Background
Recent advances in our understanding of the pathophysiology of rheumatoid arthritis (RA) have led to the development of new biologic treatments designed to act against a precise therapeutic target: these include tumor necrosis factor-alpha (TNFα), interleukin-1-receptor antagonist (IL1-Ra, anakinra), cytotoxic T-lymphocyte-associated protein 4 (CTLA4-Ig or abatacept (ABA)), CD20 expressed on B cells (rituximab) and IL-6 receptor (tocilizumab) [
1]. All these biologic agents have proven efficacy in stopping joint inflammation and structural damage in association with the anchor drug methotrexate (MTX) [
1]. No one molecule has proven its clinical superiority over others in terms of efficacy [
2‐
6]. In addition, no response to these treatments is obtained in approximately 30% of patients with RA, and the response to all these medications is highly variable from one patient to another. Nevertheless, prescription of biologics agents remains highly empirical [
7‐
9]. Presently, we are still unable to predict the clinical efficacy of these treatments in a given patient because of the heterogeneity of RA and the subgroups of patients susceptible to responding better to one molecule than to another. Given the increasing number of available molecules, it is crucial to identify predictive markers in order to optimize drug prescription only to those patients susceptible to responding and thus avoiding side effects.
In the current context of personalized medicine, large-scale analysis of gene expression to predict drug response is a relevant and original approach, which has already shown its utility in cancer or kidney transplantation [
10‐
12]. Gene expression profiling is clearly a powerful method for the identification of biomarkers and the development of personalized medicine in the field of rheumatology [
13‐
16]. This approach has already allowed us to identify and validate two gene combinations able to predict response to infliximab or anakinra although in small cohorts of patients with RA [
17,
18]. The proof of concept of gene expression profiles or signatures as a predictor of good response to drugs in RA was further confirmed by other teams using other biologic agents, but in cohorts that included very few patients with RA [
19‐
21]. Thus, we conducted this present ancillary study as a follow-up to the APPRAISE trial [
22,
23]. Our main objective was to identify and validate a gene expression profile associated with a good response to ABA/MTX, and to understand the involvement of this signature in RA pathophysiology.
Methods
Patients from the APPRAISE trial
A total of 68 patients with RA from the APPRAISE trial were enrolled in this ancillary study [
22,
23]. The APPRAISE study assessed the capability of the composite power of Doppler and greyscale ultrasound score to measure the early effect and time course of response to treatment with ABA in biologic-naïve patients with active RA despite MTX therapy. APPRAISE (NCT00767325) initially including 104 patients with RA was a 24-week, Phase IIIb, open-label, multicenter, single-arm study conducted at 21 sites across Europe (Denmark, France, Germany, Hungary, Italy, Norway, Spain and the UK) [
22,
23]. Eligible patients were ≥18 years of age, had American College of Rheumatology (ACR)-defined RA according to the 1987 classification criteria for at least 6 months [
24], and had been on MTX (≥15 mg/week) for at least 3 months prior to baseline, with a stable MTX dose for at least 28 days before baseline (except in cases of intolerance to MTX). Patients were required to have active disease, defined by a baseline disease activity score in 28 joints (calculated with C-reactive protein (CRP)) (DAS28(CRP)) >3.2 or tender and ≥6 swollen joint counts, and CRP above the upper limit of normal. All patients received intravenous (IV) infusions of ABA at a weight-titered dose of 10 mg/kg at baseline (day 1), and at weeks 2, 4, 8, 12, 16, 20 and 24, in addition to stable doses of concomitant MTX (≥15 mg/week). MTX dose increases were not permitted, and dose decreases were allowed only in cases of intolerance. Oral corticosteroid use (stable dose of ≤10 mg prednisone/day) was permitted during the study. For this study, 5 ml of whole blood was collected in PAXgene RNA tube (PreAnalytiX, Qiagen) just before the first infusion and 6 months later and was stored at –80 °C until use.
Clinical evaluation and response to ABA/MTX
Several clinical characteristics were collected at baseline and 6 months later: age, gender, disease duration and MTX and corticosteroid doses. Disease activity was evaluated at all assessment visits (baseline, weeks 1, 2, 4, 6, 8, 12, 16, 20 and 24) using the DAS28(CRP) calculated from 28 tender joints, 28 swollen joints, CRP and patient global assessment (visual analog scale (VAS); 0–10 scale).
The response to ABA/MTX was evaluated at 6 months using European League Against Rheumatism (EULAR) response criteria based on DAS28(CRP) [
25]. Since we were looking for biomarkers associated with good response to ABA/MTX, patients were categorized according to their EULAR response as responders (R) (
n = 36) or non-responders (NR) (comprising moderate responders (
n = 25) and no responders (
n = 7)) [
25].
RNA preparation
Total RNAs from whole blood were extracted with PAXgene blood RNA kit according to the manufacturer’s recommendations (Qiagen PreAnalytiX GmbH, Courtaboeuf, France) and stored at –80 °C until use. Total RNA from 10 healthy donors (5 women and 5 men) was pooled and used as an internal standard reference (control pool). The quality and quantity of isolated mRNAs were assessed using the 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA) and the Nanodrop device (Thermo Scientific, Wilmington, USA). Only RNA samples with a minimal RNA integrity number of 7 were used for subsequent experiments.
Microarrays
Whole human genomic DNA microarrays were used to analyze two-colored gene expression profiling (4 × 44 K Whole Human Genome, Agilent Technologies, Les Ulis, France). Each RNA sample from patients with RA was labeled by Cyanine-5 and co-hybridized with a Cyanine-3 labeled RNA control pool according to the manufacturer’s instructions (Low Input QuickAmp Labeling Kit, Agilent Technologies, Les Ulis, France). Briefly, 100 ng of RNAs were labeled with cyanine-5 CTP (patients with RA) or cyanine-3 CTP (control pool). After hybridization reaction using a hybridization kit (Agilent Technologies, Les Ulis, France) co-hybridization was performed at 65 °C for 17 hours. After wash steps, the microarrays were scanned with a 5-μM pixel size using the DNA Microarray Scanner GB (Agilent Technologies, Les Ulis, France). Image analysis and extraction of raw and normalized signal intensities (lowess) were performed using Feature Extraction Software 10.5.1.1 (Agilent Technologies). The data were in agreement with the guidelines for minimum information about a microarray experiment and were deposited in the database of the National Center for Biotechnology Information Gene Expression Omnibus (
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE68215). The data are accessible [GEO:GSE68215]. Non-uniform spots and saturated spots or spots with intensities below the background were not taken into account. Only spots that passed these quality controls on 100% of arrays were selected for further analysis. Hierarchical clustering was performed using the Pearson coefficient metric and complete linkage to build the transcripts and sample dendrograms.
Quantitative reverse transcription-PCR (qRT-PCR)
cDNA was synthesized from 1-μg RNA samples using random primers and M-MLV enzyme (Invitrogen™, Carlsbad, USA). qRT-PCR was performed using a Lightcycler as instructed by the manufacturer (Roche™, Meylan, France). qRT-PCR reactions were performed for each sample in duplicate using SYBR-Green (Roche™, Meylan, France) and values were normalized using the geometric mean of three control genes (
18S, ACTB, B2M) defined by the geNorm algorithm [
26]. Sequences of primers (Eurogentec™, Fremont, USA) used for qRT-PCR were:
BLOC1S1 forward, 5’-AAGCAGACAGGCCAGTGGAT-3’;
BLOC1S1 reverse, 5’-CAGTGCAGTGGCAATGGTG-3’;
RNASE3 forward, 5’-CAGGAGCCACAGCTCAGAGA-3’;
RNASE3 reverse, 5’-GAGCCCTCCACACCCATAAG-3’;
COX6A1 forward, 5’-CCACTTCCAACTGGCTACGA-3’;
COX6A1 reverse, 5’-AAGCAAAGGGATGGGAGACC-3’;
PTRH2 forward, 5’-GCTGTTGGAGTTGCTTGTGG-3’;
PTRH2 reverse, 5’-AGGCTGAAACAGCAGCATGA-3’;
18S forward, 5’-GTGGAGCGATTTGTCTGGTT-3’;
18S reverse, 5’-CGCTGAGCCAGTCAGTGTAG-3’;
ACTB forward, 5’-CTGGAACGGTGAAGGTGACA-3’;
ACTB reverse, 5’-AAGGGACTTCCTGTAACAATGCA-3’;
B2M forward, 5’-TGCTGTCTCCATGTTTGATGTATCT-3’;
B2M reverse, 5’-TCTCTGCTCCCCACCTCTAAGT-3’.
Statistical and functional analysis
Comparisons of clinical and biological data between R and NR were performed at baseline using Student’s t test for continuous variables. Comparisons of R or NR before and after treatment were performed using the paired t test. Identification of clinical parameters able to predict good response to ABA/MTX was performed in two different multivariate analyses: (1) a logistic regression model with variable selection using Bayesian information criterion (BIC) and (2) linear discriminant analysis (LDA), which estimates a coefficient for each variable.
Data from transcriptomic analysis were analyzed using GeneSpring GX V.13.0 (Agilent Technologies, Les Ulis, France). The normality of log 2 ratio of gene expression was evaluated using the Shapiro–Wilk statistical test. The unpaired Student’s
t test (
p value < 0.05), with the Benjamini–Hochberg correction to check the false discovery rate (FDR), was used to determine the statistical significance of differences in gene expression levels between R and NR. Gene Ontology (GO) analysis was used to investigate the biological processes, molecular function or cellular localization enriched in the transcripts list, showing a significant fluctuation in gene expression between R and NR. The
p value was computed by standard hypergeometric distribution. The GeneSpring Single Experiment Analysis (SEA) bio-informatics tool was used for computational analysis to identify potential curated canonical pathways with setting parameters (reactome and GenMAPP for pathway source), which are enriched in the differentially expressed transcripts list, using the WikiPathways database (
http://www.wikipathways.org/index.php/Pathway:WP111). The significance of the association between the genes and the pathways was measured by Fisher’s exact test.
Discussion
In this ancillary study, we identified 87 transcripts with relative abundances, which were able to separate R and NR to ABA/MTX in 36 patients with RA before treatment (Fig.
3 and Additional file
2). Next, for the first time we validated a minimal combination of four transcripts associated with response to ABA/MTX in an independent subset of 32 patients with RA (Fig.
4). According to GO and WikiPathways analysis, we identified enrichment of 7 genes among the aforementioned 87 belonging to the ETC pathway, suggesting a specific signature of response (Fig.
5). Indeed, patients with RA with pre-silencing of the ETC pathway before treatment were those most susceptible to responding to ABA/MTX 6 months later.
In the current context of personalized medicine for the management of RA, the optimization of drug prescription is crucial. Hence, drug prescription should be tailored exclusively only to patients susceptible to responding to the drug. The identification of biomarkers to predict of drug response is paramount. To date, there are only two biomarkers associated with good response to ABA/MTX in RA, anti-CCP positivity and low baseline number of CD8
+CD28
– T cells [
27‐
30]. Furthermore, there is no clinical and/or biological biomarker used in routine practice able to predict response to ABA/MTX prior to treatment initiation. Indeed, in our study, two statistical approaches to multivariate analysis were unable to identify predictive clinical or biological parameters of response to ABA/MTX, even if disease duration, CRP, global patient assessment of disease and DAS28(CRP) were significantly different between R and NR at baseline (Table
1 and Fig.
2). It would be interesting to establish a relationship between response to ABA and anti-citrullinated protein antibodies (ACPA) and/or rheumatoid factor status, given that ACPA positivity is associated with better response to ABA in RA. Unfortunately, we were unable to include ACPA or rheumatoid factor status in our analysis because these data were not collected during the APPRAISE trial and were not available. The difficulties we faced in predicting drug response with clinical and/or biological variables led us to use a more global functional genomic approach without a priori having to identify a signature associated with good response to ABA/MTX.
The transcriptomic approach based on the whole genome allowed us to identify signatures able to separate R from NR to several drugs (infliximab, tocilizumab, rituximab) used in RA [
17‐
21,
31]. For ABA, one study previously measured the type-I interferon-regulated transcripts from peripheral blood mononuclear cells in patients with RA, independently of response to ABA [
32]. Another study identified gene sets associated with remission according to clinical disease activity index (CDAI) in ABA-treated patients with RA [
33]. In our study, even if there were differences at baseline in disease duration, CRP, patient assessment of disease and DAS28(CRP) between R and NR, which might have influenced gene expression, we identified and validated a gene signature associated with response to ABA/MTX. This signature included 87 genes variously involved in the ETC, proteasome, interferon and RNA processes, etc. This combination of 87 genes is also specific to ABA as it was not able to predict TNFα blocking agent response. Next, we validated this signature with an independent subset of 32 patients with RA by means of a minimal combination of four genes with expression levels measured by qRT-PCR, which are more easily usable in routine practice than microarrays. This signature is the optimized combination of genes associated with drug response with good accuracy, because it correctly predicted the future response in 81% of patients with RA (75% sensitivity, 85% specificity, 85% negative predictive value and 75% positive predictive value) prior to treatment (Fig.
4). Each gene taken separately was unable to predict response to ABA/MTX (data not shown) but all genes together were associated with response to ABA/MTX with good accuracy. Nevertheless, further validation studies in independent cohorts are essential before considering this signature as a predictive biomarker and of use in clinical practice.
Of the four genes in the minimal combination, the
RNAseIII gene codes for the RNAseIII enzyme that specifically cleaves double-stranded RNA and is involved in the processing of ribosomal RNA precursors of some mRNAs [
34]. Biogenesis of lysosomal organelles complex-1, subunit 1 (
BLOC1S1) codes for the protein BLOC1S1, also known as GCN5L1, and is an essential component of the mitochondrial acetyltransferase machinery and modulates mitochondrial respiration via acetylation of ETC proteins [
35].
COX6A1 codes for the mitochondrial protein cytochrome c oxidase subunit 6A1 (COX6A1) located in complex IV. This is the last enzyme in the mitochondrial ETC that drives ATP synthesis [
36].
PTRH2 codes for the peptidyl-tRNA hydrolase 2, which is a mitochondrial protein released from mitochondria to the cytoplasm during apoptosis.
Out of these 4 transcripts, 3 (
BLOC1S1,
COX6A1 and
PTRH2) and 13 probes out of 87 code for proteins located in mitochondria. In addition, some of them are involved in the ETC pathway, suggesting implication of mitochondrial metabolism in response to ABA/MTX. Moreover, GO analysis and SEA are in agreement as these two analyses point to the involvement of ETC in the response to ABA/MTX. Interestingly, we found that seven transcripts with levels that are significantly lower in R than in NR at baseline were similarly regulated in patients in remission and patients not in remission defined by clinical disease activity index (CDAI) in a recent study [
33]. The ETC is a series of five complexes anchored to the inner membrane of mitochondria that transfers electrons via redox reactions, which drives ATP synthesis, generating reactive oxygen species (ROS) and subsequent oxidative stress [
37]. Redox balance in mitochondria is a critical component in T cell activation and proliferation [
38]. The production of ROS by the ETC complex III leads to production of large amounts of ATP to enhance activity of proliferating T cells after TCR cross-linking [
39‐
42]. In our study, ETC genes were downregulated in R compared to NR, suggesting potentially low ROS production and potentially less T cell activation before treatment. Also, a previous study suggested it would be interesting to determine whether ROS production in T cells might be a predictor of clinical response to ABA [
43]. Our results suggest the possible involvement of ROS in the pathophysiological processes of ABA response, but further functional studies are necessary to confirm this hypothesis.
This 87 mRNA signature also included
RASSF5, which was significantly upregulated in R compared to NR.
RASSF5, also known as
RAPL, is the effector of Rap1, which plays a central role in T cell response through TCR and co-stimulation signals. Indeed, a model was proposed in which inactivation of Rap1 plays a central role in establishing oxidative stress and can influence T cell response in RA [
43,
44]. These data suggest less oxidative stress in future responders to ABA while NR present with high expression of genes from the ETC pathway, showing oxidative stress. A reduction in the expression level of ETC genes seems to increase the sensitivity of patients with RA to ABA/MTX. This model has already been highlighted in esophageal adenocarcinoma and colorectal cancer treated by chemotherapy [
45,
46]. Indeed,
ATP5J and
COX7A2 included in our combination were also found to be downregulated and associated with response to chemotherapy, respectively in colorectal cancer and esophageal adenocarcinoma [
45,
46]. As in cancer, pre-treatment conditions targeting the mitochondrial metabolism might be a determinant of susceptibility to ABA/MTX [
47].
After 6 months of treatment with ABA, we showed that five genes (COX7B, COX11, UQCRC2, NDUFU3, NDUFS1) involved in the ETC pathway were significantly upregulated in R, while these genes were invariant in NR to ABA/MTX. So, whereas ABA does not seem to affect the oxidative state in NR, it seems to modulate oxidative stress in R, as genes from the ETC pathway were upregulated under ABA. This drug restores the expression of genes involved in redox balance.
Acknowledgements
The authors would like to thank Manuela Le Bars and Corine Gaillez from Bristol-Myers Squibb for designing the study and providing study logistics. They also wish to thank the principal investigators of the APPRAISE study. APPRAISE principal investigators are Silvano Adami, Vivi Bakkenheim, Hilde Berner Hammer, Stefano Bombardieri, Maria-Antonietta D'Agostino, Paul Emery, Liana Euller-Ziegler, Gianfranco Ferraccioli, Maurizio Galeazzi, Philippe Gaudin, Walter Grassi, Annamaria Iagnocco, Herbert Kellner, Thierry Lequerré, Ingrid Möller, Esperanza Naredo, Mikkel Østergaard, Fredeswinda Romero, Istvan Szombati, Lene Terslev, Jacqueline Uson, Esther Vicente, Olivier Vittecoq and Richard Wakefield. The authors are grateful to Nikki Sabourin-Gibbs, Rouen University Hospital, for her help in editing the manuscript.