A subset of eight healthy volunteers and eight patients with an established diagnosis of SoJIA was used as a training set to identify the classifier genes (Fig. 2 A, left). The 50 most significantly differentially expressed transcripts (Table I) were then evaluated within the same set of patients in a leave-one-out cross-validation scheme. Using this strategy, 100% of the healthy and 88% of the SoJIA samples were classified accurately (seven were predicted accurately, and the class of the remaining sample was indeterminate). The ability of these transcripts to classify a test set composed of eight new healthy and eight independent SoJIA patients was tested. Using this approach, 100% of the patients and controls were accurately classified (Fig. 2 A, right).

View larger version (66K): [in this window] [in a new window] [Download PPT slide]   Figure 2. Class prediction. (A) (Left) Eight healthy and eight SoJIA samples obtained from our initial study group were used as a training set to generate a list of classifier genes displaying the best ability to discriminate patients from healthy controls. In our training set, 100% of healthy and 88% of SoJIA patients were classified accurately. (Right) Those classifier genes were then tested on a test set (eight Healthy and eight SoJIA), and 100% of patients were classified accurately. Expression values were normalized per gene to the healthy group. Samples and genes were arranged by hierarchical clustering. Transformed expression levels are indicated by color scale, with red representing relative high expression and blue indicating relative low expression. The list of the genes from this figure is shown in Table I. (B) Specificity of the SoJIA signature. The 50 best classifier genes from A were used to classify a test set of 39 healthy controls, 16 SoJIA, 31 S. aureus, 15 S. pneumoniae, 30 E. coli, 18 influenza A, 38 SLE patients, and 6 PAPA syndrome patients. The number of samples within each disease group predicted as SoJIA is represented on top of the figure. Genes were arranged by hierarchical clustering. Transformed expression levels are indicated by color scale, with red representing relative high expression and blue relative low expression.

Lack of specificity of the SoJIA signature
Children with SoJIA present with severe systemic symptoms (fever and rash) that usually precede the development of arthritis from weeks to years. Thus, the main differential diagnosis at presentation is an infectious disease. The 50 best SoJIA classifier transcripts described above were next tested for their ability to discriminate SoJIA patients from patients with infections (31 patients with Staphylococcus aureus, 15 patients with Streptococcus pneumoniae, 30 patients with Escherichia coli, and 18 patients with influenza A). As controls for noninfectious disease and steroid treatment, we included a group of 38 pediatric systemic lupus erythematosus (SLE) patients and 6 patients with PAPA syndrome, an IL-1–mediated autoinflammatory disease. These 50 genes were also dysregulated in patients with inflammatory conditions, as 45% of S. aureus, 47% of S. pneumoniae, 36% of E. coli, 5% of Influenza A, 29% of SLE, and 33% of PAPA syndrome patients were incorrectly classified as SoJIA (Fig. 2 B). Some of the genes included in the 50 best classifiers were mildly dysregulated in patients with established SoJIA and persistent arthritis, but not systemic symptoms (SoJIA 2) and in asymptomatic patients (SoJIA 3) (Fig. S1A, available at http://www.jem.org/cgi/content/full/jem.20070070/DC1). These patients, however, were correctly discriminated from the patients with systemic disease (Fig. S1 B).

Thus, the comparison of transcripts from SoJIA PBMCs and healthy controls is insufficient to yield SoJIA-specific signatures. Furthermore, the blood transcription patterns of SoJIA patients during the systemic phase of the disease are closer to those of patients with infections than to those of SoJIA patients in later (arthritic) stages of the disease.

Identification of a specific SoJIA signature
To identify a diagnostic SoJIA signature, the transcription profiles of SoJIA patients with systemic symptoms were directly compared with all the other infectious/inflammatory conditions. However, a large proportion of the predictor genes identified using this approach was found expressed similarly in SoJIA patients and healthy controls (unpublished data). Because it is difficult to control potentially confounding factors such as age or sex when comparing more than two groups of patients, we adopted a different strategy (see Fig. S2, available at http://www.jem.org/cgi/content/full/jem.20070070/DC1, and reference 28 for details). First, we performed statistical comparisons between each group of patients (16 SoJIA, 16 S. aureus, 10 S. pneumoniae, 10 E. coli, 10 influenza A, and 16 SLE) and their respective control groups composed of age- and gender-matched healthy donors. The p-values obtained from each comparison were then subjected to selection criteria that permitted the identification of genes significantly changed in SoJIA patients versus their control group, and not in any of the other disease versus control groups.

A nonstringent statistical group comparison (nonparametric Mann-Whitney rank test; P < 0.01) between 16 SoJIA and 10 healthy control samples yielded 4,311 differentially expressed transcripts (Fig. 3 A). 88 of these transcripts were expressed with a p-value of >0.5 in all other diseases compared with their corresponding healthy control groups (Fig. 3 B and Table II). More than 50% of these genes (47/88) encode proteins with unknown function. Among those encoding known proteins, several are involved in microtubule/cytoskeleton reorganization, ubiquitination, cellular transport, apoptosis, metabolism, transcription, protein biosynthesis, and posttranslational protein modification (Table II). The gene tree corresponding to these differentially expressed transcripts in individual patients and controls is displayed in Fig. 3 C. Only 1 of these 88 best classifiers (AK2) overlapped with the 50 genes that best discriminated SoJIA patients from healthy controls (Tables I and II), confirming the lack of specificity of our initial approach.

View larger version (55K): [in this window] [in a new window] [Download PPT slide]   Figure 3. SoJIA-specific signature. (A) Genes expressed at statistically different levels in the SoJIA patient group compared with healthy volunteers (P < 0.01; Wilcoxon-Mann-Whitney test) were selected (4,311 probe sets). p-values were similarly obtained from patients suffering from S. aureus, S. pneumoniae, E. coli, influenza A, and SLE. Each of these cohorts was compared with the corresponding control group. p-values are represented according to color scale: turquoise, low p-value; pink, high p-value. (B) 88/4,311 transcripts were found expressed at statistically different levels in the SoJIA patient group compared with healthy controls (P < 0.01; Wilcoxon-Mann-Whitney test), but not in all of the other groups compared with their healthy controls (P > 0.5; Wilcoxon-Mann-Whitney test). (C) The 88 genes from B were hierarchically clustered in the 107 samples used in the meta-analysis. Expression values of those 88 genes were normalized per gene to the healthy group. (D) The 12 most significant genes (P < 0.0001 in SoJIA group) were used as predictors for an independent test set of 10 healthy patients, 9 patients with febrile syndromes to rule out the diagnosis of SoJIA (2 of these patients were later on found to suffer from other diseases), 15 S. aureus, 5 S. pneumoniae, 20 E. coli, 8 influenza A, 22 SLE, and 6 PAPA syndrome patients. Expression values of those 12 genes were normalized per gene to the healthy group. Genes were arranged by hierarchical clustering. Transformed expression levels are indicated by color scale, with red representing relative high expression and blue indicating relative low expression. The list of the 88 and 12 genes shown in B–D is displayed in Table II. (E) The 12 gene signature displayed in D is shown in 16 healthy controls, 23 samples from 19 patients with established diagnosis of SoJIA while presenting systemic symptoms, and 11 patients with established diagnosis of SoJIA in whom the systemic symptoms had subsided but arthritis persisted. Only one patient with confirmed clinical diagnosis of SoJIA did not display the signature while having systemic symptoms (*).

12 genes can be used to diagnose SoJIA
A more stringent analysis (P < 0.0001) permitted us to identify 12/88 genes highly differentially expressed in SoJIA compared with healthy controls, but not differentially expressed (P > 0.5) in all other disease groups compared with their respective controls. These 12 genes are included in Table II.

The ability of this set of 12 genes to identify patients with SoJIA was then validated using independent groups of patients and controls. A training set composed of 10 healthy and 16 SoJIA samples was used to predict sample class for an independent test set composed of a random group of 10 healthy patients, 9 patients with fever of more than 10 d in duration and negative bacterial cultures (suspected to have SoJIA), and 15 S. aureus, 5 S. pneumoniae, 20 E. coli, 8 influenza A, 22 SLE, and 6 PAPA syndrome patients (Fig. 3 D). This analysis allowed us to accurately classify six out of seven patients fulfilling SoJIA clinical diagnostic criteria. The only SoJIA patient not accurately classified (Sys 99 in Table S1) had fever, rash, and arthritis at the time of blood draw and eventually responded well to treatment with steroids and methotrexate. Two additional patients who were not classified as SoJIA (Sys 85 and Sys 91) were subsequently diagnosed with different diseases. Overall, 93% of S. aureus, 100% S. pneumoniae, 95% of E. coli, 91% of SLE, and 100% of influenza A and PAPA syndrome samples were correctly discriminated from SoJIA (Fig. 3 D). Thus, the "normalization" of each patient group to healthy control values and the comparison of significances rather than expression levels permitted us to extract a signature unique to the majority of SoJIA patients.

The expression of the 12 transcripts was next tested in the 11 SoJIA patients described above, in whom the systemic symptoms had subsided but continued to display arthritis. As shown in Fig. 3 E, 11/11 patients did not differentially express these genes compared with healthy controls, further indicating that this signature is specific to the systemic phase of the disease.

Whether the dysregulated expression of these 12 genes is related to the pathogenesis of SoJIA remains to be determined. Overall, seven of these genes encode uncharacterized proteins. Among those encoding proteins with known function, the most significantly up-regulated transcript (average, 8.2-fold), chloride intracellular channel 2 (CLIC-2), belongs to the ubiquitous glutathione transferase structural family. CLIC-2 is one of only a few cytosolic inhibitors of cardiac ryanodine receptor 2 channels and may suppress their activity during diastole and stress (29). Interestingly, CLIC-2 is the only transcript encoding a protein with potential link to IL-1 secretion, as chloride has been shown to play an important role in maintaining the P2X7 receptor (P2X7R) in a conformationally restrained state. This in turn limits the coupling of this receptor to signaling pathways that regulate caspase 1 and IL-1b signaling cascade (30).

To determine if a specific cell type is preferentially contributing to the overrepresentation of these transcripts, their expression was analyzed in B cells, T cells, monocytes, and neutrophils from healthy donors and SoJIA patients, as well as myeloid DCs (mDCs) and plasmacytoid DCs (pDCs) from healthy donors. Interestingly, 5/12 transcripts showed the highest expression levels in mDCs (not depicted). Within the cells that we analyzed, expression of CLIC-2 is restricted to this cell type (Fig. 4, A and B). However, whether the 12 transcript signature derives from a cell type not normally present in the blood (i.e., a bone marrow precursor) needs to be ruled out.

View larger version (19K): [in this window] [in a new window] [Download PPT slide]   Figure 4. Transcription of CLIC-2 in blood cell subsets. (A) Total RNA was extracted from the PBMCs of 16 healthy donors and 16 SoJIA patients; B cells, T cells, and monocytes from 3 healthy donors and 4 SoJIA patients, and neutrophils from 6 healthy donors and 7 SoJIA patients. Amplified cRNA was hybridized on Affymetrix HG-U133 chips. Raw intensity values from each chip were first pre-scaled to the 500 target intensity value in Affymetrix Microarray suite before being imported and analyzed in GeneSpring 6.1. (B) Total RNA was extracted from the PBMCs of 16 healthy donors and 16 SoJIA patients, mDCs from 9 healthy donors, and pDCs from 6 healthy donors. Amplified cRNA was hybridized to Affymetrix HG-U133 chips. Raw intensity values from each chip were pre-scaled to the 500 target intensity value in Affymetrix Microarray suite before being imported and analyzed in GeneSpring 6.1.

View larger version (24K): [in this window] [in a new window] [Download PPT slide]   Figure 5. Anakinra effect on the specific SoJIA signature. Transcription levels (raw values) corresponding to the genes that best differentiate SoJIA patients from healthy controls (from Fig. 1 and Table S2) were analyzed before and after the initiation of Anakinra in eight patients. The 10 genes with most significant p-values are displayed (p-values were calculated using a paired, two-tailed t test).

	DISCUSSION

Top ABSTRACT RESULTS DISCUSSION MATERIALS AND METHODS REFERENCES

 
SoJIA is the only form of JIA in which systemic symptoms precede the appearance of joint inflammation. Because systemic symptoms and current laboratory tests are nonspecific, a major remaining challenge is how to establish the prompt diagnosis of the disease during the systemic phase to avoid lengthy hospitalizations and initiate effective therapy. Here, we show that gene expression patterns in blood leukocytes may be used, together with careful clinical assessment, to discriminate SoJIA during the systemic phase of the disease from other febrile conditions.

Patients with SoJIA display a very striking pattern of leukocyte gene transcription when compared with healthy controls. These differences might reflect changes in blood cell composition, i.e., the presence of a cell population not normally found in the blood, or fluctuations in numbers of blood-specific cell populations rather than real transcriptional changes. Active SoJIA patients, for example, display increased platelet and leukocyte numbers compared with healthy controls. Erythroid precursors, which are not normally present in peripheral blood, are also found in these patients, and erythroid-specific transcripts are among the most significantly up-regulated in SoJIA blood.

To control for changes in cell composition and find a SoJIA-specific signature, we included in our analysis 138 samples from pediatric patients with systemic inflammation, including bacterial and viral infections, autoimmune (SLE), and autoinflammatory (PAPA) diseases. Some of these patients (i.e., bacterial infections and PAPA syndrome) display alterations in blood cell numbers similar to those of SoJIA patients. As age, gender, and time of day when the blood is drawn have been described to influence blood gene expression patterns (35), our samples were matched for age and gender with controls and most of them were obtained within similar time frames.

There was a significant degree of overlap between the signatures of SoJIA patients and all the other inflammatory disease groups included in this study, especially Gram (+) bacterial infections, further stressing the need for multi-cohort studies when searching for disease-specific biomarkers. The overlap could be due to dysregulated cytokine production and/or signaling cascades that may be shared by some of these diseases. IL-1, for example, is important in the cascade of defense mechanisms against many bacterial infections (36, 37). Increased IL-1b production has also been described in autoinflammatory diseases including PAPA syndrome. Indeed, PAPA patients display a mutation in the PSTPIP1 gene (22) that exerts a dominant-negative effect on the activity of pyrin and leads to increased IL-1b production by peripheral blood leukocytes (16). Interestingly, the herein described blood transcription patterns of SoJIA patients during the systemic phase of the disease are closer to those of patients with infections than to those of SoJIA patients who have resolved the systemic phase but continue displaying active arthritis. This could be explained if different pathogenic events, i.e., different cytokine cascades, were responsible for the different phases of the disease. Alternatively, the same cytokine alteration in a more localized environment (i.e., the joint) may not give rise to a blood cell signature.

To identify a specific blood signature that would permit the accurate differential diagnosis of SoJIA patients during the systemic phase of the disease, we designed an analysis of significance across multiple febrile inflammatory diseases and control groups. One of the advantages of this analysis is that it permits us to normalize each disease group to its own matched control group, therefore avoiding biological (i.e., age, gender) or technical (i.e., array runs) confounding factors. Using this approach, a SoJIA-specific signature composed of 88 genes was identified. This signature is very stable over time, as we could identify it in two samples from a patient obtained >2 yr apart (unpublished data). Using a more stringent analysis (P < 0.0001 in SoJIA and P > 0.5 in all other groups), 12 highly significant genes permitted to accurately classify the disease in 18/19 SoJIA patients. Furthermore, it allowed us to rule out SoJIA in two febrile patients who later developed arthritis but proved to suffer from different diseases. It also allowed us to discriminate systemic infections with Gram (+) and Gram (–) bacteria in 48/50 patients. Perhaps more interesting, six out of six patients with PAPA syndrome were not classified as SoJIA. How this signature will perform in discriminating SoJIA patients from other autoinflammatory diseases where IL-1 is also dysregulated is currently being investigated.

Most of the genes included within the SoJIA-specific signature encode proteins that remain uncharacterized. Of those that encode known proteins, we did not find any obvious components of the IL-1 pathway. As discussed above, this might be expected as IL-1 also plays an important role in the inflammatory response against some of the infectious diseases included in this study. IL-1–related genes would have therefore been counter-selected. CLIC-2, the most up-regulated of the 12 transcripts that compose the SoJIA-specific signature, might be involved in the regulation of IL-1 secretion. This gene encodes a chloride channel, and chloride levels control the conformation of P2X7R (30). This in turn limits the coupling of this receptor to signaling pathways that regulate caspase 1 and IL-1b signaling cascade. Interestingly, expression of CLIC-2 within PBMCs seems to be restricted to mDCs (38). IL-1 amplifies DC function, and IL-1 production is induced when monocytes are co-cultured with alloreactive T cells and autologous DCs through a cell contact–dependent mechanism (39). Studies are currently underway to determine whether these genes are up-regulated in SoJIA blood mDCs or in a different cell population not normally present in the blood of healthy and infected children.

Even though the clinical symptoms were controlled in the majority of SoJIA patients treated with an IL-1 receptor antagonist, expression of only a fraction (389/873) of the dysregulated transcripts returned to control levels in treated patients. Several explanations might be put forward to explain this observation. IL-1 could be downstream of a factor present in SoJIA serum that is not blocked by Anakinra. Indeed, the residual SoJIA signature might be a tool to identify such an IL-1–inducing factor. The IL-1 antagonist effect of Anakinra may be sufficient to silence the clinical symptoms, but not to erase the molecular signature. Finally, our limited patient sample before and after Anakinra treatment may not be enough to give statistical power to this analysis, and more patients may need to be studied to draw firm conclusions.

The small number of genes identified in this study as SoJIA-specific might help in the diagnosis of patients with febrile conditions included under the term "fever of unknown origin." They might also allow to promptly initiate specific therapy in SoJIA patients even before arthritis develops, thus avoiding the need for additional therapies and the development of long-term disabilities. The identification of genes whose expression is restored back to normal levels upon successful IL-1 blockade might also help identify predictors of response to therapy in longitudinal studies.

SoJIA patients are heterogeneous, however, regarding severity of symptoms and disease course (38, 40). Further studies will therefore be required to confirm the value of blood gene transcription profiling in establishing the diagnosis and predicting the response to IL-1 blockade in larger patient cohorts.

	MATERIALS AND METHODS

Top ABSTRACT RESULTS DISCUSSION MATERIALS AND METHODS REFERENCES

 
Patient information.
Blood samples were obtained from 21 patients with clinical diagnosis of SoJIA (4 patients were tested twice at an interval of 1–2 yr) during the systemic phase of the disease (median age, 6 yr; range, 2–17 yr). In two patients, subsequent clinical course ruled out this diagnosis. Samples were from 11 patients with SoJIA in the arthritis phase of the disease, 14 asymptomatic patients, and 8 patients with SoJIA treated with Anakinra (6 yr, 2–19 yr). These were also analyzed together with 94 patients hospitalized with active infections: 30 patients with E. coli infection (1 yr; 2 wk–16 yr), 31 patients with S. aureus infection (7 yr; 3 mo–18 yr), 15 patients with S. pneumoniae (3.1 yr; 2 mo–16 yr), 18 with influenza A infections (1.5 yr; 3 wk–16 yr), 38 patients with SLE (12 yr; 5–18 yr), and 6 patients with PAPA syndrome (32.6 yr; 11–68 yr). Patients were divided in training and test sets according to age and treatment (Table S1). Subjects were recruited at Texas Scottish Rite Hospital and Children's Medical Center of Dallas. Healthy controls included patients with hypermobility attending the rheumatology clinic and healthy children undergoing routine surgical procedures with no history of asthma, acute infections, or inflammatory conditions. Most of the samples were obtained between 9:00 a.m. and 12:00 p.m. The study was approved by the Institutional Review Boards of UT Southwestern Medical Center, Texas Scottish Rite Hospital, and Baylor Health Care System, and informed consent was obtained from all patients (legal representatives and patients over 10 yr of age). Bacterial and viral infections were confirmed by standard bacterial cultures, direct fluorescent antigen testing, and viral cultures. Patients with infections were recruited once a confirmed microbiologic diagnosis was established. Respiratory viral cultures were performed in 60 of 73 (82%) patients with bacterial infections. The clinical and demographic characteristics of these patients are summarized in Table S1 and have been reported elsewhere (41). PAPA syndrome patients were diagnosed based on clinical data and sequencing the PSTPIP1 gene.

RNA and microarray sample preparation from PBMC and blood cell subpopulations.
All blood samples were obtained in EDTA purple-top tubes (BD Vaccutainer). Fresh PBMCs were isolated via Ficoll gradient. Cells were lysed in RLT lysis buffer containing ß-mercaptoethanol (QIAGEN).

pDCs and mDCs were purified from a healthy donor's buffy coats. Ficoll-enriched PBMCs were depleted of lineage+ cells with CD3, CD14, CD19, CD16, CD56, and glycophorin A microbeads (Miltenyi Biotec). After staining with lineage cocktail-FITC, CD11c-allophycocyanin, and CD123-PE (BDBiosciences), and HLA-DR-QR (Sigma-Aldrich) mAbs, cells were sorted on a FACSVantage (Becton Dickinson) to at least 99% purity.

Monocytes, B cells, and T cells were positively selected from PBMCs using CD14 microbeads, CD19 microbeads, or CD3 microbeads (Miltenyi Biotec) and MS column (Miltenyi Biotec) according to the manufacturer's instructions to at least 95% purity. Neutrophils were isolated from venous blood of healthy volunteers by dextran (GE Healthcare) sedimentation of erythrocytes and density gradient centrifugation of leukocytes. The resulting cell populations contained <2% contaminating cells.

Total RNA was isolated using the RNeasy kit (QIAGEN) according to the manufacturer's instructions, and the RNA integrity was assessed by using an Agilent 2100 Bioanalyzer (Agilent). For PBMCs and neutrophils, 5 µg of total RNA was used to generate double-stranded cDNA containing the T7-dT (24) promoter sequence (Operon) as a template for in vitro transcription single-round amplification with biotin labels, using the Enzo R BioArrayTm HighYieldTM RNA Transcript Labeling kit (Affymetrix, Inc.). For the other cell subtypes, the same protocol was used starting with 50 ng of total RNA and performing two rounds of amplification. Biotinylated cRNA targets were purified using the Sample Cleanup Module (Affymetrix, Inc.) and hybridized to human U133A and B GeneChips (Affymetrix, Inc.) according to the manufacturer's standard protocols. Arrays were scanned using a laser confocal scanner (Agilent).

Microarray data analysis.
For each Affymetrix U133A and B GeneChip, raw intensity data were normalized to the mean intensity of all measurements on that array and scaled to a target intensity value of 500 (TGT) in Affymetrix Microarray Suite 5.0. Data were then further analyzed using GeneSpring software version 7.0. Data were normalized to a set of healthy controls (sex- and age-matched). An Affymetrix flag of "present" in at least 75% of samples of each cohort designated the filter of reliable intensity measurement from each individual gene chip. The combined two lists (17,231 probes) were used as quality control for statistical tests, class prediction, and clustering algorithms subsequently performed on the data. Class comparison was performed using nonparametric ranking statistical analysis test (Mann-Whitney) applied to quality control genes. In the vertical direction, hierarchical clusters of genes were generated using the Spearman correlation. Class prediction was performed using a supervised learning algorithm, K-Nearest Neighbors Method, that assigns a sample to predefined classes in three steps: (a) identification of genes that have strong correlations to parameters (predefined classes) of a training set of samples; (b) determination of an estimate of prediction error rates of training set by a leave-one-out cross-validation method; and (c) validation with an independent test set to obtain a true prediction error rate. In step 1, the Fisher exact test is used to identify genes by their degree of correlation to the predefined class (by user) of the training set of samples. Genes are then ranked by their predictive strength (negative natural log of p-value) that represents the probability of obtaining the observed number of samples from each class above or below the ideal pattern by chance. In step 2, samples from the training set are clustered using the k-NN method, where neighbors are identified by representing gene expression as vectors and placing samples according to the Euclidean distance. Each gene's discriminative ability is considered equally regardless of its value determined by Fisher's exact test (i.e., each classifier "votes" for a cohort and each vote is equal). After each gene evaluates the sample, the votes are summed to determine classification of the sample. Leave-one-out cross validation estimates the prediction error rate (or accuracy) by the systematic removal of one donor from the training set to use as a test sample. This process is repeated until all the donors have been "tested." A p-value ratio cutoff of 0.2 was used in all discrimination analyses. A p-value ratio of 0.2 (equivalent to 1/5) indicates that the algorithm will make a prediction if the p-value (probability that the test sample is predicted as belonging to one class by chance) of the first best class is at least five times smaller than the p-value of the next best class. If the actual p-value ratio is less than the cutoff, a prediction will be made; if the ratio is higher, no prediction will be made. Setting the p-value cutoff to 1 will force the algorithm to always make a prediction but may result in more prediction errors. Thus, each gene will cast a vote for each sample in the dataset if the p-value from the predicted class is at least five times smaller than the other class. Given this, it is possible to have a tie-in class prediction (each predictor casts an equal vote), resulting in unclassified samples. In step 3, an independent sample set is evaluated, as in step 2, except for the leave-one-out cross-validation estimates.

For the comparison between PBMCs and cell subtypes, Affymetrix U133A and B GeneChip raw intensity data were scaled to a target intensity value of 500 (TGT) in Affymetrix Microarray Suite 5.0. Data were then imported into Genespring and analyzed without any further normalization steps.

RT-PCR.
RNA samples were DNase treated with TURBO DNA-free kit (Ambion), and total RNA for RT-PCR analysis was further amplified due to low yields of total RNA. 5 µg of each RNA sample was converted to cDNA using the High Capacity cDNA Archive kit (Applied Biosystems) in the PerkinElmer GeneAmp PCR System 9600. Quantitative PCR was performed on selected targets using pre-developed primers and probe TaqMan Gene Expression Assays (Applied Biosystems) on the ABI Prism 7700 Sequence Detection System. Expression results were calculated as the difference in cycle threshold relative to the median of four healthy volunteers for each target confirmed.

Online supplemental material.
Fig. S1 shows the differential gene expression profiles in PBMCs from healthy controls and SoJIA patients at three different stages of disease. Fig. S2 shows the analysis scheme that was followed to identify a SoJIA-specific signature. Fig. S3 shows the validation of the discriminative value of selected genes by RT-PCR. Table S1 shows the patient clinical data, and Table S2 shows a list of 873 genes distinguishing SoJIA patients from healthy controls. The online supplemental material is available at http://www.jem.org/cgi/content/full/jem.20070070/DC1.