Research ArticleKidney Disease

Fetal Urinary Peptides to Predict Postnatal Outcome of Renal Disease in Fetuses with Posterior Urethral Valves (PUV)

See allHide authors and affiliations

Science Translational Medicine  14 Aug 2013:
Vol. 5, Issue 198, pp. 198ra106
DOI: 10.1126/scitranslmed.3005807


Bilateral congenital abnormalities of the kidney and urinary tract (CAKUT), although are individually rare diseases, remain the main cause of chronic kidney disease in infants worldwide. Bilateral CAKUT display a wide spectrum of pre- and postnatal outcomes ranging from death in utero to normal postnatal renal function. Methods to predict these outcomes in utero are controversial and, in several cases, lead to unjustified termination of pregnancy. Using capillary electrophoresis coupled with mass spectrometry, we have analyzed the urinary proteome of fetuses with posterior urethral valves (PUV), the prototypic bilateral CAKUT, for the presence of biomarkers predicting postnatal renal function. Among more than 4000 fetal urinary peptide candidates, 26 peptides were identified that were specifically associated with PUV in 13 patients with early end-stage renal disease (ESRD) compared to 15 patients with absence of ESRD before the age of 2. A classifier based on these peptides correctly predicted postnatal renal function with 88% sensitivity and 95% specificity in an independent blinded validation cohort of 38 PUV patients, outperforming classical methods, including fetal urine biochemistry and fetal ultrasound. This study demonstrates that fetal urine is an important pool of peptides that can predict postnatal renal function and thus be used to make clinical decisions regarding pregnancy.


Bilateral congenital abnormalities of the kidney and urinary tract (CAKUT) are developmental renal diseases including urinary tract malformations, obstructive uropathy, and hyper/hypodysplasia. Bilateral CAKUT are, although individually rare diseases, the main cause of chronic renal failure in children worldwide (1). Prognosis is, in general, good in unilateral CAKUT. However, bilateral CAKUT are associated with a wide spectrum of outcomes ranging from extremely severe phenotypes with terminal renal failure, absence of amniotic fluid, and pulmonary hypoplasia leading to prenatal death to normal phenotypes with a standard renal function. Prenatal counseling in bilateral CAKUT is based on fetal ultrasound, allowing the clinician to assess the quantity of amniotic fluid and the appearance of the renal parenchyma, and on the concentration of fetal urine analytes, such as sodium and β2-microglobulin (β2M). However, several studies have shown that neither ultrasound (2, 3) nor the fetal analytes investigated to date (sodium, β2M, calcium, chloride, osmolality, and total protein) have sufficient accuracy to be used with confidence in the prediction of poor postnatal renal function and thus termination of pregnancy (TOP) (4). This is exemplified by a recent study on severe bilateral CAKUT in which, of the 10 refused TOPs, 5 had normal serum creatinine at a median age of 29 months (5).

These recent data show that all currently available, clinically used methods are unsatisfactory in predicting postnatal function and may lead to unjustified offering of TOP or continuation of the pregnancy resulting in early (in utero, at birth, or within the first months) end-stage renal disease (ESRD). Adequate prenatal counseling is of utmost importance in these diseases to discriminate between patients with early ESRD and patients who will not progress to ESRD before 2 years of age. Renal replacement therapy (dialysis or transplantation) in infants is still associated with high morbidity when initiated before the age of 2 (6).

Using proteome and peptidome (that is, the low–molecular weight proteome) analysis of postnatal urine, we and others have identified and validated urinary markers of diseases of the kidney and urinary tract (711) that outperform classical clinical biomarkers such as creatinine clearance and albuminuria (12, 13), ultrasound, and renograms (8, 14). Similar to postnatal urine, fetal urine potentially contains biomarkers of disease. However, the fetal urinary proteome or peptidome has not been explored. In contrast to postnatal urine, only small volumes of fetal urine are available (up to 1 ml). Currently, for reproducible analysis of the urinary proteome, even using the latest state-of-the-art mass spectrometers, at least 5 to 10 ml of urine is needed (1517), and therefore, proteome analysis of fetal urine is currently not feasible. In contrast, it is possible to analyze the fetal urinary peptidome because it only requires small amounts of urine [several hundred microliters (18, 19)].

In this prospective study, we aimed to improve the use of fetal urine analysis for postnatal renal function prediction by applying urinary peptidomics to identify peptides that can discriminate between early ESRD and absence of ESRD before the age of 2 in bilateral CAKUT. We used urine of fetuses diagnosed with posterior urethral valves (PUV), the prototypic bilateral CAKUT (20), and identified several fetal urinary peptides associated with early ESRD in this disease. A combination of 12 of these fetal urinary peptides into a classifier allowed the prediction of postnatal renal function with high sensitivity and specificity in a new and blinded patient cohort of 38 patients with PUV, outperforming currently used imaging- and biochemistry-based clinical methods.


Study setup and capillary electrophoresis–mass spectrometry analysis of fetal urinary peptides

Fetal urine samples from PUV patients obtained from 26 different French centers were divided into two cohorts (Fig. 1A): one discovery cohort (n = 28; table S1) and one blinded cohort for prospective validation (n = 38; table S2). All neonatal death cases were included in the blinded validation cohort to obtain clinical hard endpoints of ESRD. Capillary electrophoresis coupled with mass spectrometry (CE-MS) analysis using 150 μl of fetal urine per patient allowed the detection of a total of 4199 different peptides annotated with a unique tag consisting of a mass/migration time in all samples (Fig. 1B).

Fig. 1 Study design and CE-MS analysis in fetal urine of patients with PUV.

(A) This study was divided in two separate phases: a discovery phase, where the fetal urine proteome of 28 PUV patients (15 noESRD, 13 ESRD; 66 total samples) was analyzed, resulting in a 12-peptide classifier called 12PUV, and a validation phase, where the 12PUV model was tested in an independent cohort of PUV patients (n = 38) by a blinded analyst. The predictive value of 12PUV in this validation cohort was then confirmed by renal function ≥2 years of age. (B) Representation of 4199 peptides, detected in all 66 fetal urine samples by CE-MS. Each peptide was identified by a unique identifier on the basis of the migration time (min) and specific mass (kD), with a peak height representing the relative abundance. (C) In the discovery phase, 26 fetal urine peptides were identified as differentially secreted between PUV patients with early ESRD and PUV patients not progressing to ESRD before the age of 2 (noESRD). (D) Cross-validation score of an SVM peptide model called 12PUV consisting of 12 of the 26 peptides obtained from the analysis of the discovery cohort. ***P < 0.0001, Mann-Whitney test for independent samples.

Identification of fetal urinary peptides predictive of postnatal renal function

For the identification of fetal urine peptides predictive of postnatal renal function, the PUV discovery cohort was divided into two groups (Fig. 1A): early ESRD (n = 13) and absence of ESRD before the age of 2 (noESRD, n = 15). ESRD patients were selected from terminated pregnancies with a severe renal phenotype confirmed by autopsy (table S1). Six of the ESRD patients displayed severe hypodysplasia leading to Potter’s sequence (sample IDs: C41-1-2, C41-1-3, C41-2-7, C41-2-8, C41-3-1, and C41-3-2 in table S1); five displayed bilateral dysplasia with diffuse glomerular and tubular cysts (C41-1-6, C41-1-8, C41-2-1, C40-3-8, and C41-2-5 in table S1); one displayed severe cortical thinning because of major urine-induced dilation (C41-1-5 in table S1); and for one patient, no fetopathology was available, but this patient displayed extremely high β2M levels (35 mg/liter, C44-1-6 in table S1), which is nearly threefold higher than the published cutoff value (4). Normal renal function or moderate renal failure before the age of 2 is defined by noESRD (table S1).

Comparing fetal urine from ESRD and non-ESRD patients led to the identification of 26 differentially excreted peptides between these two groups (Fig. 1C, Table 1, and table S3). The selection of these peptides was corrected for multiple testing to avoid the selection of peptides associated by chance to one of the disease groups using the false discovery rate concept described by Benjamini and Hochberg (21) and as recommended by the clinical proteomics guidelines (22, 23). Twenty of the 26 peptides could be sequenced. These 20 peptides were reduced to a support vector machine (SVM) model with 12 peptides—which we called “12PUV”—by a leave-one-out procedure whereby the model was run for all peptide candidates minus one. Peptides that did not influence the accuracy of the model in the total cross-validation of the training data were left out of the final model. Scoring the patients from the discovery cohort with the resulting 12PUV classifier clearly separated ESRD from noESRD patients (Fig. 1D).

Table 1 The 26 differentially excreted fetal urinary peptides between PUV patients with and without early ESRD in the discovery cohort.

Data were obtained from 28 PUV patients. Peptides with different abundance in fetal urine between PUV patients with noESRD or ESRD were chosen after correction for multiple testing. P values were determined using Wilcoxon rank-sum test followed by adjustment for multiple testing using the Benjamini and Hochberg method. The 12 peptides used for modeling were chosen using a leave-one-out cross-validation procedure. An extended version of this table with more detailed information on the 26 peptides can be found in table S3. Abbreviations in sequence: p, hydroxyproline; k, hydroxyllysine; m, hydroxylmethionine.

View this table:

Validation of 12PUV in a separate, blinded cohort

In the next step, following the recommendations for biomarker identification and qualification in clinical proteomics (23), the 12PUV model was prospectively validated in a separate, blinded study using fetal urine from 38 PUV patients not in the discovery cohort (table S2). These fetal urine samples were analyzed by CE-MS, scored using the 12PUV model (table S2), and then given a prediction as to whether the patient would develop ESRD by 2 years of age. With a positive 12PUV score (>0), patients were predicted to develop ESRD by 2 years of age. These predictions were compared to the clinical data ≥2 years after birth, at TOP, or at neonatal death.

The 12PUV classifier predicted postnatal renal function (noESRD versus ESRD) with a sensitivity of 88% [95% confidence interval (CI), 66 to 98%], a specificity of 95% (95% CI, 80 to 100%) (Table 2), and an area under the curve (AUC) of 0.94 (95% CI, 0.82 to 0.99) (Fig. 2A). Given the relatively low number of peptides in the 12PUV SVM classifier, we also analyzed the possibility of using simpler classifiers, including a linear and a k-nearest neighbors (KNN) classifier. With the 12 peptides, the AUCs of both models were very similar to the SVM model (Fig. 2A, inset). The linear model was not significantly different from the SVM model (P = 0.53), whereas the KNN model did not perform as well as the SVM-based classifier (P = 0.032) in a pairwise comparison of receiver operating characteristic (ROC) curves (Fig. 2A, inset, and table S2). These results suggest robustness of the selected peptides because they were mostly insensitive to the different models applied. For the remainder of the study, we used the SVM-based classifier.

Table 2 Sensitivity, specificity, and predictive values of the 12PUV classifier.

The 12PUV was compared to ultrasound parameters and fetal urine biochemistry in the validation cohort (n = 38). PPV and NPV calculations were based on a prevalence of 0.42 of ESRD in patients with PUV (24, 25). *P < 0.05; **P < 0.01; ***P < 0.001 compared to 12PUV using the McNemar test (42) for comparison of sensitivities and specificities and the test statistics proposed in (43) for comparison of PPV and NPV.

View this table:
Fig. 2 Blinded validation of urinary peptide classifier 12PUV in a separate PUV patient population.

Predictions of ESRD were made with 12PUV for 38 patients with PUV, and follow-up ≥2 years later was used to confirm initial prediction. (A) ROC curve for the 12PUV model. Inset: ROC curves obtained when using a linear or KNN model based on the 12 peptides, instead of the SVM-based 12PUV model, for classification. (B) Box-whisker plot for classification of the patients with and without ESRD in the validation set according to the 12PUV score. (C) Representative figure of the abundance of the 26 fetal urinary peptides in PUV patients with and without ESRD in the validation cohort. Data were obtained in Toulouse, France. (D) Change of CE-MS setup did not affect the 12PUV classifier. All 38 fetal urine samples of the validation cohort (n = 38; table S2) were also analyzed on a different CE-MS setup (Glasgow, Scotland) from that used for data in (B). Data are presented in a ROC curve. Inset: A representative graph of the abundance of 26 fetal urinary peptides in PUV patients with and without ESRD in the validation cohort analyzed in Glasgow. ***P < 0.0001, Mann-Whitney test for independent samples.

The 12PUV model predicted seven of eight neonatal death cases correctly, showing the efficacy of the model to detect patients with severe PUV and ESRD. The misclassified patient (C41-3-4 in table S2) died at a relatively late stage after birth (6 weeks). The 12PUV also predicted ESRD in seven of eight TOP cases. The misclassified patient (C41-1-4 in table S2) displayed a Potter’s sequence. The distribution of the 12PUV scores for the validation cohort showed significant separation of the two patient populations (Fig. 2B). This was associated with a difference in the abundance of the original 26 fetal urinary peptides between ESRD and noESRD PUV patients in the validation cohort (Fig. 2C).

We also analyzed the samples of the validation cohort on a different CE-MS setup to study potential issues on the portability of the classifier (that is, issues related to machine calibration and normalization). The same 12PUV classifier worked similarly well on the other CE-MS setup (Fig. 2D) without a significant difference (P = 0.72, pairwise comparison of ROC curves).

Comparison with the clinical methods for evaluation of postnatal outcome

We compared the sensitivity and specificity of the 12PUV classifier with the currently used clinical methods for evaluation of postnatal outcome in the validation cohort. This included fetal urine biochemistry (sodium and β2M) and characteristics of the fetal kidney (for example, presence of cysts, hyperechogenicity, presence of dysplasia, and aspect of the renal cortex) and amniotic fluid as measured by ultrasound. The 12PUV classifier significantly outperformed these traditional parameters in both sensitivity and specificity in our study (Table 2). In contrast, in the validation cohort, fetal urinary sodium or β2M, depending on the extreme cutoffs observed in literature (4), had either a high sensitivity or a high specificity [β2M cutoff >2 mM: sensitivity, 100% (95% CI, 83 to 100%); specificity, 45% (95% CI, 27 to 65%), or β2M cutoff >13 mM: sensitivity, 31% (95% CI, 13 to 55%); specificity, 95% (95% CI, 80 to 100%)] (Table 2), but never both together. In addition, most of the ultrasound parameters had both low to moderate sensitivity and specificity except for dysplastic multicystic kidneys, which displayed low sensitivity but high specificity [sensitivity, 31% (95% CI, 13 to 55%); specificity, 100% (95% CI, 87 to 100%)] (Table 2). The incidence of ESRD in PUV has been estimated to range from 24 to 50% (24, 25). The percentage of ESRD in the validation PUV cohort in our study (42%) lied within this range. Using this prevalence of 42%, we calculated a positive predictive value (PPV) of 93% and a negative predictive value (NPV) of 91% for the 12PUV classifier (Table 2); again, this outperformed the PPV/NPV of other methods.

In a severe disease such as PUV, a >90% success rate for the identification of patients at risk is acceptable. In a cohort of 38 (the size of the validation cohort), this translates into a target benefit-to-harm ratio [(true positives + true negatives):(false positives + false negatives)] of 35:3. The derived target sensitivity, specificity, PPV, and NPV were not significantly different from those of the 12PUV classifier (table S4). This observation supports the potential clinical suitability of the 12PUV classifier.

Next, we investigated whether we could further improve the 12PUV classifier–based prediction by combining with clinical, “gold standard” methods. We combined the 12PUV classifier using logistic regression, one by one, with β2M, urinary sodium, and ultrasound parameters and performed a pairwise comparison of the ROC curves (Fig. 3). None of the combinations resulted in significant improvement of the 12PUV classifier (P = 0.36 to 0.78, pairwise comparison of ROC curves). Overall, these data support recent (meta) studies that ultrasound parameters and fetal urine analytes (β2M and sodium) have insufficient clinical accuracy to predict poor postnatal renal function in bilateral CAKUT (24), and suggest that the fetal urinary peptide classifier is potentially a new tool with high sensitivity and specificity.

Fig. 3 Combination of the 12PUV classifier with classical clinical parameters.

(A to C) ROC curves of the combination of the 12PUV classifier using data from the validation cohort (table S2), one by one, with β2M or urinary sodium (A), and ultrasound parameters (B and C) by logistic regression. AAF, absence of amniotic fluid; ANCMD, absence of normal corticomedullary differentiation; DMK, dysplastic multicystic kidneys; HK, hyperechogenic kidneys; HKCT, hypoplastic kidneys with corticomedullary thickening; Na, sodium; Olig, oligohydramnios.

The 12PUV peptides and molecular pathological changes

Fetal urinary β2M and sodium decrease significantly with gestation owing to the development of functional kidneys (26). The 12PUV classifier was found to be independent of gestational age (r = −0.1684; P = 0.18, Pearson correlation), suggesting that these peptides do not reflect fetal renal function in PUV patients, but instead molecular pathological changes associated with ESRD. Nineteen of the differentially secreted peptides represent fragments of different types of collagen (I, II, IV, V, VII, XI, and XVI collagen) (Table 1 and table S3). However, in contrast to what is observed in postnatal urine of chronic renal diseases (27), the abundance of collagen fragments was increased in fetal urine of PUV patients with severe ESRD. Decreased urinary collagen fragment secretion in renal diseases may reflect increased intrarenal extracellular matrix accumulation, that is, fibrosis, which is a hallmark of chronic renal disease. We hypothesize that increased collagen fragment secretion in fetal urine of PUV patients with ESRD represents degradation of mostly conjunctive tissue due to disruption of nephrogenesis (visible as dysplasia or hypoplasia in fetopathology). At this point, however, this is speculation that requires follow-up.

One fetal urine peptide indicator of ESRD in PUV patients was identified as a fragment of the XLαs variant of the G protein (heterotrimeric guanine nucleotide–binding protein) α subunit (GNAS) involved in signal transduction of seven-transmembrane receptors. The GNAS complex locus expresses at least eight identified gene products, some of which are paternally expressed (for example, XLαs), whereas others are exclusively maternally expressed (for example, NESP55) (28). In contrast to the collagen peptides, the abundance of this peptide was decreased in fetal urine in most of the PUV patients with ESRD (P < 0.0001, Mann-Whitney test for independent samples) (Fig. 4, A and B). We confirmed this decrease by analyzing the expression of NESP55 by enzyme-linked immunosorbent assay (ELISA) in fetal urine. Opposite to what we observed for XLαs, the average NESP55 concentration of this maternally expressed protein was significantly higher in the ESRD than in the noESRD group (Fig. 4C; P < 0.0001, Mann-Whitney test for independent samples).

Fig. 4 Abundance of components of the GNAS complex locus in the fetal urine of all 66 PUV patients.

(A) Reduced abundance of GNAS, in contrast to the three collagen fragments, in fetal urine of PUV patients with ESRD. (B) Relative amount of GNAS peptide as detected by CE-MS normalized to housekeeping peptides (as described in Materials and Methods). (C) Relative amount of fetal urine NESP55 as detected by ELISA. ***P < 0.0001, Mann-Whitney test for independent samples.


We have identified and validated 12 fetal urinary peptides as a composite indicator of a rare renal developmental disease. These peptides, called 12PUV, predicted postnatal renal function in patients with PUV with high sensitivity and specificity, outperforming currently used clinical methods. The main advantage of the 12PUV classifier is that it provided high sensitivity and specificity at the same time, whereas with the currently used clinical methods, there is always a payoff on sensitivity when specificity is high, and vice versa (3, 4) (Table 2 and table S4). In a severe disease such as PUV, we believe that a >90% success rate for the identification of patients at risk is acceptable. Accordingly, in the validation cohort, this translated into a target benefit-to-harm ratio of 35:3. The derived target sensitivity, specificity, PPV, and NPV were not significantly different from those of the 12PUV classifier (table S4), supporting the potential clinical suitability of the 12PUV classifier. In contrast, for currently used clinical methods, either sensitivity or specificity and PPV or NPV were lower than the target values. Therefore, we believe that our method, using modern MS-based tools, is a much needed improvement over the current state of the art in a field that only progressed little over the last 2 decades (3, 4). Although there are currently no treatments for PUV and in utero repair of the valves does not improve outcome (29), the 12PUV classifier can contribute to truly informed prenatal counseling by preventing unnecessary TOP. In addition, predicting the postnatal outcome can be of high value for appropriate clinical measures such as long-term planning for implementation of renal replacement therapy (for example, dialysis or kidney transplantation).

PUV is the prototypic CAKUT, which is a heterogeneous group of fetal pathologies leading to up to 45% in utero mortality. In addition, among the CAKUT patients that survive the neonatal period, up to 30% develop ESRD, necessitating dialysis and/or transplantation (4). In fetuses, conventional biomarkers of renal function, such as creatinine or urea clearance, cannot be used because these molecules cross the placenta and are cleared by the mother, thus reflecting maternal renal function (30). Many studies have investigated the prognostic potential of other fetal urine metabolites as biomarkers of fetal and postnatal renal function, including sodium, calcium, chloride, and β2M concentrations or osmolality [for review, see (4)]. However, no individual analyte has been shown to yield significant clinical accuracy to predict postnatal renal outcome (4). In the past decade, we and others have identified and validated new, unconventional urinary biomarkers of diseases of the kidney and urinary tract, such as diabetic nephropathy, vasculitis, polycystic kidney disease, renal Fanconi syndrome, or unilateral ureteropelvic junction obstruction, in both children and adults, using proteome and peptidome analysis of postnatal urine (79, 11, 14, 31). Here, we applied peptidomics on fetal urine to identify biomarkers of postnatal ESRD. In contrast to fetal urine β2M and sodium, these peptides most likely do not merely represent fetal renal function, but the actual molecular pathological changes within the kidneys.

One advantage of the urinary peptidome versus that of serum or plasma is that it is stable: it can be kept for a few hours at room temperature and frozen for years at −20°C without significant modification of the peptide content (3234). In addition, it has been shown that several cycles of freezing and thawing of postnatal urine do not modify the peptide content (7). Moreover, we have shown that it is possible to analyze samples in a local setting rather than a centralized laboratory because validation of the 12PUV classifier on two different CE-MS setups did not significantly differ (Fig. 2, A and D).

A limitation of the study is the relatively low sample number, which is due to the rarity of the disease. However, the use of stringent statistical procedures (for example, correction for multiple testing) still resulted in high sensitivity and specificity, indicating the potential validity of the 12PUV classifier. To be considered as a biomarker-based urinary classifier, the 12PUV classifier will need to be tested prospectively in several independent cohorts. Using the point estimates of a sensitivity of 88% and a specificity of 95% for 12PUV, an 80% power and α = 0.05, and a prevalence of 0.42 of ESRD in the PUV patient population, we calculated according to (35) that, to show superiority of 12PUV over, for example, β2M (β2M cutoff of 2 mM with 100% sensitivity and 45% specificity), we would need 58 ESRD and 58 noESRD PUV patients. For a β2M cutoff of 13 mM (sensitivity of 31% and specificity of 95%), a total number of 50 ESRD and 50 noESRD subjects will be required in such a follow-up study.

Another issue that can be seen as limitation is that the same MS-based technology was used for validation and discovery. However, an antibody-based method to analyze the peptides in the 12PUV classifier will be difficult to develop because some of the markers differ by only one amino acid from nonmarkers. In addition, in view of the current state of the art, CE-MS peptidomics is currently the only method to analyze—in a reproducible manner—the peptidome of small quantities of biofluids, such as the volume available for fetal urine (up to 1 ml). Although targeted methods are being developed for biofluids, such as selected reaction monitoring, these still need relatively large amounts of biofluids [for example, 10 ml of urine (17)]. A final limitation of the study is that the CE-MS technology focused on peptides, which yielded only limited information on the pathophysiology of the disease, because it is currently not known whether the observed differences in fetal urinary peptide abundance originate from altered peptide cleavage of proteins or altered abundance of the proteins themselves or a combination thereof. Once methods are available to study the proteome of small quantities of fetal urine (not just the peptidome), this should lead to improved insight in the pathophysiology of PUV.

Implementation of such an MS-based test to analyze the fetal urine peptidome in a clinical setting can be foreseen. Fetal urine can be obtained in the clinic, frozen, and then sent to a central laboratory for CE-MS analysis. This is exactly the path followed by the samples used for the current study: fetal urine was obtained in 26 different centers and analyzed in a central laboratory with a turnaround time of ~2 days once a sample is received. For improved patient acceptance and to reduce the invasiveness of the procedure, the next step will be to search for the 12PUV classifier in amniotic fluid. Fetal urine is a major source of amniotic fluid in the second half of pregnancy (36), and thus, several of the fetal urinary peptides will potentially be present in amniotic fluid. Indeed, the presence of the 12PUV classifier will be analyzed in the context of a large-scale European project focusing, among others, on the detection biomarkers of rare kidney diseases, including CAKUT from amniotic fluid ( However, their abundance is potentially modified compared to fetal urine owing to the secretion of large volumes of fluid by the fetal lungs in this period (36) and to rapid movement of water and solutes between amniotic fluid and fetal or maternal blood (37). Finally, with the CE-MS method described here for fetal urine, one could imagine application in other diseases. There are more than 4000 peptides in fetal urine, which may represent an important pool of potential biomarkers for the other developmental diseases where early prediction could majorly affect treatment and clinical decision-making.

Materials and Methods

Study design

The objective of the study was to determine the presence of peptides in fetal urine that could serve as biomarkers to predict postnatal renal function (noESRD versus ESRD) in patients with PUV. Renal function at 2 years of age was chosen as the clinical endpoint because high morbidity is observed in renal replacement therapy before the age of 2 (6). Sixty-six patients with PUV were recruited and divided into a discovery (n = 28) and a blinded cohort for prospective validation (n = 38) (Fig. 1A). Patients were randomly assigned to the discovery or validation cohorts, except for severe cases of TOP that were included in the discovery cohort and neonatal death cases that were included in the blinded cohort to obtain a hard endpoint of ESRD for validation purposes. The fetal urinary peptidome was analyzed by CE-MS. In parallel, classical clinical data were obtained including fetal urine biochemistry (Na and β2M) and fetal ultrasound–based parameters. A model of 12 fetal urinary peptides was developed on the basis of the urinary peptidome data obtained in the discovery study. The capacity of this model to predict ESRD was tested in the blinded cohort and compared to the performance of the classical clinical parameters.

Patient recruitment and patient description. The decision to recommend fetal urine sampling was made by the clinician in charge of the patient independent of the research protocol. In accordance with French law, written consent for fetal urine sampling and for laboratory testing has been obtained from each woman since 1997. Clinical and fetal urine data from the PUV patients are in tables S1 and S2. PUV was confirmed in all cases either by fetopathology in TOP and neonatal death or by a voiding cystourethrogram in case of live birth. In accordance with the French law, TOP was considered at the parents’ request in the case of severe anomalies, such as severe oligohydramnios, bilateral renal dysplasia, aneuploidy, or severe associated malformations. Prenatal ultrasound findings obtained at the time of fetal urine sampling were recorded. Postnatal renal function was evaluated with postnatal serum creatinine values.

The “noESRD” group is defined by PUV patients with normal serum creatinine (≤51 μM) and PUV patients with moderate renal failure at >2 years of life (ranging from 52 μM at 2 years and 6 months to 206 μM at 6 years and 8 months). For five infants, we only obtained renal function between 13 and 18 months; however, values ranged from 24 to 200 μM, and we considered that these children would not progress to ESRD before the age of 2 (tables S1 and S2, postnatal creatinine columns).

The ESRD group consisted of PUV patients with early ESRD: either PUV patients for whom pregnancy was terminated (TOP) or neonatal death (up to 6 weeks of age). TOP was based on severe phenotype: oligohydramnios or absence of amniotic fluid associated with hyperechogenic or cystic kidneys at prenatal ultrasound examination and validated by fetopathology.

Sample collection and preparation. Fetal urine was collected from the bladder of the fetus under ultrasound guidance in 26 different French centers and stored at −20°C. Samples were prepared in Toulouse (France). The samples of the validation cohort, prepared in Toulouse, were analyzed on a CE-MS setup both in Toulouse and in Glasgow (Scotland). Because of the small volume obtained by in utero sampling, it was necessary to adapt the postnatal urine protocol to a volume of 150 μl, similar to the method used for rodent urine (18). Briefly, immediately before preparation, fetal urine aliquots were thawed, and 150-μl aliquots were diluted with the same volume of 2 M urea and 10 mM ammonium hydroxide containing 0.2% SDS. Subsequently, samples were ultrafiltered with a Centristat 20-kD cutoff centrifugal filter device (Sartorius) to eliminate high–molecular weight compounds. The filtrate was desalted with a NAP-5 gel filtration column (GE Healthcare Life Sciences) to remove urea and electrolytes, and thereby to decrease matrix effects. The sample was lyophilized in a Savant SpeedVac SVC100H connected to a VirTis 3L Sentry freeze dryer (Fischer Scientific) and stored at 4°C until use. Shortly before CE-MS analysis (Supplementary Methods), the samples were resuspended in 10 μl of high-performance liquid chromatography–grade water.

Data processing. Mass spectral ion peaks representing identical molecules at different charge states were deconvoluted into single masses with MosaiquesVisu software (38). The software automatically examined all mass spectra from a CE-MS analysis for signals with a signal-to-noise ratio of at least 4 present in three consecutive spectra. Furthermore, the isotopic distribution was assessed, and charge was assigned on the basis of the isotopic distribution, as well as conjugated masses, with a probabilistic clustering algorithm. This operation resulted in a list wherein all signals that could be interpreted are defined by mass/charge, charge, migration time, and signal intensity (ion counts). Time-of-flight MS data were calibrated with Fourier transform ion cyclotron resonance MS data as reference masses applying linear regression. CE migration time was calibrated by local regression with more than 1700 reference signals.

Normalization of the amplitude of the fetal urine peptides was based on sequenced endogenous “housekeeping” peptides (table S5) that varied little among all samples in discovery cohort (±18%). On the basis of the sequences, most of the housekeeping peptides found in fetal urine were similar to those observed in postnatal urine, therefore allowing the application of a CE-MS procedure for sample analysis, data processing, and normalization similar to the procedure used for postnatal urine (34, 39). The amplitudes were normalized with linear regression on the basis of comparison between average amplitudes of 29 housekeeping peptides (table S5) and amplitudes detected in individual samples. All data, independent of CE-MS setup, were normalized in the same manner with the same reference values (that is, housekeeping peptides).

Peptide sequencing

Candidate peptides for the 12PUV and other native peptides from fetal urine were sequenced with MS/MS and searched against human entries in the UniProt database, as described in the Supplementary Methods.

Peptide identification and statistical analysis

For the identification of potential fetal urinary peptide biomarkers, the normalized levels of fetal urinary peptides were compared between the noESRD and ESRD patient groups. Only peptides that were detected with a minimal frequency of 70% in at least one of the diagnostic groups were considered for statistical analysis. Unadjusted P values were calculated for the comparison between the noESRD and ESRD patient groups with the Wilcoxon rank-sum test followed by adjustment for multiple testing with the method described by Benjamini and Hochberg (21). Only peptides with a corrected P < 0.05 were considered significant.

The number of differentially expressed peptides was reduced to an SVM model with 12 peptides (12PUV) by a leave-one-out procedure whereby the model was run for all peptide candidates minus one. Peptides that did show influence in accuracy of the model in the total cross-validation of the training data were left out of the final model. The SVM model (type C-SV) was generated with the cran R-package e1071 (40) using Gaussian basis radial functions as the kernel. This enabled transforming the logarithmic amplitudes of selected peptides into a higher-dimensional space to obtain a linear hyperplane as separation surface. The parameters of the kernel function for the 11-dimensional hyperplane were found with the tune.svm function from the same package that uses a greedy search in the two-dimensional space spanned by the penalty parameter (C) and the kernel with (γ). Using a 10-fold cross-validation on the training data, we found the optimal parameters to be C = 51.2 and γ = 0.00256.

For setting up the linear model, the average signal intensity for one of the 12PUV peptides in the ESRD group was compared to the average intensity for the same peptide in the noESRD group. To avoid artificial weighting of specific peptides in the set due to the difference in observed signal intensities for the two groups investigated, the relative distance between the within-group averages was compared with the between-group distance of these averages always being set to two. This relative distance of signal intensities between the two groups was calibrated with the following formula:(UkiGrandmean)2|x¯ESRDx¯noESRD|where Uki is the log-transformed signal intensity of the ith peptide in the kth sample, Grandmean is the average of the mean intensity of all potential peptides, x¯ESRD represents the mean observed signal intensity of the peptide from all samples in the ESRD group, and x¯noESRD is the mean signal intensity of the peptide in the noESRD group (the values used for the calculation are listed in table S6). The score for each subject was subsequently obtained by summing the 12 Uki.

The KNN algorithm classified the samples based on closest training examples in the 12-dimensional feature space spanned by the 12PUV. By cross-validating the training data, we found that using five neighbors (k = 5) was optimal for the data in this study. A subject was classified by a majority vote of its neighbors, with the subject being assigned to the class most common among its five nearest neighbors (either ESRD or noESRD). The proportion of the votes for the winning class was then used as the score for a given subject.

The CIs for sensitivity, specificity, as well as NPV and PPV were computed assuming that data were obtained by binomial sampling according to the formula described in (41). For computing NPV and PPV, a prevalence of 0.42 was assumed on the basis of the prevalence of ESRD in PUV patients described in the literature (24, 25). Comparison of sensitivities and specificities of the classical clinical parameters to the 12PUV classifier was performed with a simple McNemar test (42). To estimate the significance of differences in NPVs and PPVs between the 12PUV test and the classical clinical parameters, the test statistics proposed in (43) was used. All the analyses were conducted in R (Team 2013) (40). Statistical analyses were carried out, and models were developed by J.S., H.M., and M.D.

The sensitivity and specificity of the previously defined biomarker models and 95% CIs were calculated with ROC plots (MedCalc version, MedCalc Software). Logistic regression and pairwise comparison of ROC curves were also carried out in MedCalc according to (44). Unblinding of the validation cohort and assessment of efficacy were performed by J.K. and J.P.S.

Supplementary Materials


Table S1. Clinical data from the discovery cohort (28 patients with PUV).

Table S2. Clinical data from the validation cohort (38 patients with PUV).

Table S3. Detailed information on the 26 PUV biomarkers.

Table S4. Comparison of sensitivity, specificity, and predictive values of the different clinical predictors used in the study and the target clinical test.

Table S5. Fetal urine peptides used for normalization and their matching with postnatal urinary housekeeping peptides.

Table S6. Values used for the linear model.

References (45, 46)

References and Notes

  1. Acknowledgments: We thank the COST action EuroKUP and C. Rumpf for redesign of the figures. Funding: The research presented in this manuscript was supported by the FP7 programs “Improvement of tools and portability of MS-based clinical proteomics as applied to chronic kidney disease” (Protoclin, PEOPLE-2009-IAPP, GA 251368) and “European Consortium for High-Throughput Research in Rare Kidney Diseases” (EURenOmics, GA2012-305608) and the Agence Nationale pour la Recherche (Beyond Markers, ANR-07-PHYSIO-004-01) program. CE-MS equipment was funded by the Fondation pour la Recherche Médicale “Grands Equipements pour la Recherche Biomédicale” and the CPER2007-2013 program. S.D., F.B., and J.-L.B. acknowledge support by INSERM and the “Direction Régional Clinique” (CHU de Toulouse) under the Interface program. J.P.S. was supported by INSERM and the “Direction de la Recherché Médicale et Innovation” (CHU de Toulouse) under the “Contrat Hospitalier de Recherche Translationnelle” program. Author contributions: J.K., C.C., and B.B. performed wet laboratory experiments. J.S. and H.M. analyzed CE-MS data and developed models. M.D. and J.S. performed statistical analyses. A.S. performed CE-MS analysis at the Glasgow site. C.L., P.Z., and W.M. performed sequence analysis. J.K., J.S., H.M., J.-L.B., B.M., S.D., F.M., and J.P.S. designed the study, interpreted the data, and wrote the manuscript. F.M., S.D., and F.B. handled and interpreted the clinical data. Competing interests: H.M. is the chief executive officer and founder and J.S., M.D., and P.Z. are employees of Mosaiques Diagnostics and Therapeutics. The other authors do not declare competing interests. Data and materials availability: All data are available as supplementary material to the manuscript. Fetal urine can be made available under a materials transfer agreement.
View Abstract

Stay Connected to Science Translational Medicine

Navigate This Article