A canonical fibroblast serum response signature discriminates between normal, RA and OA synovial tissues

RA FALS display an agressive phenotype, which involves increased invasiveness and resistance to apoptosis [13]. Because of this, the expansive synovial tissue called ‘pannus’ at the cartilage-bone interface has been suggested to share many hallmarks of cancer. However, there is no evidence of increased cell proliferation in RA compared to normal and OA FLS. Invasiveness, apoptosis and proliferation in RA FLS derived from patients share the same properties than their in vivo counterparts [1]. We therefore formulated the hypothesis that a fibroblast serum-response signature, linked to cancer clinical outcome, may be able to discriminate RA and OA derived synovial fibroblasts. Consistent with the current understanding of the biology of synovial FLS, we also hypothesize that migration and cell death but not proliferation may be the key difference between RA and OA derived FLS. We first tested this hypothesis by developing a statistical model predictive of disease type based on the original transcriptional signature developed by Chang et al [6]. We were able to map 95% of the genes originally described on a public domain dataset representing whole synovial tissue from normal, RA and OA individuals [14]. The hypothesis was indeed consistent with the prediction results as a K-nearest neighbour (KNN) based model was able to predict disease state with good accuracy. Normal samples were predicted with a sensitivity of 61% and a specificity of 95%. RA samples were predicted with a sensitivity of 88% and a specificity of 89% and OA samples were predicted with a sensitivity of 80% and a specificity of 81% (S1 Fig.). This observation therefore supported our hypothesis that fibroblast pathways involved in wound healing and neoplasia are relevant to the persistence of inflammation in RA and OA.

A distinct hierarchy of anatomical site and serum response distinguishes fibroblasts derived from arthritis patients

Since the ‘core serum response’ signature (CSR) developed by Chang et al was based on fibroblasts derived from normal skin, lung and gum we reasoned that it may not fully represent the responses of synovial fibroblast populations in diseased joints. We therefore set out to experimentally define a serum response signature based on fibroblasts derived from both RA and OA patients. In order to take account of the documented effect of site on fibroblast transcriptional profiles and to address the possibility of multiple site involvement in systemic disease, we compared the serum response of fibroblasts of synovial origin to those of bone marrow and skin from the same individual. We set out to identify differentially expressed genes in at least one of the fibroblast groups. By using a SAM multiclass approach we could identify a signature of 3633 genes (FDR < 3% followed by a two-fold cut-off). Microarray gene expression was correlated with qPCR data for validation purposes (S2 Table). Principal component analysis (PCA) was used to visualise the relationship between the genes (Fig. 1). This unsupervised analysis showed that fibroblasts cluster primarily on the basis of anatomical location. Furthermore all matched low and high serum samples were separated on the first principal component (PC1, x-axis). By contrast, minimal clustering on the basis of disease was seen. These results provide the first evidence of a molecular hierarchy of transcriptional control, where site of origin is the dominant factor, followed by response to serum stimulation and disease status.

Within anatomical locations, RA and OA fibroblasts are molecularly distinct and differ in their response to serum stimulation

While the PCA described above provides initial evidence for the existence of a molecular hierarchy of transcriptional control, it does not quantify the effect and it does not identify the specific genes linked to anatomical location, serum response and disease type. We therefore set out to directly quantify the differences in gene expression across all three levels of organisation and performed the following direct comparisons:

1. Anatomical location (Bone Marrow vs. Skin vs. Synovial tissue)
2. Serum Response (High vs. Low Serum)
3. Disease type (RA vs. OA)

Fig. 2A summarized the results of this analysis showing that as expected anatomical location has the highest number of differentially expressed genes followed by serum response and finally disease (3763, 435 and 180 differentially expressed genes, respectively at < 5% and a 2-fold cut-off). We then focused within each anatomical location and performed a more detailed analysis. More precisely, within each tissue and within each disease group we identified genes differentially expressed between low and high serum (Fig. 2B). We found that in both RA and OA the synovial tissue has the highest number of differentially expressed genes (738 and 1419 genes respectively). Bone marrow tissues both expressed similar number of genes (RA: 469, OA: 558) while skin showed a much higher number of genes in RA (633) than OA (276) (Fig. 2B). We then identified within each anatomical location and within each serum level the number of differentially expressed genes between RA and OA samples (Fig. 2C). Here interestingly a large difference in the low serum but not the high serum condition could be observed in the synovial tissue (302 vs. 0 differentially expressed genes respectively). Skin as well as bone marrow both expressed similar numbers of differentially expressed genes (96±2) (Fig. 2C). The fold changes of differentially expressed genes were broadly similar between groups indicating that numbers of genes are a useful approximation of the degree of difference observed between transcriptional profiles. For each anatomical location we also performed a multi-class SAM analysis to identify genes that change in at least one of the different serum and disease types. We then visualised the resulting gene-lists using a PCA (Fig. 3). This clearly demonstrated that bone marrow and skin RA and OA fibroblasts responded to serum following parallel trajectories along the first principal component (PC) and were separated by disease predominantly according to the second PC (Figs. 3A and 3B), while RA and OA synovium fibroblasts separated along the second PC but converged in response to serum (Fig. 3C). In order to investigate the serum response overlap between bone marrow and skin fibroblasts, we extracted the most contributing genes within the first principle component, and found that 120 and 285 genes contributed to the first principle component of bone marrow and skin PCA analyses respectively, sharing 36 genes. This exclusive skin/bone marrow overlap suggested a specific synovial serum response.

Serum response signatures defined by anatomical location and disease type define distinct functional profiles

Having defined the serum response signatures in each anatomical location and within each disease group we asked whether the serum response programs overlap significantly Samples derived from all three different anatomical locations of RA patients exhibited a greater overlap in serum response profile than OA samples, which exhibited more divergent, site-specific serum responses (11% vs 3% in common, supplementary S2 Fig.). Next, in order to address the biological significance of our findings we used the downstream effect analysis pipeline within the Ingenuity Pathway Analysis (IPA) knowledge management system. This algorithm assesses the probability that the observed transcriptional response is consistent with a phenotypic change. The rationale behind this approach is as follows: The IPA database contains experimentally observed causal relationships that provide a direction of effect between genes and functional terms. Given that generally a number of genes may be linked to any given function, a statistical approach to quantify this effect was developed by the IPA Team. Functional terms that have significantly more “increased” or “decreased” predictions are found by comparing the submitted gene lists with a random selection with equal probability. Fig. 4 represents a summary of the resulting functional terms linked to the serum response of anatomical locations and disease types. Most of the genes submitted by each comparison can be classified into functions relating to migration and invasion of tumor cell lines, proliferation of cells and cell viability (See S3 Table for more details). However there are a number functions specific to each anatomical site and disease type. Particularly bone marrow and synovial tissue show a number of overlapping functional terms such as invasion, synthesis of DNA and apoptosis. Moreover, bone marrow RA specific functions are one of the largest set of functions associated to a number of genes (423 genes associated to 30 increased and 104 genes to decreased functions) including differentiation, angiogenesis, vasculogenesis, activation of tumor cell lines and apoptosis of tumor cell lines (decreased). Within the synovial tissue, RA samples are characterized by proliferation of fibroblasts, adhesion of tumor cell lines and necrosis and cell death (decreased 145 genes in 2 functions). Similarly to RA, OA also regulate genes associated with apoptosis but with a much larger number of genes (246 genes associated to 2 functions). Interestingly, OA synovial samples are the only sample type with functions relating to sensitization of cells which may have an impact on their ability to respond to serum.

Analysis of the synovium fibroblast response to serum identifies OA-specific repression of a tissue remodelling transcriptional signature

A key question in arhritis research is whether RA and OA fibroblasts are different in their ability to respond to grow-factor signals. As shown in Fig. 3C, the transcriptional state of RA and OA derived fibroblasts converged in high serum culture conditions. We therefore set out to further characterize the difference in response between RA and OA in low serum conditions within the synovium and to assess whether these changes in expression follow serum stimulation. In the low serum condition, 302 genes were differentially expressed between synovial OA and RA fibroblasts. Genes significantly down-regulated in OA synovial fibroblasts during low to high serum transition and not included in classical GO classifiers included tumour suppressor and pro-apoptotic genes PTEN, SLIT3, TRERF1 (TReP-132), SHOX, SH3D19, ubr5 and EDD; genes with roles in cytoskeletal rearrangement supervillin, paxillin, and genes involved in ubiquitination and signalling included PIAS1, crebbp and the IGFBP family of genes. Genes regulated at higher levels in RA than OA fibroblasts included Secretogranin-II and tensin 3. Of the 302 genes, 98% were not differentially expressed in bone marrow or skin fibroblasts, confirming a synovium-specific response (Fig. 5A). Fig. 5B shows the relative expression in low and high serum of differentially expressed genes in synovial fibroblasts. The majority of genes (88%) were down-regulated in response to high serum in OA synovial fibroblasts. This relationship was also seen to a lesser extent in skin and bone marrow OA fibroblasts, while RA fibroblasts showed no convincing regulation of this gene list in response to serum. To define the biological relevance of the serum mediated synovium-specific disease signature, we performed a network analysis using Ingenuity Pathway Analysis software. This identified three inter-linked gene networks representing the control of actin remodelling through insulin-like growth factor associated proteins (IGFBP5, IGFBP6, IGFBP7 and GnRH1) (Fig. 6A), control of tissue remodelling by β3 integrin (Fig. 6B), and regulation of apoptosis through CD44 (Fig. 6C), each representing facets of the fibroblast phenotype in OA. This last observation shows that molecular signatures characterizing RA fibroblast serum response not only discriminate between RA and OA FLS but that the functional profile is related primarily to adhesion and migration pathways. This is consistent with our initial hypothesis.

Analysis of synovial tissue datasets from public repository

We retrieved a dataset from the NCBI GEO database which contained synovial tissue biopsy samples from RA, OA and normal patients (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE12021) [14]. The dataset was normalized using robust multi-array average (RMA) methodology; genes with expression close to the detection level were removed. We mapped cDNA common serum response (CSR) genes to Affymetrix® Ids using the ID converter facility in DAVID [37, 38]. Expression of successfully mapped genes was then determined in the GSE12021 dataset. To determine the ability of the CSR to classify synovial tissue from different diseases, tissue datasets were classified by CSR genes using a k-nearest neighbour (KNN) classification algorithm, with validation by a leave one out cross-validation (LOOCV) procedure.

Patients and fibroblast culture

Fibroblasts were isolated as previously described [39] from synovium, bone marrow and skin tissue samples taken at the time of knee or hip replacement surgery from 12 rheumatoid arthritis patients meeting the 1987 ACR criteria [40] and 6 osteoarthritis patients diagnosed on the basis of characteristic x-ray findings and the absence of features suggestive of inflammatory arthritis (S1 Table). Only one hip sample was present in either disease group. Fibroblasts were maintained in fibroblast medium (consisting of 81.3% RPMI 1640, 10% FCS, 0.81x MEM non-essential amino acids, 0.81 mM sodium orthopyruvate, 1.62 mM glutamine, 810 U/ml penicillin and 81 μg/mL streptomycin) at 37°C in a humidified 5% CO2 atmosphere. Ethical approval for the use of this material was given by the ‘NRES Committee West Midlands—The Black Country’ and all patients gave written informed consent (LREC reference 5735). RA disease activity was quantified using the DAS28 disease activity score [41].

RNA extraction and microarray hybridization

Cells were seeded in duplicate at the required density using 10% FCS. Once attached, the medium was removed and the cells washed in serum free medium before replacing with fibroblast medium, which was identical but contained either 10% or 0.1% FCS [6]. mRNA was extracted from all fibroblasts at passage 5, under identical conditions. In order to remove confluence of cells as a confounding factor, proliferation rates were measured, and appropriate numbers of cells seeded to achieve 75% confluent cultures at harvest. Fibroblasts were retrieved from culture by trypsin digestion and washed in fibroblast medium then PBS before resuspending in 200 μl of RLT RNA protection buffer (Quiagen). 102 fibroblast samples in all were prepared according to Affymetrix® manufacturer instructions using 50 ng total RNA, which was amplified using Nugen™ Ovation™ Biotin RNA Amplification prior to hybridization with U133 plus 2.0 Affymetrix® Genechips. All chips passed QC criteria and visual inspection of Expressionist thumbnail images. The data is available under GEO accession number GSE56409.

Array validation using quantitative PCR

Synovial fibroblasts from 3 OA and 3 RA patients were cultured as above. Cells were trypsinised and RNA extraction performed using an RNeasy kit (Qiagen). RNA from 1×10⁵ fibroblasts was reverse transcribed using Superscript Vilo (Life Technologies). 1 μl of the resulting cDNA was added to the real-time PCR master mix (Applied Biosystems) and TaqMan primers/probes were used to measure the relative quantities of SLIT3, PIAS1 and TNS3 (VIC labelled) with GAPDH (FAM labelled) as control using an ABI 7900HT machine (Applied Biosystems).

Dataset and data processing

Out of the 108 samples retrieved six samples did not meet quality control and hence did not undergo further analysis. These included 1x Bone Marrow RA low serum, 1x Bone Marrow RA high serum, 1x Skin OA low serum, 1x Synovial OA low serum, 1x Synovial OA high serum and 1x Skin OA high serum sample. Gene expression data from the Affymetrix® arrays were processed and normalized using RMA methodology. Data were processed to remove genes with expression close to the detection level.

Identification of differentially expressed genes

Statistical analysis for microarrays (SAM) [42] as implemented by the R package (samr) was used to identify differentially expressed genes between groups. A cut-off of FDR < 3% and an additional cut-off of two fold change was applied on all differentially expressed gene sets. Initially we identified differentially expressed genes for each of the three levels of organisation (anatomical location, serum response and disease type). Here all samples belong to the given level were compared in either a “Multiclass” comparison (anatomical location) or a “Two class unpaired” comparison (serum response and disease type) using the SAM algorithm. The number of differentially expressed genes were then filtered to the previously mentioned thresholds. To identify genes differentially expressed in each tissue and within serum and disease types, we first subset the data into the three tissues. Then for each serum level we identify differentially expressed genes between RA and OA. Similarly for each disease type we identify differentially expressed genes between high and low serum. Apparent discrepancies in total gene numbers in tables compared to the text result from the common phenomenon of genes with multiple Affymetrix® IDs (See S1 Dataset for more details).

K-nearest neighbour classification

To test whether the CSR signature is able to distinguish between normal, RA and OA samples retrieved from the public domain we applied a K-nearest neighbour algorithm as implemented by the knn.cv function in the ‘class’ library in the statistical environment R. This methodology uses a leave-one-out-crossvalidation. For each sample in the training set the k nearest neighbours are identified and the classification decided by majority vote. We have tested 2 < k < 10 and observed that the accuracy was independent of the value of k (standard deviation of the prediction accuracy over the values of k was < 0.05% in all comparisons).

Data reduction techniques

Genes differentially expressed between all sample groups were visualised using Principal component analysis (PCA), a technique designed to represent highly multi-dimensional objects in a lower dimensionality space (Fig. 2). The first principal component (PC) on the x-axis represents the majority of the variance information, whilst the second PC on the y-axis, is orthogonal to the first PC and holds less variance. All data were mean centred prior to application of PCA using the prcomp function in the statistical environment R [43].

Ingenuity Pathway Analysis (IPA) of differentially expressed gene sets

Gene sets derived from differential gene expression and Venn diagram analyses were analysed using the IPA application (Palo Alto, http://www.ingenuity.com) that uses the Ingenuity Pathways Knowledge Base (IPKB) to explore biological interaction networks. Each gene identifier is mapped to a global molecular network within IPKB. An algorithm then generates networks based on the connectivity of focus genes, limited to a maximum of 35 genes for simpler visualisation. For the downstream effect analysis gene lists and calculated fold changes were uploaded to the IPKB. For each biological function within the IPKB an activation z-score is calculated by the software. Here the molecular network within IPKB is interrogated and experimentally observed causal relationships between genes and function extracted. These connections provide a literature-derived direction of effect which is used in combination with the foldchange to calculate the activation state of the given functional terms. As multiple genes may be connected to a function a statistical approach is necessary to calculate its significancy as compared to a randomly derived variable. This variable is obtained by randomly choosing 1 or -1 with equal probability representing an “increased” or “decreased” state respectively. For each number of genes affecting the functional term the number of “increased” (N₊) and “decreased” (N₋) predictions is summed up for the random. This however assumes that there is only 1 possible downstream effect to a given function. To correct for this the IPAKB uses a weighted z-score statistics in which: (1) where M_activating and M_inhibiting represent the number of findings underlying an edges “activation” or “inhibition” respectively. This is then used to correct for the direction of change within the z score algorithm. The z-score can therefore be calculated as: (2)