Nineteen patients (15 females, four males; mean age, 50.8 ± 11.0 years; mean duration of disease, 15.8 ± 9.4 years; all Caucasian) who met the American College of Rheumatology criteria for RA 21 were studied; for details, refer to Table 1. More than three different disease-modifying antirheumatic drugs had failed to control disease activity before etanercept was administered. The study was approved by the ethics committee of the University of Magdeburg (71/99) and all patients were asked for written consent.

Each patient was given a standard dose of 2 × 25 mg etanercept per week subcutaneously. Disease-modifying antirheumatic drugs and steroids remained unchanged in all patients for the first week of TNF-blocking therapy. Blood samples were taken at 7:00 a.m. before treatment (time t₀; baseline), and at 72 hours after the first application of etanercept (time t₁). Comedication was given after blood was taken.

Patients were assessed for overall disease activity using the 28-joint-count Disease Activity Score (DAS28) as described elsewhere 22. Patients were categorized according to the European Leage against Rheumatism (EULAR) recommendations 3 months after the start of treatment, considering an improvement of the DAS28 >1.2 a good response. X-ray scans were read by two independent experienced physicians, but the sequence of the X-ray scans was not blinded. After reviewing X-ray scans of hands and feet, the responder group was further characterized by the absence of new bone erosions after a time interval of at least 9 to 12 months of follow up.

Peripheral blood mononuclear cells from 25 ml blood were separated on a Ficoll density gradient 23. Using a FACSCalibur Flow Cytometer (Becton Dickinson, San Diego, CA, USA) the populations of CD3⁺, CD14⁺, CD19⁺ and CD56⁺ cells were determined to ensure comparability of peripheral blood mononuclear cell fractions of individual patients in the course of the study. Extraction of total RNA was performed using the Qiagen RNeasy kit (Qiagen, Hilden, Germany) including a DNA digest on-column according to the manufacturer's instructions.

Affymetrix^® microarray technology (Human Genome U133A gene chip) was used to analyse the expression levels of about 18,400 transcripts interrogated by more than 22,000 probe sets. The Human Genome U95A gene chip was applied to verify array data with selected patients. Labelling and microarray processing was performed according to the manufacturer's protocol. The scanning was carried out with 3 μm resolution, 488 nm excitation and 570 nm emission wavelengths employing the GeneArray Scanner (Affymetrix, St. Clara, CA, USA). The microarray data were stored according to the MIAME standard and are available from ArrayExpress 24 (accession number E-MTAB-11).

Expression levels of a subset of genes were measured by quantitative real-time RT-PCR performed with TaqMan assay reagents according to the manufacturer's instructions on a 7900 High Throughput Sequence Detection System (Applied Biosystems, Foster City, CA, USA) using predesigned primers and probes (GAPDH Hs99999905_m1, ICAM1 Hs00164932_m1, TNFAIP3 Hs00234713_m1, IL1β Hs00174097_m1, PDE4B Hs00277080_m1, PPP1R15A Hs00169585_m1, NFKBIA Hs00153283_m1, CCL4 Hs00237011_m1, IL-8 Hs00174103_m1, ADM Hs00181605_m1).

To calculate the gene expression change of selected genes, the ΔΔC_T method was used. According to this method, the threshold cycle values (C_T) for specific mRNA expression in each sample were normalized to the C_T values of GAPDH mRNA in the same sample. This provides ΔC_T values that were used to calculate the changes of gene expression levels. Thereby, for each gene, the gene expression change in the first 3 days (ΔΔC_T) is defined by the difference of the ΔC_T value at day 3 (t₁) and the ΔC_T value before treatment (t₀).

The microarray data were preprocessed using the Microarray Suite, version 5.0 (MAS5.0; Affymetrix, Santa Clara, CA, USA) in the default configuration, and were analysed by a set of algorithms.

First, an algorithm for calculation of a score J to rank differentially regulated genes. Basically, the J score introduced here is a t statistic, which compares the logarithm of the expression ratios t₁/t₀ (signal log ratios) between responders and nonresponders. Thereby, the confidence intervals of the signal log ratios provided by MAS5.0 are used. In this way, the J score considers interindividual differences as well as measurement errors. A higher J score represents a more significant differential regulation. J > 0 was used as the cutoff point to define genes as differentially regulated.

Second, an algorithm for learning of classifiers used for prediction of the therapy outcome on evaluation of the fold change of pairs and triplets of genes (Support-Vector Machine algorithm together with cross-validation by the leave-one-out method).

Finally, an algorithm for inference of hypothetic gene regulatory networks (modified LASSO algorithm).

These three algorithms are described in detail in Additional file 1.

Methods of multiple testing to control the type I error rates taking into account the large multiplicity (more than 22,000 probe sets) were not applied. This feature was circumvented by validating expression patterns of a selected set of genes (ICAM1, TNFAIP3, IL1B, PDE4B, PPP1R15A, NFKBIA, CCL4, IL8, ADM).

Before the start of treatment, all RA patients presented with a high disease activity reflected by a DAS28 (mean ± standard deviation) of 5.7 ± 0.7. Within 3 months of TNFα-blocking therapy, the disease activity decreased significantly looking at all patients as a group (DAS28 = 3.8 ± 2.1) (Table 1).

Twelve patients (patients 3, 8 to 13, and 15 to 19) were characterized by a good therapy response, as indicated by a significant reduction of the DAS28 >1.2 without progression of bone erosions as shown by X-ray scans of hands and feet. Three out of seven nonresponders (patients 4, 5 and 7) showed mild progression of bone erosion by X-ray reviewing. One patient (patient 6) was considered a nonresponder despite a good DAS28 response due to a progressive joint destruction as demonstrated by the X-ray scan. None of the clinical characteristics at baseline was significantly associated with the clinical outcome (Table 2)

Application of Affymetrix DNA-chip technology to monitor changes in the expression profile of about 14,500 known genes in peripheral blood mononuclear cells during anti-TNFα therapy reflected a differential response by our patients as evinced by changes in the DAS28 greater than 1.2. Forty-two genes represented by 46 probe sets (Table 3) were found to be differentially regulated in therapy responders and nonresponders. The majority (40 probe sets representing 36 genes) was stronger downregulated or lesser upregulated in responders compared with nonresponders.

The mean of expression signals at t₀ averaged over the responders (n = 12) and over the nonresponders (n = 7) did not differ significantly in these genes, with the exception of SCN2B with P < 0.05 (Additional file 1, Table S3a). A subset of 23 genes (represented by 27 probe sets) were approved to be differentially expressed according to the permutation test, with a significance level α = 0.05.

All 1,035 gene pairs resulting from the 46 preselected probe sets of differentially expressed genes were examined according to their ability to clearly discriminate responders and nonresponders. For each gene pair, a set of classifiers was constructed and evaluated by cross-validation using the leave-one-out method. Seven gene pairs (Table 4) produced a prediction accuracy Q > 89%. Baseline levels of the selected gene pairs were not reliable in predicting the outcome as reflected by Q_t0log values between 42.1% and 79.0% (Additional file 1, Table S4a). The classification performance was also insufficient when using expression levels at t₁ (Q_t1log). Figure 1 shows a representative example of a discriminating gene pair (Q = 90.5%). Only one of the 19 patients (patient 16 – preclassified to be a clinical responder) matches with the pool of nonresponders. Owing to a DAS28 score that remained reasonably high, patient 16 eventually resembles a nonresponder according to EULAR criteria.

Finally, the separation strength of classification could be further improved by taking triplets of differentially regulated genes. Thereto, 15,180 triplets as combinations of the 46 selected probe sets were computed. Ten triplets were identified to express a prediction accuracy >95%. Figure 2 shows a three-dimensional plot of one representative triplet gene set as presented in Table 4.

Expression levels of a subset of genes were measured by quantitative real-time PCR for each patient and were compared with Human Genome arrays U133A and U95A (patients 1 to 11). As shown in Table 5, high correlations between the datasets obtained by three different methods of gene expression analysis were found.

In eight out of 20 genes selected for real-time quantitative RT-PCR (NFKBIA, CCL4, IL8, IL1B, PDE4B, TNFAIP3, PPP1R15A and ADM), the means of the gene expression change differed significantly for responders and nonresponders at significance level α < 0.05, as shown in Table 6. For all these genes, the means of the gene expression changes measured by quantitative real-time RT-PCR averaged over the seven nonresponders are positive, whereas those averaged over the 12 responders are negative or less positive than for the nonresponders.