Serum metabolomic profiling predicts synovial gene expression in rheumatoid arthritis

The Assessment of rituximab’s immunomodulatory synovial effects (ARISE) clinical trial (registered at ClinicalTrials.gov, NCT00147966) has been described in detail [3]. Briefly, the study enrolled people between 18 and 70 years of age with an established diagnosis of RA and a positive serum test for rheumatoid factor (RF). Patients had to have active disease (defined as a tender joint count > 8/68, a swollen joint count > 6/66, and either early morning stiffness > 45 min in duration, or an elevation in erythrocyte sedimentation rate (ESR) > 28 mm/h or C-reactive protein (CRP) > 1.5 mg/dL), despite the concomitant use of methotrexate (MTX) at a dose of > 12.5 mg/week for at least 12 weeks. One of the involved joints had to be a knee or a wrist that could be appropriately examined by arthroscopy. Concomitant use of non-steroidal anti-inflammatory drugs and oral prednisone at doses of 10 mg/day or less were permitted, provided dosing was stable for at least 4 weeks before the study. Patients previously treated with tumor necrosis factor (TNF-α) inhibitors were permitted to enroll in the study provided they had been off therapy for > 2 months for etanercept and > 3 months for adalimumab or infliximab. Patients meeting eligibility criteria underwent baseline arthroscopic synovial biopsy of an affected knee or wrist. Nineteen patients for whom both baseline synovial biopsy gene expression data and baseline serum metabolomics data were available were analyzed in the current study. Clinical disease parameters, including disease activity score (DAS), health assessment questionnaire (HAQ), pain, joint swelling and tenderness, ESR, RF, and anti-cyclic citrullinated peptide (anti-CCP) are described in Additional file 1: Table S1.

Synovial RNA was extracted from pools of six tissue fragments and complementary DNA (cDNA) was synthesized. TaqMan PCR was performed using predeveloped reagents (Applied Biosystems, Foster City, CA, USA) as described previously [3]. Gene expression, utilizing quantitative reverse transcriptase (RT)-PCR, was performed to measure inflammatory mediators and B cell survival factors, including IgM (heavy chain), IgG (heavy chain), IgKappa (light chain), CD3E, TNF-α, interleukin (IL)-1β, IL-6, IL-8, matrix metalloproteinase (MMP)-1, MMP-3, B lymphocyte stimulator (BLyS), stromal cell-derived factor 1 (SDF1), and a proliferation-inducing ligand (APRIL). Synovial gene expression for the 19 patients analyzed in this study are summarized in Additional file 1: Table S2.

TNF-α, MMP3 IL-1β, and IL-6 from serum were evaluated by DuoSet enzyme-linked immunosorbent assays following the manufacturer’s protocol (R&D systems).

Frozen serum was obtained from the Division of Rheumatology, Allergy, and Immunology at the University of California (UC) San Diego School of Medicine (San Diego, CA, USA). Lipid and protein fractions were removed via ultrafiltration (Nanosep 3 K OMEGA, Pall Corporation, Ann Arbor, MI, USA) at 4 °C. The filtered biofluid was used for NMR analysis. An aliquot of 160 μL of filtered serum was mixed with 20 μL D₂O and 20 μL of phosphate buffer (100 mM final concentration) containing TMSP-d4 (0.1 mM final concentration) and sodium azide (0.05% (w/v) final concentration). The prepared samples were centrifuged to remove any remaining particulates and a 180 μL aliquot was transferred to a 3 mm NMR tube (Norell, Landisville, NJ, USA) prior to acquisition. NMR spectra were acquired with a 16.4 T (700 MHz) Bruker Avance spectrometer (Bruker BioSpin Corp., Billerica, MA, USA) equipped with a 5 mm TCI cryogenically cooled probe and autosampler at 30 °C. Following acquisition, spectra were processed using NMRlab and MetaboLab [16]. Metabolite assignment and quantification was performed using several databases [16]. Metabolite assignment and quantification was performed using Chenomx NMR Suite (Chenomx Inc., Edmonton, AB, Canada), the Birmingham Metabolite Library [17], and the Human Metabolome Database [18]. The NMR results were recently published [19] and are summarized in Additional file 1: Table S3 for the 19 patients analyzed in this study.

The data, consisting of 19 patient samples measured across 18 synovial markers and 49 metabolites, were processed using Python. Hierarchically clustered heatmaps were generated for correlation between synovial markers and metabolites separately. Hierarchical clustering and visualization was performed using the scientific computing package SciPy, and the visualization package Seaborn (https://​seaborn.​pydata.​org/​). Dendrograms were divided into flat clusters using a cophenetic distance metric. Linear regression was performed between each cytokine-metabolite pair, controlling for patient age and gender using the ordinary least squares (OLS) method from the Python package StatsModels. Normally distributed independent variables were standardized so they had a mean of zero and a standard deviation of one. Discriminant analyses were performed to determine coefficients for linear combinations of variables that assigned cluster membership to individual cases. Basic descriptive statistics used to describe the patient population and discriminant analysis were performed using the SPSS software version 15.0.

We first analyzed whether synovial markers clustered into different groups (Fig. 1). IL-6, MMP1, and MMP3 are strongly correlated among themselves but are inversely correlated with TNF-α, which interestingly, is strongly correlated with CD3E. MMP1 and MMP3 are also inversely correlated with another cluster that includes IL-1β and IL-8. In addition, there was a big cluster comprising B and plasma cell markers, and growth factors, including SDF1, APRIL, CD138, CD19, CD79A, IgG and IgM heavy chains, and IgKappa.

We also characterized the blood metabolites. As shown in Additional file 1: Table S3 and Fig. 2a, most of the metabolites were downregulated compared to reference values, suggesting that these metabolites might be consumed by the inflamed synovium due to an increase in its metabolic demand. A few metabolites were upregulated compared to reference values, including glycolytic metabolites such as lactate and pyruvate. This likely reflects the increased bioenergetic and biosynthetic demands of sustained inflammation. Choline metabolism has recently been strongly related to inflammation [20]. Dietary intake of choline, through two circulating metabolites, trimethylamine (TMA) and trimethylamine N-oxide (TMAO), are mechanistically linked to cardiovascular inflammation [20]. Interestingly TMA was also elevated in our patients with RA. In addition, 3-hydroxybutyrate, a ketone body, and select amino acids, such as leucine, threonine, tyrosine, and aspartate were upregulated in patients with active RA. We also analyzed whether blood metabolites could be clustered in groups. Metabolites primarily clustered into groups according to their biological function or chemical classification (Fig. 2b). As expected, the group of metabolites that were elevated in patients, namely lactate, methylmalonate, xanthine, and 3-hydroxybutyrate, were inversely correlated with the most of metabolites.

Linear regression was analyzed between each synovial marker-metabolite pair; age and gender were controlled for by including these factors as covariates in the model. The regression coefficients for each cytokine-metabolite pair were used to form a clustered heatmap to lend insight into which groups of synovial markers were correlated with which groups of metabolites. We observed significant clustering structures in the data (Fig. 3a). The color bar along the top of Fig. 3 preserves the synovial marker clusters from Fig. 1. Interestingly the clusters of synovial markers almost correspond to the clusters observed within synovial markers (Fig. 1), suggesting that cytokine clusters in the synovial tissue have a similar metabolite signature in blood. The most striking difference is seen in the cluster comprising SDF1, APRIL, CD138, CD19, CD79A, and IgG and IgM heavy chain, and IgKappa in Fig. 1, which is seen split into two groups in Fig. 3. One metabolite signature correlates with CD19, CD79A and IgG heavy chain, markers of B cells; and the other metabolite signature correlates with SDF1, APRIL, CD138 and IgM heavy chain, markers related to plasma cell biology. Of interest BLyS, had a different metabolite profile than the rest of plasma cell biomarkers. Metabolite regression p values are displayed in Fig. 3b, where the row and column order are preserved from Fig. 3a and Additional file 2: Figure S1.

As observed in Fig. 3, metabolites can be grouped into five clusters (Fig. 4) that were further analyzed using the MetaboAnalyst [21, 22] web tool for functional enrichment of these groups of metabolites. Both pathway significance and pathway impact were assessed using this tool (Additional file 3: Figure S2).

We then determined the most strongly correlated or anti-correlated serum metabolites for each synovial marker, using linear regression, and controlling for both age and gender. We also included Benjamini-Hochberg false discovery rate (FDR)-adjusted p values to correct for multiple testing. As shown in Fig. 5, the synovial markers TNF-α and CD3E were negatively correlated with several metabolites in serum. The significant polar metabolites were mapped to known metabolic pathways using MetaboAnalyst 3.0 [22, 23]. and ranked by their overall p values (Fig. 5c). Additional file 4: Figure S3, Additional file 5: Figure S4, Additional file 6: Figure S5, and Additional file 7: Figure S6 show correlation between metabolites and the remaining synovial marker clusters.

We then explored whether or not one or more metabolites in serum could discriminate between high or low levels of synovial marker gene expression. At present, no factors have been identified that fully explain or predict response to RA therapy [24], but pre-treatment differences at baseline between patient groups have been identified, including synovial tissue TNF expression and an increased number of synovial macrophages and T cells in patients who subsequently exhibited clinical improvement after initiation of anti-TNF therapy [25]. Therefore, we used stepwise discriminant function analyses to discriminate TNF-α or CD3E levels. Multivariate and cross-validation classification using the “leave-one-out” classification method was used for these calculations. We defined high or low marker levels according to their synovial marker gene expression mean. This stepwise discriminant analysis is presented in Fig. 6. For TNF-α discriminant analysis, three metabolites namely glutamine, TMA, and dimethylsulfone were sufficient to correctly classify 94.7% of TNF-α levels. There was canonical correlation of 0.821 and Wilks’ lambda of 0.326 when these three variables were used, with high significance (p < 0.001; Fig. 6a). For CD3E discriminant analysis, two metabolites namely carnitine and methionine were sufficient to correctly classify 89.6% of CD3E levels. There was canonical correlation of 0.765 and Wilks’ lambda of 0.414 when these three variables were used, with high significance (p < 0.001; Fig. 6b).