Synovial membrane cytokine expression is predictive of joint damage progression in rheumatoid arthritis: A two-year prospective study (the DAMAGE study cohort)

Authors


Abstract

Objective

The primary aim of this prospective 2-year study was to explain the wide variability in joint damage progression in patients with rheumatoid arthritis (RA) from measures of pathologic changes in the synovial membrane.

Methods

Patients underwent clinical measurements and joint damage assessments by magnetic resonance imaging (MRI) and radiography at enrollment and at year 2. Synovial membrane was obtained by knee biopsy and assessed histologically by hematoxylin and eosin staining. Interleukin-1β (IL-1β), IL-10, IL-16, IL-17, RANKL, tumor necrosis factor α (TNFα), and interferon-γ (IFNγ) messenger RNA (mRNA) expression was determined by quantitative reverse transcription–polymerase chain reaction. The relationship of synovial measurements to joint damage progression was determined by multivariate analysis.

Results

Sixty patients were enrolled. Histologic features had no relationship to damage progression. Multivariate analysis by several different methods consistently demonstrated that synovial membrane mRNA levels of IL-1β, TNFα, IL-17, and IL-10 were predictive of damage progression. IL-17 was synergistic with TNFα. TNFα and IL-17 effects were most pronounced with shorter disease duration, and IL-1β effects were most pronounced with longer disease duration. IFNγ was protective. These factors explained 57% of the MRI joint damage progression over 2 years.

Conclusion

We have demonstrated for the first time in a prospective study that synovial membrane cytokine mRNA expression is predictive of joint damage progression in RA. The findings for IL-1β and TNFα are consistent with results of previous clinical research, but the protective role of IFNγ, the differing effects of disease duration, and IL-17–cytokine interactions had only been demonstrated previously by animal and in vitro research. These findings explain some of the variability of joint damage in RA and identify new targets for therapy.

Rheumatoid arthritis (RA) is characterized by progressive irreversible joint damage resulting in increasing disability (1, 2). However, joint damage accumulation is highly variable, differing up to 10-fold between individuals after 8–12 years of disease (3, 4). The mechanisms that produce these widely varying results are as yet poorly defined in humans. Rheumatoid joint damage is thought to arise primarily from processes in the inflamed synovial membrane (5, 6). This tissue becomes the site of a persistent immune response with accumulation of many immune cells, producing a large array of proteolytic enzymes, cytokines, chemokines, growth factors, and oxy-radicals, which may affect bone and cartilage structure (7, 8). In recent years there has been progress in the delineation of synovial pathologic processes in RA by animal and in vitro research (9, 10). Despite this increased information, the specific synovial pathways that mediate joint damage in patients are poorly defined. These pathways are ideal targets for therapy, as demonstrated by treatments that block the cytokine tumor necrosis factor α (TNFα), showing significant reductions in joint damage progression (11).

To better explain why accumulated joint damage differs greatly, we need large, well-designed, prospective studies. The relationships will inevitably be multifactorial and interactive and call for clinically informed statistical analysis. In one of the few illustrative studies, patients who responded to treatment both clinically and with reduced radiographic progression showed a reduction in many synovial cell types and TNFα protein identified by immunostaining (12). Other reports suggest that synovial membrane in vitro interleukin-1β (IL-1β) protein levels (13) and matrix metalloproteinase 2 levels (14) correlate with joint damage progression. These studies have been handicapped by suboptimal design and analysis. We report the findings of a prospective study of a large cohort of RA patients to explain the variability of joint damage by measures of synovial pathologic changes.

PATIENTS AND METHODS

Patient assessments.

Patients classified as having RA according to the 1987 revised criteria of the American College of Rheumatology (formerly, the American Rheumatism Association) (15) were recruited from outpatient clinics. We studied patients with varying disease duration and severity who fell across a spectrum of disease duration–dependent quantile curves of radiographic damage previously defined in >200 patients studied at our center (16). The study received approval by the Ethics of Human Research Committee of St. George Hospital, and prior informed consent was obtained from all patients.

Disease activity was assessed by the 28 swollen joint count (SJC), 28 tender joint count (TJC), erythrocyte sedimentation rate (ESR), C-reactive protein (CRP) level, DR4 genotyping, and self-reported patient function (by the Health Assessment Questionnaire [HAQ] [17] and by pain on a visual analog scale). The primary outcome measure was joint damage progression assessed by magnetic resonance imaging (MRI) of the dominant-hand metacarpophalangeal (MCP) joints from baseline and year 2. The secondary measure was joint damage progression assessed by radiographs of hands and wrists using the Larsen score (18). Patients took a variety of disease-modifying drugs throughout the study. Biologic therapies were not available at this time, and treatment was adjusted according to the usual care of their rheumatologists.

Imaging measures: MRI and radiographs.

MRI (Signa Horizon 1.5T unit; General Electric, Milwaukee, WI) was performed at baseline and at 24 months (RS). T1-weighted spin-echo axial and coronal images of the dominant second to fifth MCP joints were obtained with the following parameters: repetition time 480 msec, echo time 18 msec, slice thickness 3-mm coronal images with 0.3-mm interslice gap, 4-mm axial images with 0.4-mm interslice gap, field of view 120 × 120 mm, matrix 230 × 256 pixels. The MR images were scored by the same observer (PAB) in sequence using the Outcome Measures in Rheumatology Clinical Trials V RA MRI Score criteria (19). The reader was blinded to all other study data. Radiographs of hands and wrists were performed using a standard projection. Radiographs were read using the Scott modification of the Larsen scoring method (20) on 2 occasions in paired and known chronologic sequence by a single reader (JPE) who was blinded to all other study data.

Synovial biopsy samples.

Synovial biopsy samples were obtained (by IJP) from predetermined sites (suprapatellar pouch, medial perimenisceal, and cartilage–pannus junction) within knee joints at baseline during arthroscopy, with the patient under local anesthetic, as previously described (21). Synovial biopsy samples were either immediately snap-frozen in liquid nitrogen and stored at −80oC for subsequent RNA extraction or placed in 10% formalin for histology.

Synovial histology.

Using standard procedures, histologic sections on hematoxylin and eosin–stained slides were scored by an experienced histopathologist (CSL) who was blinded to all other study data. The histopathologist used a previously validated scoring system that assessed lining layer thickness, vascularity, sublining fibrosis, and cellular infiltrates and patterns (perivascular, diffuse, or focal aggregates) (22).

Synovial membrane messenger RNA (mRNA) analysis.

We quantified the lymphokines interferon-γ (IFNγ), a Th1 cytokine, and IL-17, another Th1 cytokine with potent in vitro bone- and cartilage-degrading properties that has synergy with IL-1β and TNFα and is implicated in joint damage. In addition, we quantified the proinflammatory cytokines IL-1β and TNFα, the potent osteoclast stimulator activator RANKL, and IL-10 and IL-16, which have putative antiinflammatory actions (KMJ). Total RNA was extracted from snap-frozen synovial biopsy samples in parallel, using the Chomczynski technique, then quantitated and stored at −80°C before reverse transcription–polymerase chain reaction (PCR) analysis of samples in parallel, using previously reported conditions (21).

Quantitation was performed using an external standard curve, in a modification of a technique previously reported by Klimiuk et al (23). Standard curves were generated from PCR product purified using QIAquick PCR purification kits (Qiagen, Victoria, Australia). Dilutions of PCR product were distributed into single-use aliquots and stored at −30oC. The standard curve was constructed by plotting the adjusted density of the PCR product bands of the standard curve dilutions against the original amount of purified PCR product (fg) added to the PCR on a log-linear graph using GraphPad Prism version 2.01 (GraphPad Software, San Diego, CA). A nonlinear regression one-site binding (hyperbola) analysis was applied to create a standard curve. PCR product band data for each synovial sample were converted to the original amount of cytokine complementary DNA (in fg) per 0.13 μg total mRNA, the amount of total RNA used in each PCR reaction, using the equation of the standard curve run in duplicate in the same PCR. The standard curves were highly reproducible, with midrange intraassay coefficients of variation of 2.8% for IL-1β, 5.7% for TNFα, 11.9% for IFNγ, 6.8% for IL-10, 12.4% for IL-17, 5.3% for RANKL, and 20.8% for IL-16.

Statistical analysis.

All univariate, bivariate, and multivariate statistical analyses were performed (by MNL) using STATA 7.0 (24). Results are expressed as the mean ± SD if normally distributed or as the median if otherwise. We constructed graphic representations with scatter plots of all bivariate analyses of baseline and followup variables to evaluate distributional characteristics, nonlinearity, and influential outliers. Checks for normality were performed, and if there were significant deviations, log and other transformations were applied. Parametric statistical methods were preferred to nonparametric methods unless assumptions of the former could not be met. Inferential statistical tests were considered significant when P values were less than 0.02.

Multivariate regression methods were used to determine the predictive relationships between joint damage progression and measures of synovial membrane cytokine expression. Unlike bivariate analysis, multivariate analysis controls for the effect of other covariates, thereby potentially revealing relationships that are missed with bivariate analysis. Furthermore, by adding interaction terms to the analysis, we can look for effect modification. More importantly, the results from many multivariate procedures, but particularly from multiple linear (ordinary least squares [OLS]) regression, are invalid if the mathematical and statistical assumptions of the method are not looked for and met. Multiple linear regression is not robust to nonlinearity, nonconstant variance, residuals not normally distributed, influential cases, and other problems common to biologic data, particularly if the sample is small. We therefore carried out detailed regression diagnostics to check for these potential confounders. Although our data set is the largest yet reported, it is still small. Because of these limitations of the data, we reanalyzed the data using a method that had fewer assumptions about the characteristics of the data.

For our initial analysis we used multiple linear OLS regression with standard regression diagnostics to investigate normality, linearity, homoscedasticity, and independence of residuals. Raw and standardized residuals were investigated with statistical (Shapiro-Francia test) and graphic normality (quantile–quantile plots) procedures. Other graphic diagnostics included residual versus fitted plot, leverage versus squared residual plot, and added variable plots. Collinearity was examined with the variance inflation factor (VIF). We did not impute missing data. We created specific interaction terms based on known biologic synergisms of cytokines and to investigate whether disease duration modified the effect of synovial cytokines on damage progression. The final regression model was one that 1) only included variables with significant coefficients, 2) was not overfitted, and 3) maximized goodness of fit assessed using the adjusted R2, the proportion of the variation that is explained by the model, adjusted for the number of terms in the model. To assess the robustness of the model generated by multiple linear OLS regression, we repeated the analysis using a method that had fewer assumptions about the characteristics of the data, namely, robust regression (25). For the purpose of comparing the regression output of cytokine variables with that of variables that are more familiar to clinicians, we also examined traditional clinical and laboratory variables with our regression procedures.

RESULTS

Patients.

Sixty patients with RA (40 women and 20 men) were recruited (mean ± SD age 58.8 ± 13.2 years [range 32.8–83.7 years], disease duration 0.1–29.7 years, median 4.1 years [interquartile range 9]). Thirty-four percent of patients had a disease duration of <2 years, 70% (42 of 60) were seropositive for rheumatoid factor (RF), and 58% (29 of 50) were HLA–DR4 positive. Sixty-seven percent of patients (40 of 60) were taking disease-modifying antirheumatic drugs (DMARDs), 50% (30 of 60) were taking glucocorticoids, and 35% (21 of 60) were taking both DMARDs and glucocorticoids, while 18% (11 of 60) were taking neither.

Disease activity and damage values at baseline varied widely (see Table 1). Since many patients had low disease activity, shown by the wide range of most values, the mean scores for disease activity were generally lower than those found in clinical trials of active RA. At baseline, 47% of patients (28 of 59) had a Larsen score of zero, and 7% (4 of 59) had MRI scores of zero. Adequate baseline synovial membrane results plus good-quality MRI and radiographic images at both baseline and year 2 were available for 43 enrolled patients for MRI and for 52 enrolled patients for radiographs. Reasons for incomplete data sets were patient death (2 patients), inadequate RNA samples (4 patients), incomplete MRI sets (11 patients), and incomplete radiographic imaging sets (5 patients). Damage progression analysis used data from 43 patients with complete MRI and synovial membrane data and from 48 patients with complete radiographic and synovial membrane data.

Table 1. Summary imaging, synovial, demographic, clinical, and laboratory findings in patients with rheumatoid arthritis*
 No. of subjectsMean ± SDMinimumMedianMaximum
  • *

    MRI = magnetic resonance imaging; IL-1β = interleukin-1β; TNFα = tumor necrosis factor α; IFNγ = interferon-γ; TJC = tender joint count; SJC = swollen joint count; ESR = erythrocyte sedimentation rate; CRP = C-reactive protein; HAQ = Health Assessment Questionnaire; VAS = visual analog scale.

MRI score at baseline, scale 0–80587.3 ± 10.80465
MRI score at 2-year followup, scale 0–80479.2 ± 13.50480
MRI progression472.4 ± 3.3−1115
Radiographic score at baseline, scale 0–1505911.7 ± 19.40496
Radiographic score at 2-year followup, scale 0–1505216.3 ± 21.908.5101
Radiographic progression523.2 ± 6.20031
Cytokines, fg/0.13 μg total mRNA     
 IL-1β56196.2 ± 135.40182.8749.4
 TNFα561,198.7 ± 737.833.4945.63,189.3
 IL-10561,348.2 ± 1,060.165.3981.15,988.3
 IL-175427.4 ± 70.300372.1
 RANKL56363.5 ± 852.20157.25,951.3
 IFNγ5614.1 ± 16.407.178.1
 IL-16514,772.2 ± 2,428.3557.94,182.611,404.6
Histology (modified Rooney score), scale 0–405516 ± 4.4516.324.3
Age, years6058.8 ± 13.232.85983.7
Disease duration, years606 ± 5.90.14.129.7
TJC, 0–28 joints6010 ± 7.20828
SJC, 0–28 joints608.2 ± 5.408.523
ESR, mm/hour6036.4 ± 25.602690
CRP level, mg/liter6017.6 ± 21.806.688
HAQ score, scale 0–3601.2 ± 0.701.32.8
VAS pain score, 0–100 mm6044.8 ± 28.1047.5100

Cytokine mRNA results.

All patients (100%) expressed TNFα at least at one site, with IL-1β expressed in 55 of 56 patients (98%). IL-10 was expressed in all patients and IFNγ in 48 of 56 patients (86%). IL-17 was expressed in 15 of 54 patients (28%), RANKL in 52 of 56 patients (93%), and IL-16 in 51 of 51 patients (100%). For each patient, the mean knee synovial cytokine expression was calculated (Table 1) and used in all subsequent analyses. Analysis of the influence of DMARD or glucocorticoid therapy on the expression of each cytokine investigated showed no differences between groups receiving different medication with or without glucocorticoids.

Associations of synovial cytokine expression with measures of joint damage and disease activity by correlational (bivariate) analyses.

All correlations are reported in Tables 2 and 3. Although the correlation between damage assessed by MRI and damage assessed by radiography at baseline was weak, damage progression assessed by both modalities was strongly correlated. Baseline damage score and damage progression by MRI correlated with the expression of synovial membrane IL-1β mRNA (Table 2).

Table 2. Spearman correlation coefficients (and P values) for correlations of imaging measures with baseline synovial, clinical, and laboratory measures*
 MRI damageRadiographic damage
At baselineProgressionAt baselineProgression
  • *

    Statistically significant results are shown in boldface. RF = rheumatoid factor (see Table 1 for other definitions).

MRI damage at baseline1.00   
MRI progression0.34 (0.02)1.00  
Radiographic damage at baseline0.32 (0.01)0.37 (0.01)1.00 
Radiographic progression0.23 (0.11)0.62 (0.00)0.44 (0.00)1.00
IL-1β0.40 (0.00)0.34 (0.03)0.14 (0.32)0.05 (0.76)
TNFα−0.01 (0.95)0.10 (0.53)−0.20 (0.14)−0.06 (0.67)
IL-10−0.11 (0.42)0.12 (0.45)−0.04 (0.79)0.18 (0.22)
IL-170.05 (0.75)0.17 (0.30)−0.02 (0.86)0.08 (0.61)
IFNγ0.17 (0.23)0.04 (0.82)−0.25 (0.07)−0.20 (0.17)
RANKL0.06 (0.64)0.06 (0.69)−0.07 (0.62)−0.01 (0.95)
IL-16−0.20 (0.17)0.04 (0.80)−0.05 (0.74)0.06 (0.68)
Histology (modified Rooney score)0.22 (0.12)0.06 (0.71)−0.01 (0.96)−0.03 (0.84)
ESR0.38 (0.00)0.45 (0.00)−0.05 (0.70)0.32 (0.02)
CRP level0.20 (0.14)0.26 (0.07)0.32 (0.01)0.10 (0.46)
RF titer0.28 (0.03)0.42 (0.00)0.21 (0.11)0.42 (0.00)
TJC−0.04 (0.78)0.15 (0.32)−0.18 (0.16)−0.09 (0.52)
SJC0.00 (0.99)0.11 (0.46)−0.06 (0.63)−0.09 (0.54)
HAQ score0.18 (0.17)0.08 (0.59)0.00 (0.99)0.03 (0.85)
VAS pain score0.33 (0.01)0.08 (0.59)0.17 (0.19)0.05 (0.72)
Table 3. Spearman correlation coefficients (and P values) for correlations of baseline synovial cytokine and histology measures with clinical and laboratory measures*
 IL-1βTNFαIL-10IL-17RANKLIFNγIL-16
  • *

    Statistically significant results are shown in boldface. RF = rheumatoid factor (see Table 1 for other definitions).

IL-1β1.00      
TNFα0.22 (0.11)1.00     
IL-100.07 (0.63)0.42 (0.000)1.00    
II-170.40 (0.00)0.21 (0.13)0.29 (0.030)1.00   
RANKL−0.03 (0.81)0.10 (0.47)0.05 (0.71)0.23 (0.09)1.00  
IFNγ0.67 (0.00)0.42 (0.00)0.24 (0.07)0.52 (0.00)0.11 (0.41)1.00 
IL-16−0.02 (0.88)0.29 (0.04)0.32 (0.02)0.45 (0.00)0.23 (0.10)0.09 (0.53)1.00
Histology (modified Rooney score)0.58 (0.00)−0.11 (0.44)−0.07 (0.67)0.40 (0.00)−0.11 (0.44)0.52 (0.00)0.00 (0.98)
RF titer0.03 (0.81)0.11 (0.41)−0.01 (0.97)−0.07 (0.63)0.01 (0.93)−0.17 (0.20)0.00 (0.98)
ESR0.46 (0.00)−0.03 (0.81)0.02 (0.89)0.38 (0.00)−0.04 (0.78)0.35 (0.01)0.01 (0.92)
CRP level0.39 (0.00)−0.08 (0.54)−0.02 (0.87)0.52 (0.00)0.00 (0.98)0.19 (0.15)0.16 (0.26)
TJC−0.09 (0.51)0.11 (0.43)0.08 (0.58)0.06 (0.66)0.28 (0.04)0.19 (0.15)0.16 (0.26)
SJC0.03 (0.83)0.32 (0.02)0.19 (0.16)−0.01 (0.96)0.21 (0.12)0.08 (0.54)0.01 (0.95)
HAQ score0.30 (0.03)0.05 (0.69)−0.32 (0.01)0.13 (0.36)−0.14 (0.29)0.26 (0.05)−0.14 (0.34)
VAS pain score0.02 (0.90)−0.01 (0.96)−0.24 (0.07)−0.08 (0.59)−0.01 (0.92)0.01 (0.94)0.03 (0.82)

Damage progression by MRI and radiography correlated with both the ESR and RF titer. Damage at baseline by MRI correlated with the ESR and pain score. Radiographic damage at baseline correlated with the CRP level and disease duration.

Synovial histologic features correlated strongly with levels of IL-1β mRNA and with both IL-17 and IFNγ mRNA. The baseline ESR and CRP level correlated with IL-1β and IL-17 mRNA levels. The SJC correlated with expression of TNFα mRNA, but there were no correlations with the TJC.

Predictors of MRI damage progression.

To determine the predictors of damage progression on MRI, we first used a simple multivariate analysis with all cytokine variables entered. We then performed steps in which variables were either sequentially removed or added to assess their importance to the equation, followed by the development of interaction terms between IL-1β, TNFα, and IL-17, based on known in vitro properties, together with another interaction term of disease duration. The latter equation was also analyzed with rigorous regression diagnostic tests to check for properties of specific items in the equation that would make them invalid to remain in the analysis. This sequential process produced the final equation of 5 items, which contains the most important predictors of joint damage progression that also pass the regression diagnostic tests for validity.

1. Multiple linear regression with all synovial cytokine variables entered.

For the multivariate analysis with all variables entered into the model including MRI damage progression, the adjusted R2 was 0.282 (i.e., the model explained 28% of the variance). After controlling for all other factors, only the regression coefficients for IL-1β (positive coefficient) and IFNγ (negative coefficient, i.e., more IFNγ associated with less damage progression) made a statistically significant contribution (Table 4).

Table 4. Regression output of backward and forward selection stepwise multiple regression, with 7 cytokines entered in model*
MRI progressionCoefficientStandard error§P
  • *

    Statistically significant results are shown in boldface. See Table 1 for definitions.

  • n = 39, F[7,31] = 3.13.

  • Probability > F = 0.0127.

  • §

    R2 = 0.414.

  • Adjusted R2 = 0.282.

IL-1β0.0150.0040.000
TNFα0.00030.0010.739
IL-170.0080.0080.352
RANKL0.0010.0010.295
IFNγ−0.1020.0390.013
IL-100.0010.0000.102
IL-16−0.00010.0000.516
Constant−0.3851.4990.799

2. Multiple linear regression, using backward and forward stepwise regression (critical level of P < 0.1 for new variable entry and of P ≥ 0.2 for variable removal), with all synovial cytokine variables entered.

The next model kept 3 cytokines (IL-1β, IL-10, and IFNγ), yielding a marginally higher adjusted R2 (n = 39, R2 = 0.315) (Table 5) since fewer variables were included. IL-1β (positive coefficient) and IFNγ (negative coefficient) had P values less than 0.01.

Table 5. Regression output of backward selection stepwise multiple regression of MRI damage progression, with 7 cytokines entered in model*
MRI progressionCoefficientStandard error§P
  • *

    Statistically significant results are shown in boldface. See Table 1 for definitions.

  • n = 39, F[3,35] = 6.82.

  • Probability > F = 0.001.

  • §

    R2 = 0.369.

  • Adjusted R2 = 0.315.

IL-1β0.0150.0040.000
IL-100.0010.0000.068
IFNγ−0.090.0320.008
Constant−0.5071.0490.632

3. Backward and forward stepwise regression entering interaction terms and application of regression diagnostics.

IL-17 interactions with TNFα and IL-1β and disease duration interactions with each cytokine were generated. Entering all main effects and all interaction terms, the adjusted R2 was 0.75 and the model kept 10 variables (clearly overfitted for the number of subjects). Variables with positive coefficients were IL-10, RANKL, IL-17–TNFα interaction, and certain disease duration–cytokine interactions (disease duration–IL-1β and disease duration–IL-17–IL-1β), whereas variables with negative coefficients were IL-17–IL-1β interaction and disease duration–TNFα, disease duration–IFNγ, disease duration–IL-17, and disease duration–IL-17–TNFα interactions. All terms in the model were significant. Regression diagnostics showed that there were no omitted variables and that the model had constant variance (Table 6). Examination of DFBETA for each variable revealed 3 influential cases for all variables except IL-10. The regression was repeated with each case singly removed and with all 3 cases removed. RANKL, the disease duration–IFNγ interaction, and the IL-17–IL-1β interaction were no longer significant and were removed from the model (adjusted R2 = 0.63). The VIF revealed significant multicollinearity among the disease duration–IL-17 interaction terms. These terms were removed to produce the final 5-term model.

Table 6. Optimal model of multiple ordinary least squares regression of MRI damage progression: regression diagnostics*
CollinearityVIF1/VIF
  • *

    In addition to the test for collinearity, diagnostics performed included the Ramsey test for omitted variables (probability > F = 0.153), the Cook-Weisberg test for heteroscedasticity (probability > χ2 = 0.825), and the Shapiro-Francia test for normality of raw residuals (probability > 2 = 0.202). VIF = variance inflation factor; DD = disease duration (see Table 1 for other definitions).

DD–IL-17 interaction9.750.103
IL-17–TNFα interaction9.510.105
DD–TNFα interaction1.470.681
DD–IL-1β interaction1.240.809
IL-101.130.884

4. Final model.

The final model containing 5 items is shown in Table 7. It has positive coefficients for IL-10, IL-17–TNFα interaction, and disease duration–IL-1β interaction and negative coefficients for disease duration–TNFα interaction and disease duration–IL-17 interaction. This indicates that the effect of TNFα and IL-17 on MRI damage progression is attenuated with increasing disease duration, in contrast to IL-1β, which has a more important role in disease of longer duration. The model is not overfitted since it has only 5 terms with an adjusted R2 of 0.574, and regression diagnostics were met. To evaluate the consistency of our results, the analysis was repeated using robust multiple regression, and the results were identical.

Table 7. Optimal model of multiple ordinary least squares regression of MRI damage progression: regression coefficients*
MRI progressionCoefficientStandard error§P
  • *

    DD = disease duration (see Table 1 for other definitions).

  • n = 41, F[5,35] = 11.79.

  • Probability > F = 0.000.

  • §

    R2 = 0.628.

  • Adjusted R2 = 0.574.

IL-100.0010.0000.002
IL-17–TNFα interaction0.000040.0000.001
DD–IL-1β interaction0.0010.0000.000
DD–TNFα interaction−0.00020.0000.005
DD–IL-17 interaction−0.0090.0030.002
Constant0.51300.6330.423

Predictors of radiographic damage progression.

In analyses of the predictors of damage progression by radiography, the adjusted R2 was 35% (n = 43, IL-10 and IFNγ significant) using OLS regression and 40% with stepwise regression (IL-1β, IL-10, and IFNγ significant). Backward and forward selection stepwise regression after entry of all cytokines, disease duration, and their interactions gave an adjusted R2 of 0.65 keeping 11 variables. Although 7 of the terms were identical to those found for MRI damage progression, regression diagnostics showed that the model had omitted variables (P = 0.008) and that the raw residuals were not normally distributed (P = 0.002). The fitted values versus residuals plot demonstrated that radiographic progression was a disguised proportion because 64% of cases had zero radiographic progression. Further modeling and attempts to transform the radiographic progression variable did not solve these problems. Therefore, using stepwise logistic regression, radiographic progression was treated as a binary variable. The model retained 7 terms, of which 5 had P values less than 0.05 (IFNγ, IL-1β, disease duration–TNFα interaction, IL-17–TNFα interaction, disease duration–IFNγ interaction). IFNγ was protective with an odds ratio (OR) of 0.34 (95% confidence interval 0.14–0.82, P = 0.016). The disease duration–TNFα interaction OR was also <1.0 (P = 0.012). The model pseudo R2 was 63%.

Multivariate analysis of standard clinical and laboratory measures.

We compared synovial mRNA measures with standard clinical measures as predictors of joint damage variation. In the model with MRI damage progression, the adjusted R2 was 0.16. After controlling for all other factors, only the regression coefficient for the ESR (coefficient of 0.06) was significant. In the model with radiographic progression, the adjusted R2 was slightly higher, explaining 21% of the variance, and both the ESR (coefficient of 0.16) and CRP level were significant. The stepwise regression model yielded an adjusted R2 higher in the model with MRI damage progression (n = 47, adjusted R2 = 0.21) including the ESR and SJC in the model, but only the coefficient of the ESR was significant (P < 0.001). In the stepwise regression model with radiographic progression, the ESR, CRP level, and HAQ score remained in the model (n = 52, adjusted R2 = 0.23), but only the ESR and CRP level were significant.

DISCUSSION

This study has shown, for the first time in humans, 4 important findings previously demonstrated only in animal and in vitro studies. These are 1) the role of IL-17 in producing joint damage; 2) the synergistic interaction between IL-17 and TNFα; 3) the modifying effect of disease duration and 2 key proinflammatory cytokines, TNFα and IL-1β, the former important in early disease and the latter in established disease; and 4) the role of IFNγ in reducing joint damage progression.

We have also shown for the first time that quantitative assessment of mRNA expression of synovial cytokines explains at least half of the MRI joint damage progression in RA in a prospective study. These synovial membrane cytokine measures are superior to standard clinical measures in explaining joint damage progression variation.

With 60 patients prospectively entering the cohort and undergoing multiple clinical examinations, laboratory tests, imaging analyses, and biopsies over a 2-year period, the study generated a complex data set with unavoidable missing data. Multivariate analysis statistics are the most appropriate techniques for this complex data set. To validate our results, we subjected the data to rigorous assessment including several multivariate analyses. It is encouraging that irrespective of imaging modality (MRI or radiography) and method of analysis, we found a consistent explanation of damage progression.

The finding that IL-1β and TNFα contribute to joint damage in RA could be expected from known in vitro activities, animal model findings, and specific blocking treatments in patients with RA (10, 11, 26). Our findings modify these results by suggesting that the effect of these cytokines on damage progression depends on the stage of the rheumatic disease. Synovial TNFα expression has a greater effect in early disease on both MRI and radiographic damage progression, a finding consistent across all statistical analyses. Similarly, the IL-17–TNFα–disease duration interaction had an inverse relationship with damage progression. Conversely, IL-1β appeared to have a greater role in damage progression with increasing disease duration. These findings are similar to the situation in the animal model of collagen-induced arthritis, in which TNFα plays an important role in early disease and IL-1β becomes more critical later (26). Synergy between TNFα and IL-17 in relation to inflammation and cartilage and bone damage has been demonstrated in vitro (27), but has not yet been demonstrated in humans. This synergy could provide an important target for specific therapy.

In contrast, the finding that IL-10 is positively associated with damage progression may initially seem counterintuitive. Many studies suggest that IL-10 acts as an antiinflammatory cytokine in RA (28, 29). These studies were performed in animal models or in vitro assays and gave no indication that IL-10 mediated bone resorption, either directly or indirectly. Recombinant IL-10 used therapeutically in RA yielded no clinical improvements. We found that IL-10 measurements correlated with those of IL-1β and TNFα, both at the mRNA level (weakly) in the present study and as secreted protein in in vitro synovial membrane cultures (Kirkham BW, Juhasz KM: unpublished observations).

There are at least 3 possible explanations for our findings. If IL-10 levels correlate with macrophage activation, they may be a surrogate for other factors probably derived from activated macrophages, such as IL-18, that produce joint damage. Another more interesting explanation is suggested by results of studies of IL-10 genotyping in RA, which show that the genotype associated with high IL-10 production is associated with more rapid joint damage in a cohort of RA patients followed up prospectively (30). Results of that study suggested that high IL-10 production may increase RF and anti–citrullinated protein autoantibody production and that these factors may contribute to joint damage. Finally, IL-10 may be neither a surrogate nor a pathogenic factor, but its up-regulation could reflect an ineffective counterbalance to TNFα and other proinflammatory cytokines. Irrespective of the explanation, IL-10 proved to be the most consistent synovial marker predictive of damage progression.

IFNγ repeatedly demonstrated a protective role against joint damage progression in the analyses. This finding supports the reported findings of in vitro studies showing that IFNγ has a potent negative effect on RANKL-induced osteoclastogenesis by degrading the RANK adapter protein TNF receptor–associated factor 6 (31). Other explanations also include the finding that in adjuvant arthritis, inhibition of IFNγ in later disease exacerbates the disease, and our previous work shows high levels of synovial IFNγ expression in this phase (32). Similar results are found in the immune-mediated model of experimental allergic encephalomyelitis, and it has been hypothesized that this reflects T regulatory cell activity (33). IFNγ may also mediate macrophage maturation associated with a switch of cytokines from IL-1β and TNFα to less inflammatory cytokines such as IL-1 receptor antagonist.

RANKL acts as a potent osteoclast activator producing bone erosions in adjuvant arthritis (2). The limited predictive power of RANKL in our study may show that osteoprotegerin, its endogenous inhibitor, modifies RANKL-induced joint damage, in addition to the interaction with IFNγ above. Additionally, RANKL expression in our samples had a dichotomous distribution suggesting heterogeneity in the RA population.

Our study correlated synovial pathologic changes in the knee joint with damage progression in joints of the hand and wrist. Our results, as well as those of a previous study showing similar synovial pathologic changes between the 2 sites (34), suggest that RA is a systemic disease. Our choice of cytokines was constrained by available techniques that could accurately quantify mRNA expression in the small samples obtained by biopsy. In investigating the relationship of cytokines and joint damage, cytokine protein or mRNA levels may be used. Protein measurements resolve the issue of posttranslational modification of mRNA (i.e., the protein is closer to the effector molecule); however, mRNA is reliably quantifiable and can act as a signal for processes contributing to joint damage.

The design of our study has several features about which we should comment. In selecting subjects for the study, we identified patients with variable joint damage progression to include both patients with slowly progressive joint disease and those with rapidly progressive joint disease. We also wished to study “typical” RA patients, the majority of whom receive conventional DMARD therapy. Although these treatments slow radiographic damage, it is clear that progression often continues. The conventional DMARD treatments all produce general synovial immune suppression, with no drug having specific effects to alter the relationship of synovial activity to damage outcome. Cytokine-targeted therapies will require specific studies of their mechanisms of action.

IL-1β levels correlated in the bivariate analysis with MRI scores at baseline (i.e., joint damage status) and also with MRI joint damage progression (i.e., the difference between the baseline and year 2 scores). The other cytokines found to be important in joint damage progression in the multivariate regression analysis did not correlate with baseline damage scores. This study emphasizes that the identification of synovial processes involved in joint damage requires prospective studies to measure change in joint damage for each individual. It also shows the strength of multivariate analyses. Unlike bivariate analysis, multivariate analysis controls for the effect of other covariates, thereby potentially revealing relationships that are missed with bivariate analysis. Furthermore, by adding interaction terms to the analysis, we can look for effect modification. However, there are clear risks to these procedures. Interaction terms should only be generated if they are based on sound biologic foundations. Otherwise, spurious relationships are likely.

More importantly, the results from many multivariate procedures, but particularly from multiple linear (OLS) regression, are invalid if the mathematical and statistical assumptions of the method are not looked for and met (35). Although our data set is the largest yet reported, it is still small. We powered the study for 60 cases, but a number of patients could not or would not undergo a repeat MRI at followup. Because of these limitations of the data, we believed it was important to reanalyze the data using methods that had fewer assumptions about the characteristics of the data. If we showed consistency of results across all statistical approaches, this would increase our confidence in the validity of the results. In fact, this consistency was shown. Although our data set is the largest yet reported, its small size requires us to limit the number of terms included in our final model.

The main outcome measures of joint damage used in this study were change in MRI scores of the dominant hand, measuring joint erosions and bone defects (19), and the modified Larsen score of the hands and wrists. The 2 measures correlated only weakly at baseline for joint damage status measurements, showing the much higher sensitivity of MRI for detecting erosions, particularly in patients with early disease. In contrast, we have shown that these 2 systems are comparable for detecting changes in joint damage progression over a 2-year period (36), and our main synovial pathologic outcomes correlated with both measures. Longer studies would show larger changes in radiographic scores, but increasing the time between synovial sampling and study end point introduces increasing uncertainty about other factors that may influence the outcome between synovial sampling and final outcome measurement.

Nevertheless, lack of radiographic progression in our cohort of patients was also a problem. Nearly 70% of patients had zero progression on hand and wrist radiographs scored using the Larsen method; consequently, there was not enough variation in our radiographic outcome to discriminate among the synovial cytokine variables. The radiographic progression variable behaved more like a proportion than a continuous variable. Although we were able to generate a regression equation, it failed a number of regression diagnostics. The large number of zero scores prevented our use of log and other data transformations, and it also prevented optimal performance of logistic regression because there were too few patients with disease that progressed compared with those with disease that did not. A nonstatistical solution requires rescoring the radiographs using the Sharp method, since it may be more responsive than the Larsen method.

Our findings provide an important advance in our understanding of RA pathogenesis. The application of informed statistical analysis to sophisticated laboratory measurements has advanced our understanding of clinical progression in RA. Although we are confident of our findings, we recognize that the process of translational research must be iterative—from the laboratory to the patient and back again. We therefore believe that our findings may prove to be most useful in generating powerful hypotheses for further testing.

Ancillary