Baseline metabolic profiles of early rheumatoid arthritis patients achieving sustained drug-free remission after initiating treat-to-target tocilizumab, methotrexate, or the combination: insights from systems biology

From the U-Act-Early strategy trial, patients were selected who achieved sDFR, defined as being drug-free for ≥ 3 months until end of the study, after initiating tocilizumab, step-up methotrexate, or tocilizumab plus methotrexate therapy. As controls, we selected patients who never achieved a drug-free status at any time point during the study. A detailed description of the study design has been reported previously [20]. Briefly, DMARD-naïve patients with early RA were randomized (1:1:1) to one of the three strategy arms and treated to the target of sustained remission, defined as disease activity score assessing 28 joints (DAS28) < 2.6 with ≤ 4 swollen joints for ≥ 24 weeks. Tocilizumab was administered intravenously every four weeks at a dose of 8 mg/kg with a maximum of 800 mg. Methotrexate (oral) was given every week with a starting dose of 10 mg and was increased to 30 mg (or maximum tolerable dose) with steps of 5 mg every four weeks until remission was reached. If the treatment target was achieved, medication was tapered stepwise and finally discontinued, if remission persisted.

Baseline serum samples were measured on oxidative stress, amines, and oxylipins MS platforms, which have been applied previously and are validated [21‐23]. The oxidative stress platform covers various isoprostane classes, signaling lipids from the sphingosine and sphinganine classes and their phosphorylated forms, as well as three classes of lysophosphatidic acids: lysophosphatidic acids, alkyl-lysophosphatidic acids, and cyclic-lysophosphatidic acids (all in the chain length species range of C14–C22). The amine platform covers amino acids and biogenic amines and the oxylipin platform covers classical and non-classical eicosanoids from different polyunsaturated fatty acids. In total, 263 metabolites were measured for each sample on the different platforms: 57 signaling lipid mediators, 128 oxylipins, and 78 amines. Serum samples were thawed on ice and vortexed before preparation procedures following inner standard protocols [22, 23]; extra samples were pooled for internal quality control (QC). For each platform, QC samples were added and the relative standard deviation (RSD) per metabolite was calculated. Only those metabolites that complied with the acceptance criteria (RSD < 15% for amines; RSD < 30% for oxidative stress and oxylipins) were selected for further data analysis. Additional information about the metabolite profiling on the different platforms is provided in Additional file 1. All metabolite analyses were performed by the Biomedical Metabolomics Facility Leiden department at Leiden University.

For peak determination and integration, signaling lipid mediators profiled by the oxidative stress platform were pre-processed by LabSolutions (Shimadzu, Version 5.65); peak-picking of oxylipins was performed with Agilent MassHunter Quantitative Analysis software (Agilent, Version B.05.01) and amines with MultiQuant Software for Quantitative Analysis (AB SCIEX, Version 3.0.2). For all metabolites, raw data correction was accomplished using selected internal standards by calculating the ratio of peak area of the target compound to the peak area of assigned internal standard from which a response ratio for each analyte was obtained. QC samples were used for evaluating the quality of the targeted compounds according to the in-house written protocol and the data were hereafter ready to be used for statistical analyses.

Baseline clinical characteristics are described as mean (standard deviation [SD]), median (interquartile range [IQR]) or as proportions (%); between-group differences (sDFR versus controls) were tested within each strategy arm using independent t test, Mann–Whitney U test, or Pearson χ² test, respectively. A linear mixed model with a random intercept and baseline DAS28, week of visit, and group (sDFR versus controls) as fixed effects was built to evaluate, within the strategy arms, differences in disease activity over time. As metabolite concentrations are influenced by a variety of factors, we performed principal component analyses (PCA) to identify possible confounders. The following parameters were considered: age; body mass index, gender, ethnicity, disease duration, smoking, alcohol consumption, seropositivity for rheumatoid factor (RF) or cyclic citrullinated peptide (CCP), erythrocyte sedimentation rate (ESR), and C-reactive protein (CRP). Thereafter, supervised partial least square discriminant analyses (PLSDA) were performed for each class (lipids, amines, and oxylipins) to identify relevant metabolites within each strategy arm. Several multivariate discrimination techniques currently exist but the main advantage of PLSDA is the handling of collinearity and noisy data (i.e. more observations than samples), both common in metabolomics experiments [24]. Data were first normalized (natural log-transformed) and then standardized (z-score) to ensure that all metabolite scores are comparable by giving them equal weight. The variable importance on projection (VIP) was used for metabolite selection; this measure accumulates the importance of each variable, whereas a higher VIP score shows that it is more relevant to predict the outcome [25]. Metabolites with VIP > 1 in the first component were considered important as the squared sum of all VIP values is equal to “1”, i.e. the average VIP. Thereafter, the Mann–Whitney U test (sDFR versus controls) was performed within these selected metabolites to identify those who are most relevant (p < 0.10) within each strategy arm (not corrected for multiple testing), which were subsequently used for pathway analysis in the Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. Pathways were considered relevant when p ≤ 0.05. In addition, integrative analyses between the previously identified transcripts [18] and proteins [19] and relevant metabolites were performed by calculating statistically significant (p ≤ 0.05) transcript–protein and protein–metabolite correlations (Pearson correlation coefficients [PCC]). These network analyses were visualized using VisANT 5.0 software [26, 27]. All other analyses were performed using the web-based tool MetaboAnalyst version 4.0 [28] and R version 3.4.3 (R Foundation for Statistical Computing, Vienna, Austria).

PCA revealed no clear confounders (data not shown) and therefore metabolite concentrations were not corrected for clinical characteristics in further analyses. PLSDA identified 35 metabolites (15 signaling lipid mediators, 14 oxylipins, 6 amines) in the tocilizumab plus methotrexate arm, 33 metabolites (9 signaling lipid mediators, 15 oxylipins, 9 amines) in the tocilizumab arm, and 33 metabolites (11 signaling lipid mediators, 13 oxylipins, 9 amines) in the methotrexate arm. Of these metabolites, 19, 13, and 12, respectively, were subsequently selected for further pathway analyses (Table 2). When comparing the metabolites between the strategy arms, one (LPA c16:1) showed overlap in the tocilizumab plus methotrexate and tocilizumab arm; two (12,13-DiHODE and LPA c16:0) in the tocilizumab plus methotrexate and methotrexate strategy, and one (3-Methylhistidine) in the tocilizumab and methotrexate strategy.

Pathway overviews for each strategy arm are shown in Additional file 2. The three most relevant KEGG pathways in the tocilizumab plus methotrexate arm were “histidine metabolism” (p < 0.001), “sphingolipid metabolism” (p = 0.004), and “arachidonic acid metabolism” (p = 0.037). Within the “histidine metabolism” pathway, a significant lower concentration of histamine (p = 0.002) was found in the sDFR group when compared to controls (Fig. 2). In the “arachidonic acid metabolism” pathway, production of PGD2 and 12,13-DiHODE, 9,10,13-TriHOME, and 9,12,13-TriHOME through lipoxygenase was higher in those who achieved sDFR. Within the “sphingolipid metabolism” pathway, significantly lower levels of sphinganine (Spha c18:0, p = 0.007) and sphingosine (Sph c18:1, p = 0.012) were found in the sDFR (versus controls) group, which are both related to ceramide generation. In the tocilizumab arm, three most relevant pathways were identified: “arachidonic acid metabolism” (p = 0.018); “lysine degradation” (p = 0.023); and “cysteine and methionine metabolism” (p = 0.030, Fig. 2). In the “arachidonic acid metabolism” pathway, higher concentrations of prostaglandin E2 (PGE2), 8-isoprostaglandin E2 (8-iso-PGE2), prostaglandin A2 (PGA2), 8-isoprostaglandin A2 (8-iso-PGA2), 8,9-DiHETrE, and 5,6-DiHETrE (p ≤ 0.08) were found in the sDFR group, compared to controls, implying a more active role of prostaglandins and isoprostanes, which are critical signaling molecules in various inflammatory disease including RA [29‐31]. Furthermore, in the “lysine degradation” pathway, significantly higher concentrations were found in the sDFR group (versus controls) of L-pipecolic acid (p = 0.026) and in the “cysteine and methionine metabolism” pathway, slightly but not statistically significantly lower concentrations were found in the sDFR group for cystathionine (p = 0.052) and homocysteine (p = 0.09). Most metabolites in the methotrexate arm were associated with the “arginine and proline” and “histidine metabolism” pathways (p = 0.022 and p = 0.025, respectively). When compared to controls, lower oxylipin levels of 14,15-DiHETE, 19,20-DiHDPA, and 12,13-DiHODE were found in the sDFR group (p ≤ 0.10), which indicates fewer active cytochromes P450 (CYP450) and lipoxygenase-based fatty acids metabolism, while the observed increased L-proline (p = 0.040) and L-arginine (p = 0.08) levels suggest a more active role of this pathway (Fig. 2). In the “histidine metabolism” pathway, decreased levels of 3-methylhistidine (p = 0.06) and higher levels of anserine (p = 0.06) were observed in the sDFR group (versus controls). Other changes of amine levels included higher concentrations of lysine (p = 0.032) and lower concentrations of 5-hydroxy-tryptophan (p = 0.08). Another significant pathway in the methotrexate arm was “aminoacyl-tRNA biosynthesis” (p = 0.027).

Figure 3 shows significant transcript–protein and protein–metabolite correlations within the three strategy arms. The average PCC between the biomarkers in the tocilizumab plus methotrexate arm was 0.54; these were 0.48 and 0.57 in the tocilizumab and methotrexate arm, respectively. Biomarkers showing > 10 connections within the networks were considered most important (i.e. signature biomarker) as they show the highest connectivity and therefore contribute most to the pathway analyses. The signature biomarkers in the tocilizumab plus methotrexate arm were the protein chemokine (C-C motif) ligand 5 (CCL5), as it was significantly correlated to seven metabolites and four transcripts, and the protein tumor necrosis factor receptor 1 (TNF-R1), being significantly correlated to three metabolites and 14 transcripts. In the tocilizumab arm, the proteins TIMP metallopeptidase inhibitor 1 (TIMP-1, 13 with transcripts) and thyroid peroxidase (TPO, 12 with transcripts) showed the highest number of correlations and in the methotrexate arm the protein granulocyte colony-stimulating factor (G-CSF, 4 with metabolites; 26 with transcripts).