Lack of replication of genetic predictors for the rheumatoid arthritis response to anti-TNF treatments: a prospective case-only study

Anti-tumor necrosis factor (anti-TNF) therapies have revolutionized the treatment of rheumatoid arthritis (RA) [1, 2]. Three drugs of this type, infliximab, etanercept, and adalimumab, have been used with success in hundreds of thousands patients with RA around the world. New drugs targeting TNF are in development or have been recently approved [3]. The beneficial effects of these drugs include a better quality of life; control of inflammation, stiffness, and pain; and slowing progression to joint erosions and deformity. It seems also that they are able to decrease cardiovascular risk and overall mortality of patients with RA [4, 5]. However, there is a significant percentage of patients who do not obtain these advantageous effects [1–3]. In some of these patients, this lack of response is primary, from the start of the treatment, whereas others develop resistance to treatment after a period of initial response. Unfortunately, there are no useful predictors to forecast what the clinical response of a specific patient will be. This has led to an unsatisfactory trial-and-error approach in the selection of drugs, meaning that some patients will miss an effective treatment at a critical window of opportunity [6] and that health service resources will be wasted. In response to this challenge, multiple lines of research are looking for predictors of response to anti-TNF therapies among patient clinical features, synovial tissue biomarkers, blood proteins, or genetic variants [7–10]. Very promising, though preliminary, findings have been reported in this last field. Sixteen single-nucleotide polymorphisms (SNPs) with an important association with response to treatment were identified in a recent genome-wide association study (GWAS) [7]. In our view, the most remarkable aspect of these findings was the marked effect size of each SNP, with levels very rarely found in genetic studies of complex traits. All showed an odds ratio (OR) of more than 3.5 in the comparison between patients with good response and non-responders. Some of these SNPs showed effect sizes of an OR of more than 10. If confirmed, these effects, together with minor allele frequencies of more than 12%, will allow the prediction of response to anti-TNF treatments with great accuracy at the level of the individual patient [11]. The limitation of this study was that only 89 patients were included, and even very significant results in a study of this size are uncertain. Our objective has been to provide the necessary replication to these exciting findings with the expectation that at least a few of them will be confirmed. This will be a first step before proceeding to prospective clinical studies to assess their utility in clinical practice.

The aim of our study was to replicate the strong association of 16 SNPs with response to anti-TNF therapy reported in a recent GWAS [7]. Therefore, we used the same variables and type of analysis. Data from the 151 patients with RA are shown in Table 1. Some of the characteristics of our study population were different from those of the patients analyzed in the GWAS [7]. Specifically, our patients showed a lower percentage of rheumatoid factor positivity, higher positivity for anti-CCP (anti-cyclic citrullinated peptide) antibodies, and higher baseline HAQ and DAS28 levels. There were 70.2% of patients with high disease activity at baseline as assessed by a DAS28 of greater than 5.1. In spite of this high activity, there were 40.4% and 43.7% of good responders at 3 and 6 months, respectively, and only 22.5% and 21.8% of non-responders at 3 and 6 months, respectively. The percentages of responders and non-responders were similar in the two studies. In contrast, the proportion of patients treated with each of the three anti-TNF drugs was different (Table 1). In our cohort, most patients were treated with infliximab (66.9%), followed by etanercept (23.2%) and adalimumab (9.9%). We also checked that there was a good correlation between the variable used as primary outcome in our analysis, relDAS28, and the EULAR response classification (Figure 1), allowing for consistency in the analyses.

The relationship between the SNP genotypes and response to anti-TNF treatment at 3 months was evaluated by linear regression analysis between the genotypes and the continuous variable relDAS28. There was no association of any of the 16 SNPs with relDAS28 at 3 months (Table 2). Secondary analyses showed very similar results. Comparison of non-responders with good responders according to the EULAR criteria at 3 months did not show any significant association (Table 2). The most extreme OR (1.9, 95% CI 1.0 to 3.6) corresponded to rs437943 in the CNTDE1 locus but compared poorly with the previously reported OR (4.6, 95% CI 1.8 to 12.3). In addition, there was no association of relDAS28 with any of the 16 SNPs at 6 months or of the classification in responders and non-responders (Additional file 2). Finally, analysis of patients treated with infliximab, which represented 66.9% of our study, did not show any significant association between response and the SNPs (Additional file 2). Because of the small number of patients in the etanercept or adalimumab subgroups, no separate analyses of response to treatment were done. We also visually explored whether joint consideration of the 16 SNPs was able to discriminate between the different groups of patients according to their response to treatment, but patients with different responses did not show any clustering in identifiable groups in this analysis (Figure 2).

To interpret the above results, it was critical to assess whether our study had enough statistical power to replicate the previously reported associations. Power for the weakest association in the GWAS, which corresponds to rs928655 (P = 3 × 10^-5), was larger than 95% for a P value of 0.002. It is important to remark that ORs from the GWAS are very likely heavily biased upwards as a consequence of the winner's curse affecting any GWAS and especially those of small size [17, 18]. Therefore, this power estimate is valuable only in the context of the reported ORs taken at face value.

There is a great need of good predictors for RA response to the anti-TNF treatments [1–3, 9]. The development and approval of new effective drugs for RA add to this urgency [3]. The recent GWAS from Liu and colleagues [7] was especially remarkable because it showed such strong associations that, according to published simulations [11], they could be used for prediction in individual patients. This is a characteristic that has not been found in any of the previous studies. However, the size of the study implied that results should be replicated before they could be taken at face value, as already acknowledged by the authors. We have tried to provide here the needed replication in the expectation that some of them will be confirmed and that validation in prospective studies will soon follow.

Unfortunately, in spite of the moderately larger sample size of our study and the corresponding very good power to detect this type of strong association, none of the associations was replicated. These results make it very unlikely that any of the 16 SNPs could have an association as strong as suggested by the previous GWAS [7]. It is possible that the differences between the patients with RA in the two studies could have had an effect on the lack of replication, but these differences were not large enough to completely explain the very divergent results. In addition, patients in the GWAS were predominantly of European ancestry as were all of the patients in our study. Therefore, it seems more likely that the original strong associations were due to random variation of allele frequencies in a study including more than 300,000 SNPs and to the heavy bias characteristic of GWASs of small sample size [17, 18]. This possibility was already considered by us before beginning this study, but we judged that some SNPs would be replicated given that they showed low P values, five of them with P values of less than 10^-6 [7], and low P values are the best indication of the reproducibility of results [19].

Our negative results imply that we still do not have any strong evidence supporting a significant role of genetic variation in the response to anti-TNF treatments. In addition, our results imply that none of the SNPs in our study will be useful as individual predictors of response to anti-TNF therapy, but do not exclude a weaker association.

We thank Carmen Pena-Pena for her excellent technical assistance and Yolanda Lopez-Golan for her help in recruiting patients. MS-G is the recipient of an FPU predoctoral bursary of the Spanish Ministry of Education. MC is the recipient of an 'Isabel Barreto' bursary of the government of Galicia. This project was supported by an unrestricted grant from Roche Spain and by grants PI080744 and PI09/90744 from the Instituto de Salud Carlos III (Spain) with participation of funds from FEDER (European Union).