A discrete cluster of urinary biomarkers discriminates between active systemic lupus erythematosus patients with and without glomerulonephritis

Two independent cohorts were used for this study. The discovery cohort consisted of 85 active SLE patients satisfying four or more of the revised 1997 ACR classification criteria for SLE [28] and 24 age- and sex-matched healthy controls (HC). Patients were recruited from the University Health Network and Mount Sinai Hospitals. Sixty patients had ALN, confirmed by renal biopsy performed within 2 weeks of urine sampling for 88 % of patients (mean 7.5 days from sample accrual), with the remainder having active disease (score >0 on the clinical SLE Disease Activity Index-2000 (SLEDAI-2 K) components [29]) and no clinical evidence of LN. The validation cohort consisted of 33 patients with ALN (with ≥1 of the renal SLEDAI-2 K indices scoring positive and requiring changes in therapy), 16 patients with ANLN, and 30 patients with a prior history of biopsy-proven LN who were in remission with no urinary abnormalities (except one patient with fixed proteinuria (0.82 g/day) not requiring treatment).

All urine samples were spun to remove cellular debris and frozen at –80 ° C. To avoid repeated freeze/thaws, samples were thawed once on ice, sub-aliquoted, re-frozen at –80 ° C, and then individual aliquots thawed immediately prior to use. The urinary concentrations of 128 analytes were measured by coupled bead assay (Luminex using MILLIPLEX® Map Kits (EMC Millipore Corporation), Eve Technologies Inc.) Further information regarding the sensitivity and dynamic range of the assays can be found on the company website (http://​www.​emdmillipore.​com/​). For the majority of assays, the urine samples were run undiluted except: KIM-1 and renin (diluted 1/2); albumin, beta-2-microglobulin, clusterin, cystatin C, and osteopontin (diluted 1/50); and TIMP-1 and TIMP-2 (diluted 1/5). For the discovery phase, all analytes were measured in duplicate for ALN patient samples, with a single sample on each of two separate plates, and averaged. ANLN and HC samples were measured in singles and randomly assigned to one of the two plates with equivalent numbers for each group per plate. As the duplicates run on separate plates were very reproducible, samples in the validation phase were run in singles on a single plate. All results were normalized to urinary creatinine prior to analysis.

ISN-RPS histopathological class [30] and activity and chronicity scores [31, 32] were determined by an individual renal pathologist (CA-C), blinded to the results of the urine protein determination.

Normalized protein data were loaded into the R statistical environment (v3.1.1) for analyses. Data were scaled using the standard deviation for each variable and hierarchical clustering performed using divisive analysis (DIANA) with a Pearson’s correlation as a similarity metric. The Adjusted Rand Index, available in the mclust package (v4.4), was used to corroborate the clustering. Normalized abundance data were correlated with clinical variables across all patients using Spearman's correlation, followed by false discovery rate (FDR) adjustment of the p values to correct for multiple testing. A Venn diagram was generated using the VennDiagram package (v1.6.9) to visualize the overlap of significantly correlated genes among clinical variables [33]. Data were log₂ transformed and linear modeling was performed with the limma package (v3.20.9) in R to identify proteins with significant differences in abundance between groups. An empirical Bayes moderation of the standard error [34], followed by FDR correction, was employed [35]. Coefficients (i.e., log₂ fold-changes relative to control samples) that were determined to be significantly different from 0 following FDR correction (p _adj < 0.01) were carried forward. A Venn diagram was generated (as described above) to visualize overlap of significantly altered proteins among clinical variables, with hypergeometic testing performed to determine if overlap was greater than expected by chance alone. Finally, results (magnitude, direction, and significance of change) were visualized in a dotmap using the lattice (v0.20-29) and latticeExtra (v0.6-26) packages for R.

A receiver operating characteristic (ROC) analysis was performed to evaluate the ability of various analytes to discriminate between active patients with or without nephritis, or with and without proliferative changes on renal biopsy. A total of 9 analytes were evaluated using data from 85 active lupus patients for the analysis of nephritis and 8 analytes on 60 patients with paired renal biopsies for analysis of proliferative nephritis. The pROC package (v1.8) for R (v3.2.1) was used to calculate the true positive (sensitivity) and false positive (1 – specificity) rates across various analyte level thresholds, along with the area under the curve (AUC). ROC curves were created using the lattice (v0.20-33) and latticeExtra (v0.6-26) packages for R.

To assess the ability of selected analytes alone or in combination with conventional biomarkers to discriminate between proliferative and non-proliferative/chronic nephritis, all possible combinations of conventional biomarkers (C3, anti-dsDNA, albuminuria) and the top performing univariate analytes associated with the activity index (vWF, IP-10, PDGF-BB, IL-16, adiponectin) were evaluated. Samples were divided into four balanced folds (such that each fold contained a similar number of active proliferative (1) and non-proliferative (0) nephritis patients). For each combination of analytes, a linear model was fit using disease status (1/0) as the response to be predicted by the indicated analytes (i.e., response ~ analyte.1 + analyte.2 …). Each model was trained using data from three of the four folds, with the fourth fold used for testing. This process was repeated four times such that each fold was used for testing only once. Mean sensitivity, specificity, and accuracy were calculated across the four repetitions.

Urinary concentrations of 128 analytes (see Additional file 1: Table S1) were determined in SLE patients with ALN or ANLN, and HC, and normalized to urinary creatinine. Both SLE groups were comparable with respect to age, sex, and disease activity, although patients with ALN were on higher mean doses of prednisone and more immunosuppressive medications as compared with ANLN (Table 1). Hierarchical clustering (Fig. 1a) demonstrated that distinct groups of analytes identified ALN patients versus both HC and patients with ANLN (Adjusted Rand Index = 0.138, a measure of agreement between clusters with 0 indicating no agreement and 1 complete agreement), with no specific protein cluster discriminating between the latter two groups. These results suggested that a discrete number of urinary proteins preferentially identified patients with ALN.

To explore the relationship between individual urinary analytes and the three clinical states, linear modeling was performed. No statistically significant difference in urinary protein expression was seen when comparing all SLE patients (ALN and ANLN) to HC (Fig. 1b, left column). In contrast, statistically significant differences in analyte levels (p _adj < 0.01) were noted between the two active SLE populations with 44 proteins preferentially elevated in patients with ALN (Fig. 1b, right column, and Table 2). Differences between the two patient populations were striking, with 13 proteins having >4-fold (2 log₂ fold-change in the table) difference in the mean urinary concentration between patients with ALN versus ANLN, five of which (adiponectin, PAI-1, IL-15, PF4, and TIMP-1) demonstrated >8-fold increase in abundance (Fig. 2a). The concentration of these analytes in patients with ANLN was statistically indistinguishable from levels detected in HC, supporting their candidacy as LN-specific biomarkers. In further support of the concept that these biomarkers are specific for active renal disease in SLE, there was no association between the SLEDAI and urinary analyte levels in ANLN patients, and any significant associations between the SLEDAI and urinary analyte levels in the total patient cohort (or ALN subset) were lost when the renal components of the SLEDAI were removed from the calculation.

MCP-1, TWEAK, NGAL, and sVCAM have been proposed as potential LN activity-specific biomarkers in adult and pediatric SLE patients [7‐9, 11‐23, 25‐27]. Our findings confirmed that these urinary analytes are elevated in the urine of ALN as compared to ANLN patients (Fig. 2b and Table 2). However, the differences between the two patient populations were not as marked as those noted for many of our newly identified urinary analytes. These results suggest that our newly identified proteins may have a greater capacity to discriminate ALN from ANLN. To assess this possibility we contrasted the receiver operating characteristics (ROC) for the five analytes with the highest fold-difference between ALN and ANLN, with previously proposed urinary and serum (C3, anti-dsDNA) biomarkers, as well as established markers of renal dysfunction (Table 3; see Additional file 1: Figure S1). With the exception of sVCAM, the AUCs were higher and the sensitivities and specificities at optimal cut-offs were improved as compared to existing biomarkers. Furthermore, for many of the analytes, the AUC and sensitivities approached those seen for albuminuria, with several of the analytes demonstrating improved specificity over this conventional measure.

Notably, within the subset of patients with ALN, there was a poor correlation between the levels of these analytes and albuminuria (ρ ranging from 0.14 for TIMP-1 to 0.49 for IL-15), eGFR (ρ ranging from –0.04 for adiponectin to –0.57 for TIMP-1) or SLEDAI (ρ ranging from 0.16 for TIMP-1 to 0.38 for adiponectin). Representative plots showing the analytes with the strongest correlation for each of these parameters are shown in Fig. 2c.

As all of the ALN patients in the discovery cohort had a renal biopsy performed close to the time of their urine sampling (see Additional file 1; biopsy characteristics summarized in Additional file 1: Table S2), a Spearman correlation analysis followed by correction for multiple testing was performed to further delineate the relationship between urinary analytes and renal disease activity. Ten proteins (adiponectin, PAI-1, IL-16, wVF, IP-10, TIMP-1, eotaxin, sgp130, HGF, and PDGF-BB) were found to significantly correlate with renal biopsy activity score (q < 0.01; see Additional file 1: Table S3), with no analytes correlating with chronicity scores. Of the individual components of the activity index, only glomerular hyaline thrombi/wire loops and interstitial monocytes were not independently associated with these urinary biomarkers, with the strongest association being with the cellular crescents sub-score (ρ = 0.39 to 0.63) for the majority of the biomarkers.

Given that the activity score is driven primarily by histopathological features associated with proliferative lesions [31, 32], we examined whether any of these activity index-associated proteins discriminated between active proliferative (ISN/RPS III (A or A/C), IV (A or A/C)) and other renal lesions including non-proliferative (ISN/RPS I, II, V) or chronic lesions (ISN/RPS III (C), IV (C), V (C) or VI). As shown in Fig. 3a, eight of these proteins were found to significantly discriminate between proliferative and non-proliferative or chronic lesions (see Additional file 1; ROC analysis shown in Additional file 1: Table S4), and similar findings were observed when just the subset of class III/IV patients with chronic lesions was examined. For four of these analytes, the increases detected in ALN were driven solely by patients with active proliferative lesions, indicating that some urinary biomarkers may be specific for proliferative lesions. In contrast, urinary elevations of albumin, IL-15, and PF4 did not differentiate between proliferative and non-proliferative/chronic lesions on biopsy, but appeared to effectively discriminate patients with non-proliferative/chronic lesions from patients without LN (Fig. 3b). Notably, none of the urinary analytes discriminated between class II or V and chronic renal lesions, nor were there urinary analytes that could discriminate between class III and IV LN. There were no significant differences in the levels of the urinary analytes between class III/IV patients with or without class V changes. Although there were trends to higher levels of the activity-correlated urinary proteins in class IV as compared to class III nephritis and in class IV-G as compared to class IV-S nephritis, these did not achieve statistical significance when corrected for multiple comparisons. There was no association between corticosteroid dose at the time of urine sampling and the levels of any of the urinary analytes.

To investigate whether measurement of urinary analytes associated with the activity index results in an improved ability to discriminate between active proliferative and non-proliferative or chronic nephritis over conventional biomarkers, such as anti-dsDNA, C3, and albuminuria, we performed a multivariate analysis assessing all combinations of these conventional biomarkers together with the top five performing urinary analytes from the univariate analysis. The most accurate combination of conventional biomarkers was C3 and dsDNA (accuracy of 0.7675, sensitivity 0.875, specificity 0.5925). A number of different combinations of two to four urinary analytes together with anti-dsDNA had improved specificities and accuracies over this combination (see Additional file 1: Table S5), with the top performing combination being anti-dsDNA + vWF + IP-10 + PDBG-BB + adiponectin (accuracy 0.8325, sensitivity 0.87, specificity 0.7675). These findings support the potential clinical utility of measurement of these activity-specific biomarkers in the diagnosis of active proliferative nephritis.

To validate the findings in the discovery cohort, we measured the levels of the significantly increased urinary analytes from this cohort in a second independent cohort of 33 ALN and 16 ANLN SLE patients (see demographics in Table 1). In this smaller cohort, 18 of the original analytes identified replicated, including all five of the urinary analytes that previously demonstrated >3 log₂ fold-increase in LN (Table 2 and Fig. 4).

A key requisite of an activity-specific LN biomarker is normalization during renal remission. Therefore the levels of the replicated urinary LN biomarkers in the ALN patients were contrasted with those in 30 patients with previous LN that were in remission. Of the 18 replicated urinary analytes, all but two (albumin and IL-8) normalized in patients in remission (Table 2 and Fig. 4). These findings support the potential clinical utility of these analytes as biomarkers of active renal involvement in SLE.