Identification and Validation of Urinary Biomarkers for Differential Diagnosis and Evaluation of Therapeutic Intervention in Anti-neutrophil Cytoplasmic Antibody-associated Vasculitis

Abstract

Renal activity and smoldering disease is difficult to assess in anti-neutrophil cytoplasmic antibody-associated vasculitis (AAV) because of renal scarring. Even repeated biopsies suffer from sampling errors in this focal disease especially in patients with chronic renal insufficiency. We applied capillary electrophoresis coupled to mass spectrometry toward urine samples from patients with active renal AAV to identify and validate urinary biomarkers that enable differential diagnosis of disease and assessment of disease activity. The data were compared with healthy individuals, patients with other renal and non-renal diseases, and patients with AAV in remission. 113 potential biomarkers were identified that differed significantly between active renal AAV and healthy individuals and patients with other chronic renal diseases. Of these, 58 could be sequenced. Sensitivity and specificity of models based on 18 sequenced biomarkers were validated using blinded urine samples of 40 patients with different renal diseases. Discrimination of AAV from other renal diseases in blinded samples was possible with 90% sensitivity and 86.7–90% specificity depending on the model. 10 patients with active AAV were followed for 6 months after initiation of treatment. Immunosuppressive therapy led to a change of the proteome toward “remission.” 47 biomarkers could be sequenced that underwent significant changes during therapy together with regression of clinical symptoms, normalization of C-reactive protein, and improvement of renal function. Proteomics analysis with capillary electrophoresis-MS represents a promising tool for fast identification of patients with active AAV, indication of renal relapses, and monitoring for ongoing active renal disease and remission without renal biopsy.

Systemic vasculitides are a heterogeneous group of disorders with inflammation of the blood vessel wall as their common hallmark. These disorders often pose difficulties with regard to diagnosis and monitoring of disease activity both at the initial presentation and during follow-up. In one subgroup of small vessel vasculitides, the advent of anti-neutrophil cytoplasmic antibodies (ANCAs)¹ in the 1980s not only provided a new pathogenetic concept but also a diagnostic marker (1, 2). In this group, the granulomatous ANCA-associated vasculitis (GAAV) (previously named Wegener granulomatosis) and the microscopic polyangiitis (MPA) share several common features, including pauci-immune focal crescentic necrotizing glomerulonephritis and often a pulmonary capillaritis (3). Because of the association with ANCA, these diseases (together with the Churg-Strauss syndrome) are sometimes collectively referred to as ANCA-associated vasculitis (AAV). Recently circulating endothelial cells have emerged as an important marker correlating with severity and activity of the systemic vasculitic disease, and their clinical use in small vessel AAV has been demonstrated (4, 5). Regarding renal involvement, which is found in up to 80–90% of the patients with AAV, activity is defined by kidney biopsy with pauci-immune necrotizing glomerulonephritis. Renal involvement may occur or recur at every point of the disease and the follow-up even if other organ involvement is controlled by immunosuppressive therapy. Early detection is important as renal prognosis depends on early administration of immunosuppressive treatment (6), and scarring and relapses increase the risk for terminal renal failure, which itself is a risk factor for patient survival. As kidney biopsy is invasive and the risk of bleeding increases with chronic renal damage, surrogate markers, such as rising creatinine, increasing proteinuria, and most importantly erythrocytes and erythrocyte cast in the urinary sediment, are used. However, these markers have limitations. Microhematuria might persist despite remission, proteinuria might increase despite improvement in renal function, and other renal diseases can also develop (7). Therefore, new markers for renal disease activity are eagerly awaited.

Recently proteome analysis of urine has presented itself as a promising tool in the definition of chronic renal diseases (8, 9). We have developed an analytical platform for human urine analysis using capillary electrophoresis (CE) coupled on line to an ESI-TOF mass spectrometer (8, 10, 11). This approach permits the rapid analysis of the low molecular weight urinary proteome/peptidome in a single step and has enabled identification and validation of several urinary biomarkers in patients with different renal diseases (11–14).

In this study, we aimed toward identification of biomarkers for active AAV and response to immunosuppressive treatment. The results indicate that urinary proteome analysis enables differential diagnosis and monitoring of renal disease activity of AAV.

EXPERIMENTAL PROCEDURES

Patients, Procedures, and Demographics

Informed consent was obtained from all patients and healthy controls after local ethics committee approval. These studies were also performed in accordance with the Helsinki Declaration.

Urine samples were collected in the morning, after voiding the first urine of the day, from 18 patients with active AAV with a pauci-immune necrotizing glomerulonephritis (eight GAAV and 10 MPA) and 19 patients with inactive AAV all with previous biopsy-proven pauci-immune necrotizing glomerulonephritis and comparable therapy. Diagnosis of GAAV (previously named Wegener granulomatosis) and MPA was made according to the American College of Rheumatology criteria and the Chapel Hill consensus conference definition (15). Five of the 18 patients with active AAV were treated with immunosuppressive drugs; four were treated with steroids prior to the urinary analysis. All renal biopsies were analyzed by a triple diagnostic procedure including immunohistology as well as electron microscopy. Clinical characteristics of all patients with AAV are summarized in Table I.

Table I.

Characteristics of patients with active AAV and patients in long term remission

	Active AAV patients (n = 18)	AAV patients in remission (n = 19)
Age (years)	64 (range, 19–74)	67 (range, 43–83)
Gender	9 female, 9 male	13 female, 6 male
Disease	9 GAAV, 9 MPA	15 GAAV, 4 MPA
Creatinine (μmol/liter)	167.5 (range, 65–501)	119 (range, 77–299)
Proteinuria (g/24 h)	0.8 (range, 0.1–2.1)	0.1 (range, 0.1–0.6)

Sensitivity and specificity were evaluated in 40 blinded samples from individuals with different, biopsy-proven active glomerular diseases: 10 with active AAV, 29 with other glomerular diseases (membranous glomerulonephritis, n = 9; IgA nephropathy, n = 4; focal segmental glomerular sclerosis, n = 6; minimal change disease, n = 2; lupus nephritis, n = 4; others, n = 4), and one normal control. In 10 of the active AAV patients (median age, 61.5 years; range, 19–69 years; five female), repeated urine samples were collected prior to and 1, 3, and 6 months after initiation of immunosuppressive treatment with prednisolone and intravenous cyclophosphamide pulses together with the documentation of creatinine, urinary sediment, proteinuria, and C-reactive protein (CRP). Clinical activity was expressed by the Birmingham vasculitis activity score (BVAS) (16).

In addition, CE-MS data from 225 patients with non-vasculitis kidney-related diseases and 200 healthy controls were utilized. Further controls for specificity were samples from patients with cytomegalovirus infection after renal transplantation (influence of a viral infection), nephrolithiasis, bladder cancer, and renal cancer (patients with other causes of microhematuria) and patients with hypertension and diabetes type II without microhematuria (influence of blood pressure). Samples and patients' demographic data are described in recent studies (12, 13, 17–22). All samples were midstream urine collected following a standardized collection protocol as also outlined previously (23). In short, second urine of the day was collected and frozen immediately after collection without the addition of any preservatives.

Sample Preparation and CE-MS Analysis

All urine samples for CE-MS analyses were from spontaneously voided urine and were stored at −20 °C for up to 3 years until analysis. Of note, we were unable to detect significant storage-related degradation in samples that were stored at −20 °C for >10 years when comparing these with similar samples that were stored for only several weeks. For proteomics analysis, a 0.7-ml aliquot of urine was thawed immediately before use and diluted with 0.7 ml of 2 m urea, 10 mm NH₄OH containing 0.02% SDS. To remove higher molecular mass proteins, samples were filtered using Centrisart ultracentrifugation filter devices (20-kDa-molecular mass cutoff; Sartorius, Goettingen, Germany) at 3000 relative centrifugal force until 1.1 ml of filtrate was obtained. This filtrate was applied onto a PD-10 desalting column (Amersham Biosciences) equilibrated in 0.01% NH₄OH in HPLC grade H₂O (Carl Roth GmbH & Co. KG) to remove urea, electrolytes, and salts and to enrich polypeptides present. Finally all samples were lyophilized, stored at 4 °C, and suspended in HPLC grade H₂O shortly before CE-MS analysis as described previously (20).

CE-MS analysis was performed as described previously (20, 24) using a P/ACE MDQ capillary electrophoresis system (Beckman Coulter, Fullerton, CA) on line coupled to a Micro-TOF MS instrument (Bruker Daltonics, Bremen, Germany). Data acquisition and MS acquisition methods were automatically controlled by the CE system via contact closure relays. Spectra were accumulated every 3 s over a range of m/z 350–3000 thomson. Accuracy, precision, selectivity, sensitivity, reproducibility, and stability are described in detail elsewhere (20, 25). In short, the detection limit is in the range of 1 fmol depending on the ionization properties of the individual peptide. In a urine sample, the detection limit (in the crude sample before processing) is in the range of 100–1000 fmol/ml. In 15 consecutive analyses the standard deviation was between 2.5 and 4%, and intermediate precision was 5.5%. Stability and variability over time were also assessed and are available in the supplemental documentation.

Data Processing and Cluster Analysis

Mass spectral ion peaks representing identical molecules at different charge states were deconvoluted into single masses using MosaiquesVisu software (26). Migration time and ion signal intensity (amplitude) were normalized based on 29 collagen fragments that serve as internal standards (25). These internal polypeptide standards are the result of normal biological processes and appear to be unaffected by any disease state studied to date (greater than 5000 samples analyzed to date) (23). The resulting peak list characterizes each polypeptide by its molecular mass (kDa), normalized migration time (min), and normalized signal intensity. All detected polypeptides were deposited, matched, and annotated in a Microsoft structured query language (SQL) database, allowing further analysis and comparison of multiple samples (patient groups). To establish identity of polypeptides observed in different samples, a linear function was used that allowed, depending on the mass of the polypeptide, a 50-ppm absolute mass deviation for peptides of 800 Da that increased linearly to 100-ppm absolute mass deviation for peptides with a maximum mass of 20 kDa. A similar linear function was used when comparing CE migration times, allowing a 5% absolute deviation. CE-MS data of all individual samples can be accessed in supplemental Table 3.

Disease-specific polypeptide patterns were generated using both support vector machine (SVM)-based MosaCluster software (14) and a method using a linear combination of log-transformed data (27). To perform the linear combination, all normalized signal intensity values for possible biomarkers below 1 were substituted with a value of 1. The average signal intensity for a specific biomarker over all cases was compared with the average intensity for the same biomarker over all controls. To avoid artificial weighting of specific biomarkers in the set because of the difference in observed signal intensities for case and control, the relative distance between the two averages (case and control) was always set to 2. This relative distance of signal intensities between the disease and control samples was provided using the formula (A_kⁱ − mean_averages)(2/|x̄_case − x̄_control|) where A_kⁱ is the log-transformed signal intensity of the ith biomarker in the kth sample in either the test set or the blinded set, mean_averages is the average of the mean intensity of all possible markers for test set samples, x̄_case represents the mean observed signal intensity of the possible biomarker from all vasculitis samples, and x̄_control represents the mean signal intensity of the possible biomarker from the combined control samples (apparently healthy individuals and patients with chronic renal disease other than vasculitis). This linear classification procedure is in some way a simplified version of the nearest centroid classification approaches. Variants of these approaches are those that include shrinkage as implemented in the significance analysis of microarrays and prediction analysis of microarrays classifiers (28) as well as those without shrinkage like the classical Fishers diagonal linear discriminant analysis and the classification to nearest centroids procedures (29).

Statistical Methods, Definition of Biomarkers, and Sample Classification

Estimates of sensitivity and specificity were calculated based on tabulating the number of correctly classified samples. Confidence intervals (95%) were based on exact binomial calculations and were carried out in MedCalc version 8.1.1.0 (MedCalc Software, Mariakerke, Belgium). The receiver operating characteristic (ROC) plot was obtained by plotting all sensitivity values (true positive fraction) on the y axis against their equivalent (1 − specificity) values (false positive fraction) for all available thresholds on the x axis (MedCalc Software). The area under the ROC curve (AUC) was evaluated as it provides a single measure of overall accuracy that is not dependent upon a particular threshold (30). The reported unadjusted p values were calculated using the natural logarithm-transformed intensities and the Gaussian approximation to the t distribution. Statistical adjustment for multiple testing was performed by adjusting according to the strong control of the familywise error using the Bonferroni procedure (31). In addition, the p values according to the false discovery rate (FDR) adjustments of Benjamini and Hochberg (32) were calculated (case versus control).

Sequencing

Candidate biomarkers were sequenced using LC-MS/MS analysis as recently described in detail (13, 33) using instruments with electron transfer dissociation (ETD) capability (34–36). Spectral data were searched against the International Protein Index human non-redundant database using the open mass spectrometry search algorithm (free from the National Center for Biotechnology Information) using an e-value cutoff of 1.00e−2. All matched sequences were manually validated. All sequences obtained from human urine can be accessed as described previously (23).

RESULTS

The samples used and the flow of data are schematically shown in Fig. 1.

Fig. 1.

Usage of samples and flow of information. A, identification and validation of diagnostic biomarkers and biomarker models. 18 cases of AAV were compared with 425 controls (healthy individuals and patients with different chronic renal diseases), resulting in the definition of 113 potential biomarkers. Of these, 58 could be sequenced, and 18 were used in two biomarker models, one based on linear combination and one that was SVM-driven. All 113 potential biomarkers and the two biomarker models were evaluated in a test set of 40 blinded samples that consisted of 10 samples from patients with AAV and 30 controls. B, identification and validation of biomarkers and biomarker models for therapy assessment. CE-MS data from 18 urine samples from patients with active AAV were compared with data from urine samples from 19 patients with AAV in complete remission. In parallel, the data from the 18 samples from active AAV were also compared with data from urine samples from 200 healthy controls. 166 and 226 biomarkers that changed significantly with active AAV (p value <0.05 upon adjustment for multiple testing) could be defined in the two comparisons, respectively; 122 of these were significant in both analyses. Of these 122 potential biomarkers associated with disease activity, 47 could be sequenced. Using linear combination or SVM, respectively, biomarker models indicative of disease activity were established based upon these 47 urinary peptides. These two biomarker models were subsequently tested in a set consisting of samples longitudinally collected (at months 0, 1, 3, and 6) from 10 patients where therapy of AAV was initiated. CMV, cytomegalovirus.

Identification of Urinary Biomarkers for Differential Diagnosis of Vasculitis

Data from CE-MS analysis of 18 patients with active ANCA-associated vasculitis were compared with data obtained from healthy volunteers (n = 200) and focal segmental glomerulosclerosis (FSGS) (n = 30), diabetic nephropathy (n = 78), IgA nephropathy (IgAN) (n = 57), minimal change disease (MCD) (n = 25), and membranous glomerulonephritis (MNGN) (n = 35) patients. Clinical data of these patients have been described in several recent studies (12, 17, 27, 37). This resulted in six data sets in the control groups: one from healthy volunteers and five from disease controls. In total, 425 samples were utilized as controls with the aim to define biomarkers that enable differentiation between vasculitis and healthy controls as well as other renal diseases. The compiled data from the CE-MS analysis of the seven groups are shown in Fig. 2. A minimum frequency of 0.3 in cases or controls was required for any feature to be accepted as a potential biomarker. Upon application of the student's t test adjusted for multiple testing according to the stringent method of Bonferroni (31), 113 possible biomarkers with an adjusted p value of 0.05 or lower could be defined. A graphic depiction of the distribution of these 113 potential biomarkers in the seven different groups is shown in Fig. 3; all data, including the AUC values for the individual biomarkers and the p value after adjustment of the FDR according to the method of Benjamini and Hochberg (32), are given in supplemental Table 1.

Fig. 2.

Compiled protein patterns of the CE-MS analysis of urine samples from patients and controls examined in this study. Shown are compiled patterns consisting of all samples from patients with active vasculitis and each of the six control groups. (NC, apparently healthy normal control; DN, diabetic nephropathy.) The molecular mass on a logarithmic scale (0.8–25 kDa; indicated on the left) is plotted against normalized migration time (18–45 min; indicated on the bottom). Signal intensity is encoded by peak height and color. Although the difference in peptides present between apparently healthy controls and the six different chronic renal diseases is evident, distinct biomarkers that can distinguish between AAV and the other chronic renal diseases are not easily identified.

Fig. 3.

Distribution of only the 113 peptides that reveal significant differences between AAV and controls (p value <0.05 after adjustment for multiple testing) shown in the patients with active vasculitis and patients with other chronic renal disease or healthy controls. The molecular mass on a logarithmic scale (0.8–25 kDa; indicated on the left) is plotted against normalized migration time (18–45 min; indicated on the bottom). Signal intensity is encoded by peak height and color. All statistically significant biomarkers from supplemental Table 1 are shown. (NC, normal control; DN, diabetic nephropathy.) When examining only these 113 potential biomarkers, differences between AAV and the other chronic renal diseases become evident.

58 of the 113 initially defined potential biomarkers could be sequenced as also shown in supplemental Table 1. The most prominent observation is the appearance of C- and N-terminal hemoglobin fragments, whereas no fragments from the middle of the hemoglobin protein could be identified yet. Further up-regulation of α₁-antitrypsin and albumin fragments could be observed, whereas several collagen fragments appear to be underrepresented in comparison with the controls.

Limiting the available data to only the sequenced biomarkers, we attempted to combine these into a vasculitis-specific biomarker model. As the active vasculitis group consisted of only 18 patients, we reduced the number of biomarkers used in the panel without decreasing accuracy. To achieve this lower number of biomarkers included in the model, a take-one-out method was used whereby the model was run for all biomarkers minus a specific biomarker for each of the 58 initial possibilities. Biomarkers that did not influence the accuracy, specificity, and sensitivity of the model upon self-validation when taken out were left out of the final panel; this ensured that only the top scoring biomarkers were included in the final panel, which consisted of 18 biomarkers (Table II).

Table II.

18 biomarkers utilized for the discriminatory models between patients with vasculitis and controls (apparently healthy or patients with other chronic renal diseases)

Shown are the protein/peptide identification number in the data set (ID), mass (in Da), normalized migration time (CE-T; in min), the unadjusted (unadj) p values, p values adjusted using Bonferroni (Bonferroni) and Benjamini-Hochberg (BH), the AUC value in the ROC analysis (AUC), and the frequency and mean (median) amplitude based on data sets where the biomarker could be detected (omitting all 0 values) in the two groups of the training set, AAV and control (Contr). In addition sequence (lowercase p represents hydroxyproline), original protein, and the position of the first and last amino acids of the peptide in the respective protein sequence are given. Hemoglb, hemoglobin; Antitryp, antitrypsin; glycoprot, glycoprotein; Coll, collagen.

ID	Mass	CE-T	unadj p value	Bonferroni	BH	AUC	AAV	Mean (median)	Contr	Mean (median)	Sequence	Protein	Start	Stop
15800	1068.51	21.75	3.533920e−05	0.02887212	2.077131e−04	0.772	0.22	2.31 (2.52)	0.76	2.43 (2.47)	GEYKFQNAL	Serum albumin	423	431
25219	1213.61	27.2	9.220984e−07	0.000753354	1.295982e−05	0.880	0.83	3.24 (3.26)	0.13	2.53 (2.43)	VDEVGGEALGRL	Hemoglb β	21	32
30174	1292.59	28.28	1.697524e−05	0.01386877	1.185365e−04	0.797	0.78	2.48 (2.49)	0.15	2.78 (2.81)	TIDEKGTEAAGAM	α1-Antitryp	363	375
35339	1378.61	28.82	2.200089e−05	0.01797473	1.461360e−04	0.889	1	2.80 (2.81)	0.97	3.52 (3.55)	ApGEDGRpGPpGPQ	Coll α-1 (II)	511	524
38879	1439.66	29.82	1.009797e−06	0.000825004	1.309531e−05	0.825	0.94	3.06 (3.05)	0.6	2.61 (2.49)	TIDEKGTEAAGAMF	α1-Antitryp	363	376
40075	1448.78	20.27	8.613008e−07	0.000703682	1.256576e−05	0.876	0.83	3.53 (3.48)	0.14	2.90 (2.81)	VVAGVANALAHKYH	Hemoglb β	134	147
51120	1629.85	24.89	1.164354e−19	9.51277e−17	1.358967e−17	0.907	1	3.44 (3.62)	0.17	3.56 (3.59)	PMSIPPEVKFNKPF	α1-Antitryp	32	45
56033	1705.89	20.73	3.940837e−06	0.003219664	3.617600e−05	0.872	0.78	2.98 (3.01)	0.04	2.66 (2.71)	VLAHHFGKEFTPPVQ	Hemoglb β	114	128
60320	1797.93	23.9	3.645637e−06	0.002978485	3.423546e−05	0.867	0.78	4.70 (4.81)	0.1	3.48 (3.45)	WGKVNVDEVGGEALGRL	Hemoglb β	16	32
62557	1847.95	21.2	5.068519e−06	0.00414098	4.452667e−05	0.867	0.78	3.08 (3.22)	0.05	2.83 (2.81)	VLAHHFGKEFTPPVQAA	Hemoglb β	114	130
63772	1873.96	21.43	5.528687e−07	0.000451693	8.856740e−06	0.895	0.83	3.45 (3.52)	0.06	2.98 (2.96)	AAHLPAEFTPAVHASLDK	Hemoglb α	111	128
66483	1923.97	21.6	1.098620e−05	0.008975724	8.467664e−05	0.816	0.78	3.00 (2.97)	0.14	3.11 (2.98)	MGVVSLGSPSGEVSHPRKT	α2-HS-glycoprot	321	339
70633	2013.89	31.76	2.443416e−05	0.01996271	1.571867e−04	0.726	0.11	2.25 (2.25)	0.59	2.05 (2.06)	AGpPGPPGppGTSGHpGSpGSpG	Coll α-1 (III)	176	198
89325	2356.15	19.52	1.345930e−24	1.09962e−21	2.749062e−22	0.934	1	3.35 (3.35)	0.22	2.91 (2.91)	DAHKSEVAHRFKDLGEENFK	Serum albumin	25
90840	2389.24	22.4	7.466781e−10	0.00000061	3.050180e−08	0.860	0.94	4.40 (4.50)	0.3	3.85 (3.73)	MIEQNTKSPLFMGKVVNPTQK	α1-Antitryp	398
94543	2478.25	23.03	2.907457e−07	0.000237539	5.278650e−06	0.899	0.83	3.67 (3.69)	0.06	2.94 (2.69)	AAHLPAEFTPAVHASLDKFLASVS	Hemoglb α	111	134
98089	2559.18	19.41	8.924616e−06	0.007291412	7.079040e−05	0.845	0.89	3.78 (3.87)	0.44	2.88 (2.85)	DEAGSEADHEGTHSTKRGHAKSRP	Fibrinogen α	605	628
118597	3021.35	23.42	8.417339e−07	0.000687696	1.254231e−05	0.841	0.17	2.87 (2.92)	0.82	2.95 (3.03)	DGVSGGEGKGGSDGGGSHRKEGE EADAPGVIPG	CD99 antigen	97	129

Two algorithms were used to combine the 18 biomarkers into a disease-specific model. A model using SVMs resulted in correct classification of 424 of 443 of the samples in the training set upon self-validation, giving an accuracy of 95.7% with sensitivity and specificity of 100 and 95.5%, respectively. 94.4% sensitivity and 95.1% specificity were obtained upon complete cross-validation using an ideal cutoff of 0.2. Alternatively linear combination was used. This model resulted in 94.4% sensitivity and 97.4% specificity at a cutoff of 4.8. ROC curves from both models are shown in Fig. 4A.

Fig. 4.

Performance of the proteomic biomarker models based on 18 urinary peptides specific for vasculitis. A, ROC analysis of the classification results obtained upon complete cross-validation in the training set consisting of 18 cases and 425 controls. The numerical values of the classification obtained using either linear (lin) combination of the 18 biomarkers (left panel) or the SVM-driven model based on the same 18 biomarkers (right panel) were examined. The 95% confidence interval is indicated by the dashed line. B, performance of the biomarker models in the test set of the 40 blinded samples. Upon unblinding, 10 were found to be AAV patients, 29 harbored other chronic renal diseases, and one sample was from a healthy control (see also Table III).

The biomarker models were subsequently validated in a blinded data set. To this end, 40 samples were analyzed using CE-MS and evaluated using both the “vasculitis-specific” SVM-based model as well as the linear combinations (for results see Table III). Blinded analysis with these two classification algorithms correctly identified nine of the 10 active AAV samples as active AAV. Of 30 patients without active AAV (nine MNGN, six FSGS, four IgAN, four proliferative lupus nephritis, two MCD, two membranoproliferative glomerulonephritis, two glomerular sclerosis, and one normal control), 27 or 26 patients (SVM model, four falsely classified; linear model, three falsely classified) were correctly identified as not having AAV. Of the false positive scoring patients, the majority (two or three, depending on the algorithm used) had a progressive IgA nephropathy.

Table III.

Classification result of the biomarker models in the validation set consisting of 40 blinded samples using the cutoff values obtained from the results of the complete cross-validation in the training set (0.2 for the SVM model and 4.8 for the linear model, respectively)

Given are the sample identification number (ID), the clinical diagnosis, and the classification result using the SVM model and the linear (Lin) model. Samples scoring positive for vasculitis in either model are indicated in bold; correct positive scores are labeled in bold italic. LN, lupus nephritis; MPGN, membranoproliferative glomerulonephritis; GS, glomerular sclerosis; NC, normal control.

Sample ID	Diagnosis	SVM model	Lin model
21310	NC	−2.168	−20.73
21092	MCD	−2.147	−18.56
20915	FSGS	−2.015	−17.19
21144	GS	−1.938	−19.28
21309	LN	−1.834	−18.78
21139	LN	−1.798	−20.43
21312	MNGN	−1.734	−9.24
20918	FSGS	−1.682	−16.44
20970	FSGS	−1.676	−10.20
21068	MNGN	−1.599	−11.58
21224	AAV	−1.528	−12.20
21279	FSGS	−1.528	−10.39
20972	MPGN	−1.493	−8.47
20980	MNGN	−1.471	−12.53
21070	MNGN	−1.45	−10.46
21313	MPGN	−1.421	−5.75
20917	FSGS	−1.366	−15.55
21287	MNGN	−1.282	−16.22
21278	MNGN	−1.27	−10.99
21482	FSGS	−1.203	−9.84
21014	MNGN	−1.195	−7.61
20903	MCD	−0.818	−12.42
21157	MNGN	−0.296	−0.55
21102	IgAN	−0.292	−2.60
20971	MNGN	0.009	2.79
20976	LN	0.131	1.41
21217	GS	0.187	1.54
9619	AAV	0.259	11.14
21197	AAV	0.274	9.22
20933	IgAN	0.359	4.52
21311	AAV	0.903	8.70
21277	AAV	1.029	8.51
21065	IgAN	1.316	16.61
20932	LN	1.427	14.41
21195	IgAN	1.606	21.65
21192	AAV	2.153	22.29
21199	AAV	2.211	21.10
15701	AAV	2.225	21.81
21488	AAV	2.388	24.11
21143	AAV	2.58	29.97

As evident from Table III and Fig. 4, classification based on the linear combination or SVMs gave similar results; the linear combination appeared to show slightly higher accuracy, but the difference was insignificant. The misclassified patients were generally the same irrespective of model or algorithm. ROC curves from the blinded application of both models onto the 40 samples are shown in Fig. 4B.

As the data on the blinded cohort suggested that the biomarkers identified may lack specificity in differentiating IgAN from vasculitis, we investigated the performance of the biomarker model in a set of 18 IgA patients matched on a case to case basis concerning creatinine and proteinuria (compared with the 18 active vasculitis patients). Of these samples, 14 scored negative (correct) and four scored positive (false) for vasculitis, resulting in a specificity of 78% in the linear model, and 13 scored negative (correct) and five scored positive (false) for vasculitis in the SVM model.

We also assessed the significance of the tentatively identified 113 biomarkers for AAV in the test set. As the number of cases and controls was much smaller in the test set, we did not adjust for multiple testing. Of the 113 biomarkers, 101 were found above the required frequency threshold of 0.3 in the test set. Of these, 55 were also found to be significant (p value below 0.05) in the test set. An additional 12 biomarkers showed a trend (p value between 0.05 and 0.1), and 34 did not show any significant correlation with vasculitis in the test set. The unadjusted p values of the biomarkers in the test set are all given in supplemental Table 1.

To further examine the specificity of the biomarker panel regarding the influence of virus infections on the one hand and microhematuria and hypertension on the other, we examined additional patient groups. Urine samples obtained from 51 patients with cytomegalovirus infection after renal transplantation (22) scored negative in all but one patient. Further controls were urine samples from patients with microhematuria due to nephrolithiasis (20) (30 of 33 negative) or bladder cancer (20) (95 of 110 negative) and patients with renal cancer (20) (111 of 113 negative). Urine from 11 patients with proliferative lupus nephritis was investigated, and only one patient was positive. 17 urine samples from patients with hypertension and type II diabetes without microalbuminuria all scored negative for the vasculitis pattern (21). To investigate whether the slightly reduced specificity observed in the cohort of the samples from patients with bladder cancer could be attributed to specific biomarkers, we investigating only the 15 samples that scored positive in this group. We were unable to correlate a specific biomarker with the false positive scoring of these patients but rather the combination of fragments from hemoglobin, α₁-antitrypsin, fibrinogen, and α₂-HS-glycoprotein. Overall the data indicate a high specificity of the biomarker pattern used.

Biomarkers for Disease Activity

We subsequently sought to identify biomarkers that could potentially be used to address response to therapy/activity of disease. 18 samples from patients having active vasculitis and 19 samples from patients who had undergone treatment for the disease and were in stable clinical remission for more than 18 months were used to identify potential biomarkers. Comparison of the data sets resulted in 166 possible biomarkers when applying an adjusted p value cutoff of 0.05. As potentially useful biomarkers should also display significant differences between active vasculitis and apparently healthy normal controls, we also compared the 18 patients with active disease with the 200 normal controls using the same stringent statistics. This resulted in the identification of 266 potential biomarkers; 122 of these were found to be significant in both comparisons. The complete data on these 122 potential biomarkers are given in supplemental Table 2. 47 of these potential biomarkers could be sequenced. As these biomarkers should enable assessment of disease activity, we decided not to reduce their number further but instead use all 47 sequenced biomarkers to enable a maximum distance between disease and control with respect to the data space. Upon complete cross-validation, this panel of biomarkers gave 37 (100%) correct identifications in the training set. As SVMs do not give any level of confidence and may not be ideally suited to address linear changes as expected in progression, we also combined the 47 potential biomarkers in a linear model, which enabled classification of the training set with 100% accuracy.

In 10 patients, measurements were done before and 1, 3, and 6 months after initiation of immunosuppressive treatment. With successful treatment, clinical activity expressed by the BVAS, CRP, creatinine, and proteinuria parameters declined after 6 months as compared with the value at month 0: BVAS, p = 0.002 (median at 0, 1, 3, and 6 months, 19.5, 6, 0, and 0, respectively), CPR, p = 0.002 (median at 0, 1, 3, and 6 months, 117, 15, 3, and 1.5 mg/liter, respectively), creatinine, p = 0.098 (median at 0, 1, 3, and 6 months, 193, 128, 118, and 131 μmol/liter, respectively), and proteinuria, p = 0.0098 (median at 0, 1, 3, and 6 months, 1.0, 1.2, 0.6, and 0.4 g/24 h, respectively). ANCA titer declined significantly from a median of 1:32 (range, 1:8 to 1:4096) at month 0 to 1:16, 1:8, and 0 at months 1, 3, and 6, respectively (p = 0.002). Prednisolone doses decreased to a median of 50 mg at month 1, 15 mg at month 3, and 10 mg at month 6. After 6 months all but one patient (still receiving cyclophosphamide pulses) were on maintenance treatment with azathioprine or mycophenolate mofetil. The urinary proteome patterns changed gradually from active to inactive. This change was delayed in two patients, 2255 and 2259. These two patients experienced a transient rise in creatinine that was attributed to an infection or a subsequent relapse, respectively. Although these data further support the value of proteome-based monitoring, no definitive conclusion can be drawn from two patients only.

The assessment of all samples utilizing an SVM-based model or a linear combination is given in Fig. 5. These results are indicative of decreasing disease activity with a decline in BVAS, CRP, and renal parameters with increasing time of immunosuppressive treatment. To examine whether these results merely reflect changes in proteinuria, microhematuria, or renal function, we investigated the correlation between the scores obtained from the two biomarker models and proteinuria, microhematuria, and creatinine. Although the distribution of data in the comparison of proteinuria and creatinine with the proteomic patterns clearly indicated a random distribution and no correlation at all, we were able to find some correlation between erythrocytes in urine and proteomics scores. Because both proteomic pattern and microhematuria are expected to change as a response to treatment, it is not clear whether these two variables are dependent. We examined all 22 hemoglobin fragments. In general, the reduction was >10-fold after the 1st month of therapy. After 3 months, most fragments were generally below detection limit in the samples. Certainly in general, erythrocyturia/hematuria only is not reflected by the proteomic pattern as patients with other diseases who have microhematuria do not score positive with the vasculitis-specific proteomic pattern.

Fig. 5.

Scoring AAV activity based on proteome analysis during therapy in a longitudinal assessment. Shown is the scoring of the samples obtained at the indicated period of time after initiation of treatment in the models based on SVM (A) or linear combination of the biomarkers (B). The allocation of the 10 patients (indicated by their internal identification tag) to the different symbols used in the graph is shown on the right.

DISCUSSION

In this study we show that urinary proteome analysis with CE-MS permits differentiation of patients with active AAV versus healthy individuals and patients with other chronic renal diseases. A panel of biomarkers permits distinguishing between patients with active AAV and patients in remission. Initiation of immunosuppressive treatment results in a change of the pattern from active to inactive disease, correlating with clinically decreasing vasculitis activity and the achievement of remission.

Both the cross-sectional and longitudinal studies resulted in the identification of >100 potential biomarkers based on statistical evaluation even when adjusted for multiple testing. As a consequence, we limited ourselves to only the biomarkers that could be identified. Using the 18 biomarkers in the SVM-based model and in the linear model produced similar results. However, in all cases the linear model insignificantly outperformed the SVM-based model. This is likely due to the fact that SVM-based models tend to overfit data especially when the training sets used are not overly large and may therefore not reflect the true diversity of a real population. Consequently the use of SVM-derived models should be limited to large data sets that, because of their diversity, are less vulnerable to overfitting. This assumption could be further confirmed in experiments involving larger sets of cases and controls where SVM and other multiparametric machine learning algorithms consistently outperformed linear combination, whereas linear combination gave better results if only a smaller subset of cases and controls was used.²

To enable assessment of the validity of the single biomarker independently of any classification algorithm involved, we also tested whether the prediction of validity (based on statistics) was correct for each of the 113 individual biomarkers. As a rough estimate, we expected that six biomarkers (5%) would turn out to be not significant, whereas the others should again show significance. As evident, the expectations were not fully met by the data set. Only 55% of the assessed biomarkers again showed significant differences (66% if a trend was also valued acceptable). However, 34% revealed no significant association with AAV. These findings further support an observation that we made in several previous studies: even adjustment for FDR and for multiple testing results in an overestimation of the actual quality of the biomarkers in the training set likely due to unknown bias introduced. These results further imply that any proteomics results must be confirmed in a blinded test set; mere statistical testing (even if performed properly) is insufficient to establish association of a biomarker for disease with good confidence.

It should be noted that the majority of the false positive results from both models were from samples that, upon unblinding, were shown to be from patients with severe IgA nephropathy, which may be interpreted as a limited form of Schönlein-Henoch vasculitis. Patients with severe IgA nephropathy may present quite similarly to patients with ANCA-associated vasculitis regarding the kidney with a rapidly declining renal function. In a matched comparison, four or five (depending on the model used) of 18 patients with IgA nephropathy scored positive. Using samples from patients with bladder cancer, nephrolithiasis, or renal cancer, we showed that the reduced specificity observed in IgA nephropathy is not merely due to microhematuria as these additional cohorts with microhematuria did not show an increase in false positive results.

No correlation between creatinine, proteinuria, and proteomics scores for vasculitis could be shown. However, erythrocyturia correlated with the proteomics scores (results are shown in supplemental Fig. 1), which is not unexpected as microhematuria increases in active renal vasculitis and is linked to vasculitis activity. A missing influence of renal function on a specific pattern is also supported by our previous study in IgA nephropathy (18) where no systematic differences in frequency of the discriminating polypeptides could be shown when patients with creatinine <120 and ≥120 μmol were compared. These IgA nephropathy patients had a degree of renal insufficiency and proteinuria not different from the 18 AAV patients of this study (18). The results also indicate that proteome analysis delivers an additional parameter that can be used for clinical evaluation that is independent of creatinine levels and may enable a more accurate assessment of disease and disease state.

58 of the biomarkers used within the differential diagnosis pattern and 47 biomarkers used in the disease state classification panel could be sequenced. In both of the models, the most frequently observed peptides were proteolytic products of hemoglobin, including multiple fragments from both the α and β subunits. The existence of these fragments is to be expected as microhematuria is a characteristic finding in vasculitis. Of the 37 fragments of the hemoglobin α and β chains that could be identified to date in human urine, 21 were found to be statistically significant biomarkers in this study. Remarkably only C- and N-terminal fragments could be defined as biomarkers, whereas fragments from the core of the molecules, although present in urine, showed no significant value as biomarkers. The existence of specific fragments may be a result of the specifically released proteases in ANCA-associated vasculitis after activation of neutrophils by ANCA (38). Consequently these fragments, which are indicators of such specific protease activity, appear suitable in differentiating vasculitis from other pathological situations where hematuria is observed but results in different fragments because of divergent proteolytic activity. This is also reflected by the data on patients with microhematuria-associated diseases like nephrolithiasis and bladder and renal cancer where >90% of the samples scored negative for vasculitis.

In addition to these hemoglobin fragments, we found several fragments from albumin and α₁-antitrypsin increased in active vasculitis. Of the 21 α₁-antitrypsin peptides identified, 14 serve as biomarkers for vasculitis. α₁-Antitrypsin has a central role in controlling tissue damage by inhibiting proteases including elastase and proteinase 3, which are both released after activation of primed leukocytes by ANCA in vitro (38). Without inhibition, proteinase 3 and elastase are capable of inducing endothelial cell apoptosis and cytokine and tissue factor production (39–41). Moreover proteinase 3 and elastase have been shown in kidney biopsies of ANCA-associated pauci-immune necrotizing vasculitis (42), and elevated levels of proteinase 3 and elastase could be demonstrated in patients with ANCA-associated vasculitis (43, 44). Thus these results point to a role of α₁-antitrypsin as a mediator of activity in ANCA-associated vasculitis.

Of the 16 fragments from albumin identified to date in human urine, five, all starting at the same position at the N terminus, could be identified as biomarkers in this study. As also observed in diabetic nephropathy (27), specific collagen fragments were found reduced in comparison with the controls. Although collagen itself is very common, the ability to use collagen fragments as specific biomarkers for multiple different disease states indicates that these fragments are the result of variations in the in vivo protease activity and are therefore directly linked to the disease activity. As outlined in a recent review (45), it is likely that the urinary peptides reflect to a substantial degree turnover of extracellular matrix. (Patho)physiological changes in such turnover, e.g. in the case of fibrosis but also in inflammation, will display as indicative alterations of several urinary peptides. When comparing the changes observed in the kidney in diabetic nephropathy and vasculitis, the inflammatory component appears as one hallmark of vasculitis, whereas in contrast fibrosis is much more commonly found in diabetic nephropathy. As a consequence, we were able to detect the significant decrease of several collagen fragments in diabetic nephropathy (indicating reduced degradation of collagen, which results in an increase of extracellular matrix) that was by far less pronounced in the samples from vasculitis patients reported here. At the same time, the up-regulation of peptides from proteins involved in inflammatory processes appears to be a hallmark of urine from patients with active vasculitis. Candidates for those proteins are α₁-antitrypsin (see above), CD99 (a cell surface glycoprotein involved in leukocyte migration and T-cell adhesion, which are both processes that play a role in the pathogenesis of ANCA-associated vasculitis), and hemoglobin fragments only found in this patient group pointing to the specific proteases active in ANCA-associated vasculitis.

In light of other recent data sets on urinary protein fragments (12, 14, 20, 27), it appears that the combination of peptides that alone are not specific for a particular disease with proteolytic activity results in the generation of a specific panel of degradation products that can be utilized to identify the disease with high specificity and also assess disease progression/therapeutic benefits. Similar findings (the presence of specific fragments of albumin and α₁-antitrypsin in chronic renal disease) have recently been reported by Candiano et al. (46).

A major medical need in patients with ANCA-associated vasculitis and renal involvement is assessment of the activity of the disease. Here urinary proteome analysis may be of significant value. If the disease is in remission, immunosuppressive treatment may be reduced. This is important for the patients to minimize the risk of side effects. Microhematuria may be used as a parameter for activity, but its value is limited as it may persist in remission. Repeated biopsies are generally not indicated because of the associated risk (mostly of severe bleeding). Moreover as the disease is focal, active lesions might be missed. ANCA has been used as a marker of active disease, but the results were disappointing (47). Circulating endothelial cells will need time to decrease (48) probably as repair mechanisms lead to release of cells during the healing process. CRP might be increased as a result of an ongoing infection or severe arteriosclerosis (49). Endothelial microparticles may represent a promising tool, but there are not enough data to draw a final conclusion (50, 51). The data presented here suggest that urinary proteomic biomarkers may be an excellent tool to overcome the above mentioned shortcomings. They indicate that urinary proteome analysis does not only enable non-invasive diagnosis and differential diagnosis of AAV but also allows non-invasive monitoring of disease activity in the kidney.