ARTHRITIS & RHEUMATOLOGY Vol. 67, No. 2, February 2015, pp 344–351 DOI 10.1002/art.38947 © 2015, American College of Rheumatology
Modular Analysis of Peripheral Blood Gene Expression in Rheumatoid Arthritis Captures Reproducible Gene Expression Changes in Tumor Necrosis Factor Responders Michaela Oswald,1 Mark E. Curran,2 Sarah L. Lamberth,2 Robert M. Townsend,3 Jennifer D. Hamilton,4 David N. Chernoff,5 John Carulli,6 Michael J. Townsend,7 Michael E. Weinblatt,8 Marlena Kern,1 Cassandra M. Pond,1 Annette Lee,1 and Peter K. Gregersen1 analysis. Strikingly, nonresponders exhibited very little change in any modules, despite exposure to TNF blockade. These patterns of change were highly consistent across all 3 cohorts, indicating that immunologic changes after TNF treatment are specific to the combination of both drug exposure and responder status. In contrast, modular patterns of gene expression did not exhibit consistent differences between responders and nonresponders at baseline in the 3 study cohorts. Conclusion. These data provide evidence that using gene expression modules related to inflammatory disease may provide a valuable method for objective monitoring of the response of RA patients who are treated with TNF inhibitors.
Objective. To establish whether the analysis of whole-blood gene expression is useful in predicting or monitoring response to anti–tumor necrosis factor (antiTNF) therapy in patients with rheumatoid arthritis (RA). Methods. Whole-blood RNA (using a PAXgene system to stabilize whole-blood RNA in the collection tube) was obtained at baseline and at 14 weeks from 3 independent cohorts, consisting of a combined total of 240 RA patients who were beginning therapy with anti-TNF. We used an approach to gene expression analysis that is based on modular patterns of gene expression, or modules. Results. Good and moderate responders according to the European League Against Rheumatism criteria exhibited highly significant and consistent changes in multiple gene expression modules after 14 weeks of therapy, as demonstrated by hypergeometric
The development of tumor necrosis factor (TNF) inhibitors for the treatment of rheumatoid arthritis (RA) and other inflammatory disorders has been a seminal advance for the field of rheumatology (1). However, therapeutic approaches to RA remain a substantial clinical
Supported by the NIH (grant N01-AR1-2256), the Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP) Consortium (Biogen Idec, Bristol Myers-Squibb, Crescendo Bioscience, Genentech, Janssen Research and Development, Regeneron Pharmaceuticals, and SanofiAventis), and Janssen Research and Development. 1 Michaela Oswald, PhD, Marlena Kern, DNP, Cassandra M. Pond, BS, Annette Lee, PhD, Peter K. Gregersen, MD: Feinstein Institute for Medical Research, Manhasset, New York; 2Mark E. Curran, PhD, Sarah L. Lamberth, PhD: Janssen Research and Development, Springhouse, Pennsylvania; 3Robert M. Townsend, PhD: Bristol-Myers Squibb, Princeton, New Jersey; 4Jennifer D. Hamilton, PhD: Regeneron Pharmaceuticals, Tarrytown, New York; 5David N. Chernoff, MD: Chernoff Consulting, LLC, Sausalito, California; 6John Carulli, PhD: Biogen Idec, Cambridge, Massachusetts; 7Michael J. Townsend, PhD: Genentech, South San Francisco, California; 8 Michael E. Weinblatt, MD: Brigham and Women’s Hospital, Boston, Massachusetts. Drs. Curran and Lamberth own stock or stock options in Janssen, a subsidiary of Johnson & Johnson. Dr. R. M. Townsend owns stock or stock options in Bristol-Myers Squibb. Dr. Hamilton owns stock or stock options in Regeneron Pharmaceuticals. Dr. Chernoff
has received consulting fees from Crescendo Bioscience and Adamas Pharmaceuticals (more than $10,000 each) and is listed on patents for the use of multiple biomarker protein panels in autoimmune diseases and on patents for the use of aminoadamantanes for treatment of neurobiologic disorders. Dr. Carulli owns stock or stock options in Biogen Idec. Dr. M. J. Townsend owns stock or stock options in Roche. Dr. Weinblatt has received consulting fees and/or honoraria from Biogen Idec, Sanofi-Aventis, Crescendo Bioscience, BristolMyers Squibb, Genentech, and Regeneron Pharmaceuticals (less than $10,000 each). Dr. Gregersen has received an honorarium from Biogen Idec (less than $10,000) and research grants from the BATTER-UP Consortium. Address correspondence to Peter K. Gregersen, MD, Robert S. Boas Center for Genomics and Human Genetics, Feinstein Institute for Medical Research, 350 Community Drive, Manhasset, NY 11030. E-mail: [email protected]. Submitted for publication June 24, 2014; accepted in revised form October 30, 2014. 344
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
challenge, since ⬃30% of RA patients do not respond to TNF blockade, and many others obtain only partial reduction in joint inflammation (2). In addition, the determination of the clinical response to drug therapy is an inexact science, largely relying on both the subjective perceptions of the patients as well as the physician’s assessments of disease activity, both of which exhibit considerable variability. The integration of laboratory measures of inflammatory activity, such as the C-reactive protein level or the erythrocyte sedimentation rate, is standard but nonspecific, and these measurements often do not parallel the clinical parameters (3). The recent development of multiparameter methods for determining disease activity may be helpful (4,5), but these biomarkers have not yet been widely used in formal studies of clinical drug response or been demonstrated to benefit patient management. Thus, there is a substantial unmet need for improvement in both the prediction of drug response in order to individualize therapy and the development of quantitative laboratory methods to determine clinical activity of disease, which in turn, will aid in the development of predictive biomarkers and enable more precise treatment concepts such as “treat to target” (6). Expression microarrays have been used by a number of groups in the development of biomarkers of drug response (7–15). The early applications of this technology, however, have been compromised by a number of issues (16). Use of small sample sizes for initial feature set definition lead to overfitting of discriminative models, and such studies are further compromised by the semiquantitative nature of disease activity phenotypes. Critically, there has been rather limited replication of specific predictive gene expression patterns between the studies conducted to date. Although transcriptomics is a maturing technology, there are various approaches to expression analysis, and sample collection methods are often not standardized. Many studies examine gene expression in purified peripheral blood mononuclear cells. This can lead to experimental noise due to extracorporeal changes in gene expression, especially if RNA purification is not carried out immediately. More recently, use of the PAXgene Blood RNA system or other technologies that stabilize whole-blood RNA in the collection tube has helped to reduce this source of variability. Of course, this does not permit the analysis of expression in specific separated cell subsets of interest. Overall, none of these studies have provided convincing, replicated results using a standardized protocol for sample collection, microarray platform, and method of analysis (16).
345
In the current study, we attempted to address these issues by examining peripheral blood gene expression using a standard approach to blood sample collection (PAXgene) and using standard Illumina array platforms in 3 independent, prospectively collected cohorts of RA patients beginning therapy with TNF inhibitors. For our primary approach to analysis, we also took advantage of an empirically derived set of “modules” of gene expression developed by Chaussabel and colleagues (17,18). The results reveal that patients who have a clinical response to TNF inhibition exhibit a common pattern of changes in these gene expression modules after 14 weeks of therapy, whereas clinical nonresponders do not have major changes in modular gene expression. In contrast, we were not able to identify modular gene expression changes at baseline that identify patients who will or will not respond to TNF blockade. These results suggest to us that global gene expression patterns in peripheral blood can be a useful adjunct for determining whether clinical response has occurred. However, we expect that successful prediction of response to therapy is likely to require a focus on particular cell subsets, along with the integration of genetic factors and additional biomarker data. PATIENTS AND METHODS Patient cohorts. Three prospectively collected cohorts of patients with RA were studied, as summarized in Table 1. The Autoimmune Biomarkers Collaborative Network (ABCoN) cohort was enrolled as part of the National Institute of Arthritis and Musculoskeletal and Skin Diseases– supported contract to develop new approaches to biomarkers for RA and lupus. A total of 50 ABCoN patients had complete data available for study. Patients were treated with infliximab (n ⫽ 20), etanercept (n ⫽ 21), or adalimumab (n ⫽ 9). At 14 weeks, the following distribution of European League Against Rheumatism (EULAR) responses was observed: 14 had a good response, 21 had a moderate response, and 15 were nonresponders. A second patient cohort was taken from the phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate Therapy (GOFURTHER) (19). All patients were started on 2 mg/kg of golimumab intravenously, and clinical assessment of response was determined at 14 weeks. A total of 72 patients had gene expression data available for analysis. The EULAR responses in these patients were as follows: 29 had a good response, 37 had a moderate response, and 6 were nonresponders. A third patient cohort of 118 RA patients was taken from a multicenter cohort called Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP). These patients were treated with a variety of anti-TNF agents, including infliximab (n ⫽ 23), etanercept (n ⫽ 31), golimumab (n ⫽ 9), adalimumab (n ⫽ 41), and certolizumab pegol (n ⫽ 14). EULAR response rates for the BATTER-UP cohort were as follows: 42 had a
346
OSWALD ET AL
Table 1. Clinical characteristics of the rheumatoid arthritis study patients from the ABCoN, GO-FURTHER, and BATTER-UP cohorts*
Age, mean ⫾ SD years No. female/male No. (%) CCP⫹ and/or RF⫹ DAS, mean ⫾ SD EULAR response group, no. of patients Good responder Moderate responder Nonresponder TNF inhibitor, no. of patients Infliximab Etanercept Golimumab Adalimumab Certolizumab pegol Race/ethnicity, no. of patients White Black Asian Hispanic Other No. of days from visit 1 to visit 2 Range Mean ⫾ SD
ABCoN (n ⫽ 50)
GO-FURTHER (n ⫽ 72)
BATTER-UP (n ⫽ 118)
55.62 ⫾ 11.7 39/11 43 (86) 5.24 ⫾ 0.79
50.36 ⫾ 10.91 53/19 72 (100) 6.01 ⫾ 0.84
53.3 ⫾ 13.7 96/22 74 (62.7) 5.12 ⫾ 1.1
14 21 15
29 37 6
42 37 39
20 21 0 9
0 0 72 0
23 31 9 41 14
44 2 2 2 0
72 0 0 0 0
100 13 2 7 3
91–160 103.4 ⫾ 11.9
95–110 99.3 ⫾ 1.94
70–182 105.1 ⫾ 19.49
* ABCoN ⫽ Autoimmune Biomarkers Collaborative Network; GO-FURTHER ⫽ phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate; BATTER-UP ⫽ Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients; CCP ⫽ citrullinated peptide; RF ⫽ rheumatoid factor; DAS ⫽ Disease Activity Score; EULAR ⫽ European League Against Rheumatism; TNF ⫽ tumor necrosis factor.
good response, 37 a moderate response, and 39 were nonresponders. Microarray analysis. All subjects had blood drawn into PAXgene tubes at baseline (before starting therapy) and again at 14 weeks. RNA was isolated using a Qiagen QIAcube system, following the manufacturer’s protocol for PAXgene Blood RNA part 1, automated protocol. Extracted samples were eluted in 80 l of elution buffer (BR5) and were subsequently run on an Agilent 2100 Bioanalyzer, for integrity, using an RNA 6000 NanoChip. Samples with RNA integrity numbers ⬎6.5 were diluted to 30 ng/l in a total of 11 l of RNase-free water. Samples were amplified using an Illumina RNA Total Prep Amplification kit (Life Technologies). Then, 750 ng of complementary RNA was resuspended in 5 l of RNase-free water for analysis on a Human HT-12v4 Expression BeadChip, and 1.2 g was resuspended in 10 l for analysis on a WG-6v3 BeadChip (all from Illumina). All samples were processed according to the manufacturer’s instructions. The Illumina WG-6v3 microarray was used for samples in the ABCoN cohort, while the HT-12v4 microarray was used for samples in the GO-FURTHER and BATTERUP cohorts. Raw data were exported from GenomeStudio and further analyzed with the R programming language. All 3 data sets were background corrected using the R/Bioconductor package Lumi (20). Data were further transformed using a variance stabilization transformation and quantile normalized. The moderated T statistic from the R/Bioconductor package Limma (21) was used for differential expression analysis. We analyzed modules of coexpressed genes that have previously demonstrated coordinated regulation of expression across a set of diseases (whole blood), as described by
Chaussabel et al (17). A total of 28 modules were analyzed, each containing 22–325 gene probes. Because modules were originally defined using Affymetrix probe sets, these probes were mapped to Entrez Gene IDs, which in turn, were mapped to Illumina probes, as shown in Figure 1. Group comparisons (e.g., modular expression at baseline versus 14 weeks; responders versus nonresponders) were carried out using a moderated
Figure 1. Flow chart summarizing the experimental approach and data analysis in this study of rheumatoid arthritis patients receiving anti–tumor necrosis factor therapy. Gp ⫽ group.
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
T statistic comparing groups and categorizing all individual probes as up-regulated or down-regulated, using a significance cutoff of P ⬍ 0.05. For each module, we compiled a list of all significantly changed probes (up or down) at the P ⬍ 0.05 cutoff for significance. We then determined what fraction of the probes was significantly up-regulated or down-regulated within a given module. When comparing groups (e.g., responders versus nonresponders), the significance of the observed difference in distribution of up-/down-regulated probes in a module was calculated using a hypergeometric test (equivalent to a 2 ⫻ 2 table analysis using Fisher’s exact test).
RESULTS Consistent changes in modular gene expression after 14 weeks of anti-TNF therapy in clinical responders. As originally reported by Chaussabel et al (17), 28 gene expression modules can be observed in whole blood across a variety of inflammatory and infectious diseases. Where possible, these modules were labeled after the fact, according to the identity of the gene transcripts present in these modules (e.g., B
347
cells, myeloid lineage, T cells, interferon-inducible, etc.). However, many modules could not easily be categorized and were designated as “undetermined.” We used these designations as well as the original module numbers (M1.1–1.8, M2.1–2.10, and M3.1–3.9) in our study. We initiated analysis by investigating whether changes in modular gene expression occurred after TNF blockade, and if so, whether the changes differ among responders and nonresponders at 14 weeks. The results for the 3 data sets are summarized separately for EULAR responders (good and moderate responders combined) and EULAR nonresponders in Figure 2. For each module, the direction of change in module expression is shown for each data set. For the responder groups, there was a consistent and statistically significant increase in modules containing transcripts related to plasma cells (M1.1), B cells (M1.3), major histocompatibility complex and ribosomal proteins (M1.7), and T cells (M2.8), as well as a variety of “undetermined”
Figure 2. Summary of changes in transcriptional modules (M1.1–1.8, M2.1–2.11, and M3.1–3.9) at 14 weeks compared with baseline in responders (R) and nonresponders (NR) to anti–tumor necrosis factor (anti-TNF) therapy. The 3 sets of bars for each module indicate the 3 cohorts studied: the Autoimmune Biomarkers Collaborative Network (ABCoN) cohort (A), the phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate (GO-FURTHER) cohort (G), and the Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP) cohort (B). Modules are designated according to the original nomenclature (see ref. 17), with descriptions accompanying selected modules. The height of the bars for the responders shows the fraction of probes significantly changed (up or down; y-axis shows fractional change for responders only) within each module. Solid bars ⫽ P ⬍ 0.0005; darkly shaded bars ⫽ P ⬍ 0.005; lightly shaded bars ⫽ P ⬍ 0.05; open bars ⫽ P not significant. MHC ⫽ major histocompatibility complex; Undet. ⫽ undetermined.
348
OSWALD ET AL
Figure 3. The distribution of significantly changed (P ⬍ 0.05) probes within the modules between baseline and 14 weeks (see Patients and Methods for details). The responders and nonresponders are shown for each of the 3 data sets: A, ABCoN, B, GO-FURTHER, and C, BATTER-UP. Colors indicate probes in modules with significant changes up (red; sig_up) or down (blue; sig_down) or with no significant modular changes (green; not_sig) (P ⬍ 0.005 by hypergeometric test). Many of these modules exhibit changes with extremely high significance. For example, P values for module 2.5 are 10⫺37, 10⫺82, and 10⫺68 in the ABCoN, GO-FURTHER, and BATTER-UP cohorts, respectively. See Figure 2 for other definitions.
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
349
Figure 4. Comparison of differences in transcriptional modules between responders and nonresponders in the 3 study cohorts at baseline, before initiation of tumor necrosis factor blockade: ABCoN (A), GO-FURTHER (G), and BATTER-UP (B). Modules are designated according to the original nomenclature (see ref. 17), with descriptions accompanying selected modules. The height of the bars shows the fraction of probes that are different between responders and nonresponders. Solid bars ⫽ P ⬍ 0.0005; darkly shaded bars ⫽ P ⬍ 0.005; lightly shaded bars ⫽ P ⬍ 0.05; open bars ⫽ P not significant. See Figure 2 for definitions.
modules (M1.8, M3.4, M3.7, M3.8, and M3.9). In addition, there was a consistent down-modulation of transcripts in several myeloid lineage modules (M1.5 and M2.6) and in platelets (M2.1), as well as in inflammation 1 (M3.2) and inflammation 2 (M3.3). These results stand in marked contrast to the lack of consistent modular changes across the 3 data sets for nonresponders. Indeed, in general, nonresponders showed a very small fraction of probes that significantly changed in comparisons of data at 14 weeks versus baseline within any of the 28 modules. In order to provide a more detailed representation of these modular changes and their magnitude, we plotted for each data set and responder group the fold changes for all probes achieving a significant difference (P ⬍ 0.05 by moderated t-test) within each module. Figures 3A–C show these results for the ABCoN, GOFURTHER, and BATTER-UP cohorts, respectively. Within the responders from the 3 data sets, there was striking consistency in the pattern of changes in the modules. Significant changes (P ⬍ 0.005 by hypergeometric test) in probe distribution, whether up or down, are indicated in the figures. Indeed, some of the changes achieved remarkably compelling levels of statistical significance. For example, the P values associated with a reduction in myeloid modules M1.5 and M2.6 after TNF blockade ranged from 10–20 to 10–87, and reductions in “undetermined” modules M3.2 and M3.3 ranged from 10⫺12 to 10⫺85 across the various data sets. These dramatic changes were only seen in the responder group, with minimal change in the nonresponder group.
No consistent differences in modules between responders and nonresponders at baseline. While substantial changes in modular gene expression were seen in responders after TNF blockade, the baseline pattern of module expression did not show consistent differences among the 3 cohorts (Figure 4). While some of the data sets suggested the presence of meaningful differences between responders and nonresponders at baseline in 1 cohort (e.g., several “undetermined” modules were relatively less prominent in responders in the ABCoN cohort), these changes were not replicated in the other 2 cohorts. This finding emphasizes the importance of having replication data sets in which to validate the initial results and demonstrates that an improved understanding of RA patient heterogeneity will be required for the development of an accurate predictor of drug response. DISCUSSION In this study, we demonstrated a consistent change in whole-blood gene expression in clinically responding RA patients after exposure to TNF-blocking agents for 14 weeks. These changes were not observed in patients who do not exhibit a clinical response to treatment, as determined using the standard EULAR criteria. The data are based on a consideration of sets of genes that are expressed together in a modular pattern across different inflammatory conditions, as first demonstrated by Chaussabel et al (17). These data provide compelling evidence that these gene expression changes
350
are not simply a result of exposure to TNF blockade, but rather, they reflect a change in both the immunologic and disease state in patients who are responding to TNF therapy. Interestingly, we have recently had an opportunity to examine a small number of patients who did not receive study drug (placebo group; 7 responders and 16 nonresponders). There was no difference between responders and nonresponders in this group (data not shown), suggesting that the changes are specific to the physiologic effects of drug therapy in the setting of a clinical response. A larger study will be required to confirm these findings. In contrast to the gene expression changes observed in responders over time, the modules did not predict future response to TNF blockade when these modular expression patterns were interrogated at baseline, before starting therapy, as shown in Figure 4. The development of useful biomarkers for the management of RA has been extremely challenging. At present, there is no clinical parameter or laboratory test that can distinguish likely responders to anti-TNF blockade prior to initiation of therapy. Given the fact that at least 25% of patients will not respond to anti-TNF therapy, while at most, one-third will have a robust clinical response, there is clearly a need to develop such markers. Several small studies have proposed biomarkers to predict such a response, but none have been validated for any biologic agent (16). It is important to note that published studies suffer from reporting primary results and have not had multiple independent cohorts available for validation and replication of results. A primary component of our present work is replication of findings in multiple independent cohorts. In this context, our use of modular approaches to whole-blood microarray analysis identified suggestive feature sets for prediction of response in individual cohorts, but these marker sets failed to replicate or yield a convincing gene expression biomarker for future drug response. This emphasizes the biologic complexity of the RA response phenotype. If the scientific community is to succeed in the goal of predicting response to treatment, it is likely that multiple parameters will need to be combined, including genetic and genomic data, along with cellular and biochemical data that more precisely reflect the underlying TNF dependence of the current inflammatory state and drivers of inflammation in a given patient. Our current results are likely to have utility in the context of determining the presence of reduced disease activity in response to drug therapy. This is an important issue, since the current measures, such as the American College of Rheumatology criteria and the EULAR
OSWALD ET AL
criteria, are highly dependent on primarily subjective patient-based and physician-based data points. The current results suggest that analysis of modular gene expression can enhance the accuracy of a responder versus nonresponder designation. For example, while the responder and nonresponder groups showed consistent differences in the pattern of change in these modules, individuals may have a more or less robust change. In the future, we will generate an individually applicable “modular change score.” A diagnostic biomarker of this type may be useful in assigning likelihood for membership in the responder or nonresponder group, allow for decisions on change from current therapy, and allow us to track the timing and durability of the response. As part of such an effort, we plan to incorporate other current measures of disease activity, such as the proteomic panel used in the Vectra DA test (4,22). Development of an improved RA disease status scoring method will require careful exploration of combinations of modules and other laboratory measures in order to optimize and validate such a method. Fortunately, an additional validation data set of over 100 additional subjects in the BATTER-UP study will shortly be available to permit a robust replication of the current findings as well as new combinatorial approaches to individual categorization of response at a molecular level. Our preliminary findings suggest that some patients who have a clinical response may not show strong evidence of a biologic response, possibly due to a placebo effect. Likewise, patients with a biologic response may exist within the “nonresponder” group, again underlining the heterogeneity in the mechanisms of disease. Methods to better classify patients quantitatively with regard to biologic response will facilitate clinical treatment decisions and enable the practice of “individualized medicine” (23), including such concepts as treating to target (6). Finally, the outcome data and molecular interpretation presented herein are restricted to anti-TNF treatment. In future studies, it will be interesting to examine the effect of alternative mechanisms of action on transcript modules. In addition to these considerations, the need for physician-based data with which to categorize patient response is a major barrier to carrying out large studies of biomarkers of drug response. Without physicianbased data, studies of drug response, even when prospective, are justifiably viewed with great skepticism. This means that any prospective study investigating potential new predictive biomarkers is logistically difficult and extremely expensive. The current results suggest that it may be possible to combine changes in modular gene expression with solely patient-based data
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
to reach a convincing categorization of drug response. If this can be achieved, it will dramatically enhance opportunities to recruit large cohorts of patients starting new drugs without the traditional expenses associated with physician-led enrollment sites.
7.
8.
ACKNOWLEDGMENTS We are grateful to the Boas Family and Eileen Ludwig Greenland for general research support of Dr. Gregersen’s laboratory over many years. The authors thank the patients who have participated in these studies. AUTHOR CONTRIBUTIONS All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be published. Dr. Gregersen had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study conception and design. Curran, Lamberth, R. M. Townsend, Hamilton, Carulli, Weinblatt, Kern, Gregersen. Acquisition of data. Curran, Lamberth, Hamilton, Chernoff, Carulli, M. J. Townsend, Weinblatt, Kern, Pond, Lee, Gregersen. Analysis and interpretation of data. Oswald, Curran, Lamberth, Chernoff, Carulli, Weinblatt, Kern, Lee, Gregersen.
ROLE OF THE STUDY SPONSOR Janssen Research and Development had no role in the study design or in the collection, analysis, or interpretation of the data, the writing of the manuscript, or the decision to submit the manuscript for publication. Publication of this article was not contingent upon approval by Janssen Research and Development.
9.
10. 11.
12.
13.
14.
15.
ADDITIONAL DISCLOSURES
16.
Authors Curran and Lamberth are employees of Janssen Pharmaceuticals. Author R. M. Townsend is an employee of BristolMeyers Squibb. Author Hamilton is an employee of Regeneron Pharmaceuticals. Author Carulli is an employee of Biogen Idec. Author M. J. Townsend is an employee of Genentech.
17.
REFERENCES 1. Feldmann M, Maini RN. Anti-TNF␣ therapy of rheumatoid arthritis: what have we learned? Annu Rev Immunol 2001;19: 163–96. 2. Scott DL, Wolfe F, Huizinga TW. Rheumatoid arthritis. Lancet 2010;376:1094–108. 3. Kay J, Morgacheva O, Messing SP, Kremer JM, Greenberg JD, Reed GW, et al. Clinical disease activity and acute phase reactant levels are discordant among patients with active rheumatoid arthritis: acute phase reactant levels contribute separately to predicting outcome at one year. Arthritis Res Ther 2014;16:R40. 4. Centola M, Cavet G, Shen Y, Ramanujan S, Knowlton N, Swan KA, et al. Development of a multi-biomarker disease activity test for rheumatoid arthritis. PLoS One 2013;8:e60635. 5. Ringold S, Lu L, Wallace CA, Sasso EH, Eastman PS. A7: initial assessment of multi-biomarker disease activity assay in JIA [abstract]. Arthritis Rheumatol 2014;66 Suppl S3:S10–11. 6. Smolen JS, Aletaha D, Bijlsma JW, Breedveld FC, Boumpas D, Burmester G, et al. Treating rheumatoid arthritis to target:
18. 19.
20. 21.
22.
23.
351
recommendations of an international task force. Ann Rheum Dis 2010;69:631–7. Toonen EJ, Gilissen C, Franke B, Kievit W, Eijsbouts AM, den Broeder AA, et al. Validation study of existing gene expression signatures for anti-TNF treatment in patients with rheumatoid arthritis. PLoS One 2012;7:e33199. Van Baarsen LG, Wijbrandts CA, Gerlag DM, Rustenburg F, van der Pouw Kraan TC, Dijkmans BA, et al. Pharmacogenomics of infliximab treatment using peripheral blood cells of patients with rheumatoid arthritis. Genes Immun 2010;11:622–9. Oliveira RD, Fontana V, Junta CM, Marques MM, Macedo C, Rassi DM, et al. Differential gene expression profiles may differentiate responder and nonresponder patients with rheumatoid arthritis for methotrexate (MTX) monotherapy and MTX plus tumor necrosis factor inhibitor combined therapy. J Rheumatol 2012;39:1524–32. Julia A, Erra A, Palacio C, Tomas C, Sans X, Barcelo P, et al. An eight-gene blood expression profile predicts the response to infliximab in rheumatoid arthritis. PLoS One 2009;4:e7556. Bienkowska JR, Dalgin GS, Batliwalla F, Allaire N, Roubenoff R, Gregersen PK, et al. Convergent Random Forest predictor: methodology for predicting drug response from genome-scale data applied to anti-TNF response. Genomics 2009;94:423–32. Koczan D, Drynda S, Hecker M, Drynda A, Guthke R, Kekow J, et al. Molecular discrimination of responders and nonresponders to anti-TNF␣ therapy in rheumatoid arthritis by etanercept. Arthritis Res Ther 2008;10:R50. Lequerre T, Gauthier-Jauneau AC, Bansard C, Derambure C, Hiron M, Vittecoq O, et al. Gene profiling in white blood cells predicts infliximab responsiveness in rheumatoid arthritis. Arthritis Res Ther 2006;8:R105. Mesko B, Poliska S, Vancsa A, Szekanecz Z, Palatka K, Hollo Z, et al. Peripheral blood derived gene panels predict response to infliximab in rheumatoid arthritis and Crohn’s disease. Genome Med 2013;5:59. Meugnier E, Coury F, Tebib J, Ferraro-Peyret C, Rome S, Bienvenu J, et al. Gene expression profiling in peripheral blood cells of patients with rheumatoid arthritis in response to antiTNF-␣ treatments. Physiol Genomics 2011;43:365–71. Smith SL, Plant D, Eyre S, Barton A. The potential use of expression profiling: implications for predicting treatment response in rheumatoid arthritis. Ann Rheum Dis 2013;72:1118–24. Chaussabel D, Quinn C, Shen J, Patel P, Glaser C, Baldwin N, et al. A modular analysis framework for blood genomics studies: application to systemic lupus erythematosus. Immunity 2008;29: 150–64. Gregersen PK, Oswald M. The power of a modular approach to transcriptional analysis [editorial]. Arthritis Rheumatol 2014;66: 1418–20. Weinblatt ME, Bingham CO III, Mendelsohn AM, Kim L, Mack M, Lu J, et al. Intravenous golimumab is effective in patients with active rheumatoid arthritis despite methotrexate therapy with responses as early as week 2: results of the phase 3, randomised, multicentre, double-blind, placebo-controlled GO-FURTHER trial. Ann Rheum Dis 2013;72:381–9. Du P, Kibbe WA, Lin SM. Lumi: a pipeline for processing Illumina microarray. Bioinformatics 2008;24:1547–8. Smyth GK. Limma: linear models for microarray data. In: Gentleman R, Carey V, Dudoit S, Irizarry R, Huber W, editors. Bioinformatics and computational biology solutions using R and Bioconductor. New York: Springer; 2005. p. 397–420. Curtis JR, van der Helm-van Mil AH, Knevel R, Huizinga TW, Haney DJ, Shen Y, et al. Validation of a novel multibiomarker test to assess rheumatoid arthritis disease activity. Arthritis Care Res (Hoboken) 2012;64:1794–803. Topol EJ. Individualized medicine from prewomb to tomb. Cell 2014;157:241–53.
Modular Analysis of Peripheral Blood Gene Expression in Rheumatoid Arthritis Captures Reproducible Gene Expression Changes in Tumor Necrosis Factor Responders Michaela Oswald,1 Mark E. Curran,2 Sarah L. Lamberth,2 Robert M. Townsend,3 Jennifer D. Hamilton,4 David N. Chernoff,5 John Carulli,6 Michael J. Townsend,7 Michael E. Weinblatt,8 Marlena Kern,1 Cassandra M. Pond,1 Annette Lee,1 and Peter K. Gregersen1 analysis. Strikingly, nonresponders exhibited very little change in any modules, despite exposure to TNF blockade. These patterns of change were highly consistent across all 3 cohorts, indicating that immunologic changes after TNF treatment are specific to the combination of both drug exposure and responder status. In contrast, modular patterns of gene expression did not exhibit consistent differences between responders and nonresponders at baseline in the 3 study cohorts. Conclusion. These data provide evidence that using gene expression modules related to inflammatory disease may provide a valuable method for objective monitoring of the response of RA patients who are treated with TNF inhibitors.
Objective. To establish whether the analysis of whole-blood gene expression is useful in predicting or monitoring response to anti–tumor necrosis factor (antiTNF) therapy in patients with rheumatoid arthritis (RA). Methods. Whole-blood RNA (using a PAXgene system to stabilize whole-blood RNA in the collection tube) was obtained at baseline and at 14 weeks from 3 independent cohorts, consisting of a combined total of 240 RA patients who were beginning therapy with anti-TNF. We used an approach to gene expression analysis that is based on modular patterns of gene expression, or modules. Results. Good and moderate responders according to the European League Against Rheumatism criteria exhibited highly significant and consistent changes in multiple gene expression modules after 14 weeks of therapy, as demonstrated by hypergeometric
The development of tumor necrosis factor (TNF) inhibitors for the treatment of rheumatoid arthritis (RA) and other inflammatory disorders has been a seminal advance for the field of rheumatology (1). However, therapeutic approaches to RA remain a substantial clinical
Supported by the NIH (grant N01-AR1-2256), the Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP) Consortium (Biogen Idec, Bristol Myers-Squibb, Crescendo Bioscience, Genentech, Janssen Research and Development, Regeneron Pharmaceuticals, and SanofiAventis), and Janssen Research and Development. 1 Michaela Oswald, PhD, Marlena Kern, DNP, Cassandra M. Pond, BS, Annette Lee, PhD, Peter K. Gregersen, MD: Feinstein Institute for Medical Research, Manhasset, New York; 2Mark E. Curran, PhD, Sarah L. Lamberth, PhD: Janssen Research and Development, Springhouse, Pennsylvania; 3Robert M. Townsend, PhD: Bristol-Myers Squibb, Princeton, New Jersey; 4Jennifer D. Hamilton, PhD: Regeneron Pharmaceuticals, Tarrytown, New York; 5David N. Chernoff, MD: Chernoff Consulting, LLC, Sausalito, California; 6John Carulli, PhD: Biogen Idec, Cambridge, Massachusetts; 7Michael J. Townsend, PhD: Genentech, South San Francisco, California; 8 Michael E. Weinblatt, MD: Brigham and Women’s Hospital, Boston, Massachusetts. Drs. Curran and Lamberth own stock or stock options in Janssen, a subsidiary of Johnson & Johnson. Dr. R. M. Townsend owns stock or stock options in Bristol-Myers Squibb. Dr. Hamilton owns stock or stock options in Regeneron Pharmaceuticals. Dr. Chernoff
has received consulting fees from Crescendo Bioscience and Adamas Pharmaceuticals (more than $10,000 each) and is listed on patents for the use of multiple biomarker protein panels in autoimmune diseases and on patents for the use of aminoadamantanes for treatment of neurobiologic disorders. Dr. Carulli owns stock or stock options in Biogen Idec. Dr. M. J. Townsend owns stock or stock options in Roche. Dr. Weinblatt has received consulting fees and/or honoraria from Biogen Idec, Sanofi-Aventis, Crescendo Bioscience, BristolMyers Squibb, Genentech, and Regeneron Pharmaceuticals (less than $10,000 each). Dr. Gregersen has received an honorarium from Biogen Idec (less than $10,000) and research grants from the BATTER-UP Consortium. Address correspondence to Peter K. Gregersen, MD, Robert S. Boas Center for Genomics and Human Genetics, Feinstein Institute for Medical Research, 350 Community Drive, Manhasset, NY 11030. E-mail: [email protected]. Submitted for publication June 24, 2014; accepted in revised form October 30, 2014. 344
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
challenge, since ⬃30% of RA patients do not respond to TNF blockade, and many others obtain only partial reduction in joint inflammation (2). In addition, the determination of the clinical response to drug therapy is an inexact science, largely relying on both the subjective perceptions of the patients as well as the physician’s assessments of disease activity, both of which exhibit considerable variability. The integration of laboratory measures of inflammatory activity, such as the C-reactive protein level or the erythrocyte sedimentation rate, is standard but nonspecific, and these measurements often do not parallel the clinical parameters (3). The recent development of multiparameter methods for determining disease activity may be helpful (4,5), but these biomarkers have not yet been widely used in formal studies of clinical drug response or been demonstrated to benefit patient management. Thus, there is a substantial unmet need for improvement in both the prediction of drug response in order to individualize therapy and the development of quantitative laboratory methods to determine clinical activity of disease, which in turn, will aid in the development of predictive biomarkers and enable more precise treatment concepts such as “treat to target” (6). Expression microarrays have been used by a number of groups in the development of biomarkers of drug response (7–15). The early applications of this technology, however, have been compromised by a number of issues (16). Use of small sample sizes for initial feature set definition lead to overfitting of discriminative models, and such studies are further compromised by the semiquantitative nature of disease activity phenotypes. Critically, there has been rather limited replication of specific predictive gene expression patterns between the studies conducted to date. Although transcriptomics is a maturing technology, there are various approaches to expression analysis, and sample collection methods are often not standardized. Many studies examine gene expression in purified peripheral blood mononuclear cells. This can lead to experimental noise due to extracorporeal changes in gene expression, especially if RNA purification is not carried out immediately. More recently, use of the PAXgene Blood RNA system or other technologies that stabilize whole-blood RNA in the collection tube has helped to reduce this source of variability. Of course, this does not permit the analysis of expression in specific separated cell subsets of interest. Overall, none of these studies have provided convincing, replicated results using a standardized protocol for sample collection, microarray platform, and method of analysis (16).
345
In the current study, we attempted to address these issues by examining peripheral blood gene expression using a standard approach to blood sample collection (PAXgene) and using standard Illumina array platforms in 3 independent, prospectively collected cohorts of RA patients beginning therapy with TNF inhibitors. For our primary approach to analysis, we also took advantage of an empirically derived set of “modules” of gene expression developed by Chaussabel and colleagues (17,18). The results reveal that patients who have a clinical response to TNF inhibition exhibit a common pattern of changes in these gene expression modules after 14 weeks of therapy, whereas clinical nonresponders do not have major changes in modular gene expression. In contrast, we were not able to identify modular gene expression changes at baseline that identify patients who will or will not respond to TNF blockade. These results suggest to us that global gene expression patterns in peripheral blood can be a useful adjunct for determining whether clinical response has occurred. However, we expect that successful prediction of response to therapy is likely to require a focus on particular cell subsets, along with the integration of genetic factors and additional biomarker data. PATIENTS AND METHODS Patient cohorts. Three prospectively collected cohorts of patients with RA were studied, as summarized in Table 1. The Autoimmune Biomarkers Collaborative Network (ABCoN) cohort was enrolled as part of the National Institute of Arthritis and Musculoskeletal and Skin Diseases– supported contract to develop new approaches to biomarkers for RA and lupus. A total of 50 ABCoN patients had complete data available for study. Patients were treated with infliximab (n ⫽ 20), etanercept (n ⫽ 21), or adalimumab (n ⫽ 9). At 14 weeks, the following distribution of European League Against Rheumatism (EULAR) responses was observed: 14 had a good response, 21 had a moderate response, and 15 were nonresponders. A second patient cohort was taken from the phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate Therapy (GOFURTHER) (19). All patients were started on 2 mg/kg of golimumab intravenously, and clinical assessment of response was determined at 14 weeks. A total of 72 patients had gene expression data available for analysis. The EULAR responses in these patients were as follows: 29 had a good response, 37 had a moderate response, and 6 were nonresponders. A third patient cohort of 118 RA patients was taken from a multicenter cohort called Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP). These patients were treated with a variety of anti-TNF agents, including infliximab (n ⫽ 23), etanercept (n ⫽ 31), golimumab (n ⫽ 9), adalimumab (n ⫽ 41), and certolizumab pegol (n ⫽ 14). EULAR response rates for the BATTER-UP cohort were as follows: 42 had a
346
OSWALD ET AL
Table 1. Clinical characteristics of the rheumatoid arthritis study patients from the ABCoN, GO-FURTHER, and BATTER-UP cohorts*
Age, mean ⫾ SD years No. female/male No. (%) CCP⫹ and/or RF⫹ DAS, mean ⫾ SD EULAR response group, no. of patients Good responder Moderate responder Nonresponder TNF inhibitor, no. of patients Infliximab Etanercept Golimumab Adalimumab Certolizumab pegol Race/ethnicity, no. of patients White Black Asian Hispanic Other No. of days from visit 1 to visit 2 Range Mean ⫾ SD
ABCoN (n ⫽ 50)
GO-FURTHER (n ⫽ 72)
BATTER-UP (n ⫽ 118)
55.62 ⫾ 11.7 39/11 43 (86) 5.24 ⫾ 0.79
50.36 ⫾ 10.91 53/19 72 (100) 6.01 ⫾ 0.84
53.3 ⫾ 13.7 96/22 74 (62.7) 5.12 ⫾ 1.1
14 21 15
29 37 6
42 37 39
20 21 0 9
0 0 72 0
23 31 9 41 14
44 2 2 2 0
72 0 0 0 0
100 13 2 7 3
91–160 103.4 ⫾ 11.9
95–110 99.3 ⫾ 1.94
70–182 105.1 ⫾ 19.49
* ABCoN ⫽ Autoimmune Biomarkers Collaborative Network; GO-FURTHER ⫽ phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate; BATTER-UP ⫽ Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients; CCP ⫽ citrullinated peptide; RF ⫽ rheumatoid factor; DAS ⫽ Disease Activity Score; EULAR ⫽ European League Against Rheumatism; TNF ⫽ tumor necrosis factor.
good response, 37 a moderate response, and 39 were nonresponders. Microarray analysis. All subjects had blood drawn into PAXgene tubes at baseline (before starting therapy) and again at 14 weeks. RNA was isolated using a Qiagen QIAcube system, following the manufacturer’s protocol for PAXgene Blood RNA part 1, automated protocol. Extracted samples were eluted in 80 l of elution buffer (BR5) and were subsequently run on an Agilent 2100 Bioanalyzer, for integrity, using an RNA 6000 NanoChip. Samples with RNA integrity numbers ⬎6.5 were diluted to 30 ng/l in a total of 11 l of RNase-free water. Samples were amplified using an Illumina RNA Total Prep Amplification kit (Life Technologies). Then, 750 ng of complementary RNA was resuspended in 5 l of RNase-free water for analysis on a Human HT-12v4 Expression BeadChip, and 1.2 g was resuspended in 10 l for analysis on a WG-6v3 BeadChip (all from Illumina). All samples were processed according to the manufacturer’s instructions. The Illumina WG-6v3 microarray was used for samples in the ABCoN cohort, while the HT-12v4 microarray was used for samples in the GO-FURTHER and BATTERUP cohorts. Raw data were exported from GenomeStudio and further analyzed with the R programming language. All 3 data sets were background corrected using the R/Bioconductor package Lumi (20). Data were further transformed using a variance stabilization transformation and quantile normalized. The moderated T statistic from the R/Bioconductor package Limma (21) was used for differential expression analysis. We analyzed modules of coexpressed genes that have previously demonstrated coordinated regulation of expression across a set of diseases (whole blood), as described by
Chaussabel et al (17). A total of 28 modules were analyzed, each containing 22–325 gene probes. Because modules were originally defined using Affymetrix probe sets, these probes were mapped to Entrez Gene IDs, which in turn, were mapped to Illumina probes, as shown in Figure 1. Group comparisons (e.g., modular expression at baseline versus 14 weeks; responders versus nonresponders) were carried out using a moderated
Figure 1. Flow chart summarizing the experimental approach and data analysis in this study of rheumatoid arthritis patients receiving anti–tumor necrosis factor therapy. Gp ⫽ group.
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
T statistic comparing groups and categorizing all individual probes as up-regulated or down-regulated, using a significance cutoff of P ⬍ 0.05. For each module, we compiled a list of all significantly changed probes (up or down) at the P ⬍ 0.05 cutoff for significance. We then determined what fraction of the probes was significantly up-regulated or down-regulated within a given module. When comparing groups (e.g., responders versus nonresponders), the significance of the observed difference in distribution of up-/down-regulated probes in a module was calculated using a hypergeometric test (equivalent to a 2 ⫻ 2 table analysis using Fisher’s exact test).
RESULTS Consistent changes in modular gene expression after 14 weeks of anti-TNF therapy in clinical responders. As originally reported by Chaussabel et al (17), 28 gene expression modules can be observed in whole blood across a variety of inflammatory and infectious diseases. Where possible, these modules were labeled after the fact, according to the identity of the gene transcripts present in these modules (e.g., B
347
cells, myeloid lineage, T cells, interferon-inducible, etc.). However, many modules could not easily be categorized and were designated as “undetermined.” We used these designations as well as the original module numbers (M1.1–1.8, M2.1–2.10, and M3.1–3.9) in our study. We initiated analysis by investigating whether changes in modular gene expression occurred after TNF blockade, and if so, whether the changes differ among responders and nonresponders at 14 weeks. The results for the 3 data sets are summarized separately for EULAR responders (good and moderate responders combined) and EULAR nonresponders in Figure 2. For each module, the direction of change in module expression is shown for each data set. For the responder groups, there was a consistent and statistically significant increase in modules containing transcripts related to plasma cells (M1.1), B cells (M1.3), major histocompatibility complex and ribosomal proteins (M1.7), and T cells (M2.8), as well as a variety of “undetermined”
Figure 2. Summary of changes in transcriptional modules (M1.1–1.8, M2.1–2.11, and M3.1–3.9) at 14 weeks compared with baseline in responders (R) and nonresponders (NR) to anti–tumor necrosis factor (anti-TNF) therapy. The 3 sets of bars for each module indicate the 3 cohorts studied: the Autoimmune Biomarkers Collaborative Network (ABCoN) cohort (A), the phase III clinical trial of Golimumab in Patients with Active Rheumatoid Arthritis Despite Methotrexate (GO-FURTHER) cohort (G), and the Biomarkers of Anti-TNF␣ Therapy Efficacy in Rheumatoid Arthritis to Define Unresponsive Patients (BATTER-UP) cohort (B). Modules are designated according to the original nomenclature (see ref. 17), with descriptions accompanying selected modules. The height of the bars for the responders shows the fraction of probes significantly changed (up or down; y-axis shows fractional change for responders only) within each module. Solid bars ⫽ P ⬍ 0.0005; darkly shaded bars ⫽ P ⬍ 0.005; lightly shaded bars ⫽ P ⬍ 0.05; open bars ⫽ P not significant. MHC ⫽ major histocompatibility complex; Undet. ⫽ undetermined.
348
OSWALD ET AL
Figure 3. The distribution of significantly changed (P ⬍ 0.05) probes within the modules between baseline and 14 weeks (see Patients and Methods for details). The responders and nonresponders are shown for each of the 3 data sets: A, ABCoN, B, GO-FURTHER, and C, BATTER-UP. Colors indicate probes in modules with significant changes up (red; sig_up) or down (blue; sig_down) or with no significant modular changes (green; not_sig) (P ⬍ 0.005 by hypergeometric test). Many of these modules exhibit changes with extremely high significance. For example, P values for module 2.5 are 10⫺37, 10⫺82, and 10⫺68 in the ABCoN, GO-FURTHER, and BATTER-UP cohorts, respectively. See Figure 2 for other definitions.
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
349
Figure 4. Comparison of differences in transcriptional modules between responders and nonresponders in the 3 study cohorts at baseline, before initiation of tumor necrosis factor blockade: ABCoN (A), GO-FURTHER (G), and BATTER-UP (B). Modules are designated according to the original nomenclature (see ref. 17), with descriptions accompanying selected modules. The height of the bars shows the fraction of probes that are different between responders and nonresponders. Solid bars ⫽ P ⬍ 0.0005; darkly shaded bars ⫽ P ⬍ 0.005; lightly shaded bars ⫽ P ⬍ 0.05; open bars ⫽ P not significant. See Figure 2 for definitions.
modules (M1.8, M3.4, M3.7, M3.8, and M3.9). In addition, there was a consistent down-modulation of transcripts in several myeloid lineage modules (M1.5 and M2.6) and in platelets (M2.1), as well as in inflammation 1 (M3.2) and inflammation 2 (M3.3). These results stand in marked contrast to the lack of consistent modular changes across the 3 data sets for nonresponders. Indeed, in general, nonresponders showed a very small fraction of probes that significantly changed in comparisons of data at 14 weeks versus baseline within any of the 28 modules. In order to provide a more detailed representation of these modular changes and their magnitude, we plotted for each data set and responder group the fold changes for all probes achieving a significant difference (P ⬍ 0.05 by moderated t-test) within each module. Figures 3A–C show these results for the ABCoN, GOFURTHER, and BATTER-UP cohorts, respectively. Within the responders from the 3 data sets, there was striking consistency in the pattern of changes in the modules. Significant changes (P ⬍ 0.005 by hypergeometric test) in probe distribution, whether up or down, are indicated in the figures. Indeed, some of the changes achieved remarkably compelling levels of statistical significance. For example, the P values associated with a reduction in myeloid modules M1.5 and M2.6 after TNF blockade ranged from 10–20 to 10–87, and reductions in “undetermined” modules M3.2 and M3.3 ranged from 10⫺12 to 10⫺85 across the various data sets. These dramatic changes were only seen in the responder group, with minimal change in the nonresponder group.
No consistent differences in modules between responders and nonresponders at baseline. While substantial changes in modular gene expression were seen in responders after TNF blockade, the baseline pattern of module expression did not show consistent differences among the 3 cohorts (Figure 4). While some of the data sets suggested the presence of meaningful differences between responders and nonresponders at baseline in 1 cohort (e.g., several “undetermined” modules were relatively less prominent in responders in the ABCoN cohort), these changes were not replicated in the other 2 cohorts. This finding emphasizes the importance of having replication data sets in which to validate the initial results and demonstrates that an improved understanding of RA patient heterogeneity will be required for the development of an accurate predictor of drug response. DISCUSSION In this study, we demonstrated a consistent change in whole-blood gene expression in clinically responding RA patients after exposure to TNF-blocking agents for 14 weeks. These changes were not observed in patients who do not exhibit a clinical response to treatment, as determined using the standard EULAR criteria. The data are based on a consideration of sets of genes that are expressed together in a modular pattern across different inflammatory conditions, as first demonstrated by Chaussabel et al (17). These data provide compelling evidence that these gene expression changes
350
are not simply a result of exposure to TNF blockade, but rather, they reflect a change in both the immunologic and disease state in patients who are responding to TNF therapy. Interestingly, we have recently had an opportunity to examine a small number of patients who did not receive study drug (placebo group; 7 responders and 16 nonresponders). There was no difference between responders and nonresponders in this group (data not shown), suggesting that the changes are specific to the physiologic effects of drug therapy in the setting of a clinical response. A larger study will be required to confirm these findings. In contrast to the gene expression changes observed in responders over time, the modules did not predict future response to TNF blockade when these modular expression patterns were interrogated at baseline, before starting therapy, as shown in Figure 4. The development of useful biomarkers for the management of RA has been extremely challenging. At present, there is no clinical parameter or laboratory test that can distinguish likely responders to anti-TNF blockade prior to initiation of therapy. Given the fact that at least 25% of patients will not respond to anti-TNF therapy, while at most, one-third will have a robust clinical response, there is clearly a need to develop such markers. Several small studies have proposed biomarkers to predict such a response, but none have been validated for any biologic agent (16). It is important to note that published studies suffer from reporting primary results and have not had multiple independent cohorts available for validation and replication of results. A primary component of our present work is replication of findings in multiple independent cohorts. In this context, our use of modular approaches to whole-blood microarray analysis identified suggestive feature sets for prediction of response in individual cohorts, but these marker sets failed to replicate or yield a convincing gene expression biomarker for future drug response. This emphasizes the biologic complexity of the RA response phenotype. If the scientific community is to succeed in the goal of predicting response to treatment, it is likely that multiple parameters will need to be combined, including genetic and genomic data, along with cellular and biochemical data that more precisely reflect the underlying TNF dependence of the current inflammatory state and drivers of inflammation in a given patient. Our current results are likely to have utility in the context of determining the presence of reduced disease activity in response to drug therapy. This is an important issue, since the current measures, such as the American College of Rheumatology criteria and the EULAR
OSWALD ET AL
criteria, are highly dependent on primarily subjective patient-based and physician-based data points. The current results suggest that analysis of modular gene expression can enhance the accuracy of a responder versus nonresponder designation. For example, while the responder and nonresponder groups showed consistent differences in the pattern of change in these modules, individuals may have a more or less robust change. In the future, we will generate an individually applicable “modular change score.” A diagnostic biomarker of this type may be useful in assigning likelihood for membership in the responder or nonresponder group, allow for decisions on change from current therapy, and allow us to track the timing and durability of the response. As part of such an effort, we plan to incorporate other current measures of disease activity, such as the proteomic panel used in the Vectra DA test (4,22). Development of an improved RA disease status scoring method will require careful exploration of combinations of modules and other laboratory measures in order to optimize and validate such a method. Fortunately, an additional validation data set of over 100 additional subjects in the BATTER-UP study will shortly be available to permit a robust replication of the current findings as well as new combinatorial approaches to individual categorization of response at a molecular level. Our preliminary findings suggest that some patients who have a clinical response may not show strong evidence of a biologic response, possibly due to a placebo effect. Likewise, patients with a biologic response may exist within the “nonresponder” group, again underlining the heterogeneity in the mechanisms of disease. Methods to better classify patients quantitatively with regard to biologic response will facilitate clinical treatment decisions and enable the practice of “individualized medicine” (23), including such concepts as treating to target (6). Finally, the outcome data and molecular interpretation presented herein are restricted to anti-TNF treatment. In future studies, it will be interesting to examine the effect of alternative mechanisms of action on transcript modules. In addition to these considerations, the need for physician-based data with which to categorize patient response is a major barrier to carrying out large studies of biomarkers of drug response. Without physicianbased data, studies of drug response, even when prospective, are justifiably viewed with great skepticism. This means that any prospective study investigating potential new predictive biomarkers is logistically difficult and extremely expensive. The current results suggest that it may be possible to combine changes in modular gene expression with solely patient-based data
WHOLE-BLOOD GENE EXPRESSION ANALYSIS AND RESPONSE TO ANTI-TNF IN RA
to reach a convincing categorization of drug response. If this can be achieved, it will dramatically enhance opportunities to recruit large cohorts of patients starting new drugs without the traditional expenses associated with physician-led enrollment sites.
7.
8.
ACKNOWLEDGMENTS We are grateful to the Boas Family and Eileen Ludwig Greenland for general research support of Dr. Gregersen’s laboratory over many years. The authors thank the patients who have participated in these studies. AUTHOR CONTRIBUTIONS All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be published. Dr. Gregersen had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study conception and design. Curran, Lamberth, R. M. Townsend, Hamilton, Carulli, Weinblatt, Kern, Gregersen. Acquisition of data. Curran, Lamberth, Hamilton, Chernoff, Carulli, M. J. Townsend, Weinblatt, Kern, Pond, Lee, Gregersen. Analysis and interpretation of data. Oswald, Curran, Lamberth, Chernoff, Carulli, Weinblatt, Kern, Lee, Gregersen.
ROLE OF THE STUDY SPONSOR Janssen Research and Development had no role in the study design or in the collection, analysis, or interpretation of the data, the writing of the manuscript, or the decision to submit the manuscript for publication. Publication of this article was not contingent upon approval by Janssen Research and Development.
9.
10. 11.
12.
13.
14.
15.
ADDITIONAL DISCLOSURES
16.
Authors Curran and Lamberth are employees of Janssen Pharmaceuticals. Author R. M. Townsend is an employee of BristolMeyers Squibb. Author Hamilton is an employee of Regeneron Pharmaceuticals. Author Carulli is an employee of Biogen Idec. Author M. J. Townsend is an employee of Genentech.
17.
REFERENCES 1. Feldmann M, Maini RN. Anti-TNF␣ therapy of rheumatoid arthritis: what have we learned? Annu Rev Immunol 2001;19: 163–96. 2. Scott DL, Wolfe F, Huizinga TW. Rheumatoid arthritis. Lancet 2010;376:1094–108. 3. Kay J, Morgacheva O, Messing SP, Kremer JM, Greenberg JD, Reed GW, et al. Clinical disease activity and acute phase reactant levels are discordant among patients with active rheumatoid arthritis: acute phase reactant levels contribute separately to predicting outcome at one year. Arthritis Res Ther 2014;16:R40. 4. Centola M, Cavet G, Shen Y, Ramanujan S, Knowlton N, Swan KA, et al. Development of a multi-biomarker disease activity test for rheumatoid arthritis. PLoS One 2013;8:e60635. 5. Ringold S, Lu L, Wallace CA, Sasso EH, Eastman PS. A7: initial assessment of multi-biomarker disease activity assay in JIA [abstract]. Arthritis Rheumatol 2014;66 Suppl S3:S10–11. 6. Smolen JS, Aletaha D, Bijlsma JW, Breedveld FC, Boumpas D, Burmester G, et al. Treating rheumatoid arthritis to target:
18. 19.
20. 21.
22.
23.
351
recommendations of an international task force. Ann Rheum Dis 2010;69:631–7. Toonen EJ, Gilissen C, Franke B, Kievit W, Eijsbouts AM, den Broeder AA, et al. Validation study of existing gene expression signatures for anti-TNF treatment in patients with rheumatoid arthritis. PLoS One 2012;7:e33199. Van Baarsen LG, Wijbrandts CA, Gerlag DM, Rustenburg F, van der Pouw Kraan TC, Dijkmans BA, et al. Pharmacogenomics of infliximab treatment using peripheral blood cells of patients with rheumatoid arthritis. Genes Immun 2010;11:622–9. Oliveira RD, Fontana V, Junta CM, Marques MM, Macedo C, Rassi DM, et al. Differential gene expression profiles may differentiate responder and nonresponder patients with rheumatoid arthritis for methotrexate (MTX) monotherapy and MTX plus tumor necrosis factor inhibitor combined therapy. J Rheumatol 2012;39:1524–32. Julia A, Erra A, Palacio C, Tomas C, Sans X, Barcelo P, et al. An eight-gene blood expression profile predicts the response to infliximab in rheumatoid arthritis. PLoS One 2009;4:e7556. Bienkowska JR, Dalgin GS, Batliwalla F, Allaire N, Roubenoff R, Gregersen PK, et al. Convergent Random Forest predictor: methodology for predicting drug response from genome-scale data applied to anti-TNF response. Genomics 2009;94:423–32. Koczan D, Drynda S, Hecker M, Drynda A, Guthke R, Kekow J, et al. Molecular discrimination of responders and nonresponders to anti-TNF␣ therapy in rheumatoid arthritis by etanercept. Arthritis Res Ther 2008;10:R50. Lequerre T, Gauthier-Jauneau AC, Bansard C, Derambure C, Hiron M, Vittecoq O, et al. Gene profiling in white blood cells predicts infliximab responsiveness in rheumatoid arthritis. Arthritis Res Ther 2006;8:R105. Mesko B, Poliska S, Vancsa A, Szekanecz Z, Palatka K, Hollo Z, et al. Peripheral blood derived gene panels predict response to infliximab in rheumatoid arthritis and Crohn’s disease. Genome Med 2013;5:59. Meugnier E, Coury F, Tebib J, Ferraro-Peyret C, Rome S, Bienvenu J, et al. Gene expression profiling in peripheral blood cells of patients with rheumatoid arthritis in response to antiTNF-␣ treatments. Physiol Genomics 2011;43:365–71. Smith SL, Plant D, Eyre S, Barton A. The potential use of expression profiling: implications for predicting treatment response in rheumatoid arthritis. Ann Rheum Dis 2013;72:1118–24. Chaussabel D, Quinn C, Shen J, Patel P, Glaser C, Baldwin N, et al. A modular analysis framework for blood genomics studies: application to systemic lupus erythematosus. Immunity 2008;29: 150–64. Gregersen PK, Oswald M. The power of a modular approach to transcriptional analysis [editorial]. Arthritis Rheumatol 2014;66: 1418–20. Weinblatt ME, Bingham CO III, Mendelsohn AM, Kim L, Mack M, Lu J, et al. Intravenous golimumab is effective in patients with active rheumatoid arthritis despite methotrexate therapy with responses as early as week 2: results of the phase 3, randomised, multicentre, double-blind, placebo-controlled GO-FURTHER trial. Ann Rheum Dis 2013;72:381–9. Du P, Kibbe WA, Lin SM. Lumi: a pipeline for processing Illumina microarray. Bioinformatics 2008;24:1547–8. Smyth GK. Limma: linear models for microarray data. In: Gentleman R, Carey V, Dudoit S, Irizarry R, Huber W, editors. Bioinformatics and computational biology solutions using R and Bioconductor. New York: Springer; 2005. p. 397–420. Curtis JR, van der Helm-van Mil AH, Knevel R, Huizinga TW, Haney DJ, Shen Y, et al. Validation of a novel multibiomarker test to assess rheumatoid arthritis disease activity. Arthritis Care Res (Hoboken) 2012;64:1794–803. Topol EJ. Individualized medicine from prewomb to tomb. Cell 2014;157:241–53.
