Predictors of treatment response in rheumatoid arthritis: a narrative review

Abstract

Rheumatoid arthritis (RA) is a common autoimmune disease having diverse responses to disease-modifying antirheumatic drugs. Although treat-to-target strategies have improved disease management, therapeutic selection remains largely empirical and a substantial proportion of patients fail to achieve adequate clinical response to their therapies. Therefore, reliable biomarkers capable of predicting treatment response before therapy initiation are urgently needed to guide personalised treatment decisions. Current evidence suggests that multiple categories of biomarkers may contribute to the prediction of treatment responses across different therapeutic agents in RA, including clinical characteristics, serological markers, cytokine profiles, genetic and transcriptomic signatures, immune-cell phenotypes, synovial pathotypes and microbiome features. This review summarises current biomarker strategies for predicting therapeutic response in RA and discusses their potential utility in improving precision medicine. However, most individual biomarkers show limited predictive power, partly due to the complex and dynamic nature of RA pathogenesis. Combining biomarkers from multiple biological dimensions, together with the continued exploration of novel predictive markers, may improve the robustness and accuracy of treatment response prediction in RA.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Treatment responses to disease-modifying antirheumatic drugs vary substantially among patients with rheumatoid arthritis, yet treatment selection remains largely empirical.

• Numerous clinical and molecular biomarkers have been associated with treatment outcomes, but few have been sufficiently validated for treatment selection.

WHAT THIS STUDY ADDS

• This review critically maps pretreatment biomarkers by therapeutic class, biological compartment, clinical availability and strength of evidence.

• It distinguishes clinically available response-associated factors from promising but insufficiently validated biomarkers and primarily hypothesis-generating findings.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Current biomarkers may provide supportive clinical context but should not routinely determine treatment selection without comparative evidence of differential treatment benefit.

• Future research should prioritise standardised assays, prespecified thresholds, independent validation and prospective biomarker-guided trials assessing both clinical utility and cost-effectiveness.

Introduction

Rheumatoid arthritis (RA) is one of the most common chronic inflammatory rheumatic diseases, characterised by symmetric polyarthritis leading to erosive changes as the predominant clinical manifestations.^{1 2} In addition, prevalent extra-articular comorbidities, including pulmonary and cardiovascular diseases, may significantly impair patient prognosis and long-term clinical outcomes.^{3 4} Driven by deeper insights into RA pathogenesis, current clinical management uses three distinct disease-modifying antirheumatic drugs (DMARDs) categories: conventional synthetic DMARDs (csDMARDs), such as methotrexate (MTX) and sulfasalazine (SSZ); biologic DMARDs (bDMARDs), including tumour necrosis factor-alpha inhibitors (TNF-α inhibitors/TNFi), tocilizumab (TCZ; interleukin (IL)-6 receptor blocker), rituximab (RTX; anti-CD20 monoclonal antibody) and abatacept (ABA; T-cell co-stimulation blocker); and targeted synthetic DMARDs (tsDMARDs), primarily represented by Janus kinase inhibitors (JAKi), such as tofacitinib (TOF), baricitinib (BAR), upadacitinib and filgotinib.⁵ Current management follows the treat-to-target (T2T) approach, with stepwise treatment optimisation until remission or low disease activity is achieved.^6–8 Meanwhile, studies have demonstrated that early intervention and attainment of remission are essential to reduce the risk of long-term disability and the development of comorbidities in patients with RA.^{9 10} Meta-analyses suggest that the overall efficacy of currently available bDMARDs and tsDMARDs is broadly comparable at the population level, with response rates typically ranging from 50% to 60%.^{8 11} Notably, approximately 40% of patients fail to respond to individual biologic therapies, and 5%–20% are refractory to all.^{12 13} Persistent inflammation can damage cartilage and bone tissue, resulting in irreversible structural damage, functional loss, additional adverse effects and a reduced quality of life.¹⁴

Consequently, there is an urgent need to identify predictive biomarkers that enable personalised medicine and optimise clinical outcomes. Currently, biomarkers proposed for predicting treatment response in RA span multiple levels. These range from demographic and clinical characteristics to serological markers, cytokines, immune cell phenotypes, synovial tissue features and single-cell transcriptomic signatures.

However, the available studies differ substantially in patient populations, treatment classes, definitions of treatment response, sampling time points, assay platforms and analytical strategies. This heterogeneity limits the value of a purely quantitative synthesis and makes it difficult to draw simple conclusions from a single biomarker category. Therefore, we chose a narrative review format to provide a cross-domain conceptual synthesis. We critically interpreted the translational maturity of different biomarker types and discussed how these findings may inform clinical decision-making and future precision treatment strategies in RA.

Literature search and selection

A systematic search of PubMed/MEDLINE and Cochrane Library was performed to identify relevant English-language publications available up to July 2026. Search terms included “arthritis, rheumatoid”, “biomarkers”, “disease activity”, “prognosis”, “treatment response” and “precision medicine”, “clinical characteristics”, “age”, “serological markers”, “autoantibodies”, “cytokines”, “immune-cell phenotypes”, “genetic”, “transcriptomic”, “synovial pathotypes”, “microbiome” and “DMARDs or antirheumatic agents”. Original articles, systematic reviews, meta-analyses, randomised controlled trials and observational studies were considered. Studies were selected based on methodological quality and clinical relevance. Priority was given to systematic reviews and meta-analyses, pooled registry analyses, prospective multicentre cohorts, randomised or head-to-head treatment studies and studies providing independent validation or informative non-replication. Exploratory single-cohort studies were included selectively when they represented a distinct biological pathway, analytical platform or emerging precision-medicine strategy. The evidence was narratively synthesised, and biomarkers were categorised according to their diagnostic, prognostic, predictive and disease-monitoring utility.

Demographic information and clinical profile

RA can be classified by age of onset into elderly-onset RA (EORA) and young-onset RA (YORA).¹⁵ EORA typically occurs after the age of 65, with an acute onset, a prevalence of approximately 2% and a nearly equal sex distribution or slight male predominance.¹⁵ Conversely, YORA usually develops between ages 30 and 50, with an insidious onset, female predominance (about 3:1) and a lower prevalence (0.5–1%).¹⁵ Against this epidemiologic backdrop, a natural question is whether age and gender influence treatment response in RA.

Age as a predictor of treatment response

Some studies demonstrate that age could have an impact on the response of treatment. Younger age could be a positive predictor of remission for patients with RA mainly treated with TOF,¹⁶ TCZ¹⁷ and adalimumab (ADA),¹⁸ whereas older onset has been associated with a superior response to MTX.¹⁹ However, in a retrospective cohort of 7912 patients with RA treated with csDMARDs and/or glucocorticoids (GCs), age-related differences in Disease Activity Score (DAS) and Health Assessment Questionnaire (HAQ) improvement were minimal and not clinically meaningful, indicating comparable treatment responses across age groups.²⁰ Similarly, a systematic literature review suggested that age did not significantly affect final DAS28(C-reactive protein (CRP)) or good European Alliance of Associations for Rheumatology (EULAR) response in comparisons between TNFi and placebo.²¹ Hence, although age may shape RA phenotype, its influence on treatment response appears modest and context dependent.

Sex difference in treatment response

Sex-based disparities significantly influence RA treatment outcomes, with female sex consistently identified as a negative predictor of response, particularly to TNFi.^{18 22 23} A post hoc analysis of a randomised controlled trial (RCT) found that 24-week Clinical Disease Activity Index (CDAI) remission rates were numerically higher in men than in women across the conventional treatment, certolizumab pegol (CZP), ABA and TCZ groups. After adjustment, only the effect of TCZ relative to conventional treatment differed significantly between men and women (p_interaction=0.015) with TCZ associated with a higher probability of remission in men.²⁴ However, given that this analysis was restricted to a highly selected population of newly diagnosed and DMARD-naïve early RA (eRA) patients, its generalisability to the broader RA population remains uncertain. While men with eRA respond better to TNFi, these differences disappear during follow-up in established RA (estRA) once baseline confounders are adjusted.²⁵

Notably, when evaluating treatment response, women had worse disability, as measured by the HAQ, higher pain Visual Analogue Scale (VAS), patient global assessment and fatigue VAS scores, as well as higher tender joint counts. Although not statistically significant, these differences may still contribute to persistently higher DAS28 or CDAI scores after treatment, thereby leading some women to be classified as having a poorer treatment response.^{25 26} While, in a large registry-based cohort, women showed slightly smaller improvements in DAS28-CRP than men, mainly due to smaller reductions in swollen joint count and CRP.²⁷ This suggests that subjective symptom-related components and objective inflammation-related components should be considered separately when interpreting sex-related differences in treatment response.

On a sidenote, initial treatment choices vary by sex: data derived from the Measurement of Efficacy of Treatment in the Era of Outcome in Rheumatology, an international observational register recording daily clinical practice showed that women were more frequently started on hydroxychloroquine, either as monotherapy or in combination with MTX or GCs, whereas men were more likely to initiate MTX or SSZ.²⁸ Although women may show a statistically inferior response to initial csDMARD combinations, the clinical magnitude of this difference is generally considered negligible.²⁸

Body mass index as a predictor of treatment response

Obesity has been shown to influence therapeutic outcomes in RA, with body mass index (BMI) as the standard metric: overweight 25–30 kg/m² and obese≥30 kg/m².²⁹ Multiple studies indicate that baseline obesity is a strong independent predictor of non-remission, regardless of treatment regimen.^30–36 Nevertheless, some evidence points to drug-specific differences in the impact of obesity on treatment response. For example, among obese versus non-obese patients, infliximab (IFX)+MTX shows no statistically significant efficacy difference, whereas standard-dose csDMARDs perform suboptimally in obese individuals³³; obese patients on TNFi are less likely to achieve treatment target at 6 months, independent of route of administration.^{37 38} However, some studies indicate that TCZ efficacy is not reduced in overweight or obese patients with RA and is largely independent of weight or BMI,^{37 38} suggesting that TCZ might be a reasonable therapeutic option for this population. Similarly, several studies suggest that BMI or obesity may not substantially influence response to JAKi^{32 39} or ABA,^{32 34 40} whereas evidence for RTX remains limited and inconclusive.^{32 34 40} In contrast, a recent pooled registry analysis associated obesity with poorer response and treatment persistence among patients receiving ABA, TCZ or JAKi.⁴¹

In eRA, higher BMI is not associated with achieving DAS28 remission or good EULAR response at 6 or 12 months, nor with the likelihood of receiving combination csDMARDs; in estRA, however, obese patients are less likely to attain remission despite comparable csDMARD regimens, while overweight but not obese status shows no significant association with remission.⁴²

The divergent clinical outcomes observed across different drug classes may be largely attributed to distinct metabolic and pharmacokinetic mechanisms. The reduced response to TNFi observed in patients with RA with higher BMI is consistent with population pharmacokinetic data showing that greater body weight is associated with increased drug clearance, shorter half-life and lower serum concentrations across TNFi agents.⁴³ Furthermore, obesity may impair IFX more than ADA or etanercept (ETN) possibly due in part to increased immunogenicity, suggesting the chimeric antibody is less able to overcome obesity-related inflammatory barriers.³⁶ Study shows that in patients with RA receiving csDMARDs, baseline insulin resistance and elevated pretreatment adipokine levels predict inferior responses to TNFi, IL-6 receptor blocker and anti-CD20 therapies.³⁷ Leptin contributes to impaired TNFi efficacy but seems to have negligible influence on TCZ responsiveness.³⁷ However, in a large, untreated eRA cohort, baseline adiponectin, leptin and resistin levels are not associated with the likelihood of achieving CDAI remission,⁴⁴ which may be related to differences in disease stage, prior treatment exposure and outcome definitions across the study populations.

Notably, BMI may serve as a surrogate marker of an underlying metabolic-inflammatory state, which may further influence systemic inflammation, disease activity and therapeutic efficacy.⁴⁵ Evidence from studies on glucagon-like peptide-1 receptor agonists (GLP-1 RAs) further supports the link between metabolic modulation and inflammatory disease activity. Although GLP-1 RAs were initially developed for the treatment of diabetes and obesity, emerging studies suggest that in patients with RA, their use may be associated with improvements in disease activity, reductions in acute phase reactants such as CRP and decreased expression of inflammatory cytokines such as TNF-α.^{46 47} Meanwhile, in vitro studies using RA fibroblast-like synoviocytes have shown that GLP-1 RAs can suppress TNF-α-induced inflammatory responses, oxidative stress and activation of inflammatory pathways such as nuclear factor κ-light-chain-enhancer of activated B cells (NF-κB) and c-Jun N-terminal kinase (JNK).⁴⁶ However, these findings should be interpreted with caution, as the observed clinical benefits may be partly mediated by weight loss and overall metabolic improvement⁴⁸ rather than direct RA-specific immunomodulatory effects. Therefore, BMI should currently be regarded as a modifiable clinical factor reflecting the patient’s immunometabolic background.

Smoking as a predictor of treatment response

Of note, smoking is a robust and relatively consistent predictor of poor therapeutic response in RA. Evidence from a multicentre randomised trial underscores current smoking as an independent risk factor for non-remission in eRA.³³ Specifically, current smokers carry a higher risk of inadequate response to MTX^{49 50} and various TNFi.⁵¹ Interestingly, the association between smoking and remission may differ by sex. A Japanese observational study suggests that among men, both former and current smoking, compared with never smoking, were negatively associated with DAS28-erythrocyte sedimentation rate (ESR) remission, independent of the medication used, whereas no such association was observed in women.⁵² This apparent sex difference might partly reflect sociocultural differences in smoking patterns in Japan, where smoking is substantially less prevalent among women.

Baseline disease activity

Some studies suggest that elevated baseline disease activity acts as a positive predictor for achieving remission.^{18 53 54} Elevated baseline DAS28 scores are associated with more pronounced reductions in disease activity for ADA and TCZ.^{18 53} Similarly, a moderate-to-high swollen-to-tender joint ratio significantly correlates with a twofold to threefold higher likelihood of achieving an American College of Rheumatology (ACR) 50% response during TNFi therapy.⁵⁴ However, the predictive value of baseline disease activity remains controversial. Among patients treated with TNFi or TCZ, higher baseline DAS28 correlates with lower remission rates at 3 and 6 months.^{55 56} Furthermore, higher baseline HAQ is a strong negative predictor of functional improvement during treatment across various bDMARDs.^{18 22 53 54} Given the inconsistent evidence, systematic reviews concluded that baseline activity remains insufficient for reliably predicting remission in patients treated with bDMARDs.^{21 57}

Clinical history

Beyond baseline activity, treatment outcomes may be influenced by the patient’s clinical history, including disease duration and prior medication use. Although the predictive value of disease duration was not independently associated with TOF response,¹⁶ a disease duration>5 years was associated with lower ACR50 and ACR70 response rates with ADA.^{58 59} Prior treatment history may also serve as a predictor across therapies. An increasing count of prior DMARDs linearly correlates with diminished ACR response rates, regardless of disease duration.^{59 60} Moreover, previous biologic treatment and use of non-selective non-steroidal anti-inflammatory drugs limited DAS28 improvements,¹⁸ whereas the absence of GCs at baseline predicts superior EULAR responses for patients treated with TCZ and TNFi.^{53 54}

Overall, sex may be associated with treatment outcomes in RA, with male sex linked to a higher likelihood of remission with TCZ; however, the overall effect of sex is generally modest and insufficient to guide individualised treatment selection. Obesity and current smoking show more consistent associations with a lower likelihood of remission, particularly during treatment with MTX or TNFi. Both factors should be considered in risk counselling and comprehensive disease management, but neither has been validated as a treatment-selection biomarker. Greater previous exposure to DMARDs or biologic therapies is associated with poorer subsequent outcomes, possibly reflecting disease refractoriness and treatment complexity. These readily available clinical variables can support risk stratification and interpretation of the treatment context, but no individual factor has sufficient predictive value to independently guide the selection of a specific DMARD.

Autoantibodies, antidrug antibodies and inflammatory biomarkers

Autoantibodies as predictors of treatment

Rheumatoid factor and anti-citrullinated protein antibodies

Autoantibodies are a hallmark of RA. Among these, rheumatoid factor (RF) and anti-citrullinated protein antibodies (ACPA) are well-established diagnostic markers and readily available pretreatment biomarkers, with ACPA status routinely assessed using anti-cyclic citrullinated peptide (anti-CCP) antibody assays in clinical practice.⁶¹ Many studies have attempted to use these characteristic autoantibodies to predict therapeutic response.

Some evidence consistently suggests that ACPA/RF seropositivity predicts superior therapeutic outcomes in patients treated with either RTX or ABA,^62–68 demonstrating a concentration-dependent advantage for both agents.^{63 64 66} Similarly, baseline ACPA and RF positivity correlate with enhanced responses to ADA⁶⁶ and TCZ,⁶⁸ respectively, suggesting that ACPA/RF positivity may serve as a broad predictor of higher remission rates across the entire bDMARD spectrum.⁵⁷ Meanwhile, in a retrospective study, multivariable logistic regression identified RF negativity as an independent predictor of good response to MTX.¹⁹

Nevertheless, such correlations are not universally observed across all clinical cohorts. In real-world practice among patients with newly diagnosed RA starting MTX, baseline RF/ACPA status was not associated with early remission.⁶⁹ Similarly, meta-analyses and multicentre studies failed to establish a significant association between baseline RF or anti-CCP status and clinical response to anti-TNFα therapy.^70–72 In addition to ABA,⁶⁸ where RF status failed to predict efficacy, clinical outcomes for TCZ and TOF likewise appeared to be independent of ACPA positivity.⁶⁷ Patients with high and low RF levels experienced similar clinical responses to CZP treatment, irrespective of previous inadequate responses or intolerance to TNFis.⁷³

Taken together, the most reproducible treatment-specific associations are observed for RTX, for which RF and/or ACPA positivity is supported by meta-analytic evidence, and for ABA, for which ACPA positivity, particularly at higher titres, has been replicated across registries and trial analyses. By contrast, meta-analyses do not support RF or ACPA as reliable predictors of TNFi response, while findings for MTX, ADA, TCZ and TOF remain inconsistent or insufficiently replicated. RF and ACPA may therefore provide supportive context when considering RTX or ABA but should not be regarded as broad or stand-alone biomarkers for DMARD selection.

Emerging autoantibodies

While conventional autoantibodies such as RF and ACPA are integral to clinical practice, their limitations in predicting treatment response have spurred interest in the potential of other autoantibodies to provide more refined clinical insights. Several non-classical autoantibodies show regimen-specific associations with outcomes.

Recently, a three-antigen antibody panel targeting the Hasselt University RA antigens UH-RA.305/318/329 (anti-UH-RA.305/318/329) was associated with failure to achieve remission or low disease activity following first-line therapy.^{74 75} This association was initially reported in the Care in early RA (CareRA) trial⁷⁴ and was subsequently reproduced in the independent CAP48 early-RA cohort, including among patients with seronegative RA; however, its consistency varied across assessment time points and outcome definitions.⁷⁵ Exploratory analyses from the randomised controlled Arthritis Prevention in the Pre-clinical Phase of Rheumatoid Arthritis with Abatacept (APIPPRA)/APIPPRA Long-Term Outcome (ALTO) study further suggested that a broader autoantibody profile, including anti-carbamylated protein antibodies (anti-CarP), may help identify individuals at risk of RA who derive greater benefit from preventive intervention with ABA.⁷⁶

The reported associations of anti-CarP with ABA response,⁷⁷ ACPA and/or anti-peptidyl arginine deiminase 4 antibodies positivity with ABA treatment outcomes,⁷⁸ anti-centromere protein F antibody with response to IFX⁷⁹ and different anti-carbamylated vimentin antibodies isotypes with responses to BAR-based regimens or MTX monotherapy⁸⁰ are each derived from individual studies or post hoc analyses. These findings lack independent validation and are therefore insufficient to guide routine treatment selection. A prospective cohort study additionally suggested that antibody–antibody interactions may partly explain the inconsistent associations previously reported between autoantibodies and response to TNFi; however, these findings also remain exploratory.⁸¹ Overall, among non-classical autoantibodies, the anti-UH-RA.305/318/329 panel has shown the most reproducible association with first-line treatment failure, whereas other individual and composite antibody markers may provide supplementary guidance for personalisation beyond RF and ACPA, but are not yet sufficiently validated for routine treatment selection.

Antidrug antibodies as predictors of treatment response

Treatment-emergent antidrug antibodies are associated with lower drug concentrations and poorer clinical response, with the most consistent evidence reported for ADA and IFX.^{30 82–84} Among patients switching from IFX to ADA, those with pre-existing anti-IFX antibodies were more likely to subsequently develop antibodies against ADA than TNFi-naïve patients.⁸⁴ In a separate cohort, patients who discontinued IFX or ADA and tested positive for antibodies against the respective drug achieved clinical responses to ETN comparable to those of TNFi-naïve patients. In contrast, switchers without detectable anti-IFX or anti-ADA antibodies had significantly poorer responses to ETN than both antibody-positive switchers and TNFi-naïve patients.⁸⁵ These findings suggest that antidrug antibody status may help distinguish pharmacokinetic treatment failure from mechanistic non-response and provide context for subsequent switching decisions. However, the evidence is primarily derived from observational cohorts, and antibody-guided switching strategies have not undergone prospective validation.

Inflammatory and tissue-remodelling biomarkers as predictors of treatment response

Beyond autoantibodies, circulating markers of systemic inflammation and tissue remodelling may reflect ongoing disease activity, but their treatment-predictive value remains limited. In a post hoc analysis of the tocilizumab safety and the prevention of structural joint damage (LITHE) trial, higher baseline the C-terminal telopeptide of type I collagen-to-osteocalcin ratio showed modest discrimination of week-16 joint-count response to TCZ (area under the receiver operating characteristic curve (AUC) 0.66), without independent replication.⁸⁶ Baseline CRP was associated with TCZ response in one real-world cohort¹⁷ but was not predictive in LITHE.⁸⁶ Higher baseline myeloid-related protein 8/14 levels were associated with response to ADA, IFX and RTX across three prospective treatment cohorts within a single study⁸⁷; however, the same analyte–treatment associations have not been externally replicated, and the related S100A9 analyte did not predict ETN response in an independent cohort.⁸⁸ Lower pretreatment 14-3-3η levels were associated with 1-year DAS28-ESR remission in the TCZ subgroup of a single observational study but remain unvalidated in independent cohorts.⁸⁹ Overall, these treatment-specific and outcome-specific findings are based predominantly on single post hoc or subgroup analyses, show limited or conflicting replication and have not demonstrated incremental treatment-selection value beyond routine clinical measures.

Imaging markers

Imaging complements clinical assessment in RA by detecting synovial inflammation and structural damage. Relevant features include radiographic erosions and joint-space narrowing; grey-scale synovial hypertrophy and power Doppler (PD) signal on ultrasound; and MRI-detected synovitis, tenosynovitis, osteitis (bone marrow oedema) and erosions.^{90 91} Although ultrasound and MRI can detect subclinical inflammatory changes that may not be apparent on clinical examination, their pretreatment value for predicting drug response remains uncertain.

Two systematic reviews illustrate the limitations of baseline PD. One of them found that ultrasound and clinical examination performed similarly in predicting treatment response and remission in active RA, whereas PD-positive synovitis was more informative for predicting flare in patients already in clinical remission.⁹⁰ In a subsequent review of 13 studies, higher baseline PD activity was associated with larger reductions in composite disease-activity scores in eight studies, but with a lower probability of remission in two studies and no association with remission in three.⁹² These findings are not necessarily contradictory, because greater improvement from a higher baseline does not imply attainment of an absolute remission threshold. Differences in the joints examined, therapies, follow-up periods, and outcome definitions further limit generalisability. Baseline PD should therefore be regarded as a candidate prognostic or monitoring feature.

For MRI, a post hoc analysis of 109 patients from a phase II trial found that several month 1 and month 3 changes in Rheumatoid Arthritis Magnetic Resonance Imaging Scoring system (RAMRIS)/Rheumatoid Arthritis Magnetic Resonance Imaging Quantitative assessment system (RAMRIQ) synovitis, osteitis and erosions predicted 12-month radiographic progression, whereas contemporaneous changes in CDAI and DAS28-ESR did not.⁹³ Because the analyses pooled patients receiving TOF monotherapy, TOF+MTX or MTX monotherapy, these findings support early monitoring of structural risk rather than pretreatment selection of a particular therapy. MRI is also constrained by cost, accessibility, acquisition burden and scoring complexity, whereas conventional radiography is relatively insensitive to short-term inflammatory change.⁹¹ Overall, imaging is valuable for detecting subclinical inflammation and monitoring disease evolution, but no single imaging feature is sufficiently validated to guide the selection of a specific therapy. Future studies should standardise image acquisition and scoring and prospectively test imaging-by-treatment interactions in randomised comparative designs.

Routinely available clinical variables, including sex and autoantibody status, have been evaluated in relation to treatment outcomes in RA. Table 1 summarises clinically available biomarkers with adequate evidence, as supported by the current literature. Although some associations have been reproduced across observational cohorts or exploratory trial analyses, their magnitude and consistency vary across therapies, populations and outcome definitions, and few have been supported by formal treatment–marker interactions. These variables therefore provide mainly prognostic or contextual information and should not be used individually to select a specific DMARD. Their incorporation into multivariable models may support clinical risk stratification, but their incremental predictive value and clinical utility require prospective validation.

Table 1. Selected clinically accessible candidate markers associated with treatment outcomes in RA.

Marker	Relevant therapy	Association with outcome	Evidence appraisal/replication status	Current clinical applicability
Male sex	TCZ	Male sex may be linked to a higher likelihood of remission with TCZ24	Low-to-moderate; independent prospective replication is lacking	Sex should not currently be used to select TCZ treatment
RF and/or ACPA	RTX	RF and/or ACPA positivity was associated with better response to RTX62 63	Moderate; the evidence is mainly derived from retrospective subgroup analyses and meta-analysis without prospective biomarker-guided validation	May provide supportive context for RTX use but is not validated as a stand-alone treatment-selection biomarker.
ACPA	ABA	ACPA positivity was associated with better response to ABA64–68	Moderate; the association has been replicated across real-world registries and exploratory trial analyses	May provide supportive context for ABA use, but current evidence does not establish preferential benefit from ABA over alternative therapies.

Detailed criteria are provided in the online supplemental material.

Evidence strength was qualitatively appraised at the biomarker–therapy–outcome level using a structured framework informed by established guidance for prognostic factors, prediction models and treatment-effect modification. The appraisal considered study design, sample size and statistical robustness, consistency and independent replication, external validation, comparative biomarker-by-treatment evidence, model specification and performance, and prospective clinical utility.

ABA, abatacept; ACPA, anti-citrullinated protein antibodies; RF, rheumatoid factor; RTX, rituximab; TCZ, tocilizumab.

Cellular and immunological markers associated with therapeutic response in RA

Cytokine and circulating immune-cell profiles have been investigated as accessible markers of treatment outcome in RA. In a prospective TNFi cohort, higher pretreatment plasma granulocyte–macrophage colony-stimulating factor (GM-CSF) levels were associated with 3-month EULAR response; however, discrimination was modest and imprecise (AUC 0.73, 95% CI 0.53 to 0.90) and was based on 53 patients without external validation.⁹⁴ In a separate TNFi cohort, good responders had lower baseline B-cell activating factor (BAFF) levels, but the more robust multivariable and receiver operating characteristic (ROC) findings concerned BAFF measured after 6 months of treatment, suggesting a potential response-monitoring role.⁹⁵ In a single ETN cohort, higher baseline IL-1β and IL-17A levels were observed among responders, although only IL-1β remained independently associated with response; baseline TNF-α and IL-6 levels were not informative.⁹⁶

In treatment-naïve eRA, higher baseline CXCL13 levels were associated with 2-year DAS28-CRP remission following MTX-based T2T therapy.⁹⁷ However, the remission analysis included both MTX+placebo and MTX+ADA groups and did not establish preferential benefit from ADA.⁹⁷ In an exploratory post hoc analysis of the Assessing Very Early Rheumatoid Arthritis Treatment-2 (AVERT-2 trial, higher baseline IL-10 levels were associated with better week 52 clinical and structural outcomes with ABA+MTX versus placebo+MTX in DMARD-naïve, ACPA⁺ eRA.⁹⁸ Although supported by a large, randomised comparison, this finding emerged from the evaluation of multiple biomarkers and lacks external replication.⁹⁸ Higher baseline semaphorin 4A levels were associated with subsequent treatment failure in one prospective cohort and with poorer 3-month response to MTX or TCZ in a second cohort, suggesting a general risk-stratification.⁹⁹

Circulating immune-cell composition has also shown exploratory value. Across two biologic-naïve cohorts, an inferred innate-dominant baseline profile was associated with 3-month TNFi response, whereas an adaptive-dominant profile was associated with non-response. Concordant findings in public transcriptomic datasets and a larger registry supported reproducibility, although no comparative treatment interaction was evaluated.¹⁰⁰ In a small study, higher baseline CD3⁻CD56⁺ NK-cell proportions were associated with 3-month remission with TCZ but not with TNFi response.¹⁰¹

Overall, these findings remain predominantly exploratory. With the possible exception of the post hoc IL-10 signal, the studies did not establish treatment-specific interactions, and none of the markers has undergone prospective biomarker-guided validation or demonstrated incremental clinical utility over routinely available predictors. They should therefore be regarded as research-use stratification candidates rather than clinically actionable treatment-selection biomarkers.

Genetic, epigenetic and transcriptomic markers

Genetic factors contribute to RA susceptibility and progression, whereas epigenetic and transcriptomic alterations are reflective of disease states and the level of immune activation. Therefore, using these layers to predict treatment response is complex but may be considered relatively valuable.¹⁰²

Genetic markers

Pharmacogenetic studies have identified numerous germline variants associated with treatment outcomes in RA, with most investigations focusing on TNFi.¹⁰³ However, reported associations are heterogeneous, and few have been independently replicated or evaluated for individual-level predictive performance.

Among the comparatively better-supported findings, an updated meta-analysis associated the minor allele of FCGR3A rs396991 with a modestly poorer response to TNFi, whereas the same allele was associated with a more favourable response to RTX.¹⁰⁴ Drug-specific associations were also reported for TNF rs1799724 with ETN response and TNFRSF1B rs1061622 with IFX response, but each pooled estimate was based on only two studies.¹⁰⁴ The contrasting direction observed for FCGR3A across TNFi and RTX is hypothesis-generating but does not establish a genotype-by-treatment interaction. Meta-analytic associations have also been reported between IL6R variants and TCZ response and between TNFSF13B rs9514828 and RTX response; however, these estimates were likewise based on small numbers of studies and require independent confirmation.¹⁰⁴

For TNFi, a meta-analysis combining discovery and replication cohorts comprising 2067 patients confirmed an association between LINC02549 rs7767069 and less improvement in DAS28 after correction for multiple testing.¹⁰⁵ An interaction was also observed for LRRC55 rs717117, for which the G allele was associated with less improvement among RF⁺ patients but not among RF⁻ patients.¹⁰⁵ The latter finding should be regarded as subgroup-specific pending further replication.

Most remaining associations have arisen from single candidate-gene cohorts. These include variants in the TLR/NFKB1 and MTHFR pathways, TLR9 rs352139, drug-stratified STAT4, PSORS1C1, TRAF3IP2 and PTPN2 variants, and MTX-pathway variants in ATIC, SLC19A1 and DHFR.^106–110 Several studies involved small samples, multiple variants or subgroup analyses, and some associations were attenuated after adjustment for clinical covariates. A 3.8-kb deletion at chromosome 2q14.3 was associated with poorer TNFi response in one cohort, but it occurred in only approximately 5% of patients and has not been independently replicated.¹¹¹ Conversely, four previously reported genetic variants identified through genome-wide association studies failed replication in an independent cohort and subsequent meta-analysis, illustrating the limited reproducibility of many initial pharmacogenetic signals.¹¹²

Overall, FCGR3A rs396991 and LINC02549 rs7767069 currently have the strongest replication support, although their effect sizes and individual-level predictive value remain insufficient for treatment selection. No genetic marker has yet demonstrated prospective genotype-guided clinical utility or incremental value over routinely available predictors. These variants should therefore be considered research-use association markers rather than clinically actionable pharmacogenetic tests.

Epigenetic biomarkers

Epigenetic studies of treatment response in RA have mainly examined pretreatment blood DNA methylation and chromosome-conformation signatures.

In a biomarker analysis embedded within the tREACH trial, higher baseline global leucocyte DNA methylation was associated with less improvement in DAS28 and EULAR non-response after 3 months of MTX-containing therapy, after adjustment for baseline DAS28, erythrocyte folate levels and BMI.¹¹³ However, most participants received combination therapy, the treatment-stratified associations were not statistically significant and a subsequent study using LINE-1 methylation as a surrogate measure of global DNA methylation found no overall association with MTX response^{113 114}

For TNFi therapy, a longitudinal study found concordant pretreatment methylation differences between responders and non-responders at the pathway level in discovery and independent validation cohorts. The implicated pathways included αβ T-cell activation and differentiation, regulation of small GTPase-mediated signalling and actin-filament organisation.¹¹⁵

Nevertheless, individual whole-blood cytosine–phosphate–guanine (CpG) associations did not survive correction for multiple testing, and no predefined classifier, threshold or individual-level predictive performance was evaluated. An earlier ETN study identified five CpGs with higher methylation in non-responders,¹¹⁶ but these CpG-level signals were not reproduced in the subsequent TNFi study, including in ETN-stratified analyses.¹¹⁵

A five-marker chromosome conformation signature (CCS) involving IFNAR1, IL-21R, IL-23, CXCL13 and IL-17A classified MTX non-responders in a 59-patient development cohort and a blinded validation cohort of 19 patients.¹¹⁷ Although the CCS provides a promising proof of concept, its small validation sample and the absence of large external or prospective treatment-guided validation preclude clinical application.

Overall, current epigenetic findings support biological stratification hypotheses, but no epigenetic biomarker has yet demonstrated sufficient individual discrimination or clinical utility for treatment selection in RA.

Transcriptomic markers

Transcriptomic features reflect gene-expression states shaped by the combined effects of genetic, epigenetic and environmental factors and are closely linked to downstream cellular functions. They may therefore capture the pretreatment immune and molecular states and have been investigated as candidate biomarkers of treatment outcome in RA.

In a prospective multicentre study of patients starting TCZ, baseline expression of IFI6, MX2, OASL and MT1G differed between responders and non-responders in a training cohort of 40 patients and a qPCR validation cohort of 20 patients; the best gene combinations achieved an AUC of 0.947.¹¹⁸ However, the small validation sample, study-specific response definition and absence of subsequent external validation limit the robustness and clinical generalisability of this finding. Lower baseline SOCS1 mRNA expression was associated with poorer disease control in eRA cohorts receiving MTX or other csDMARDs and with non-response in a cohort treated with RTX.¹¹⁹ Nevertheless, discrimination was modest in the independent early-RA cohort, and the RTX association was of borderline statistical significance.

A systematic review identified several microRNAs (miRNAs) associated with outcomes during MTX, TNFi or RTX therapy, including miR-146a and miR-125b.¹²⁰ However, heterogeneity in biological samples, assays, therapies, follow-up periods and response definitions precluded identification of a reproducible common panel. More recent studies have provided additional candidates. Higher baseline plasma miR-134-5p was associated with TNFi non-response after adjustment for age and sex, but it was not a differential response signal in the discovery sequencing analysis and showed only modest, non-significant discrimination in the qPCR cohort.¹²¹ Higher cellular miR-31 and miR-125b expression was associated with 6-month remission in 59 patients with treatment-naïve eRA, but the analysis used internal validation and did not establish treatment specificity.¹²² These miRNAs should therefore remain hypothesis-generating prognostic candidates.

At the systems level, analyses of public transcriptomic datasets have proposed peripheral-blood molecular subtypes characterised by neutrophil activation, interferon signalling or other immune and mitochondrial programmes.^{123 124} Exploratory comparisons suggested a numerically higher MTX response in the interferon-driven subtype¹²³ and a more favourable pooled biologic response in an innate-lymphocyte-rich mitochondrial subtype.¹²⁴ Adaptive immune receptor repertoire analysis provides another exploratory approach. In a TNFi-treated cohort, the DHRE (Asp–His–Arg–Glu) k-mer was detected at baseline only in responders, and IGH CDR3-length distributions also differed by response status.¹²⁵ These subtype and immune-repertoire findings remain exploratory because treatment-response analyses involved small or single cohorts and no locked classifier or individual-level prediction model has undergone independent prospective validation.

Overall, TCZ-associated IFI6, MX2, OASL and MT1G gene set represents a promising but insufficiently validated treatment-associated signature. Other transcriptomic, miRNA and immune-repertoire features require further external and treatment-comparative validation, and none is currently suitable for routine treatment selection in RA.

Tissue and microenvironmental markers

Synovial markers

The synovium is a principal target tissue of inflammation and joint damage in RA, and its cellular infiltrates, histopathological features, and states of molecular pathway activation directly reflect local disease mechanisms. The normal synovium is relatively paucicellular and consists primarily of a thin intimal lining layer and a sparsely cellular sublining layer containing only small numbers of immune cells. By contrast, the synovial lining layer in RA is typically markedly thickened, largely because of an increase in CD68⁺ macrophages. The sublining layer is extensively infiltrated by inflammatory cells, including T cells, B cells, plasma cells and macrophages, and is accompanied by stromal oedema and vascular proliferation.¹²⁶ Based on the composition of the inflammatory infiltrate, RA synovial tissue can be classified into three major pathotypes (figure 1): lympho-myeloid, diffuse-myeloid and pauci-immune.¹²⁷

Figure 1. Rheumatoid arthritis synovial tissue can be categorised into three major pathotypes. (1) lympho-myeloid, dominated by lymphoid lineage infiltration (T cells, B cells, plasma cells) in addition to myeloid cells; (2) diffuse-myeloid, characterised by a predominance of myeloid-lineage cells but relative paucity of B cells and plasma cells; and (3) pauci-immune, defined by scanty immune cell infiltrates and abundant stromal cells (created by Microsoft PowerPoint).

R4RA was a biopsy-driven, multicentre, open-label, phase IV randomised controlled trial that enrolled patients with RA who had an inadequate response to TNFi. It directly evaluated whether synovial B-cell status could distinguish relative response to RTX and TCZ in patients with an inadequate response to TNFi.¹²⁸ Among patients classified histologically as B-cell poor, the primary CDAI50 endpoint did not differ significantly between TCZ and RTX. Molecular reclassification using a synovial RNA-seq B-cell signature produced clearer separation: B-cell-poor patients more frequently achieved CDAI50 with TCZ than with RTX, whereas outcomes were similar in B-cell-rich patients.¹²⁸ However, the biomarker-by-treatment interaction was statistically significant for major treatment response (CDAI-MTR), defined as both a CDAI improvement of at least 50% and achievement of low disease activity (CDAI<10.1), but not for CDAI50. Thus, RNA-seq-defined B-cell status provides a comparative treatment-response signal, but it does not yet constitute a validated treatment-selection rule.

Secondary R4RA analyses identified additional treatment-associated features. Patients with a diffuse-myeloid pathotype had a significantly higher CDAI50 response rate with TCZ than with RTX, whereas responses were similar between treatments in the lympho-myeloid and pauci-immune/fibroid groups.¹³ Baseline transcriptomic analysis further showed that RTX response was associated with B-cell, immunoglobulin, antigen-presentation and interferon-related expression programmes, whereas TCZ response was associated with lymphoid as well as myeloid-cytokine, peroxisome proliferator-activated receptor signalling and metabolic programmes. Non-response to either RTX or TCZ was associated with higher expression of extracellular-matrix and tissue-remodelling genes, with Hox and fibroblast-related features being particularly prominent in RTX non-responders.¹³ These findings are biologically coherent but arose from secondary subgroup and transcriptomic analyses; the pathotype subgroups were small, and the R4RA-derived response and refractory-state models underwent internal rather than independent validation.

Stratification of Biologic Therapies for Rheumatoid Arthritis by Pathobiology (STRAP) trial and STRAP-EU trial extended biopsy-based stratification to biologic-naïve patients randomised to RTX, TCZ or ETN.¹²⁹ The primary hypothesis was not confirmed: among B-cell-poor patients, ACR20 response was similar with RTX and with TCZ or ETN. After stratification by synovial pathotype, the analysis suggested poorer RTX response in patients with a pauci-immune pathotype, but responses were similar across treatments in the lympho-myeloid and diffuse-myeloid groups.¹²⁹ Therefore, conventional B-cell assessment and histological pathotype alone have not shown reproducible treatment-selection performance across R4RA and STRAP.

Baseline analyses further identified several treatment-associated synovial features, including B-cell-related genes associated with response to ETN and RTX, SAA2 and IL18RAP associated with TCZ response and non-response, respectively, and tissue-remodelling genes associated with non-response to ETN and RTX. Deconvolution also suggested distinct cellular profiles among ETN and TCZ responders. However, these individual features were not independently validated as stand-alone biomarkers.¹³⁰

Separate pretreatment synovial RNA-seq models for ETN, TCZ and RTX achieved nested-cross-validation AUCs of 0.763, 0.748 and 0.754, respectively. The TCZ and RTX models retained discrimination in the independent R4RA cohort, while a 524-gene nCounter implementation yielded AUCs of 0.82–0.87 in residual STRAP samples.¹³⁰ However, the models were developed post hoc using a DAS28-ESR-based outcome; the ETN model lacked external validation, and nCounter testing was conducted within the original study cohort. They therefore remain research-use treatment-response classifiers without demonstrated prospective treatment-selection utility.

Other observational studies have associated greater pretreatment synovial T-cell infiltration and higher synovial TNF-α expression with better responses among patients treated with TNFi.^{131 132} However, the relevant treatment subgroups were small, the discrimination attributable to individual markers was limited, and findings for baseline synovial TNF-α were inconsistent across cohorts. A synovial myeloid expression programme was also associated with anti-TNF response. Serum sICAM1 and CXCL13, investigated as circulating surrogates of myeloid and lymphoid synovial phenotypes, showed differential associations with response to ADA and TCZ in an exploratory analysis of the ADalimumab ACTemrA (ADACTA) trial.¹³³ Because the thresholds were not prespecified and individual-level discrimination was modest, this profile requires external validation.

Overall, conventional histological pathotypes and individual synovial markers have not demonstrated sufficiently reproducible discrimination for treatment selection. RNA-seq-defined B-cell status showed a comparative signal in R4RA but was not consistently supported by STRAP, whereas externally evaluated multigene response models for TCZ and RTX showed stronger individual-level discrimination of treatment outcomes. Clinical implementation is nevertheless limited by the invasiveness and variable yield of synovial biopsy and by the expertise, infrastructure, cost and turnaround time required for RNA-based testing. Standardised assays and prospective treatment-guided validation, including assessment of feasibility and cost-effectiveness, are therefore needed before routine clinical use.

Microbiome profiles as predictors

Several observational studies have associated pretreatment gut microbiome features with treatment outcomes in RA, particularly during oral MTX therapy. One study identified Prevotella-dominant (RA E1) and Bacteroides-dominant (RA E2) enterotypes, with a higher MTX response rate observed in the RA E1 group in a separate treatment cohort.¹³⁴ However, the response analysis was limited, and the proposed enterotype-based gut microbial human index (EGMI) score was validated primarily for enterotype classification rather than as a treatment-response model.

In treatment-naïve patients with new-onset RA, a random-forest model based on 38 microbial Kyoto Encyclopedia of Genes and Genomes (KEGG) orthologs was developed in 26 patients and achieved an AUC of 0.84 for 4 month oral MTX response in an independent cohort of 21 patients.¹³⁵ Microbial functions related to purine, folate and MTX metabolism differed between responders and non-responders, and greater ex vivo depletion of MTX by pretreatment microbiota was associated with poorer subsequent response, providing biological plausibility. Nevertheless, the cohorts were small, the response definition was study-specific, and the model has not undergone larger prospective validation.

Other observational cohorts have reported associations between baseline taxonomic or functional microbiome profiles and outcomes during csDMARD or mixed DMARD therapy.^{136 137} However, the identified features and endpoints differed across studies, and predictive performance was generally modest. Microbiome composition is also sensitive to diet, geography, medication exposure, sample handling and analytical methods and may itself change during treatment. Overall, no microbiome-based marker is currently validated for routine prediction of treatment response or treatment selection in RA.

Selected research-use biomarkers associated with treatment response in RA include novel autoantibodies, soluble and cellular immune markers, genetic and epigenetic variants, transcriptomic and synovial signatures and microbiome features. Table 2 summarises experimental biomarkers with adequate evidence, as supported by the current literature. Although some have shown replicated associations or moderate discrimination in specific treatment settings, most remain exploratory because of small cohorts, post hoc analyses, heterogeneous methods and outcomes and limited external validation. Standardised assays, locked algorithms, independent validation and prospective biomarker-guided trials are required before clinical use.

Table 2. Selected experimental biomarkers associated with treatment response in RA.

Marker	Relevant therapy	Association with outcome	Evidence appraisal/replication status	Translational interpretation
Anti-UH-RA.305/318/329 antibody panel	First-line therapy, mainly MTX/csDMARDs	Anti-UH-RA.305/318/329 positivity was associated with failure to achieve remission or LDA after first-line therapy, including in seronegative RA.74 75	Moderate; the association was reproduced in an independent early-RA cohort, but sample size was limited and performance varied across outcomes and time points.	Research-use antibody panel; requires larger prospective and treatment-guided validation together with assay standardisation.
Baseline IL-10	ABA+MTX	Among DMARD-naïve, ACPA+ patients with early RA, higher baseline IL-10 was associated with a greater likelihood of achieving week-52 efficacy outcomes and less erosive progression with ABA+MTX versus placebo+MTX.98	Low-to-moderate; supported by a large, randomised phase III trial dataset, but the biomarker analysis was exploratory and post hoc, lacking external replication and a prospectively defined cut-off.	Comparatively promising ABA-associated biomarker; requires external validation and demonstration of incremental treatment-selection value.
Baseline innate–adaptive immune-cell composition signature	TNFi	An innate-cell-enriched baseline signature was associated with good or moderate response to TNFi, whereas an adaptive-cell-enriched signature was associated with non-response at 3 months.100	Low-to-moderate; the direction was concordant across two small biologic-naïve cohorts and supported by external transcriptomic datasets and a large registry analysis, but no prospective treatment-guided validation is available.	Promising research-use immune-composition signature; requires a standardised assay, predefined scoring algorithm, cut-off and confirmation of treatment specificity.
FCGR3A rs396991	TNFi; RTX	The minor allele was associated with modestly poorer TNFi response but more favourable RTX response.104	Moderate; supported by updated meta-analysis, although the TNFi effect was small and no prospective genotype-guided validation has been demonstrated.	Research-use pharmacogenetic association; insufficient for selecting RTX over TNFi in an individual patient.
LINC02549 rs7767069	TNFi	The variant was associated with less improvement in DAS28 following TNFi treatment.105	Moderate; supported in a meta-analysis of discovery and external replication cohorts comprising 2067 patients and remained significant after multiple-testing correction; effect size was modest.	Comparatively well-replicated TNFi-response association, but individual discrimination, incremental value and prospective clinical utility remain unestablished.
Pretreatment TNFi-related pathway-level DNA methylation profiles	TNFi	Baseline methylation profiles involving αβ T-cell activation and differentiation, small GTPase-mediated signalling, and actin-filament organisation differed between responders and non-responders, with concordant pathway-level findings in discovery and independent validation cohorts.115	Low-to-moderate; individual whole-blood CpGs did not survive multiple-testing correction, and no cut-off, discrimination, or calibration was evaluated.	Research-use biological stratification profiles; requires standardised measurement, a predefined scoring algorithm, external predictive validation, and confirmation of treatment specificity.
Five-marker CCS: IFNAR1, IL-21R, IL-23, CXCL13 and IL-17A	MTX	The five-marker signature classified MTX non-responders in a development cohort and a small, blinded validation cohort.117	Low-to-moderate; proof-of-concept classification was evaluated in 59 development patients and 19 validation patients, without large external or prospective treatment-guided validation.	Promising research-use classifier; requires a locked assay, predefined thresholds, larger independent validation and demonstration of incremental clinical utility.
PBMC gene-expression set: IFI6, MX2, OASL and MT1G	TCZ	Higher baseline expression was observed in responders; selected gene combinations achieved an AUC of up to 0.947 in the small validation cohort.118	Low-to-moderate; the response endpoint was study-specific, and the model and cutoffs have not undergone larger external validation.	Promising treatment-associated signature; requires standard response endpoint, larger external validation and comparative treatment evaluation.
SOCS1 mRNA expression	MTX and other csDMARDs; RTX	Lower baseline SOCS1 mRNA expression was associated with poorer disease control after MTX or other csDMARDs in early RA and poorer RTX response in established RA.119	Low-to-moderate for prognosis; supported across early-RA and established-RA cohorts, but discrimination was modest in the independent early-RA cohort and the RTX association was borderline.	Research-use prognostic marker; insufficient for selecting RTX or a particular csDMARD.
RNA-seq-defined synovial B-cell status, based on a B-cell gene-expression module	RTX; TCZ; ETN	In R4RA, RNA-seq-defined B-cell-poor patients had a higher CDAI50 response rate with TCZ than with RTX; evidence of biomarker-by-treatment interaction was observed for CDAI-MTR but not CDAI50. A related B-cell-poor stratification did not distinguish RTX from TCZ/ETN in STRAP.128	Low-to-moderate; supported by randomised comparative analyses, but the R4RA molecular subgroup signal was endpoint-dependent and was not reproduced using a related B-cell-poor stratification in STRAP.	Research-use molecular classification; insufficient for routine selection of TCZ or ETN over RTX.
STRAP-derived synovial multigene treatment-response classifiers, with a 524-gene nCounter implementation	RTX; TCZ; ETN	RNA-seq models achieved nested-cross-validation AUCs of 0.763, 0.748 and 0.754 for ETN, TCZ and RTX. TCZ and RTX models were evaluated in R4RA, with AUCs of 0.713 and 0.786; nCounter implementation retained discrimination in residual STRAP samples.130	Moderate; post-hoc development, outcome optimisation, no external ETN validation and no completed prospective biomarker-guided trial.	Multigene research-use classifier; assay and treatment-allocation utility require prospective validation before clinical use.
Serum sICAM1/CXCL13 profile as a circulating surrogate of synovial myeloid/lymphoid phenotypes	ADA; TCZ	A high-sICAM1/low-CXCL13 profile was associated with higher ACR50 response to ADA, whereas the inverse profile was associated with higher response to TCZ in the ADACTA biomarker analysis.130	Low-to-moderate; exploratory analysis within a randomised head-to-head trial, with non-prespecified median cutoffs, multiple-testing concerns, modest individual discrimination and no independent replication.	Comparative treatment-response candidate; not validated for selecting ADA or TCZ.
Pretreatment gut metagenomic 38-KO treatment-response classifier	Oral MTX	A random-forest model developed in 26 treatment-naïve patients achieved an AUC of 0.84 for 4-month MTX response in an independent validation cohort of 21 patients; greater ex vivo MTX depletion by pretreatment microbiota was also associated with poorer response.135	Low-to-moderate; supported by a small independent validation cohort and mechanistic ex vivo findings but larger external and prospective validation is lacking.	Research-use classifier requiring a locked assay and algorithm, standardised sampling and sequencing and prospective evaluation of incremental utility and cost-effectiveness.

Detailed criteria are provided in the online supplemental material.

ABA, abatacept; ACPA, anti-citrullinated protein antibodies; ADA, adalimumab; ADACTA, ADalimumab ACTemrA trial; anti-UH-RA, antibodies against Hasselt University rheumatoid arthritis antigens; AUC, area under the receiver operating characteristic curve; CCS, chromosome conformation signature; CDAI, Clinical Disease Activity Index; CpGs, cytosine–phosphate–guanine sites; csDMARDs, conventional synthetic disease-modifying antirheumatic drugs; CXCL13, C-X-C motif chemokine ligand 13; DAS28, Disease Activity Score 28; DMARD, disease-modifying antirheumatic drug; ETN, etanercept; IL-10, interleukin 10; KO, KEGG ortholog; LDA, low disease activity; MTR, major treatment response; MTX, methotrexate; RA, rheumatoid arthritis; RTX, rituximab; SOCS1, suppressor of cytokine signalling 1; STRAP, Stratification of Biologic Therapies for Rheumatoid Arthritis by Pathobiology trial; TCZ, tocilizumab; TNFi, tumour necrosis factor inhibitors.

Discussion

Current evidence identifies numerous pretreatment biomarkers associated with outcomes during particular RA therapies (figure 2) but few have been adequately validated for treatment selection. Many studies are small, retrospective or post hoc and differ in disease stage, treatment line, prior and concomitant therapies, assays and response definitions, limiting reproducibility.

Figure 2. Evidence map of selected pretreatment factors associated with outcomes across RA therapies. Selected pretreatment clinical factors and biomarkers are grouped by therapeutic class and evidence maturity. ↑ indicates that a higher level or positive status was associated with a better treatment outcome; ↓ indicates an association with a poorer outcome or non-response. Entries without arrows represent categorical biomarkers, multidimensional signatures, profiles or classifiers that cannot be interpreted using a single higher-versus-lower direction. ‘A>B’ denotes a better observed outcome with treatment A than with treatment B in the specified biomarker subgroup. The categories reflect the current maturity of the evidence rather than validated clinical utility. None of the presented markers or models has established prospective utility for routine treatment selection. ABA, abatacept; ACPA, anti-citrullinated protein antibodies; ADA, adalimumab; AIRR, adaptive immune receptor repertoire; Anti-UH-RA, antibodies against Hasselt University rheumatoid arthritis antigens; BAFF, B-cell activating factor; CarP, carbamylated protein antibodies; CCS, chromosome conformation signature; csDMARDs, conventional synthetic disease-modifying antirheumatic drugs; CTX-I, C-terminal telopeptide of type I collagen; CXCL13, C-X-C motif chemokine ligand 13; ECM, extracellular matrix; GM-CSF, granulocyte–macrophage colony-stimulating factor; IL, interleukin; JAK, Janus kinase; miRNA, microRNAs; MTX, methotrexate; MCV, mutated citrullinated vimentin; NK, natural killer cell; OC, osteocalcin; PAD4, peptidyl arginine deiminase 4; RA, rheumatoid arthritis; RF, rheumatoid factor; RTX, rituximab; TCZ, tocilizumab; SOCS1, suppressor of cytokine signalling 1; TNFa, tumour necrosis factor-alpha.

Importantly, an association with improvement within a treated cohort does not establish which treatment is preferable; this requires comparative evidence of differential treatment benefit, ideally supported by a prespecified biomarker-by-treatment interaction. A major evidence gap is therefore the scarcity of adequately powered comparative studies across different disease stages and treatment lines. Future research should embed standardised biospecimen collection and biomarker analyses within randomised trials, followed by validation in independent and clinically diverse populations.

Routine clinical factors are readily available but generally provide limited discrimination when used alone. High-dimensional molecular approaches can improve mechanistic understanding and support multivariable classifiers but their standardisation and implementation remain challenging. Blood biomarkers facilitate repeated sampling and multicentre evaluation but may not capture local joint heterogeneity, whereas synovial profiling directly samples diseased tissue and has generated comparative signals in biopsy-driven studies such as R4RA.¹³ However, biopsy burden, sampling variability, specialist expertise and molecular-testing costs limit its scalability.

Translation requires standardised assays and locked algorithms with established analytical validity. Prediction models should then be evaluated in independent cohorts using prespecified thresholds and outcomes, with transparent reporting of discrimination, calibration and incremental value over routine predictors.¹³⁸ Clinical utility should be tested separately in prospective biomarker-guided trials by demonstrating improved disease control or fewer ineffective treatment cycles. Conceptual modelling by Wientjes et al suggested that any benefit over conventional T2T care may be modest and that cost-effectiveness depends strongly on drug prices and early tapering.⁸ Predictive performance alone is therefore insufficient; feasibility, clinical impact and economic value must also be established.

Under current recommendations, treatment decisions remain based primarily on T2T principles, disease activity, safety, comorbidities, treatment history and patient preference rather than validated biomarker-based algorithms.⁹ Biomarker development should therefore begin with a clearly defined treatment-choice question and prospectively evaluate standardised markers or locked composite models across relevant therapeutic alternatives. Until such strategies demonstrate reproducible clinical and economic benefit, most candidates should remain classified as research-use response-associated markers or classifiers rather than clinically actionable treatment-selection tools.

Conclusion

Treatment response in RA reflects interacting clinical, immune, molecular, synovial and microbiome factors. Numerous pretreatment biomarkers have been associated with outcomes during DMARD therapy but most lack reproducibility and do not distinguish differential benefit between treatments. No biomarker or model is currently sufficiently validated for routine treatment selection. Translation will require standardised assays, locked algorithms, independent validation and prospective comparative trials with prespecified biomarker-by-treatment analyses. Integrated clinical–molecular models may ultimately outperform individual markers but must demonstrate clinical benefit, feasibility and cost-effectiveness before complementing current T2T care.