics of JRA illustrates the significant potential of utilizing complementary sets of bioinformatics tools to maximize the clinical relevance of microarray data from patients with autoimmune disease, even in small cohorts.

Introduction

'Juvenile rheumatoid arthritis' (JRA), a term for the most prevalent form of arthritis in children, is applied to a family of illnesses characterized by chronic inflammation and hypertrophy of the synovial membranes. The term overlaps, but is not completely synonymous, with the family of illnesses referred to as juvenile chronic arthritis and/or juvenile idiopathic arthritis in Europe. We [1] and others [2] have proposed that the pathogenesis of rheumatoid disease in adults and children involves complex interactions between innate and adaptive immunity. This complexity lies at the core of the difficulty of unraveling disease pathogenesis. Both innate and adaptive immune systems use multiple cell types, a vast array of cell-surface and secreted proteins, and interconnected networks of positive and negative feedback [3]. Furthermore, while separable in thought, the innate and adaptive wings of the immune system are functionally intersected [4], and pathologic events occurring at these intersecting points are likely to be highly relevant to our understanding of pathogenesis of adult and childhood forms of chronic arthritis [5].

Polyarticular JRA is a distinct clinical subtype characterized by inflammation and synovial proliferation in multiple joints (four or more), including the small joints of the hands [6]. This subtype of JRA may be severe, because of both its multiple joint involvement and its capacity to progress rapidly over time. Although clinically distinct, polyarticular JRA is not homogeneous, and patients vary in disease manifestations, age of onset, prognosis, and therapeutic response. These differences very likely reflect a spectrum of variation in the nature of the immune and inflammatory attack that can occur in this disease [1].

Gene expression profiling using microarrays provides a highly parallel assay for assessing molecular pathophysiology in a comprehensive manner. It holds the potential to refine our understanding of complex disease states. However, microarray data analysis is commonly limited to a simple assessment of a single behavioral change in gene expression, genes that are up- or down-regulated on average among distinct populations. This approach has been used to identify groups of genes that are prognostically or diagnostically relevant, but the predictive power of these gene sets for autoimmune disease has proved limited [7–9]. Changes in gene behavior among individuals in diseased populations are complex and may reflect both the unique genetic makeup of individuals and distinct subclasses of disease.

In this preliminary investigation of the inflammatory genomics of JRA, we report the application of a novel bioinformatics approach to microarray data for the identification of genes whose expression behavior is modulated by disease in a complex manner at the population level. Accordingly, genes whose expression within a population changes from stable to variable are identified. This measure of gene behavior emulates at the molecular level the loss of homeostasis characteristic of disease pathogenesis. The method identified a significant number of genes relevant to the pathophysiology of polyarticular JRA distinct from those identified by standard differential gene expression analysis. In addition, we followed a subset of patients during therapy to characterize temporally dependent changes in gene expression. Using discriminant function analysis (DFA) to analyze this cohort, we identified gene expression changes characteristic of therapeutic response approximately one month before the time at which full clinical response occurred. A clinical assay could be created from this data that may predict soon after initiation of therapy which patients will respond and which will not. The predictive potential of this data is predicated on the fact that within 2 to 4 weeks after the start of therapy, gene expression in responsive patients, as measured by DFA, became more like that in healthy controls, while gene expression in nonresponsive patients became less like that in healthy controls. Moreover, the genes identified by DFA to be predictive of therapeutic response were, for the most part, known regulators and effectors of the immune system. Taken together, these data suggest that successful therapy was able to reset immune response homeostasis to a significant extent in this cohort.

Materials and methods

Patients, patient selection, preparation of clinical specimens

We studied nine children newly diagnosed with polyarticular JRA. Diagnosis was based on accepted and validated criteria endorsed by the American College of Rheumatology [10]. Children were excluded if they had been treated with corticosteroids or methotrexate, or if they had received therapeutic doses of nonsteroidal anti-inflammatory drugs for more than 3 weeks before the study. Patients with active disease ranged in age from 4 to 15 years and presented with proliferative synovitis of multiple joints and erythrocyte sedimentation rates ranging from 35 to 100 mm/hour. Control subjects (n = 12) were laboratory volunteers under 25 years of age. Leukocyte buffy coat preparations were made from peripheral blood and total RNA extracted with Trizol reagent (Invitrogen, Carlsbad, CA, USA). Fluorescent labeling of cDNA was undertaken using the Micromax TSA-labeling kit (PerkinElmer Life Sciences, Boston, MA, USA). Labeled cDNAs were hybridized with PerkinElmer Micromax human cDNA microarray containing 2,382 human genes, and arrays were scanned using an Affymetrix 428 Array Scanner (Affymetrix, Durham, NC, USA).

Five of these nine patients were followed up longitudinally (for 6–12 months) from the onset of therapy as they either responded or failed to respond to therapy. In this portion of the study, disease severity was scored for the degree of synovitis using a linear scoring system used previously in our laboratory [11]. This system is based on criteria used in clinical trials in JRA [12]. For purposes of comparison and analysis, untreated children were categorized as having active disease. Children treated for more than 6 weeks who had a ≥ 30% reduction in their disease severity score were categorized as having had a partial response to therapy, while children with < 30% reduction in their severity scores were categorized as having acute, persistent disease. Children were categorized as being fully responsive to therapy if they showed synovial thickening in ≤ 3 joints, without warmth or tenderness in those joints and with no more than 30 minutes' morning stiffness per day. These criteria for full responsiveness have been validated in previous studies we have published examining markers of inflammation in JRA [13, 14]. The patients' characteristics are summarized in Table 1.

Table 1 Data for patients with polyarticular juvenile rheumatoid arthritis

Serum cytokine levels

Serum IFN-γ levels were measured using the BioPlex system, a biometric sandwich ELISA assay from BioRad Inc (Hercules, CA, USA) in accordance with the manufacturer's instructions. Serum from four patients during periods of attack and before treatment (denoted 'patients with active disease') and from 12 healthy control subjects was collected, stored at -80°C, and assayed in duplicate.

Normalization of array data

Normalization to correct for technical variation among individual microarray hybridizations was conducted using a two-step procedure described in detail elsewhere [15]. In brief, the procedure is based on the fact that spot intensities from genes not expressed by the samples of interest constitute noise and are therefore normally distributed. The method models the signals from nonexpressed genes to a normal distribution with a mean of 0 and standard deviation (SD) of 1, using an iterative nonlinear curve-fitting procedure.

A second normalization step is then performed using the genes significantly expressed above background (> 3 SD above background). Gene expression values are log-transformed, with negative values replaced by the lowest positive logarithmic value obtained. Expression profiles of genes statistically significantly expressed above background are then adjusted to each other using a robust regression analysis. This analysis is based on the observation that the expression levels of the majority of genes do not change in compared samples, and that expression values are normally distributed around a regression line with a small proportion of differentially expressed 'outliers'. The outliers' contribution in the regression analysis is down-weighted in an iterative manner until the residuals are normally distributed as measured by deviations from the regression line calculated against the averaged profile. Expression profiles of both control and experimental groups are then scaled to the averaged profile of the control group.

The two main sources of heterogeneity in gene expression variations are the 'additive component', prominent at low expression levels, and the 'multiplicative component', prominent at high expression levels [16]. The intensity measurement y _i,j for gene i ∈ I = {i₁,..., i _n } in sample j ∈ J = {j₁,..., j _m } is modeled by the equation y _i,j = a _i,j + m _i,j × e^h+ e _i,j where a is the normal background (not dependent on gene expression), m is the expression level in arbitrary units, e is first error term (additive) – which represents the standard deviation of background – and h is the second error term, which represents the proportional error (the multiplicative component) [17, 18]. The first error term is excluded from analysis by eliminating expression values at or below background levels. The second error term is transformed from multiplicative (and therefore expression-dependent, rising with expression level [18]), into additive (expression-independent) by log-transformation of data [16] using the equation log(y) = log (m) + h, where h is the residual for log-transformed data. The independence of h from individual gene expressions is confirmed with the Kolmogorov–Smirnov normality test in our experiments. We determine h for each sample as a deviation of the gene expression ordinates from a regression line calculated against of the averaged profile for gene expressions in all samples of the control group. The majority of these deviations follow a normal distribution. Genes of the control groups whose deviations belong to this distribution are expressed at similar levels among groups; this group is therefore denoted the 'reference group'. Variations in expression among samples of the genes within this group are due principally to technical variability and normal biologic variation. The parameters of variation defined by the reference group are used to identify differentially expressed genes and hypervariable genes whose expression levels vary in a statistically significant manner from the reference group (Fig. 1). A standard F-test is used to determine if a given gene's expression is variable with respect to the reference group using Matlab software (Mathworks, Natick, MA, USA).

Figure 1

Graphical representation of hypervariable (HV) gene analysis in patients with juvenile rheumatoid arthritis (JRA) (n = 9) and a reference group (n = 12). A reference group of genes from the control group whose expression levels do not vary significantly on a population basis was identified as described in Materials and methods. Expression levels in this reference group, denoted the averaged profile, have a normal distribution. This group is represented by black lines on a plot of residuals (values representing expression level variance in the control population) vs average gene expression levels (log₁₀-transformed). Red lines represent genes whose variation in expression in healthy controls or untreated patients with acute disease was significantly greater than that of the reference group. These genes are defined as hypervariable (HV) genes.

Identification of genes differentially expressed in patients vs control group

These analyses are performed using standard statistical analysis methods in Matlab software and include:

• 1. Selection of statistically different levels of expression using the Student's t-test with the commonly accepted significance threshold of P < 0.05. Because of the large number of genes present on microarrays, a significant proportion of genes identified as differentially expressed in this manner will be false positive determinations at this threshold level.

• 2. An associative t-test, in which the replicated residuals for each gene in the experimental group are compared with the entire set of residuals from the reference group (defined above). The hypothesis that gene expression in the experimental group, presented as replicated residuals (deviations from averaged control-group profile), is distributed similarly to the several thousand members of the normally distributed set of residuals for gene expressions in the reference group is tested. The significance threshold is corrected to 1/(number of genes) to make it improbable that false positives arise. Only genes with P values below the threshold of both the Student's t-test and the associative t-test are then presented in tables as differentially expressed genes. Relative ratios of expression for genes that are differentially expressed above background in both groups are calculated.

• 3. Genes expressed distinctively above background in one group and not in another are defined as uniquely expressed genes.

Selection of hypervariable (HV) genes

To have an opportunity to evaluate inhomogeneity in gene expression variability, it is necessary to normalize this variability to make it independent of the level of gene expression. The two main sources of heterogeneity in gene expression variations – additive and multiplicative components – are excluded in our analysis by eliminating expression values at or below background levels and by log-transformation of the data. Expression deviations η are determined for each sample as a deviation of the gene expression ordinates from regression line calculated against the averaged profile for gene expressions in all samples of the control group. The majority of these deviations follow a normal distribution. The SD of this distribution is used for identification of hypervariable genes whose expression levels vary in a statistically significant manner from the reference group of stable genes as determined using an F-test (Fig. 1).

Discriminant function analysis (DFA)

DFA was used for selection of the set of genes that maximally discriminate among the groups studied. A forward stepwise DFA was performed in accordance with the manufacturer's instructions, using the statistical software package Statistica (StatSoft, Tulsa, OK, USA). In this analysis, the model for discrimination is built in a stepwise manner. Specifically, at each step all variables are reviewed to determine which will maximally discriminate among groups. This variable is then included in a discriminative function, denoted a root, which is an equation consisting of a linear combination of gene expression changes used for the prediction of group membership. An F test is used to determine the statistical significance of the discriminatory power of the selected genes. The stepwise procedure is 'guided' by a standard threshold for the F test (established by the analytical package). In general, variables will continue to be included in the model, as long as the respective F values for those variables are larger than this standard threshold. The 170 genes expressed statistically significant from background in all five groups of samples (expression levels > 3 SD over background as defined above) were used for this analysis.

The discriminant potential of the final equations can be observed in a simple multidimensional plot of the values of the roots obtained for each group. This provides a graphical representation of the similarity among the various groups. The discriminative power of each gene can also be characterized by the partial Wilks λ coefficient. This value is equal to the ratio of within-group differences in expression to within- and between-group differences in expression. Its value ranges from 1.0 (no discriminatory power) to 0.0 (perfect discriminatory power).

Biochemical function and pathway analysis

The genes in the data tables presented herein are functionally annotated. Gene functions were obtained from the Swiss-Prot Protein knowledge base (when available). This database was created and is maintained by the Swiss Institute of Bioinformatics (Biozentrum – Basel University, Basel, Switzerland). Additionally, the software package Pathway Assist (Strategene, La Jolla, CA, USA) was used to identify functional interrelationships among the genes defined as JRA-related in the analyses described above. This software uses the KEGG, DIP, and BIND databases and natural language scans of Medline to define functionally related genes. These functional relationships were then graphically represented by the software as a network.

All original programs were written using MathLab and Statistica statistical software and are available on request from http://igor-dozmorov@omrf.ouhsc.edu.

Results

Differential gene expression analysis of active disease

Statistical analysis of the difference of gene expression in samples from 9 patients and 12 healthy controls gave the following results. 1716 genes of the total number of 2382 genes in the microarray were expressed distinctively from background (P < 0.05) in both groups. Of these, 78 were statistically differentially expressed in either patients or controls. These genes passed the Student's t-test at the threshold of 0.05 and the associative t-test at the threshold of 0.0005, a stringency that results in the selection of less than one expected false positive and less than one expected false negative determination. This analysis classifies differentially expressed genes into four groups:

• 1. genes expressed at higher levels on average in untreated patients with active disease, relative to healthy controls (34 identified, Table 2A);

Table 2 Differentially expressed genes in patients with polyarticular rheumatoid polyarthritis (n = 9) and healthy controls (n = 12)

• 2. genes expressed at lower levels in treated patients with active disease, relative to healthy controls (15 identified, Table 2B).

• 3. genes whose expression was detected above background only in untreated patients with active disease (18 identified, Table 2C); and

• 4. genes whose expression was detected above background only in healthy controls (2 identified, Table 2D).

Differential gene expression analysis is a common means of identifying the genes involved in a given pathophysiology. Our analysis identified key regulators of innate immunity and inflammation including the proinflammatory mediators formyl peptide receptor 1, ICAM-1 (intercellular adhesion molecule-1), thymosin β 4, and PLA-2 (phospholipase A₂), which were up-regulated in patients, and the anti-inflammatory mediator TNF receptor 1 (TNF-R1), which was down-regulated in patients. Genes regulating the adaptive immune response were also identified, including those for β2 microglobulin, MHC class I, GTP-binding protein-HSR1 (a polymorphic microsatellite marker in the human MHC class I region), and Semaphorin/CD100 (a B-cell and dendritic-cell surface receptor that modulates cellular activation), which were all up-regulated in patients, and the gene for transcription factor 8 (a repressor of IL-2 expression), which was down-regulated in patients. These data highlight the importance of these genes in regulating the immune and inflammatory response in JRA.

Interestingly, several of the immunoregulatory genes that were up-regulated in patients are known to be induced by interferon γ (IFN-γ), including those for thymosin β4, MHC class I, and ICAM-1, suggesting that this cytokine is increased in patients. To test this hypothesis, serum IFN-γ levels were assessed by ELISA in 4 patients with active disease and in a group of 12 healthy controls. Patient serum IFN-γ levels were significantly higher than in healthy controls (P < 0.00067). Values ranged from 60 to 1,626 pg/ml in patients and from < 1.4 (the level of sensitivity of the assay) to 9.6 pg/ml in healthy controls (Fig. 2), implicating IFN-γ in the pathophysiology of polyarticlular JRA.

Figure 2

Serum IFN-γ levels in untreated patients with active juvenile rheumatoid arthritis (JRA) and healthy controls (HC). A scatter plot of serum IFN-γ concentrations in 4 patients with active disease (AD) and 13 HC is shown. The values for 11 HC that were < 1.4 pg/ml (the limit of detection of the assay) are represented by triangular symbols that appear as the lowest value in the distribution. Average values in a given population are represented as a horizontal line. Concentrations are shown in pg/ml on a log scale.

To more fully disclose the pathways relevant to JRA pathogenesis, the genes identified as differentially expressed in patients were grouped according to function using recently developed commercial software (Pathway Assist, Ariadne Genomics, Rockville, MD, USA). A subset of functionally interrelated genes was identified and this network of genes graphically represented (Fig. 3). This analysis highlighted the importance of inflammatory and immune modulation, as well as such basic cellular processes relevant to leukocyte function as apoptosis, motility, and proliferation. The network of functionally related genes generated by this software allows the connections among these basic physiologic processes to be identified, demonstrating that the pathophysiologic response of these patients is highly coordinated.

Figure 3

Functional associations of genes selected as differentially expressed in patients with juvenile rheumatoid arthritis (JRA) and normal controls. Tabular data from differential expression analysis were analyzed using Pathway Assist software. The graphical output delineating a functionally related network of genes is shown. Genes that were expressed at higher levels in JRA patients are represented as red ovals. Genes expressed at higher levels in controls are represented as blue ovals. Major biologic processes related to these genes are represented as yellow rectangles. White ovals represent genes that are functionally related to the genes used for analysis. Upon addition of these genes, several functional connections among the genes being analyzed can be observed. Green squares signify that a defined regulatory relationship exits between genes. Blue squares signify that a putative regulatory relationship between genes has been identified but not biochemically defined. +, positive regulation; -, negative regulation.

Higher variability of genes in active disease

A novel analytical method was applied to the microarray data: identification of genes whose expression is relatively unchanging in the control population and becomes HV in JRA patients with active disease. The logical basis of this approach was based on the hypothesis that the loss of homeostasis characteristic of active autoimmune disease can be used to identify genes whose expression regulates the processes involved. For example, temperature is tightly regulated in healthy controls and is relatively stable on a population level. In patients with active polyarticular JRA, low-grade fever is relatively common and temperature levels vary on a population basis to a greater degree than in healthy controls. Therefore, the genes that code for regulators of pathophysiologic processes such as temperature control, or, by analogy, inflammatory response, may likewise be expected to vary on a population level in patients more than in healthy controls.

In this analysis, 444 genes were identified as HV genes in untreated patients with active disease and stable or expressed below background in healthy controls (see Additional file 1).

Among the 122 genes identified as HV genes in both groups, 27 had a statistically significant higher level of variation in untreated patients with active disease (Table 3). Many of the genes identified as increasing in variability in these patients have a direct role in inflammation and immune regulation and are known to be involved in inflammatory arthritis. These genes provide a more concise picture of the molecular pathophysiology of JRA than is obtained in a traditional analysis of differentially expressed genes and include: IL-8, MHC class I, regulators of TNF-α (e.g. TGFβ1-induced anti-apoptotic factor 1) and granulocyte/macrophage-colony-stimulating factor (GM-CSF) (e.g. cold shock protein A), and human cartilage protein gp-39 (a major secretory product of articular chondrocytes and synovial cells). It is of note that none of these 27 genes were identified by differential expression analysis.

Table 3 Genes that distinguish patients with active juvenile rheumatoid arthritis by hypervariable gene analysis

Pathway analysis software was used to reveal the principal biologic processes revealed by these data. Interestingly, while the genes identified by HV analysis were distinct from those identified by differential expression analysis, the physiologic processes identified, such as inflammation and immune modulation, apoptosis, and cell motility, were similar (Fig. 4).

Figure 4

Functional associations of hypervariable (HV) genes in patients with juvenile rheumatoid arthritis (JRA). Tabular data from the HV gene analysis was analyzed using Pathway Assist software. The graphical output delineating a functionally related network of genes is shown. Genes that displayed increased variation in expression in JRA patients are represented as red ovals. Major biologic processes related to these genes are represented as yellow rectangles. JRA is represented as a yellow square to delineate genes directly associated with pathology. Intracellular objects upon which a given set of genes acts are represented as yellow ovals. White ovals represent genes that are functionally related to the genes used for analysis. White hexagons represent organelles functionally associated with the genes analyzed. Orange hexagons represent classes of small molecules associated with the genes analyzed. Green squares signify that a defined regulatory relationship exits between genes. Blue squares signify that a putative regulatory relationship between genes has been identified but not biochemically defined. +, positive regulation; -, negative regulation.

Discriminant function analysis (DFA)

In the above analyses, genes with behavior that varies between patients with active disease and control individuals were identified. DFA is distinct from the above analyses in that it identifies a set of genes whose expression levels, as a group, vary among populations. In this analysis, genes with the most significant power to discriminate among groups when used as variables in a linear equation, denoted a root, were identified. The groups of genes identified by DFA are statistically interrelated and may therefore be functionally interrelated. For this analysis, the following groups were used: nine untreated patients with acute disease; five of these nine patients were followed up prospectively during treatment, with partially responsive, fully responsive, and non-responsive patients defined as independent groups; and six healthy controls.

Interestingly, literature searches and pathway analysis revealed that nearly all of the 19 genes identified by this analysis are either directly or indirectly associated with immune modulation and autoimmune disease (Table 4; Fig. 5). These genes include: IL-1 receptor antagonist, an anti-inflammatory cytokine that has been recently approved by the FDA for use as a biologic therapy in inflammatory arthritis [19] and which also plays a significant role in the pathogenesis of polyarticular JRA [20]; TGFβ, another potent anti-inflammatory cytokine, which is involved in many biologic processes including immune homeostasis, regulation of apoptosis, and self-tolerance [21]; IL-8, which has been shown to be elevated in JRA patients and plays a key role in joint inflammation by recruiting neutrophils to synovial tissue and fluid [22, 23]; ferritin, highlighting the significance of anemia of chronic disease associated with JRA [24]; transcription factor octamer-binding protein (Oct-1), which is expressed in synovial cells in the majority of RA patients and may modulate synovial outgrowth [25]; CD63, a leukocyte adhesion molecule and marker of dendritic cell maturation and neutrophil activation [26–28]; and GLUT/SLC2A, a glucose transporter protein expressed by chondrocytes that is regulated by TNF-α, IL-1β, IGF-1, and a key modulator of skeletal development, chondrogenesis, and cartilage degradation in osteoarthritis [29].

Table 4 Genes that distinguish children with active juvenile rheumatoid arthritis in discriminant function analysis

Figure 5

Functional associations of genes identified by discriminant function analysis (DFA) in patients with juvenile rheumatoid arthritis (JRA) who were undergoing treatment. Tabular data from the this analysis was analyzed using Pathway Assist software. The graphical output delineating a functionally related network of genes is shown. Genes that were associated with disease or treatment response are represented as red ovals. Major biologic processes related to these genes are represented as yellow rectangles. White ovals represent genes that are functionally related to the genes used for analysis. Orange hexagons represent classes of small molecules associated with the genes analyzed, and orange circles represent specific small molecules associated with the genes analyzed. Green squares signify that a defined regulatory relationship exits between genes. Blue squares signify that a putative regulatory relationship between genes has been identified but not biochemically defined. +, positive regulation; -, negative regulation.

Fourteen of the 19 genes identified by DFA were uniquely identified by this analysis; 2 of the 19, Ribosomal protein L37 and Ferritin light chain, were also identified by differential expression analysis, and 2 of the 19, IL-8 and Oct-1, were also identified by HV gene analysis, demonstrating that these three analyses are complementary, yielding predominantly nonoverlapping results (Fig. 6).

Figure 6

An overview of the results from the three analytical methods. The numbers of genes that were identified by differential expression analysis, hypervariable (HV) gene analysis, and discriminant function analysis (DFA) are represented in a Venn diagram. The numbers of genes identified uniquely in a given analysis as relevant to patients with juvenile rheumatoid arthritis (JRA) are shown in nonoverlapping regions. The numbers of genes identified in more than one analysis are shown in overlapping regions. HC, genes expressed at higher levels in healthy controls; JRA, genes expressed at higher levels in patients.

The values of the roots obtained by DFA analysis can be used to graphically represent the differences of the gene expression values obtained for the groups analyzed. Values obtained for individuals in a given group tended to cluster (Fig. 7), suggesting that the phenotypic changes occurring in a given group during treatment are common to all individuals of that group. Moreover, this graphical representation demonstrates that as patients respond to therapy, they tend to have expression profiles that are less like those of untreated patients with active disease and more like those of healthy controls (Fig. 7). In fact, the distance a given group of treated patients moved towards the normal controls was proportional to their level of response, with fully responsive patients moving closer to healthy controls than partially responsive patients (Fig. 7). In contrast, the position on the graph of the values obtained from a patient who was nonresponsive to standard therapy, whose disease severity actually increased after initiation of therapy (values for four separate samples taken between 2 and 8 weeks after therapy are represented in Fig. 7), changed in a manner that is distinct from those responding to therapy (responders were positioned away from the active disease group towards the healthy controls in a clockwise manner, and the four values obtained from the nonresponsive patient were positioned away from the active disease group in a counterclockwise manner) (Fig. 7). This method of pharmacogenomic analysis therefore provides a means of developing a clinical assay that may predict patients' response to therapy early in the course of polyarticular JRA treatment.

Figure 7

A graphical representation of the discriminatory potential of discriminant function analysis (DFA). DFA was used to identify, in a cohort of treated patients with juvenile rheumatoid arthritis (JRA) and healthy controls, a subset of genes whose expression values can be linearly combined in an equation, denoted a root, whose overall value is distinct for a given characterized group. The expression values of the individuals in the cohort were plotted in three dimensions to visually represent the relative differences in gene expression among the distinct populations. Gene expression values were plotted on this graph for nine untreated patients with active disease. Five of these nine patients were followed up prospectively during treatment. Four patients responded to treatment and one patient was nonresponsive. The values obtained for the four responsive patients at the time of partial response (after 2–4 weeks of treatment) and full response (6–8 weeks) are shown, as are four independent values obtained from a nonresponsive patient (taken during an 8-week interval). Values for healthy controls are also represented. Values obtained for individuals from each of these groups tended to cluster. Response to therapy was reflected in the spatial relationships among groups.

Discussion

We have used three distinct statistical methods for analyzing microarray data from children with polyarticular JRA. These methods include a standard analysis of differential gene expression, a novel means of assessing genes whose behavior dynamics are modulated in populations (denoted hypervariable gene analysis), and DFA to identify the genes and molecular pathways involved in the pathogenesis of polyarticular JRA. Each method identified both well established and novel mechanisms of disease pathogenesis, and because the biologic basis of each of the methods is unique, the methods are complementary, with each highlighting a different aspect of the disease.

Analysis of differentially expressed genes demonstrated that in patients with active disease, principal modulators of both adaptive and innate immunity were up-regulated, consistent with the hyperactivation of the immune and inflammatory responses characteristic of this disease. Disease-specific up-regulation of a group of interferon-induced genes, including thymosin β4, MHC class I, and ICAM-1, was observed. Moreover, serum IFN-γ levels were dramatically increased in patients with active disease relative to healthy controls. These data provide the first broad-based molecular evidence for the existing hypothesis that JRA is a Th-1-biased disorder and demonstrate the pathologic immunomodulatory role of IFN-γ in these patients [30]. These data also suggest that therapy directed specifically at reducing IFN-γ levels may prove efficacious in treating JRA.

The fact that several genes involved in oxidative stress are among the genes most significantly up-regulated during active disease suggests that this process also plays a significant role in the pathophysiology of polyarticular JRA, as suggested by previous studies of oxidation products and free radical scavengers in patients [31]. These findings are consistent with our previous work demonstrating the importance of immune complexes (potent triggers to neutrophil superoxide release) [32, 33] in the pathophysiology of polyarticular JRA [14] and provide a logical basis to begin investigations of novel treatment modalities targeting these pathways.

This work represents the first application of hypervariable gene analysis to the study of human disease of which we are aware. This statistical measure of gene variability was designed to identify genes that have lost their normal regulation in a manner that mimics the loss of homeostasis characteristic of autoimmune disease. Regulation of genes that are involved in such processes as inflammatory cell and immune cell activity, which will vary in groups of affected individuals more than in healthy controls, would therefore be expected to become increasingly chaotic at the population level in groups of affected individuals. This hypothesis is validated by the fact that a significant number of genes identified by this analysis are involved in immune and inflammatory regulation and have been implicated as mediators of autoimmune disease. Importantly, the overall picture of disease is more accurately depicted when the genes identified by HV analysis are added to those identified as differentially expressed. For example, neutrophils play a principal role in disease pathogenesis [34–36, 1]. However, genes that regulate neutrophil function, such as those for IL-8, 47-kDa autosomal granulomatous disease protein, DNA-binding protein A variant (which regulates GM-CSF production), and cathepsin D, a potent neutrophil collagenase found to be highly expressed in the synovium of JRA, RA, and osteoarthritis patients [37–39], were identified by HV gene analysis and not by identification of genes that are up- or down-regulated in patients or controls. Moreover, the genes for YKL-40 (or cartilage protein-39), a cartilage-derived autoantigen that is up-regulated in RA and thought to be a direct target of autoimmune attack in that disease [40–43], EDG-1, a pro-angiogenic protein essential for vascular maturation [44], and Oct-1, a transcription factor expressed in RA synovial cells and which regulates gene transcription [21], were more highly variable in JRA than in controls, indicating that these proteins may play a major role in joint-specific pathology, including erosions, in polyarticular JRA, although they were not detected as significantly differentially regulated. On the basis of these findings, it is interesting to speculate that polyarticular JRA may be kept in check by the synchrony of a distinct subset of disease-specific genes whose dysregulation can predispose a patient to a flare in disease activity. These results suggest that hypervariable gene analysis will be a useful adjunct to traditional analyses of differential gene expression.

Among the most important aspects of the results reported here was the fact that DFA of microarray data may predict therapeutic response to specific therapies. The predictive potential of this analysis was based on the finding that the dynamics of gene expression in patients with active disease were modulated towards levels characteristic of healthy controls in proportion to a given individual's response to therapy. These data suggest that successful immunosuppressive therapy re-establishes some level of normality in these patients. This hypothesis is consistent with the fact that, in effectively treated patients, the disease is suppressed for some time, as if patients have re-established normal, or near-normal, homeostasis. It is also consistent with the clinical observation that disease activity can flare after periods of relative quiescence, as if this re-established homeostasis requires some trigger to be disrupted.

It is useful to note that there was uniform regression towards a normal profile, whether the response was achieved using nonsteroidal anti-inflammatory drugs alone or required more potent agents such as methotrexate, suggesting that the actions of these agents on the immune and inflammatory response in a successfully treated patient are similar. Given that treatment response can be predicted by DFA relatively early in the treatment course, these results demonstrate the potential for using pharmacogenomic analyses to identify the most efficacious treatment modality for a given patient.

We do not assert that this analysis is the 'final say' for predicting therapeutic response in polyarticular JRA. These preliminary investigations will need to be followed up longitudinally with a larger group of patients, and individual response profiles will need to be developed for specific agents (e.g. nonsteroidal anti-inflammatory drugs, methotrexate, etc). Despite these limits, these data conclusively demonstrate the power of analyzing gene expression behavior using DFA in analysis of complex diseases such as polyarticular JRA.

Conclusion

We have demonstrated that the relevance of microarray data can be maximized by applying bioinformatics tools that specifically address the nature of the data obtained. Standard differential gene expression analysis was used to illuminate specific pathways relevant to disease pathophysiology. Moreover, a novel statistical method was created specifically to exploit the fact that autoimmune disease is characterized by loss of homeostasis, and therefore that expression of immune and inflammatory genes that regulate and affect these pathophysiologic processes will vary more in patients than in healthy controls. Lastly, DFA was used to analyze prospective data from patients treated with conventional therapy, as this is the most powerful analytical tool for obtaining data from a complex cohort. These later methods proved complementary to the conventional analysis of microarray data and highlighted previously characterized aspects and identified novel aspects of disease pathophysiology. Importantly, the results of the DFA could be used to develop a clinical assay that may predict therapeutic response in patients early in the treatment course, demonstrating that molecular outcomes, when measured on a comprehensive scale, can be used as prognostically relevant biologic markers of disease activity.