Rheumatoid arthritis (RA) is complex and heterogeneous inflammatory autoimmune disease, with a multifactorial etiology.1-3 An estimated 1.3 million adults in the United States live with RA and though the disease affects both sexes, the incidence is higher in women than in men (53/100,000 vs 29/100,000 population).4 Left improperly treated, it can progress and become a debilitating disease with significant morbidity as well as increased mortality.2,5,6
RA treatment response is defined in terms of disease activity or remission scores. Commonly used are the American College of Rheumatology (ACR) and the European League Against Rheumatism (EULAR) response criteria.7-10 EULAR response criteria are based on changes in the Disease Activity Score (DAS), while the ACR improvement scores of ACR20, ACR50, and ACR70, represent the percent improvement in a standard set of indices, including clinical factors as well as laboratory markers (i.e. acute-phase reactants).11,12 A 50% response (ACR50) is needed for most patients to reach low disease activity.3 The ACR and EULAR criteria have been reported to have comparable validity.13 However, given that their components and requirements are different, and that there is known variability in patient disease assessments (PDAs) and patient-reported outcome measures (PROMs) as well as in the composite of clinical metrics used that are included in the response criteria, some patients who are classified as responders by one criterion may not be classified as responders by another criterion.9,14-16 For example, differences are seen among the various EULAR response criteria, depending on which of the composite measures is used (i.e. disease activity score using 28 joint counts (DAS28)/Erythrocyte Sedimentation Rate (ESR) vs DAS28/C-reactive protein (CRP)).17,18 Other scores, namely the clinical disease activity index (CDAI), and the simplified disease activity index (SDAI), are also commonly used in clinical care and correlate with outcomes such as progression and functional impairment.19
Conventional synthetic disease-modifying antirheumatic drugs (csDMARDs) such as methotrexate are considered first-line therapy for many patients with RA.20 However, up to 60% of patients fail to achieve treatment targets on csDMADRs.21 For them, biologic and targeted synthetic DMARDs (b/t DMARDs), are often recommended.20,22 These include tumor necrosis factor-alpha inhibitors (anti-TNFs, or TNFis), Janus Kinase tyrosine kinases (JAKs), Interleukin-6 (Il-6) inhibitors, T- and B-cell therapies. The various b/tDMARDs have, on the whole, shown similar efficacy and safety profiles; as such, many guidelines have not preferred or prioritized among them.20,23,24 Despite the fact that there are multiple classes of targeted therapies available, TNFis remain the predominant first-line b/tDMARDs in the majority of patients with RA for a variety of reasons, including long-term established efficacy and safety profiles, the comfort level of ordering physicians, and insurance policy requirements.25-29 However, up to two-thirds of patients will fail to achieve ACR50 within 6 months of therapy with their first TNFi, and more than 60% will require a third DMARD.19,30,31 After patients fail their first TNFi therapy, they are approximately 12-30% more likely to inadequately respond to their second targeted medication.32-35 Over time, up to 75% of patients may eventually reach treatment targets with multiple trial-and-error sequential approaches to treatment.19
When TNFi therapy is ineffective or results in adverse effects, patients may ‘cycle’ to another TNFi or to ‘switch’ to a drug with a different mechanism of action (MOA). The ACR conditionally recommends ‘switching’ over ‘cycling,’ based on evidence supporting greater improvement in disease activity when patients switch drug classes, rather than cycle between them.20,32,36-38 However, a claims analysis-based study found that most (~64%) patients still often cycle to a second TNFi before switching to a non-TNFi, and that patients who switch to a non-TNFi are significantly older with more comorbidities.39 These trial-and-error attempts using various drugs and drug classes, often with inadequate disease response and with drug-induced side-effects, can result in continued disease progression, as effective therapy (which is important early in RA to delay or prevent debilitating disease outcomes) is delayed.5,20
Despite the availability of multiple treatment options, there is no certain way to predict which patients will respond to the various available therapies. Certain characteristics, including obesity and sex, have been associated with a lack of response to TNFis in some studies but not in others.40-44 Moreover, up to 30% of RA patients do not have rheumatoid factor (RF) or anti-citrullinated protein antibodies (ACPA), and researchers remain divided on whether to subdivide and manage RA differently based on autoantibody status.45,46 For example, there have been conflicting reports regarding whether seronegative patients have less active disease at baseline and less radiographic progression than seropositive patients.46-49 Moreover, distinct genetic factors have been associated with seronegative patients, indicating that there may be distinct pathogenic mechanisms involved.50,51 Further, studies have found that treatment choice is influenced by the presence of autoantibodies and that patients with autoantibodies may require biologics to achieve remission more frequently than seronegative patients.47,52 Additionally, in some multivariable analyses, RF has been negatively associated with biologic therapy survival and ACPA with the inability to taper or discontinue TNFis after remission.37 However, other studies, including meta-analyses, have not found these associations; rather, they report that both seropositive and seronegative types of RA may require similar intensive treat-to-target therapies.49,53-55 Differences in study findings may be due to different classification schemes used (i.e. RA diagnosed using ACR 2010 vs older criteria) as well as on the inclusion PDAs and PROMs, which are varied and prone to inherent subjectivity. Nevertheless, there remain unanswered questions regarding the management of different sub-populations with RA. As such, there is an unmet clinical need for a test that can accurately predict which patients will or will not respond to targeted and biologic therapies.
Predictive Biomarker Tests
A number of molecular biomarker tests have been proposed that can predict response (or non-response) to certain classes or multiple classes of drugs in RA treatment.56-63 Some are based on genetic markers found in blood, some on genetic markers in combination with clinical and laboratory factors, and others on transcriptomics within the synovium. One observational cohort evaluated the differential expression and methylation of DNA to predict response to two different TNFi drugs adalimumab (ADA) and etanercept (ETN) in patients with RA. A machine learning model found divergent transcriptomic signatures in ADA and ETN responders that predicted drug response with an accuracy of up to 85-88%.57 Interestingly, in that study the majority of the patients did not respond to both drugs but had the potential to respond to either ADA or ETN. However, a prior study evaluating DNA methylation in RA found different methylation sites than those reported by Tao et al.64 The differences may have been due to the different cell types interrogated as well as different response criteria used. However, these studies were limited by small sample sizes and have not been externally validated using larger sample sizes.
The above-mentioned studies evaluated a decision regarding the choice of TNFi (ADA vs ETN). Other biomarker tests have been proposed to assist in the decision regarding which class of drug to use – i.e., whether to use a TNFi versus a drug with an alternate mechanism of action (alt-MOA, or non-TNFi). The one such test that is most widely published is a molecular signature response classifier (MSRC) from Scipher Medicine (Waltham, MA) that can be performed prior to the start of targeted or biologic therapy to predict non-response (NR) to TNFis.59,65 This test is performed on whole blood and includes 10 single-nucleotide polymorphisms (SNPs) associated with RA, 8 gene transcripts, 2 traditional laboratory tests (CRP and anti-CCP), and 3 clinical parameters (sex, body mass index (BMI), and patient disease assessment (PDA)).52,59 The raw output of the model is transformed into a continuous variable between 1 and 25 with higher numbers indicating a greater likelihood of non-response to TNFis.66
In a clinical validity study of 175 prospectively collected samples from RA patients in the CERTAIN trial, the top 23 ranked biomarkers were found to identify ACR50 non-responders with a sensitivity of 50.0%, specificity of 86.8%, and a positive predictive value (PPV) of 89.7%.59,67 The overall TNFi response rate was 30.3%, whereas patients predicted to be non-responders (NRs) by the test had a TNFi response rate of 10.3% (7/68) by ACR50.59 Conversely, lack of a NR signature did not predict response, as nearly 60% of patients with this result did not meet ACR50 response criteria when treated with a TNFi.59 Another prospective clinical study (NETWORK-004) also reported that patients with a molecular signature of NR were less likely to achieve therapy targets with TNFis than those lacking the signature with odds ratios (ORs) of 3.4–8.8 for b/tDMARD-naive (n=146) and 3.3–26.6 for TNFi-exposed patients (n=113) (notably, the OR of 26 was in previously TNFi-exposed patients).65 In this study, a NR signature was detected in nearly 45% of patients at baseline.65 Importantly, both of these studies were observational and test results were not actually used to inform treatment selection. Additionally, treatment selection was at the physician's discretion and may have been influenced by multiple clinical and non-clinical factors, as outlined above.
A prospective multi-institutional cohort study using a clinical database of RA patients (Study to Accelerate Information of Molecular Signatures [AIMS] in Rheumatoid Arthritis) evaluated outcomes in patients for who a b/tDMARD treatment decision was informed by MSRC testing.68 Patient eligibility did not consider baseline disease activity, prior biological exposure, or csDMARD use. The primary endpoint was therapeutic responsiveness defined by achievement of ACR50 at 24-weeks. According to questionnaire responses, therapy selection was informed by the test results for 73.5% (277/377) of patients.68 However, only 85 patients completed a 24-week follow-up visit. Patient responses to treatment (informed by the MSRC) at 24 weeks in predicted non-responders who received an alternate drug (alt-MOA) (n=23 patients) and in predicted non-responders who received a TNFi despite their NR signature (n=29 patients) were 34.8% and 10.3% by ACR50; by CDAI they were 56.2% and 15.4%, respectively.68 ACR50 response to TNFis in patients lacking the NR signature was only 45.8%.68 Patients with and without a molecular signature of NR who received an alt-MOA had a nearly equivalent responses (33-34%) by ACR50.68 Notably, the number of evaluable patients at 24 weeks in each sub-group was small and despite access to the test results, more patients with predicted NR to TNFis were still prescribed TNFis rather than alt-MOAs, once again highlighting the multiple variables associated with physician prescribing practices. A second interim analysis of AIMS that included a larger number of patients (N=274) with moderate or severe RA similarly showed that absolute changes in CDAI scores from baseline were improved when treatment was informed by the MSRC test.69 Finally, a comparative cohort study compared a MSRC-tested arm from AIMS with an external control arm from a United States electronic health records database.70 This study validated the reported test performance characteristics of the MSCR (i.e. PPV 88%) but again noted that physicians prescribed test-aligned therapies only 70% of the time. Despite this incomplete adherence to test results, patients in the MSRC arm were nearly 3 times more likely to achieve remission than those in the standard-of-care control arm. 70 Importantly, the authors reported that after cohort matching, 57% of patients in the MSRC-tested arm (N=489 with clinical outcomes data at 6 months) had not been included in the 2 prior interim analyses, thus providing data independent of the two prior AIMS-based cohort studies.70
There are other predictive biomarker tests that are ‘biopsy-driven,’ evaluating the transcriptomic signature of the synovium in RA patients as an indicator for therapy response and clinical outcomes.71,72 R4RA (rituximab vs tocilizumab in anti-TNF inadequate responder patients with rheumatoid arthritis) was a multicenter randomized trial evaluating RA patients with inadequate responses to TNFis.71 RNA sequencing-based identification of patients with a low or absent B cell gene expression signature in synovial tissue significantly correlated with a greater response to tocilizumab (63%) than rituximab (36%), suggesting that in patients with a low or absent B cell expression gene signature in synovial tissue, an alternative therapy may be preferred over rituximab.71 Another longitudinal study of the synovial transcriptome reported differentially expressed genes in DMARD-naïve early-RA patients versus advanced RA patients.72 Results from the synovial transcriptome studies have not yet been replicated in external validation cohorts. Moreover, studies have shown pronounced heterogeneity in the synovial tissue of RA patients, mirroring the general heterogeneity seen in this inflammatory disease.73 Finally, synovial fluid is not readily available for testing to inform therapy decisions in routine clinical practice settings. Nonetheless, it remains an active area of exploration for both the pathogenesis as well as the response to therapy in RA.
Though various candidate polymorphisms have been proposed to be associated with TNFi treatment response in RA patients, multiple genome-wide association studies (GWAS) have not identified such predictive genetic variants in a consistent or reproducible manner.74 An ‘open challenge’ comparing prediction models developed by 73 research groups found that, despite a “genetic heritability estimate of treatment non-response trait,” SNP data do not significantly contribute to the prediction of response to therapy above that which was obtained by available clinical predictors.75 Specifically, this analysis did find that certain available clinical features, including sex, age, the specific TNFi, and methotrexate use, did provide a level of prediction that performed significantly better than random.75 The authors conclude that “these results suggest that future research efforts focused on the incorporation of a richer set of clinical information—including seropositivity, treatment compliance and disease duration—may provide opportunity to leverage these methods in clinically meaningful ways. In addition, the identification of data modalities that are more effective than genetics in capturing heterogeneity in RA disease progression—whether clinical, molecular or other—may also improve predictive performance.75
Drug treatments themselves have been reported to alter the molecular profile of RA patients, both in terms of normalization of the profile (which has been associated with clinical remission), and in terms of resistance to treatments.60,76 A multi-omics analysis found that normalization and resistance are also heterogeneous, which may be explained by an imbalance of white blood cell subsets. It remains unclear, however, whether these signatures are also found in the inflamed synovium.60
Finally, the EULAR and others have reported (and cautioned) on the variability inherent in the existing big data sources, on the tolerance of poor quality data in large registries used in RA biomarker studies, on the heterogeneous methods used between studies to analyze big data, and on the lack of external validation of some of the tests.58,77-79 These issues lend themselves to the risks of ‘quantitative fallacy,’ bias, overfitting of predictive models, and the inability to generalize results.77,80 EULAR has called for comprehensive and harmonized standards, open data platforms, and interdisciplinary collaboration, such that artificial intelligence (AI) can safely and effectively be implemented in the clinical practice of RA.78
Contractor Advisory Committee (CAC)
A Contractor Advisory Committee (CAC) meeting on the topic of predictive testing in RA was held in December 2021. Similar to findings from a published survey of 248 USA-based rheumatologists, the CAC subject matter expert (SME) panelists noted that physicians would welcome predictive tests to guide targeted therapy in RA patients and find them useful if they could help minimize the trial-and-error approach of current therapy.81 However, though the panelists were aware of early and emerging data, they were not aware that any such tests had been rigorously validated for routine clinical use at the time of the CAC. Additionally, the panelists expressed the opinion that TNFis are the most commonly prescribed biologic therapies in RA for two primary reasons – (1) they were the first biologics available for RA patients and therefore physician and patient comfort levels may be higher with this class of drug over ‘newer’ therapies with similar safety and efficacy profiles and (2) insurance companies often require a trial of 1-2 TNFis before covering other targeted therapies. There was consensus that the requirement by insurance companies for patients to fail multiple TNFis prior to paying for an alternate targeted therapy is unreasonable.
Finally, during the Comment period, we received letters from our SMEs in support of the use of predictive biomarker tests for a limited RA population.