Abstract

Objectives. The overall non-response rate to biologics remains 30–40% for patients with RA resistant to MTX. The objective of this study was to predict responsiveness to the anakinra–MTX combination by peripheral blood mononuclear cell gene profiling in order to optimize treatment choice.

Methods. Thirty-two patients treated with anakinra (100 mg/day s.c.) and MTX were categorized as responders when their 28-joint DAS (DAS-28) had decreased by ≥1.2 at 3 months. Pre-treatment blood samples had been drawn.

Results. For seven responders and seven non-responders, 52 microarray-identified mRNAs were expressed as a function of the response to treatment, and unsupervised hierarchical clustering correctly separated responders from non-responders. The levels of seven of these 52 transcripts, as assessed by real-time, quantitative RT–PCR, were able to accurately classify 15 of 18 other patients (8 responders and 10 non-responders), with 87.5% specificity and 77.8% negative-predictive value for responders. Among the 52 genes, 56% were associated with IL-1β.

Conclusion. This predictive gene expression profile was obtained with a non-invasive procedure. After further validation in other cohorts of patients, it could be proposed and used on a large scale to select likely RA responders to combined anakinra–MTX.

Trial registration. Clinical Trials; NCT00213538 (http://www.clinicaltrials.gov).

Introduction

RA is a chronic, autoimmune and inflammatory polyarthritis that causes joint damage and disability. Advancement over the last few decades of our understanding of the RA pathophysiological mechanisms involved has led to the development of new treatments conceived to act against precise therapeutic targets such as TNF-α, IL-1β and IL-6 [1–4]. These new molecules called biologics include agents blocking TNF-α, such as adalimumab, etanercept and infliximab; an IL-1-receptor antagonist (IL1-Ra, anakinra); an inhibitor of the co-stimulation pathways implicated in T-lymphocyte activation [(CTLA4-Ig) or abatacept]; a mAb binding to the surface molecule CD20 expressed on B cells (rituximab); a mAb binding IL-6R (tocilizumab) [1–4]. All these agents have proven efficacy in reducing joint inflammation and, thus pain, and limiting or stopping joint destruction [1–4]. However, no response to TNF-α blocking agents or IL-1Ra is obtained in ∼30% of patients, and the responses to all these medications are highly variable from one patient to another [5]. At present, we are still unable to predict the clinical efficacy of these treatments in a given patient because of RA heterogeneity and the existence of subgroups of patients susceptible to responding better to one molecule than another. A patient’s functional outcome is so important that rheumatologists do not have the luxury of time to test each molecule in a given patient and their prescription exposes the patients to certain risks (infections, allergy, etc.) despite all the precautions taken. Consequently, rheumatologists require well-founded guidelines to orient their choice of immunotherapy for individual patients. No marker used in routine practice has been identified that might truly predict the responsiveness to biologics, even though some polymorphisms have been described [6–8]. This situation clearly demonstrates the attractiveness of pharmacogenomic approaches (including transcriptomic or proteomic approaches) that are relevant and original tools to identify markers of drug responsiveness in order to optimize the treatment prescription. Notably, we previously applied genome-wide analysis of gene expression with cDNA arrays to identify markers of responsiveness in peripheral blood mononuclear cells (PBMCs) and found a combined level of a small set of discriminative transcripts able to predict infliximab–MTX efficacy in RA patients [9]. The results of that study demonstrated for the first time the existence of a combination of genes able to predict infliximab efficacy in a given patient, before any exposure to the agent, and based on only a simple, non-invasive, peripheral blood sample. Moreover, the therapeutic response predicted by PBMC transcriptome analysis is completely innovative, not only for cancer but also for non-cancer pathologies such as RA. This field of research has also opened the window of translational, predictive and personalized medicine in rheumatology, particularly in RA.

The most recent international recommendations recognized anakinra efficacy against RA, but suggested that its specific place in the rheumatology armamentarium remains to be defined [3, 10–17]. The identification of a small subset of transcripts whose combined levels enable reliable prediction of the response to anakinra– MTX in RA patients could define the specific indication of anakinra. The objective of this study was to highlight the reliability of our tool, the reproducibility of its results and its adaptability for identifying responsiveness to another drug than infliximab and the ability to do so in different patient populations. Pertinently, the identification of an anakinra–MTX-responsiveness profile validated our original approach to identify transcript combinations able to predict RA drug responsiveness.

Patients and methods

Patients

A total of 32 patients fulfilling the ACR criteria for RA and followed at Rouen University Hospital were included in this study [18]. The criteria for patient eligibility were: MTX treatment; DAS-28 ≥ 3.2 [19]; inadequate response to at least one conventional DMARD, including MTX; and oral contraception for women of child bearing age. Exclusion criteria were: evolving infectious disease; age <18 years; pregnancy; cancer <5 years earlier; and development of hypersensitivity to anakinra. This study (no. 2003/012) was approved by our regional ethics committee (CPP Nord-Ouest 1, formerly CCPPRB Haute-Normandie, France) and all participants gave written informed consent at the time of enrolment.

For ≥1 month before the start of this study, each patient received MTX administered weekly at a stable dose and taken >48 h before any blood samples. Furthermore, some patients were also treated by a stable dose of glucocorticoids (≤10 mg/day prednisone) and/or of an NSAID; no IA steroid injection was permitted. All the patients were biologic DMARDs naïve. During the study, every patient continued to receive the same MTX dose and prednisone; anakinra (Amgen, Thousand Oaks, CA, USA) was simply added to that regimen, and given as recommended by the manufacturer and the French Drug Agency Agence Française de Sécurité Sanitaire des Produits de Santé (AFSSAPS) (100 mg/day s.c.). Before the first anakinra injection, DAS-28, ESR, serum CRP level, patient’s self-assessment of pain [0–100 mm, visual analogue scale (VAS)], duration of morning stiffness and physical function scored with the French version of the HAQ for RA were recorded [20]. After 3 months of treatment, the patients were categorized as responders when their DAS-28 had decreased by ≥1.2; all others were considered non-responders.

PBMC isolation, mRNA extraction and radiolabelling

The patients did not eat or drink anything including their drugs, such as CSs and NSAIDs, before giving the blood sample for PBMC isolation. The blood samples were obtained in the morning (between 8.00 a.m. and 9.00 a.m.) from the immunotherapy unit of the Rheumatology Department, the same day just before the first s.c. injection of anakinra. The samples were forwarded at 4°C to the laboratory for PBMC isolation, which was followed by mRNA extraction performed the same day. PBMCs were isolated from whole venous blood by Ficoll–Hypaque centrifugation and total RNAs were extracted using guanidinium thiocyanate–phenol–chloroform extraction (Fisher Bioblock Scientific, Illkirch, France), according to the manufacturer’s instructions, quality was controlled on an Agilent 2100 Bioanalyzer (Agilent Technologies Inc., Santa Clara, CA, USA) and the samples frozen at −80°C until used. An internal, arbitrary standard consisted of pooled total RNAs from the PBMCs of three healthy donors. The oligo(dT)-primed poly(A) mRNAs were incubated with [α33P]dCTP, as previously described, and the resulting, radiolabelled cDNAs were immediately used for hybridization [9, 21].

Transcriptome analysis and real-time, quantitative RT–PCR

Our home-made cDNA array covers 12 000 cDNA probes for 10 000 non-redundant genes and various negative controls. Nylon-membrane arraying of PCR-amplified probes and hybridization of [α33P]dCTP-labelled mRNAs have all been extensively described and validated previously [21]. Briefly, cDNA probes, selected based on tissue-preferred expression in liver, corresponded to genes with liver-restricted expression (10% of the probes), and genes with hepatic expression also ubiquitously expressed in some (50%) or many (40%) non-hepatic tissues [21]. All arrays were made from a single batch of cDNA probes. Every RNA sample was hybridized at least twice on separate arrays. Whenever necessary, the sequence of cDNA probes was controlled with an ABI3100 capillary sequencer (Applied Biosystems France, Courtaboeuf, France). mRNAs were subjected to quantitative RT–PCR (qRT–PCR) using a Light Cycler (Roche Diagnostics, Meylan, France) with normalization to the 18S-RNA content and all samples were run in duplicate. The primers designed with the Primer 3 Software (Howard Hughes Medical Insititute, Chevy Chase, MD, USA and National Institutes of Health, Bethesda, MD, USA) [22] are listed in supplementary table S1, available as supplementary data at Rheumatology Online.

Image analysis and data mining

The scanned 16-bit image was imported into a Linux workstation and the spots were automatically identified and analysed with the XDotsReader software, version 1.8 (COSE, Dugny, France) [23]. Spots were identified with the algorithm 2 option (i.e. spot morphology is the critical parameter) and every spot within an array was quantified with the weighted mean intensity option (i.e. the average of the spot’s pixel intensities is weighted by the distance to the centre of this spot). Due to the background homogeneity in our array, a global background over the entire image was calculated as the mean of the local background measured around each spot (global background option) and it was automatically subtracted from every spot signal. Next, the mean signal associated with negative control spots was calculated over the entire image. For all genes considered, a spot whose signal was below a given threshold (2 s.d. above the mean signal for negative controls) was considered to be non-significant (P > 0.05). Finally, the mean signal associated with negative-control spots was automatically subtracted from the signal of every significant spot.

The median absolute deviation (MAD) of all signals per image was used for normalization [24]. The MAD is defined as the median of the absolute deviations from the data’s median [MAD = mediani(|Xi − medianj(Xj)|)]. The MAD is a robust measure of statistical dispersion, in which the magnitude of the distances of a small number of outliers is irrelevant. The median of normalized duplicates per patient was used to identify expressed genes by using the supervised statistical one-class t-test with the adjusted Bonferroni’s correction.

Statistical analyses

Statistical analyses of clinical data were conducted with GraphPad InStat Software Inc. (La Jolla, CA, USA) [25]. The TIGR Multiexperiment viewer [26] was used for supervised statistical one-class t-test and unsupervised hierarchical clustering with the Manhattan distance and complete linkage options. The supervised learning classification tool was the support vector machine (SVM). Protein networks were identified with BiblioSphere [27] with different levels of stringency (B1: two genes were co-cited in the same sentence; B2: B1 level restricted to sentences with a function word; B3: B2 level restricted to sentences with order gene … function word … gene). Clinical and experimental data are reported in compliance with the recommendations for minimum information about a microarray experiments (MIAME), and the raw data have been deposited (GEO: GSE11827) in the Gene Expression Omnibus repository [28].

Results

RA patients and their responses to the anakinra regimen

After 3 months of treatment, patients were categorized as responders (R1–15) or non-responders (NR1–17) to anakinra–MTX, according to the European League against Rheumatism criteria, from the DAS-28-ESR score calculated with the four variables: the 28 swollen joint count, the 28 tender joint count, ESR and the subject global assessment of disease activity measure on a 100-mm VAS [19]. Responders included good and moderate responders. Tables 1 and 2 provide the demographic and clinical characteristics of these 32 patients at study baseline and after 3 months of the anakinra regimen. The overall mean (s.d.) disease duration was 11.5 (9) years and the DAS-28 score indicated that all these patients had highly active RA [5.4 (1)] at baseline, which fits with their resistance to one or more DMARDs. Mean baseline DAS-28 was 5.5 (0.9) and 5.2 (1.3) for responders and non-responders, respectively. Under treatment, responders’ DAS-28 significantly improved at 3 (Table 2) and 6 months (data not shown; decreased, respectively, by means of 2.3 and 2.7), while non-responders’ scores remained high (decreased, respectively, by means of 0.4 and 0.3).

Table 1

Demographic and clinical characteristics of RA patients at study baseline, according to subsets

Parameter Respondersa
 
Non-responders
 
Trainingb Validation Training Validation 
(n = 7) (n = 8) (n = 7) (n = 10) 
Age, mean (s.d.), years 49.6 (19.4) 53.9 (10.9) 56.7 (11.8) 55.7 (9.5) 
Sex (male/female), n 0/7 2/6 1/6 1/9 
RA duration, mean (s.d.), years 8 (5.7) 12.5 (8.9) 17.9 (12.7) 8.7 (6.6) 
MTX, mean (s.d.), mg/weekc 13.2 (5.7) 12.8 (6.2) 12.5 (5) 14.5 (4.5) 
Prednisone, mean (s.d.), mg/day, n 4.4 (4.3) 2.7 (3.5) 5.7 (8.4) 6.5 (6) 
Patients taking NSAIDs, n 
Patients with RF, n 
Patients with anti-CCP antibodies, n 
Parameter Respondersa
 
Non-responders
 
Trainingb Validation Training Validation 
(n = 7) (n = 8) (n = 7) (n = 10) 
Age, mean (s.d.), years 49.6 (19.4) 53.9 (10.9) 56.7 (11.8) 55.7 (9.5) 
Sex (male/female), n 0/7 2/6 1/6 1/9 
RA duration, mean (s.d.), years 8 (5.7) 12.5 (8.9) 17.9 (12.7) 8.7 (6.6) 
MTX, mean (s.d.), mg/weekc 13.2 (5.7) 12.8 (6.2) 12.5 (5) 14.5 (4.5) 
Prednisone, mean (s.d.), mg/day, n 4.4 (4.3) 2.7 (3.5) 5.7 (8.4) 6.5 (6) 
Patients taking NSAIDs, n 
Patients with RF, n 
Patients with anti-CCP antibodies, n 

In this table, all comparisons were non-significant (Mann–Whitney non-parametric test). aCategorized as described under ‘Patients and methods’ section. bTranscript levels were measured by microarray for training subset or qRT–PCR for validation subset. cMaximal tolerated dose in a given patient.

Table 2

Clinical characteristics at study baseline and after 3 months of anakinra–MTX therapy

Parameter Respondersa
 
Non-responders
 
Training
 
Validation
 
Training
 
Validation
 
Baseline 3 monthsa Baseline 3 months Baseline 3 months Baseline 3 months 
DAS-28, mean (s.d.5.5 (0.3) 3.3 (0.8)* 5.6 (1.1) 3.2 (1.2)* 4.9 (1.0) 4.4 (1.7) 5.4 (1.4) 5.0 (1.2) 
Pain, mean (s.d.), 0–100 mm VASb 77.1 (18.9) 41.4 (15.7)* 63.1 (22.2) 37.7 (23.0) 41.3 (17.7)** 40.1 (31.7) 65.0 (21.9) 56.3 (22.1) 
ESR, mean (s.d.), mm/first hour 34.6 (8.4) 14.7 (9.9)* 38 (29.1) 14.2 (13.4)* 25.6 (22.8) 19.1 (18.2) 41 (28.7) 29.2 (24.7)* 
CRP, mean (s.d.), mg/l 21.3 (11.1) 6.1 (2.7)* 26.5 (30.6) 7.6 (4)* 22.4 (29.6) 17.4 (25.2) 28.8 (25.9) 14.8 (18.3)* 
HAQ score, mean (s.d.), 0–3 scale 1.8 (0.5) 1.2 (0.6) 1.4 (0.7) 0.8 (0.6) 1.8 (0.6) 1.7 (0.7) 1.5 (0.5) 1.5 (0.4) 
Parameter Respondersa
 
Non-responders
 
Training
 
Validation
 
Training
 
Validation
 
Baseline 3 monthsa Baseline 3 months Baseline 3 months Baseline 3 months 
DAS-28, mean (s.d.5.5 (0.3) 3.3 (0.8)* 5.6 (1.1) 3.2 (1.2)* 4.9 (1.0) 4.4 (1.7) 5.4 (1.4) 5.0 (1.2) 
Pain, mean (s.d.), 0–100 mm VASb 77.1 (18.9) 41.4 (15.7)* 63.1 (22.2) 37.7 (23.0) 41.3 (17.7)** 40.1 (31.7) 65.0 (21.9) 56.3 (22.1) 
ESR, mean (s.d.), mm/first hour 34.6 (8.4) 14.7 (9.9)* 38 (29.1) 14.2 (13.4)* 25.6 (22.8) 19.1 (18.2) 41 (28.7) 29.2 (24.7)* 
CRP, mean (s.d.), mg/l 21.3 (11.1) 6.1 (2.7)* 26.5 (30.6) 7.6 (4)* 22.4 (29.6) 17.4 (25.2) 28.8 (25.9) 14.8 (18.3)* 
HAQ score, mean (s.d.), 0–3 scale 1.8 (0.5) 1.2 (0.6) 1.4 (0.7) 0.8 (0.6) 1.8 (0.6) 1.7 (0.7) 1.5 (0.5) 1.5 (0.4) 

aResponse assessed after 3 months on the anakinra–MTX regimen. bPatient’s self-assessment of pain. Significant between-group differences are noted as follows: *P < 0.05, baseline vs 3 months in this subset (paired Wilcoxon test); **P < 0.05, responders vs non-responders at baseline (Mann–Whitney test). All other comparisons were non-significant.

Within each set of responders or non-responders, patients were randomly divided into a training subset (n = 14) for transcriptome analysis and a validation subset (n = 18) for qRT–PCR. Before treatment, all variables (except VAS-rated pain), including DAS-28, were similar for responders and non-responders and the training and validation subsets. As shown in Tables 1 and 2, among the characteristics considered only VAS-rated pain differed significantly between the paired subsets, being worse for the training subset. Among responders, six are still being treated with anakinra (after an average of 43 months), three stopped anakinra because of pregnancy for one patient and side effects for two (allergy and bronchitis) and six no longer responded to anakinra after an average of 27 months of treatment.

Gene profiles of pre-treatment PBMCs correspond to treatment responsiveness

Gene profiling of PBMC transcripts was studied in the training subset from responders and non-responders (total, 14 patients). On average, 5967 (1418) transcripts were detected in PBMCs, with 85% overlap of responder and non-responder transcript identities (data not shown).

To identify precisely the transcripts that were differently regulated in responders vs non-responders, we selected every transcript whose variation between responders and non-responders reached statistical significance, as assessed by a t-test adjusted with Bonferroni’s correction; 52 different transcripts (including two redundant probes for STIP1) were selected (P < 1 × 10−4), most of which were up-regulated in non-responders and down-regulated in responders. These transcripts are listed and detailed in supplementary table S2, available as supplementary data at Rheumatology Online and Fig. 1. Identities of the corresponding microarray cDNA probes were verified by sequencing. Finally, unsupervised hierarchical clustering of the 14 patients included in the training subset was based on the levels of the 52 transcripts and resulted in the correct separation of responders and non-responders into the two major clusters shown in Fig. 1. Among these 52 transcripts, 20 transcripts with an IMAGE (integrated molecular analysis of genomes and their expression) clone number, exploitable qRT–PCR data for the subsequent validation step and higher gene expression-level divergence between responders and non-responders were selected. We verified that these 20 transcripts were able to discriminate the training-subset responders and non-responders (Fig. 2).

Fig. 1

Hierarchical clustering of RA patients as responders vs non-responders. Transcripts in PBMCs from seven responders (R) or seven non-responders (NR) who were included in two training subsets (noted subsets one in text and Tables 1 and 2) were studied by microarray. Informative transcripts as selected by a supervised statistical one-class t-test with the adjusted Bonferroni’s correction (52 transcripts) were next used for an unsupervised hierarchical clustering of the same 14 patients given as vertical column headings. The gene symbols of transcripts indicated in horizontal rows [expressed sequence tags are noted with their 5- to 6-digit IMAGE clone numbers]. Transcript levels are expressed as ratio (level in sample/level in arbitrary standard). Scale bar (log2 ratio): lower (green), higher (red) or ratio of one (black) in sample vs standard (grey squares are missing values). This figure shows 52 transcripts with two redundant probes for STIP1.

Fig. 1

Hierarchical clustering of RA patients as responders vs non-responders. Transcripts in PBMCs from seven responders (R) or seven non-responders (NR) who were included in two training subsets (noted subsets one in text and Tables 1 and 2) were studied by microarray. Informative transcripts as selected by a supervised statistical one-class t-test with the adjusted Bonferroni’s correction (52 transcripts) were next used for an unsupervised hierarchical clustering of the same 14 patients given as vertical column headings. The gene symbols of transcripts indicated in horizontal rows [expressed sequence tags are noted with their 5- to 6-digit IMAGE clone numbers]. Transcript levels are expressed as ratio (level in sample/level in arbitrary standard). Scale bar (log2 ratio): lower (green), higher (red) or ratio of one (black) in sample vs standard (grey squares are missing values). This figure shows 52 transcripts with two redundant probes for STIP1.

Fig. 2

Unsupervised hierarchical clustering of RA patients with the selected 20 transcripts. The 20 transcripts selected from the combination with 52 transcripts for further qRT–PCR application were able to discriminate the seven R and seven NR of the training subsets.

Fig. 2

Unsupervised hierarchical clustering of RA patients with the selected 20 transcripts. The 20 transcripts selected from the combination with 52 transcripts for further qRT–PCR application were able to discriminate the seven R and seven NR of the training subsets.

Then, the PBMCs obtained from 18 validation patients were used to confirm that a combination of the above transcript levels could predict responsiveness to the anakinra regimen. For this purpose, the levels of the 20 transcripts were determined by qRT–PCR and they were compared between our responder and non-responder validation subset. Using these 20 transcripts (Fig. 3A, supplementary table S2 available as supplementary data at Rheumatology Online), unsupervised hierarchical clustering of the 18 patients from the validation subset yielded two major clusters of responders and non-responders, with only three (17%) misclassified patients (R9, NR9 and NR11). Although informative, that hierarchical clustering still lacked statistical power, which led us to further evaluate the ability of the 20 selected transcripts to accurately classify patients by SVM. The results of that analysis indicated that this set of transcripts provided 80% sensitivity and 87.5% specificity for the identification of responders and non-responders (Table 3).

Fig. 3

Validation of a narrow selection of transcripts as a tool for clustering responders (R) vs non-responders (NR). Eight R and 10 NR were included in two validation subsets. In any given sample of PBMCs, the amounts of informative transcripts were determined by qRT–PCR and normalized to the corresponding 18S-RNA level. Unsupervised hierarchical clusterings obtained with 20 (A) or 7 (B) selected transcripts are shown. Transcript level expression and scale bars are as in Fig. 1. The three patients whose numbers are boxed were incorrectly classified.

Fig. 3

Validation of a narrow selection of transcripts as a tool for clustering responders (R) vs non-responders (NR). Eight R and 10 NR were included in two validation subsets. In any given sample of PBMCs, the amounts of informative transcripts were determined by qRT–PCR and normalized to the corresponding 18S-RNA level. Unsupervised hierarchical clusterings obtained with 20 (A) or 7 (B) selected transcripts are shown. Transcript level expression and scale bars are as in Fig. 1. The three patients whose numbers are boxed were incorrectly classified.

Table 3

Performance of selected transcripts to predict responsiveness to anakinra according to the number considered

Performance characteristic Number of selected transcriptsa
 
 20 
Number of NR classified as NRb 
Number of NR classified as Rb 
Number of R classified as Rb 
Number of R classified as NRb 
Chi-square test 8.1 (P < 0.01) 8.1 (P < 0.01) 
Sensitivityb80 80 
Specificityb87.5 87.5 
Positive-predictive valueb88.9 88.9 
Negative-predictive valueb77.8 77.8 
Performance characteristic Number of selected transcriptsa
 
 20 
Number of NR classified as NRb 
Number of NR classified as Rb 
Number of R classified as Rb 
Number of R classified as NRb 
Chi-square test 8.1 (P < 0.01) 8.1 (P < 0.01) 
Sensitivityb80 80 
Specificityb87.5 87.5 
Positive-predictive valueb88.9 88.9 
Negative-predictive valueb77.8 77.8 

aAs listed in Fig. 2A or B. bDetermined by SVM with 20 patients including 10 non-responders (NR) and 8 responders (R) (referred to as validation subset in text).

Next, to determine the lowest number of transcripts required for an acceptable prediction of responsiveness, we tested different combinations of minimal numbers and identities of transcripts. Using a set of only seven transcripts (Fig. 3B, supplementary table S2 available as supplementary data at Rheumatology Online), hierarchical clustering could accurately classify 15 of the same 18 patients as responders or non-responders. Finally, the results of SVM indicated that the set of seven transcripts achieved predictive values of responsiveness identical to those of the 20-transcript set described above.

The 52-transcript combination is linked to a gene network focused on IL-1β

The function of each gene retained in the combination able to predict responsiveness to anakinra–MTX was investigated. BiblioSphere enabled analyses of gene–gene and gene–transcription factor relationships based on co-citation in PubMed abstracts that corresponded to different stringency thresholds. An IL-1-targeted pathway/ network-driven approach was used to discover the putative relationships between IL-1 and selected genes. Among the 44 transcripts (for 8 genes, no information about intron or exon sequences was available) with a recognized gene symbol that were downloaded in BiblioSphere software, 34 genes passed the software filter and served as input for further examination. Review of the literature dealing with these 34 input genes and 2 co-cited genes (IL-1β, IL-1Ra) revealed strong relationships between IL-1β and 26 (76%) out of 34 with B1-level stringency, 22 (65%) out of 34 with B2-level stringency and 19 (56%) out of 34 with B3-level stringency. A gene network, including these 19 software-selected genes, was built around IL-1β (Fig. 4) with either direct connections between genes or indirect links through a transcription factor. The results presented suggest strong direct or indirect regulation of these genes by at least one IL-1β-dependent component.

Fig. 4

BiblioSphere graphical network representation of the relationships between IL-1β and the 52 transcripts. Shown are the relationships between the 34 input genes (transcripts with a gene symbol and passing the software filter) and the co-cited genes (IL-1β, IL-1Ra) with a B3 level of stringency, i.e. two genes are co-cited within the same sentence restricted with a function word, from the literature. Among these 34 input genes, the program selected 19 genes based on B3 stringency to construct the network.

Fig. 4

BiblioSphere graphical network representation of the relationships between IL-1β and the 52 transcripts. Shown are the relationships between the 34 input genes (transcripts with a gene symbol and passing the software filter) and the co-cited genes (IL-1β, IL-1Ra) with a B3 level of stringency, i.e. two genes are co-cited within the same sentence restricted with a function word, from the literature. Among these 34 input genes, the program selected 19 genes based on B3 stringency to construct the network.

Discussion

Despite an increasing body of data, the prescription of biologics to treat RA remains empirical [29]. Moreover, predictive markers of response to these agents are urgently needed because of their failure in 30–40% of the patients, their potential severe side effects and their high cost. Unfortunately, the biological markers used for RA diagnosis or prognosis are unable to predict individual responsiveness to TNF-α blocking agents [8]. In addition, the few studies that used these markers relied on the differences between protein markers measured at baseline and several weeks after treatment onset. To overcome this void, we developed a program to identify new predictive biomarkers of RA responsiveness to biologics. We combined a truly unbiased predictive approach starting before treatment was prescribed to a given patient and global approaches based on transcriptome analyses, as used by others [30–32]. We had previously applied this strategy to establish a set of transcripts from RA patients’ PBMCs able to predict responsiveness to infliximab–MTX [9]. The validity of this approach prompted us to extend it to identify transcripts from PBMCs that could indicate RA responsiveness to anakinra–MTX in patients previously unexposed to any biologics.

The clinical relevance of focusing on anakinra could appear limited but this study was begun in 2001, when infliximab and anakinra were the only two molecules available in France. Although anakinra is prescribed less and less frequently to treat RA (probably because of its lack of efficacy and limited bioavailability over the 24-h cycle), the primary objective of this investigation was to confirm that the concept of gene profiling that predicts therapeutic responsiveness is possible with another molecule and another set of patients. Moreover, in everyday practice, anakinra combined with MTX is dramatically effective in some patients, suggesting that a subset of patients might be particularly responsive to this regimen, thus supporting interest in identifying markers predictive of anakinra responsiveness.

Three months of treatment was chosen as the study end point and response was defined as a DAS-28 decline of ≥1.2, as recommended by international experts when this study was started [3]. Nevertheless, the correctly classified anakinra responders were true responders because their mean DAS-28 declined from 5.5 to 3.2, a level very close to that corresponding to low disease activity and because those responses were sustained for >2 years.

Gene expression determination in patients’ PBMCs is an easy and non-invasive procedure for evaluating diagnosis of an autoimmune disease [30, 31]. We also used an array of probes initially devoid of tissue-preferred expression in liver. Nevertheless, 90% of all tested probes provided positive signals for at least one human tissue sample, whether liver or another organ (brain, colon, stomach, kidney and muscle) [21]. Only 10% of the probes have liver-restricted expression and 5967 (1418) transcripts were detected in PBMCs, thereby enabling the examination of PBMC transcriptomes; we analysed those transcriptomes with an arbitrary collection of ∼10 000 cDNA probes [21]. Since this restrictive procedure cannot measure every transcript expressed in PBMCs, it is not intended to provide a genome-wide view of all RA-associated gene dysregulations. Yet, this approach is quite acceptable when the major task is to infer prognosis from gene profiling. Thus, our tool and strategy, previously applied to infliximab for RA patients, was shown to possess sufficient power and to be able to identify a combination of transcripts predicting the response to anakinra [9, 32].

We identified 52 transcripts whose combined expression levels in PBMCs efficiently discriminated responders and non-responders to anakinra–MTX. Among these genes, some are linked to RA and are under-expressed in anakinra responders, compared with non-responders. STIP1 encodes an extracellular co-chaperone adaptor protein for HSP70/HSP90 and was identified as an autoantigen in the sera of 111 untreated early RA patients [33, 34]. The level of leptin receptor gene expression was lower in responders than non-responders. In this setting, it should be kept in mind that leptin-deficient mice developed less severe arthritis than control mice [35]. Moreover, leptin contributes to the mechanisms of joint inflammation in antigen-induced arthritis by regulating humoral and cell-mediated immune responses [35]. Finally, ABCC5, also known as multidrug-resistance protein, is a transporter of (anti-)folates and contributes to resistance against anti-folate drugs, such as MTX [36]. Its gene was down-regulated in responders, compared with non-responders, indicating that our gene combination probably predicts the response to the combined anakinra–MTX regimen and not only to anakinra.

A pertinent observation is that, among the 34 genes passing the BiblioSphere filter, 56% were implicated in an IL-1β-dependent pathway, which led us to consider that our gene combination could mainly correspond to an IL-1β signature. Indeed, some of these genes are regulated by transcription factors, such as JUN (BST2), CEBPβ (RUNX1T1, ELF2), HIF1A (EP300), ESR1 (EMP2), CTNNB1 (CDH5, EIF3S12), TP53 (CDK8), STAT3 (LEPR), NR3C1 (ONECUT1), NFκB1 (SLC11A2, MCM3AP) and MYC (GTF2F2). These transcription factors are, however, not specific for these genes or the IL-1β pathway and are involved in different functions (cell-cycle arrest, apoptosis, angiogenesis, proliferation, etc.) and in several diseases. When IL-1 binds to its receptor and turns on the IL-1β-signalling pathway, these transcription factors are activated and regulate target-gene expression. For example, GTF2F2 is involved in both specific initiation of RNA synthesis and elongation by RNA polymerase, and is a direct functional target of Fos and Jun to initiate the gene transcription [37]. Other genes are linked to IL-1β by BiblioSphere, such as inducible co-stimulator ligand G (ICOSLG) and transthyretin (TTR), even though their potential connection with an IL-1β-signalling pathway has not yet been elucidated. Notably, IL-1β controls T-cell functions through the activation of ICOSL involved in the co-stimulation of T cells. TTR is produced by human chondrocytes, particularly after stimulation by IL-1β, and is involved in bone and cartilage metabolism [38]. Surprisingly, IL1-Ra is not among the genes differentially expressed between responders and non-responders. RA patients having an IL-1 excess exhibit an IL-1/IL-1Ra imbalance in synovium, which generates chronic inflammation. These data are not necessarily in concordance with an IL-1Ra dysregulation in circulating monocytes. In addition, the objective of the study was to find a gene expression profile with biomarkers able to better predict anakinra responsiveness. The best biomarkers are not necessarily the more expected physiopathological markers.

Next, we showed that, among these 52 transcripts, 20 evaluated by qRT–PCR achieved a predictive value of responsiveness similar to that provided by the 52 transcripts and, ultimately, that of a selected 7-transcript combination, which was as powerful as any combination of more transcripts. The observation that a specific combination of a small number of transcripts equalled or even outperformed the predictive strength of numerous transcripts was also reported for the response to HCV treatment [39]. This small but informative gene set facilitates the development of a reliable, fast and inexpensive assay, and, consequently, opens the way to implementing its use in the routine monitoring of RA patients.

The combined levels of the seven discriminative transcripts provided, for the first time, a tool that might be used to predict anakinra–MTX efficacy in patients with long-standing and highly active RA. Indeed, this tool remains to be validated on a large scale and to prove its usefulness in patients with recent-onset RA. The results reported herein further support our previous study on infliximab and, thus, the relevance of our tool to predict responsiveness to biological immunomodulators [9]. In addition, the absence of overlap between infliximab– and anakinra–gene combinations suggests that the selected transcripts shared by different RA subsets have drug specificity rather than reflecting pathophysiological mechanisms. The comparison between infliximab– and anakinra–gene combinations is possible because these combinations were identified from the same arrays performed with the same batch of PCR-amplified probes. These combinations are currently being tested on a large independent RA cohort.

The concept of using gene expression profiling to predict responses to molecular drugs has been also explored with other approaches. For instance, van der Pouw Kraan et al. [40] evaluated the relationship between baseline molecular profiles of synovial tissues from 18 RA patients determined by large-scale gene expression profiling and their clinical responses to treatment. The results of that study indicated that patients with high-level inflammation are more likely to benefit from anti-TNF-α treatment [40]. In addition, Buch et al. [41] assessed pre-treatment synovium TNF-α, IL-1α and IL-1β expressions in 51 RA patients, to determine whether the expression of these cytokines predicted infliximab response, and found that pre-treatment synovium expression of TNF-α and IL-1 did not predict the response to anti-TNF-α. These two studies and their results further demonstrated that microarrays represent an amazing approach to identify markers predictive of responsiveness and that they performed much better than targeting individual molecules, such as TNF-α and IL-1. Moreover, PBMCs are readily obtainable from a simple blood sample, thereby making unnecessary synovium biopsy, even with today’s facilitated mini-arthroscopy.

Our approach, which could be extended to other biologics (abatacept, adalimumab, etanercept, rituximab, etc.), should help the rheumatologist optimize the prescription of these agents for a given patient. However, these results should be first validated in another set of RA patients to confirm the validity of this combination in a larger population and secondly in a prospective study to define the clinical relevance of the predictive value of such biomarkers. Ultimately, we anticipate that a small series of parallel tests, each using a drug-specific combination of transcripts on a dedicated DNA chip, will allow routine selection and individualized tailoring of the most appropriate regimen for each RA patient.

graphic

Supplementary data

Supplementary data are available at Rheumatology Online.

Acknowledgements

We would like to thank Janet Jacobson for her valuable advice in editing the article. A patent application (EP no. 08305252.2) for the sets of 20 and 7 transcripts with predictive power (Fig. 3) is under consideration by Inserm.

Funding: The present study was supported by grants from the French Ministry for Research, the Arthritis Foundation, the French Society of Rheumatology and Amgen France.

Disclosure statement: The authors have declared no conflicts of interest.

References

1
Olsen
NJ
Stein
CM
New drugs for rheumatoid arthritis
N Engl J Med
 , 
2004
, vol. 
50
 (pg. 
2167
-
79
)
2
O’Dell
JR
Anticytokine therapy-a new era in the treatment of rheumatoid arthritis?
N Engl J Med
 , 
1999
, vol. 
340
 (pg. 
310
-
2
)
3
Furst
DE
Breedveld
FC
Kalden
JR
, et al.  . 
Updated consensus statement on biological agents for the treatment of rheumatic diseases, 2007
Ann Rheum Dis
 , 
2007
, vol. 
66
 (pg. 
iii2
-
22
)
4
Smolen
JS
Aletaha
D
Koeller
M
Weisman
MH
Emery
P
New therapies for treatment of rheumatoid arthritis
Lancet
 , 
2007
, vol. 
370
 (pg. 
1861
-
74
)
5
Furst
De
Breedveld
FC
Kalden
JR
, et al.  . 
Updated consensus statement on biological agents, specifically tumour necrosis factor alpha blocking agents and interleukin-1 receptor antagonist, for the treatment of rheumatic diseases, 2005
Ann Rheum Dis
 , 
2005
, vol. 
64
 
Suppl. 4
(pg. 
iv2
-
14
)
6
Pavy
S
Toonen
EJ
Miceli-Richard
C
, et al.  . 
Tumour necrosis factor alpha -308G->A polymorphism is not associated with response to TNFalpha blockers in Caucasian patients with rheumatoid arthritis: systematic review and meta-analysis
Ann Rheum Dis
 , 
2010
, vol. 
69
 (pg. 
1022
-
8
)
7
Stuhlmüller
B
Häupl
T
Hernandez
MM
, et al.  . 
CD11c as a transcriptional biomarker to predict response to anti-TNF monotherapy with adalimumab in patients with rheumatoid arthritis
Clin Pharmacol Ther
 , 
2010
, vol. 
87
 (pg. 
311
-
21
)
8
Lequerré
T
Jouen
F
Brazier
M
, et al.  . 
Autoantibodies, metalloproteinases and bone markers in rheumatoid arthritis patients are unable to predict their responses to infliximab
Rheumatology
 , 
2007
, vol. 
46
 (pg. 
446
-
53
)
9
Lequerré
T
Gauthier-Jauneau
AC
Bansard
C
, et al.  . 
Gene profiling in white blood cells predicts infliximab responsiveness in rheumatoid arthritis
Arthritis Res Ther
 , 
2006
, vol. 
8
 (pg. 
R105
-
15
)
10
Jiang
Y
Genant
HK
Watt
I
, et al.  . 
A multicenter, double-blind, dose-ranging, randomized, placebo-controlled study of recombinant human interleukin-1 receptor antagonist in patients with rheumatoid arthritis: radiologic progression and correlation of Genant and Larsen scores
Arthritis Rheum
 , 
2000
, vol. 
43
 (pg. 
1001
-
9
)
11
Cohen
S
Hurd
E
Cush
J
, et al.  . 
Treatment of rheumatoid arthritis with anakinra, a recombinant human interleukin-1 receptor antagonist, in combination with methotrexate: results of a twenty-four-week, multicenter, randomized, double-blind, placebo-controlled trial
Arthritis Rheum
 , 
2002
, vol. 
46
 (pg. 
614
-
24
)
12
Nuki
G
Bresnihan
B
Bear
MB
McCabe
D
European Group of Clinical Investigators. Long-term safety and maintenance of clinical improvement following treatment with anakinra (recombinant human interleukin-1 receptor antagonist) in patients with rheumatoid arthritis: extension phase of a randomized, double-blind, placebo-controlled trial
Arthritis Rheum
 , 
2002
, vol. 
46
 (pg. 
2838
-
46
)
13
Fleischmann
RM
Schechtman
J
Bennett
R
, et al.  . 
Anakinra, a recombinant human interleukin-1 receptor antagonist (r-metHuIL-1ra), in patients with rheumatoid arthritis: a large, international, multicenter, placebo-controlled trial
Arthritis Rheum
 , 
2003
, vol. 
48
 (pg. 
927
-
34
)
14
Schiff
MH
DiVittorio
G
Tesser
J
, et al.  . 
The safety of anakinra in high-risk patients with active rheumatoid arthritis: six-month observations of patients with comorbid conditions
Arthritis Rheum
 , 
2004
, vol. 
50
 (pg. 
1752
-
60
)
15
Le Loët
X
Nordström
D
Rodriguez
M
, et al.  . 
Effect of anakinra on functional status in patients with active rheumatoid arthritis receiving concomitant therapy with traditional disease modifying antirheumatic drugs: evidence from the OMEGA trial
J Rheumatol
 , 
2008
, vol. 
35
 (pg. 
1538
-
44
)
16
Burger
D
Dayer
JM
Palmer
G
Gabay
C
Is IL-1 a good therapeutic target in the treatment of arthritis?
Best Pract Res Clin Rheumatol
 , 
2006
, vol. 
20
 (pg. 
879
-
96
)
17
den Broeder
AA
de Jong
E
Franssen
MJ
Jeurissen
ME
Flendrie
M
van den Hoogen
FH
Observational study on efficacy, safety, and drug survival of anakinra in rheumatoid arthritis patients in clinical practice
Ann Rheum Dis
 , 
2006
, vol. 
65
 (pg. 
760
-
2
)
18
Arnett
FC
Edworthy
SM
Bloch
DA
, et al.  . 
The American Rheumatism Association 1987 revised criteria for the classification of rheumatoid arthritis
Arthritis Rheum
 , 
1988
, vol. 
31
 (pg. 
315
-
24
)
19
van Gestel
AM
Prevoo
ML
van’t Hof
MA
van Rijswijk
MH
van de Putte
LB
van Riel
PL
Development and validation of the European League Against Rheumatism response criteria for rheumatoid arthritis. Comparison with the preliminary American College of Rheumatology and the World Health Organization/International League against Rheumatism criteria
Arthritis Rheum
 , 
1996
, vol. 
39
 (pg. 
34
-
40
)
20
Guillemin
F
Braincon
S
Pourel
J
Measurement of the functional capacity in rheumatoid polyarthritis: a French adaptation of the Health Assessment Questionnaire (HAQ)
Rev Rhum Mal Ostéoartic
 , 
1991
, vol. 
58
 (pg. 
459
-
65
)
21
Coulouarn
C
Lefebvre
G
Derambure
C
, et al.  . 
Altered gene expression in acute systemic inflammation detected by complete coverage of the human liver transcriptome
Hepatology
 , 
2004
, vol. 
39
 (pg. 
353
-
64
)
22
Rozen
S
Skaletsky
HJ
Krawetz
S
Misener
S
Primer3 on the WWW for general users and for biologist programmers
Bioinformatics methods and protocols: methods in molecular biology
 , 
2000
Totowa, NJ
Humana Press
(pg. 
365
-
86
)
23
Tahi
F
Achddou
B
Decraene
C
, et al.  . 
Automatic quantitation of hybridization signals on cDNA arrays
BioTechniques
 , 
2002
, vol. 
32
 (pg. 
1386
-
97
)
24
Fundel
K
Haag
J
Gebhard
PM
Zimmer
R
Aigner
T
Normalization strategies for mRNA expression data in cartilage research
Osteoarthr Cartil
 , 
2008
, vol. 
16
 (pg. 
947
-
55
)
25
Graphpad Instat
 
http://www.graphpad.com (15 April 2010, date last accessed)
26
Tmev version 2.2
 
http://www.tm4.org (15 April 2010, date last accessed)
27
Genomatix
 
http://www.genomatix.de (15 April 2010, date last accessed)
28
Gene Expression Omnibus
 
http://www.ncbi.nlm.nih.gov/projects/geo (15 April 2010, date last accessed)
29
Scheinecker
C
Redlich
K
Smolen
JS
Cytokines as therapeutic targets: advances and limitations
Immunity
 , 
2008
, vol. 
28
 (pg. 
440
-
4
)
30
Jarvis
JN
Centola
M
Gene-expression profiling: time for clinical application?
Lancet
 , 
2005
, vol. 
365
 (pg. 
199
-
200
)
31
Olsen
NJ
Moore
JH
Aune
TM
Gene expression signatures for autoimmune disease in peripheral blood mononuclear cells
Arthritis Res Ther
 , 
2004
, vol. 
6
 (pg. 
120
-
8
)
32
Toonen
EJ
Barrera
P
Radstake
TR
, et al.  . 
Gene expression profiling in rheumatoid arthritis: current concepts and future directions
Ann Rheum Dis
 , 
2008
, vol. 
67
 (pg. 
1663
-
9
)
33
Cai
Y
Kitajima
S
Etoh
F
Kinoshita
S
Okubo
K
Hamasaki
N
Autoantibody reactive with the human general transcription factor TFIIF in sera from patients with autoimmune disorders
Clin Exp Immunol
 , 
1997
, vol. 
109
 (pg. 
488
-
94
)
34
Goëb
V
Thomas-L’Otellier
M
Daveau
R
, et al.  . 
Candidate autoantigens identified by mass spectrometry in early rheumatoid arthritis are chaperones and citrullinated glycolytic enzymes
Arthritis Res Ther
 , 
2009
, vol. 
11
 (pg. 
R38
-
51
)
35
Busso
N
So
A
Chobaz-Péclat
V
, et al.  . 
Leptin signaling deficiency impairs humoral and cellular immune responses and attenuates experimental arthritis
J Immunol
 , 
2002
, vol. 
168
 (pg. 
875
-
82
)
36
Wielinga
P
Hooijberg
JH
Gunnarsdottir
S
, et al.  . 
The human multidrug resistance protein MRP5 transports folates and can mediate cellular resistance against antifolates
Cancer Res
 , 
2005
, vol. 
65
 (pg. 
4425
-
30
)
37
O’Neill
LA
Greene
C
Signal transduction pathways activated by the IL-1 receptor family: ancient signaling machinery in mammals, insects, and plants
J Leukoc Biol
 , 
1998
, vol. 
63
 (pg. 
650
-
7
)
38
Catterall
JB
Rowan
AD
Sarsfield
S
Saklatvala
J
Wait
R
Cawston
TE
Development of a novel 2D proteomics approach for the identification of proteins secreted by primary chondrocytes after stimulation by IL-1 and oncostatin M
Rheumatology
 , 
2006
, vol. 
45
 (pg. 
1101
-
9
)
39
Chen
L
Borozan
I
Feld
J
, et al.  . 
Hepatic gene expression discriminates responders and nonresponders in treatment of chronic hepatitis C viral infection
Gastroenterology
 , 
2005
, vol. 
128
 (pg. 
1437
-
44
)
40
van der Pouw Kraan
TC
Wijbrandts
CA
van Baarsen
LG
, et al.  . 
Responsiveness to anti-tumour necrosis factor alpha therapy is related to pre-treatment tissue inflammation levels in rheumatoid arthritis patients
Ann Rheum Dis
 , 
2008
, vol. 
67
 (pg. 
563
-
6
)
41
Buch
MH
Reece
RJ
Quinn
MA
, et al.  . 
The value of synovial cytokine expression in predicting the clinical response to TNF antagonist therapy (infliximab)
Rheumatology
 , 
2008
, vol. 
47
 (pg. 
1469
-
75
)

Author notes

*Carine Bansard and Thierry Lequerré contributed equally to this work.
Deceased.

Supplementary data

Comments

0 Comments