Correlation analyses of clinical and molecular findings identify candidate biological pathways in systemic juvenile idiopathic arthritis

The study was approved by the Stanford University Administrative Panel on Human Subjects in Medical Research (protocol ID 13932). Informed consent was obtained from patients or parents or guardians before blood sample collection. Venous blood samples from all subjects were treated anonymously throughout the analysis. All JIA patients were followed at the Pediatric Rheumatology Clinic at Lucile Packard Children's Hospital. SJIA and POLY patients met amended ILAR criteria for diagnosis [1]. Thirty-one SJIA and 18 POLY individual patients participated in this study. A total of 46 SJIA samples (22 Flare and 24 Quiescence samples), and 25 POLY samples (17 Flare and 8 Quiescence samples) were analyzed. Some patients, (SJIA n = 15; POLY n = 7) contributed samples during both flare and quiescent disease states. Twelve POLY patients were rheumatoid factor (RF) negative, and six were RF positive. All samples were classified as flare (F) or quiescence (Q) based on a scheme we developed for this and other studies of JIA [10, 31, 32] (Tables 1, 2 and 3 and [33, 34]). SJIA flare samples had a systemic score of ≥ 1 and/or an arthritis score of ≥ B (≥ 5 active joints). POLY flare samples had an arthritis score of ≥ 1 (≥ 1 to 10 active joints). Arthritis severity is scored differently for SJIA and POLY patients, because the patterns of joint involvement generally are different between the two groups [34], with the exception that some SJIA patients develop POLY-like arthritis with symmetric, small joint involvement. The arthritis scoring system is based on frequency analyses of numbers of active joints in early active SJIA [34] and in active POLY [Sandborg C, frequency data not shown]. Comprehensive clinical information was collected at each patient visit, including history, physical exam and clinical laboratory values [10]. As shown in Table 4, and consistent with the known demographics of JIA [35], SJIA patients are younger than POLY patients and are gender-balanced, whereas there are more female than male POLY patients. As expected, flare (F) patients from both SJIA and POLY cohorts differ significantly from quiescent (Q) patients for variables reflecting active inflammation: erythrocyte sedimentation rate (ESR), white blood cell count (WBC), platelets (PLT) and joint count (JC, number of affected joints).

Blood samples were obtained only when there was a clinical need for blood tests. A total of 3 to 4 ml of blood was collected directly in vacutainer cell preparation tubes (CPT) with sodium citrate (Becton Dickinson, Franklin Lakes, NJ, USA). Peripheral blood mononuclear cells (PBMCs) were isolated within three hours of collection by centrifugation of CPT tubes, per the manufacturer's instructions.

Purified PBMCs were lysed in RLT reagent (Qiagen, Valencia, CA, USA) and lysate was stored at -80°C until RNA extraction. RNA was isolated using the RNeasy mini kit (Qiagen), per the manufacturer's instructions with an additional on-column DNase I (Qiagen) treatment for 40 minutes. The RNA concentration was measured by the Ribogreen assay (Molecular Probes, Grand Island, NY, USA) or by absorbance at 260 nm. The purity of RNA was assessed by the ratio of the absorbance readings at 260 and 280 nm. The integrity of the RNA samples was also checked by either agarose gel electrophoresis or with the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA).

In a pilot study, paired flare/remission PBMC samples from 14 SJIA patients were processed for RNA as described [36] and analyzed using Lymphochip cDNA microarrays (Patrick Brown, Stanford University, Stanford, CA, USA) [37, 38]. A large number of genes were identified as differentially expressed in flare versus remission samples by Significance Analysis of Microarrays (SAM) [39]. Hierarchical clustering was performed with the Cluster program and visualized using TreeView [40] (Eisen Lab, University of California, Berkeley, CA), as illustrated for a subset of genes in Additional file 1, and also in [36, 41]. The full data set, GSE37388, is released to the public on the Gene Expression Omnibus (GEO) database. From the large set, we selected genes (n = 131) representing various ontologic categories, such as signaling, transcription, inflammation and immune function. We then added other immune-related genes (n = 50) that are expressed in PBMC and implicated in JIA or RA by published reports. The genes were selected prior to analysis of any blood samples for this study, and the samples used for the microarray experiment were not re-used here. The 181 selected genes are shown on Additional file 2; we confirmed that many are immune-related using the program PANTHER 7.0 (Protein ANalysis THrough Evolutionary Relationships) Classification System (Thomas Lab, University of Southern California, Los Angeles, CA, USA), which classifies proteins by their functions, using published experimental evidence and evolutionary relationships [42] (http://​www.​pantherdb.​org/​) to categorize their biological functions. This analysis showed that the largest functional category is inflammatory chemokine and cytokine signaling pathways (14.6% of the genes), followed by interleukin signaling pathways (10.8%), apoptosis signaling pathways (9.9%) and toll receptor signaling pathways (6.2%). A full list of categories covered is shown on Additional file 3.

The kinetic RT-PCR assay was performed as described [43]. Briefly, all reactions were carried out in duplicate as a single-step RT-PCR reaction, using SYBR green chemistry. Data from duplicate reactions for each gene were averaged and normalized based on levels of expression of four housekeeping genes: eukaryotic translation elongation factor 1 alpha1 (EEF1A1), protein phosphatase 1, catalytic subunit, gamma isoform (PPP1CC), ribosomal protein L12 (RPL12), and ribosomal protein L41 (RPL41). The normalized expression level, housekeeping normalized units, of each gene was used to determine the fold change among samples. In a preliminary experiment, we found that a subset (n = 75) of our gene panel showed very limited variation in level (± 2-fold difference from the mean value) in five healthy individuals (two females and three males) over a four-month period (data not shown).

The biological pathways indicated by the group of genes associated with each clinical parameter/patient cohort subset were determined by pathway analysis with Ingenuity IPA system (Ingenuity Systems, Redwood City, CA, USA; http://​www.​ingenuity.​com). The significance of either ESR or JC related pathways was analyzed using sparse linear discriminant analysis method, as previously described [44]. Correlation between SJIA ESR-related and JC-related pathways was analyzed by Pearson correlation. To determine a threshold to extract pathways that significantly differentiate ESR and JC in SJIA, 500 simulated SJIA ESR-related and 500 simulated JC-related pathway data sets were created by permutation of canonical pathway identifications and their associated pathway P-values for SJIA ESR or JC. For each canonical pathway, the absolute P-value difference in logarithm form between SJIA ESR and JC was computed using one of the 500 simulated SJIA ESR and one of the 500 simulated JC pathway P-value data sets. This led to 500 absolute log P-value differences for each canonical pathway between SJIA ESR and JC, which later were sorted and 20%, 50% and 80% values were computed. Densities of the absolute differences between SJIA ESR and JC-related pathways for the original and the simulated data sets (20%, 50%, and 80%) were plotted using the R package. Comparison of the original data set and the 80^th percentile simulated data set determined the threshold to select significantly different pathways between SJIA ESR and JC. A similar approach was applied to the analysis of significantly different pathways between SJIA ESR and POLY ESR.

ESR was chosen as a quantitative measure of systemic inflammation for our analysis, as it typically rises in association with flares of systemic symptoms and was assessed in the largest number of samples. We also considered another measure of systemic inflammation, C-reactive protein (CRP), but few samples were assessed for CRP, precluding the use of this parameter in our analysis. The number of affected counts (joint count, JC), as defined above (Introduction), was used as a quantitative measure of arthritis.

Our samples were initially classified as flare or quiescence based on criteria that we have developed for analysis of JIA (Tables 1, 2 and 3), as previously published [10], and ESR and JC are part of these criteria. We performed a distribution analysis of ESR or JC values by disease states (flare/quiescence) using R Epicalc package (http://​cran.​r-project.​org/​web/​packages/​epicalc/​) to investigate if additional subgroups would be revealed. Visual inspection of the results show that the SJIA and the POLY flare patients could be partitioned into two groups related to their ESR values (Figure 1A): F1, with ESR values below 20, and F2, with ESR values above 20. All patients in the F1 subgroup had mild flares by our other criteria (not shown). Quiescence samples all had ESR below 20, clustering together with the F1 flare group. This analysis also showed that, in our samples, JC values in the flare and quiescence disease states are generally non-overlapping in both SJIA and POLY patients, with quiescence samples with 0 or 1 joint count, and all flare samples above zero (Figure 1B).

We analyzed the association of the 181 gene panel with ESR and JC in both SJIA and POLY samples, using the strategies delineated in Figure 2. Genes whose expression was significantly associated with ESR or JC in SJIA and POLY cohorts were identified in two ways. As described in Figure 2A, Pearson correlation analyses were performed to correlate ESR or JC values with patient expression data sets. To assess the significance of these findings, we calculated the global false discovery rate (gFDR) by 100-fold permutation of normalized kPCR data. After determining the gFDR, local FDR (lFDR) analysis can compute and assign significance measures to all features [45]. A cut-off value of lFDR ≤ 0.05 was used to select significant genes for downstream pathway analysis.

We also analyzed gene expression association using Student's t-test, as shown in Figure 2B. For ESR, based on the analysis from Figure 1A, we initially divided our samples into three groups: flare samples with ESR < 20 (F1), flare samples with ESR > 20 (F2), and quiescence (to ensure that differences between F1 and Quiescence were not overlooked). We identified genes whose mean expression value differed significantly between the F1 (ESR < 20) and F2 (ESR > 20) patient groups, but no differences in genes expressed by the F1 and the quiescence group were found. Subsequently, we grouped the flare F1 and the quiescence groups into one group for ESR analysis. For JC, no other partitioning was necessary, as shown in Figure 1B, and samples were grouped into flare and quiescence groups. As we did previously for Pearson analysis, we calculated local FDR and a value of < 0.05 was considered significant (Figure 2B). This second analysis found genes missed by correlation analysis, as the latter requires a linear relationship and captures genes with more tightly regulated expression (small differences between F and Q samples).

Pearson correlation analysis of expression data from SJIA subjects found 79 genes from our panel to be ESR-correlated and 36 genes to be JC-correlated. Student's t-test found 66 ESR-associated and 79 JC-associated genes in SJIA. This pattern differed from relationships of the expression levels of the same genes with these clinical parameters in POLY-course JIA patients: 20 ESR-correlated and no JC-correlated genes were found in POLY, and none of the genes were ESR-associated or JC-associated by Student's t-test in POLY. Combining both analyses, we found 91 ESR-related and 92 JC-related genes in SJIA, and 20 ESR-related and no JC-related genes in POLY. A list of significantly associated genes is on Additional file 4. Additional file 5 (Supplementary Figure 2A-F) diagrams the fold changes in expression of the selected genes between groups (for example, F2 versus F1 + Q) and between quartiles of ESR or JC. The probability density analysis graphically represents the normalized frequency distribution of the fold ratio of the selected two groups. This result indicates that our selected genes have significant variation between groups (limited variation = fold ratio close to 1) while showing strong association. This analysis further supports our approach (Figure 2) for the identification of significant associations. The reduced number of genes associated with these clinical parameters in the POLY cohort was not surprising given that the gene list was chosen in large part using expression data from SJIA PBMC. Indeed, this finding implies a degree of specificity of the associated genes for SJIA (see discussion). Another likely contribution to this difference might be the extent that disease-related processes are reflected in peripheral blood in the two disease types.

Using the lists of associated genes, we determined biological pathways associated with each clinical parameter/patient cohort by pathway analysis with the Ingenuity IPA system. ESR-related and JC-related pathways were then compared, to investigate whether the same biological pathways are involved in ESR and JC elevations in SJIA. As shown in Figure 3A, there is strong correlation (Pearson correlation coefficient, 0.91) between SJIA ESR and JC-related pathways (n = 189), implying that some of the same pathways play roles in the systemic and arthritic components of the disease. Shown in Figure 3B, densities of the absolute log P-value differences of all pathways between SJIA ESR and JC, for the original and the 20, 50 and 80 percentile of the simulated random data sets, were computed and plotted.

Significantly differentiating pathways between ESR and JC in SJIA revealed by this analysis are in Table 5 (top two pathways), as were pathways that were correlated comparably with both ESR and JC (Table 5). The only pathway more significantly related to SJIA ESR than to SJIA JC was the glucocorticoid (GC) receptor signaling pathway. The expression of most of the genes in this pathway was higher in samples with higher ESR compared to samples with lower ESR. In contrast, the PI3K/Akt signaling pathway was more significantly related to SJIA JC. Though the significance of the association favored JC, the expression levels of most genes in this cell survival pathway were higher in samples with higher ESR or higher JC. However, as might be expected, TP53, which encodes p53, a pro-apoptotic, negative regulator of the Akt pathway [46], was down-regulated in association with JC and ESR elevations. Consistent with these results, we have previously reported that purified monocytes have lower TP53 transcript levels and increased cellular resistance to apoptotic stimuli during SJIA flare compared to quiescence [36]. Overall, the identification of some pathways that are differentially correlated with ESR and JC raises the possibility of differences in aspects of the immunobiology of arthritis compared to systemic inflammation in SJIA, as discussed below.

A number of pathways were significantly related to SJIA ESR and JC to the same degree (Table 5). For several of these pathways, the expression of most of the associated genes was higher in samples with higher ESR or JC. These pathways include, among others, protein kinase receptor (PKR, a pattern-recognition receptor) signaling in interferon induction, T cell and B cell signaling in the pattern of rheumatoid arthritis (RA), and (macrophage) migration inhibition factor (MIF) regulation of innate immunity. Other genes in pathways associated with activating innate responses, such as lipopolysaccharide (LPS) signaling and triggering receptor expressed on myeloid cells (TREM1) signaling are also higher samples with either higher ESR or JC. Genes in other pathways showed lower expression in samples with higher ESR or JC, such as T helper cell differentiation, iCOS-iCOSL (inducible T-cell co-stimulator/ligand) signaling in T helper cells and CD40 (co-stimulatory molecule on antigen presenting cells) signaling. Notably, these down-regulated pathways are associated with adaptive immune responses. Also down-regulated in association with elevations of both SJIA ESR and JC is the pathway for crosstalk between dendritic cells and natural killer cells, which can be involved in restriction of innate responses [47].

Two genes, the DNA repair enzyme ATM and the transcription factor NFATC2 (also known as NFAT1), are in the pathway for RANK signaling in osteoclasts and are both down-regulated in association with systemic (ESR) and arthritic (JC) disease activity. The Rank/RankL pathway is an important regulator of bone remodeling [48]. An ATM deficiency has been described in CD4+ T cells from rheumatoid arthritis (RA) patients [49], associated with premature immunosenescence. However, ATM may also be involved in bone formation, and ATM deficient animals show increased numbers of osteoclasts [50]. The transcription factor NFATC2 has been identified as a negative regulator of cartilage cell growth [51]. It is also important in T cell effector function, translocating to the nucleus following T cell receptor activation and regulating expression several cytokines in CD4 T cells (reviewed in [52]). Thus, its inverse correlation with ESR and JC may be similar to the other T cell-related pathways described above. Interestingly, hyperactivation of NFATC2 in T cells is associated with decreased susceptibility to experimental autoimmune encephalomyelitis, indicating that increased NFATC2 activity may have immunomodulatory effects that down-regulate autoaggressive reactions [53].

We next asked whether some biological pathways involved in SJIA ESR elevation are also involved in POLY ESR elevation, by comparing SJIA and POLY ESR-related genes. As shown in Figure 3C, there is reduced correlation (correlation coefficient, 0.59) between SJIA and POLY ESR-related pathways (n = 119), compared to the correlation we observed between SJIA ESR- and SJIA JC-related pathways. Shown in Figure 3D, densities of the absolute logarithm P-value differences of all pathways between SJIA and POLY ESR, for the original and the 20, 50 and 80 percentile of the simulated random data sets, were computed and plotted.

Several pathways differ significantly between SJIA and POLY ESR, as quantified by absolute difference between the SJIA and POLY pathways (Table 6). These include: the role of macrophages, fibroblasts and endothelial cells in RA, IL-10 signaling, glucocorticoid receptor signaling, among others. These data suggest a greater role for these pathways in SJIA, compared to POLY. However, very few genes were associated with POLY ESR in most pathways, resulting in low significance of association with POLY (not shown). In these differentiating pathways, the (few) genes correlating with ESR were higher in POLY samples with higher ESR values, suggesting that these genes, perhaps in the context of other pathways or in the context of the identified pathways but within the joint, contribute to inflammation in polyarticular course JIA. Indeed, evidence from RA, the adult disease most similar to polyarticular JIA, implicates monocyte and macrophage activation [54] and endothelial cell dysfunction [55], both in joints and in the periphery.

As observed in the previous analysis, pathways associated with T cell responses are significantly associated with ESR in SJIA but the genes in these pathways show lower expression in samples with higher ESR in comparison to samples with lower ESR. In addition, this analysis showed that genes in B cell activating factor (BAFF) signaling, April (A proliferation-inducing ligand, TNFSF13)-mediated signaling and IL-15 signaling pathways show lower expression samples with in elevated ESR in SJIA (Table 6).

Using the previously identified JC-associated genes (Additional file 4), we then investigated whether arthritis-related gene pathways change when the disease phenotype changes from the earlier systemic and arthritic activity/flare (SAF) phase to arthritis-only activity/flare (AF) phase. Shown in Figure 4A, SJIA patients were distributed according to values of JC and systemic scores (Tables 1 and 2-) to identify SAF and AF subgroups. Figure 4B shows the JC associated genes that are significantly correlated with JC in SAF and AF subgroups. Within the SJIA SAF group, only IL-10 was identified to positively correlate with JC (P-value 0.026). In contrast, 12 genes were found to significantly correlate (negatively) with JC in AF subgroup (Figure 4B, listed in order of decreasing significance): TRAP1, IL2RG, CD40LG, PARP1, TP53, ATM, NFATC2, GZMA, CASP10, PFKFB3, IRF3 and IRF4. Canonical pathway analysis (Figure 4C) mapped 8 of the 12 JC-correlated genes to a single network with the Th2 cytokine, IL-4, at its center. These functional relationships suggest that lack of IL-4 may contribute to arthritis in the AF subgroup. The difference in JC-correlated genes/pathways between AF and SAF supports the hypothesis that different biological pathways are engaged in the chronic arthritis stage versus the more acute (or systemic symptom-associated) arthritis of SJIA.