Identification of Plasma Metabolites Prognostic of Acute Kidney Injury

Jun 3, 2015 - ... AKI across all stages, with an average accuracy of 80 ± 0.9% and an area under the receiver operating characteristic curve of 0.87 ...
0 downloads 0 Views 802KB Size
Subscriber access provided by UNIV OF CALIFORNIA SAN DIEGO LIBRARIES

Article

Identification of Plasma Metabolites Prognostic of Acute Kidney Injury after Cardiac Surgery with Cardiopulmonary Bypass Helena U. Zacharias, Jochen Hochrein, Franziska C. Vogl, Gunnar Schley, Friederike Mayer, Christian Jeleazcov, Kai-Uwe Eckardt, Carsten Willam, Peter J. Oefner, and Wolfram Gronwald J. Proteome Res., Just Accepted Manuscript • DOI: 10.1021/acs.jproteome.5b00219 • Publication Date (Web): 03 Jun 2015 Downloaded from http://pubs.acs.org on June 11, 2015

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Journal of Proteome Research is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

Identification of Plasma Metabolites Prognostic of Acute Kidney Injury after Cardiac Surgery with Cardiopulmonary Bypass

Helena U. Zacharias1, Jochen Hochrein1, Franziska C. Vogl1, Gunnar Schley2, Friederike Mayer3, Christian Jeleazcov3, Kai-Uwe Eckardt2, Carsten Willam2, Peter J. Oefner1 and Wolfram Gronwald1* 1

Institute of Functional Genomics, University of Regensburg, Josef-Engert-Str.

9, 93053 Regensburg, Germany 2

Department of Nephrology and Hypertension, Friedrich-Alexander-University

Erlangen-Nürnberg, Ulmenweg 18, 91054 Erlangen, Germany 3

Department of Anaesthesiology, Friedrich-Alexander-University Erlangen-

Nürnberg, Krankenhausstr. 12, 91054 Erlangen, Germany

Corresponding Author: Wolfram Gronwald, Institute of Functional Genomics, Josef-Engert-Str.

9,

93053

Regensburg,

[email protected],

Phone

Fax: +49-(0)941-943-5020

Abbreviated title: Plasma metabolomics in acute kidney injury

1 ACS Paragon Plus Environment

Germany,

E-mail:

+49-(0)941-943-5015,

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Abstract Acute kidney injury (AKI) is a frequent complication after cardiopulmonary bypass, but early detection of postoperative AKI remains challenging. Protein biomarkers predict AKI excellently in homogenous cohorts, but are less reliable in patients suffering from various co-morbidities. We employed nuclear magnetic resonance spectroscopy in a prospective study of 85 adult cardiac surgery patients to identify metabolites prognostic of AKI in plasma specimens collected 24 hours after surgery. Postoperative AKI of stages 1-3, as defined by the Acute Kidney Injury Network (AKIN), developed in 33 cases. A Random Forests classifier trained on the NMR spectra prognosticated AKI across all stages with an average accuracy of 80 ± 0.9% and an area under the receiveroperating characteristic curve of 0.87 ± 0.01. Prognostications were based, on average, on 24 ± 2.8 spectral features. Among the set of discriminative ions and molecules identified were Mg2+, lactate and the glucuronide conjugate of propofol. Using creatinine, Mg2+ and lactate levels to derive an AKIN index score we found AKIN 1 disease to be largely indistinguishable from AKIN 0 in concordance with the rather mild nature of AKIN 1 disease.

Keywords: acute kidney injury, AKI, metabolomics, NMR spectroscopy, prognostication, molecular signature.

2 ACS Paragon Plus Environment

Page 2 of 32

Page 3 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

Introduction Acute kidney injury (AKI) is a frequent complication following cardiac surgery with cardiopulmonary bypass (CPB), occurring in up to 40% of cases. Postoperative AKI is associated with an increased risk of mortality, exceeding 50% in those requiring dialysis, morbidity and a prolonged stay in intensive care.1 Furthermore, several studies indicate that AKI patients are more likely to have progressive renal changes even after reduction of serum creatinine to normal levels.2 The main risk factors for postoperative AKI seem to be related to both, patient characteristics (e.g., age, basal renal function) and surgical procedure (e.g., duration of CPB use).3 Several factors, including ischemia, activation of inflammatory pathways, atheroembolism, and nephrotoxins have been proposed to contribute to the pathophysiology of postoperative AKI.4 Increases in serum creatinine (SCr) levels from a measured or hypothetical baseline value and temporal decreases in urine output (UO)5,6 form the basis for diagnosis and classification of AKI according to the AKIN (Acute Kidney Injury Network), RIFLE (Risk, Injury, Failure, Loss and ESRD) and KDIGO (Kidney Disease: Improving Global Outcomes) criteria (KDIGO workgroup 2012). The main drawbacks of SCr as an indicator of acute changes in kidney function are lack of sensitivity and diagnostic delay.7 Furthermore, SCr levels are influenced by several non-renal factors including age, gender, race, intravascular volume, muscle metabolism, drugs, and nutrition.8 On the other hand, as increases in serum creatinine of ≥0.5 mg/dL (44.2 µmol/L) are associated with a greater than 18-fold increase in 30-day mortality, strategies to mitigate renal injury require early intervention before serum creatinine rises.7,9,10 Consequently, there has been an increasing interest in the identification of early urinary and/or serum biomarkers prognostic of AKI.11–14,7,15,16,1,17 In pediatric cardiac surgery patients the concentration of neutrophil-gelatinase-associated lipocalin (NGAL) in urine and serum was reported to predict AKI as early as two hours after surgery, with areas under the curve of the receiver-operating 3 ACS Paragon Plus Environment

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

characteristic (AUC-ROC) in excess of 0.90.7 However, in adult patients a number of protein biomarkers including NGAL, serum cystatin C (CysC), γglutamyltranspeptidase (GGT), alkaline phosphatase (AP), kidney injury molecule-1 (KIM-1), and interleukin-18 (IL-18) tested individually or in combination proved significantly less powerful in the early detection and prognostication of AKI with AUC-ROC values below18,19 or slightly above 0.80.9 In comparison, with an AUC-ROC value of 0.83, metabolite NMR fingerprinting of urine specimens collected 24 hours after cardiac surgery from 106 patients showed equal to superior prognostic power, with carnitine and the antifibrinolytic agent tranexamic acid being the most discriminating biomarkers aside from creatinine.17 Here, in a prospective study of 85 unselected adult patients undergoing cardiac surgery with CPB, we applied 1H-NMR spectroscopy to plasma specimens collected 24 hours after surgery to identify low molecular weight factors prognostic for AKI. Employing Random Forests (RF) based prognostication,20 a set of 24 spectral features prognostic of AKIN 1-3 AKI at 48 hours after surgery could be identified.

Experimental section Patients Of the original cohort of 106 patients, from whom urine specimens had been collected prior and 4h and 24 h after cardiac surgery with CPB use at the University Clinic Erlangen between July 2009 through August 2010,17 85 cases, for whom also frozen (-80 °C) EDTA-plasma was available 24 hours postoperatively, were included here with the patients’ written consent. Detailed information on clinical characteristics, administered medication and outcome is given in Supplemental Table S1. In total, 33 patients out of 85 were diagnosed with post-operative AKI 48h and 72h after cardiac surgery. Twenty-four, 3 and 6 patients were classified as AKIN 1 (1.5- to 2-fold increase in serum creatinine 4 ACS Paragon Plus Environment

Page 4 of 32

Page 5 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

from baseline), AKIN 2 (>2- to 3-fold increase), and AKIN 3 (>3-fold increase), respectively, after 72 h. While 32 patients had reached the same stage of AKI already after 48 h, one patient, being classified as AKIN 1 48h after surgery was re-classified as AKIN 3 72h after surgery due to a dramatic increase in serum creatinine on the third postoperative day and was persistently classified as AKIN 3 for the purpose of this analysis.

NMR spectroscopy Prior to NMR, EDTA-plasma specimens were subjected to ultrafiltration with a cut-off of 10 kD.21 Four hundred µL of ultrafiltrate were mixed with 200 µL of 0.1 mol/L phosphate buffer, pH 7.4, and 50 µL of 29.02 mmol/L 3trimethylsilyl-2,2,3,3-tetradeuteropropionate (TSP) dissolved in deuterium oxide as internal standard (Sigma-Aldrich, Taufkirchen, Germany). NMR experiments were carried out on a 600 MHz Avance III spectrometer (BrukerBioSpin, Rheinstetten, Germany). For every plasma specimen, 1D 1H and 2D 1H-13C HSQC spectra were acquired following established protocols.22 To support metabolite identification, a high-resolution 2D 1H-13C heteronuclear single-quantum correlation (HSQC), a 2D 1H-13C heteronuclear multiple bond correlation (HMBC), and a 2D 1H-1H total correlation spectroscopy (TOCSY) spectrum of a representative non-AKI plasma specimen were acquired.23

Mass spectrometry Ultrafiltered plasma samples of 5 AKI and non-AKI patients each were diluted with deionized water (1:4). Metabolic fingerprinting was performed as previously described24 with the signal-to-noise threshold set to 20 in the “find molecular feature” algorithm in CompassDataAnalysis 4.1 (Bruker Daltonics, Bremen, Germany). Feature alignment over a retention time window of 0.01-14 min was achieved using the 64-bit beta version of Profile Analysis 2.1 (Bruker

5 ACS Paragon Plus Environment

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 32

Daltonics). Reference compounds of propofol metabolites were purchased from Toronto Research Chemicals (Toronto, Canada).

Data analysis NMR data preprocessing for statistical data analysis To compensate for slight shifts in signal positions across spectra due to small variations in sample pH, salt concentration and/or temperature, the spectral region from 9.5 – 0.5 ppm of the 1D plasma spectra was evenly split into bins of 0.01 ppm employing Amix 3.9.13 (BrukerBioSpin). After exclusion of the region from 6.2 – 4.6 ppm, which contains the broad urea and water signals, and NMR signals (3.815 ppm – 3.76 ppm, 3.68 ppm – 3.52 ppm, 3.23 ppm – 3.20 ppm, and 0.75 ppm – 0.725 ppm) corresponding to residual glycerol from the ultrafiltration membrane and free EDTA, respectively, a total number of 718 spectral bins remained. The signal intensities of each bucket (feature) were summed, followed by scaling of all bucket intensities to the integral of the TSP reference signal to correct for variation in spectrometer performance. The validity of this approach has been shown previously (manuscript in preparation). To minimize heteroscedasticity, data were also log2 transformed before import into the statistical-analysis software R version 2.15.125 for further analysis.

Prognostication Method Prognostication of specimens was performed employing a Random Forests (RF) classifier18

provided

in

the

R-package

randomForest

(http://cran.r-

project.org/web/packages/randomForest)26 in combination with t-test based feature filtering. This combined strategy allows faster subsequent identification of NMR signals driving the separation of cases. It also keeps the computational model relatively sparse.

6 ACS Paragon Plus Environment

Page 7 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

Prognostications were accomplished within a nested leave-5-out cross-validation scheme, whose general outline had been described previously.27,28

Further

details regarding prognostication are given in the online Supplemental Material.

Metabolite Quantification Details about metabolite quantification including the used software routines are given in the online Supplemental Material.

Results Prognostication of AKI Figure 1 shows an exemplary subtraction spectrum obtained by subtracting the 1D 1H NMR spectrum of a non-AKI plasma specimen from that of an AKIN 3 specimen, both of which had been collected 24 hours after surgery. The subtraction of measured spectra generates a virtual NMR spectrum that highlights those spectral features that differ between the samples. It is obvious, that the total integral of spectral features upregulated in the AKIN 3 specimen compared to the non-AKI specimen is much larger than that of the downregulated features. This is mainly due to the significantly (p=0.03) higher levels of glucose, the most abundant plasma metabolite, in the AKI (9.72±2.73 mmol/L) than the non-AKI group (8.38±2.83 mmol/L) (Table 1). For RF-based prognostication of the eventual AKIN stage, log2 transformed 1D 1

H NMR spectra were split into 718 evenly spaced bins or features, excluding

chemical shifts representing water, urea, glycerol and free EDTA. Subsequent analysis of the 33 AKI and 52 non-AKI cases yielded an overall prognostication accuracy of 80.0 ± 0.9% and a corresponding area under the receiver-operatingcharacteristic (ROC) curve of 0.87 ± 0.01. On average, the RF algorithm employed 24.0 ± 2.8 of the most discriminative features as selected by a t-test based feature selection step prior to classification. As can be seen from the Supplemental Table S3 the corresponding p-values of these features showed a 7 ACS Paragon Plus Environment

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

range from 2.06e-8 to 8.55e-6. The RF parameters mtry and ntree were optimized to 3.0 ± 0.0 and 270 ± 24.5, respectively. The overall sensitivity and specificity amounted to 72.7 ± 1.9% and 84.6 ± 1.7%, respectively. Considering the AKIN stages separately, the sensitivity for AKIN 2 and 3 amounted to 100.0 ± 0.0% and 96.7 ± 6.7%, respectively, whereas it dropped to 63.3 ± 1.7% for AKIN 1 patients. The 85 plasma samples constituted a subsample of our original study of 106 individual urine specimens.17 In that study, support vector machine based prognostication employing a radial kernel function had yielded an overall prognostication accuracy of 76.0 ± 1.9%. To allow for a fair comparison between urinary and plasma data, the 1D-1H-NMR urine spectra of the 85 patients, for whom plasma specimens were available, were scaled to creatinine and log2 transformed before they were subjected to a single random forest run. The obtained overall prognostication accuracy amounted to 69.4% employing 7 features with a corresponding area under the ROC curve of 0.73. Next, permutation tests with randomly perturbed class labels29 were performed to exclude the possibility that the observed prognostication accuracies for plasma had been obtained by chance. After an initial RF run with the complete feature set of 718 features and permuted class labels, which revealed 119 as the median number of selected features, the feature selection was limited to a range of 109 to 129 features for the subsequent twenty RF runs, each of which started with a fresh permutation of the class labels and a random splitting of test and training data. Over the twenty RF runs, we received an averaged total accuracy of 55.7 ± 5.1%, a mean area under the ROC curve of 0.48 ± 0.08, and a sensitivity and specificity of 17.1 ± 7.2% and 80.2 ± 6.3%, respectively. Results for the permuted data were in all 20 runs considerably lower than for the nonpermuted data, indicating that the results for the non-permutated data were with high probability not obtained by chance (p < 0.05) and that the study was

8 ACS Paragon Plus Environment

Page 8 of 32

Page 9 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

sufficiently powered. Two exemplary ROC curves for the permuted and nonpermuted data are given in Supplemental Figures S1A and S1B, respectively. As described in the experimental section and the supplemental material, tstatistics were used for both, feature filtering and identification of spectral features that distinguish between AKI and non-AKI plasma NMR fingerprints. After correction for multiple testing by controlling the false discovery rate (FDR) at 5%, 261 significantly differential NMR features were obtained. A heat map representation of these features is displayed in Supplemental Figure S2 and a list of all significant NMR features is given in Supplemental Table S3. Their up- and down-regulation in the heat map representation is color coded in yellow and blue, respectively. The patients were arranged from left to right as follows: 45 cases correctly prognosticated not to develop AKI, 7 cases falsely prognosticated to develop AKI, 9 cases of AKIN 1 falsely prognosticated not to develop AKI, and 15, 3 and 6 cases each of AKIN 1, 2 and 3, respectively, correctly prognosticated. Rows were ordered according to increasing correlation coefficients between disease status and feature intensities. As on average 24.0 ± 2.8 of the most significant features were used by the RF algorithm, the 27 most significant NMR features are indicated with red arrows in Supplemental Figure S2. These NMR features could only partly be assigned to known metabolites due to either massive signal overlap in some regions of the 1D spectra (see Figure 1) or insufficient signal intensity. The most significant feature was a well-resolved singlet signal present at 7.285 ppm (Padj=2.1e-8), which could be identified by neither database searches nor 2D 1H TOCSY, 1H-13C HSQC, and 1H-13C HMBC spectra, respectively. Therefore, to facilitate assignment, metabolic fingerprinting was performed by means of high-resolution LC-QTOF-MS on five plasma specimens each selected from the AKI and the non-AKI group, respectively. A total of 531 features were observed in positive mode and 16 in negative mode. Features were sorted according to Student’s t-tests. After controlling the FDR at the 5% level 9 ACS Paragon Plus Environment

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

according to the method of Benjamini and Hochberg, 11 significant features remained, each of which was defined by retention time and the m/z value. By means of the Smart Formula tool (Bruker Daltonics, Bremen, Germany), molecular sum formulas were determined to search the HMDB,30 METLIN,31 and ChEBI (Chemical Entities of Biological Interest) metabolite databases.32 For the most promising hits, commercial standards were analyzed to verify identification. Furthermore, MS/MS experiments were performed on both standards and plasma specimens for additional verification. Among the most discriminating features, the propofol metabolites propofol-glucuronide and 4hydroxy-propofol-1-OH-D-glucuronide were positively identified. 1D 1H NMR reference spectra were acquired on these compounds and unambiguously verified the assignment of the NMR signal at 7.285 ppm to propofolglucuronide. Furthermore, the presence of 4-hydroxy-propofol-1-OH-Dglucuronide as another discriminating compound could be verified by the NMR data. NMR-based quantification of propofol-glucuronide showed significantly increased plasma levels (0.004 ± 0.002 mmol/L vs. 0.010 ± 0.08 mmol/L, p=0.00008) in AKI patients. However, for both the total dosage of administered propofol (2747.7 ± 1257.4 mg vs. 3313.3 ± 1896.4 mg, p=0.14) and the dosing rate (6.8 ± 3.7 mg/min vs. 5.8 ± 1.9 mg/min, p=0.09), no significant differences between non-AKI and AKI patients could be observed. However, the two groups differed significantly with regard to the duration of propofol administration (427.94 ± 185.34 min and 600.79 ± 360.69 for non-AKI and AKI patients, respectively; p=0.015) and the time elapsed between termination of propofol infusion and sample collection (1138.6 ± 201.7 min and 969.0± 362.7 min for non-AKI and AKI patients, respectively; p=0.02). This suggested that the increased plasma levels of propofol-glucuronide in AKI patients were a consequence of both prolonged administration and delayed excretion. This was confirmed by re-analysis of the urinary NMR fingerprints previously obtained for the same patients.17 At 4 hours after surgery, urinary levels of propofol10 ACS Paragon Plus Environment

Page 10 of 32

Page 11 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

glucuronide amounted to 0.67 ± 0.30 mmol/mmolcrea and 0.63 ± 0.35 mmol/mmolcrea, respectively, for non-AKI and AKI patients (p=0.55), while at 24 hours after surgery urinary levels had dropped to 0.14 ± 0.07 mmol/mmolcrea and 0.19 ± 0.10 mmol/mmolcrea, respectively, but were significantly (p=0.02) higher in the AKI group (Table 1). As can be seen from Supplemental Table S3, plasma NMR features used for prognostication correspond to compounds of both endogenous origin, such as tryptophan (Padj=1.1e-6), myo-inositol (Padj=2.3e-6), hippurate (Padj=2.5e-6), citrate (Padj=3.2e-6), and creatinine (Padj=3.9e-6), and exogenous origin, such as propofol-glucuronide (Padj=2.1e-8) and the antifibrinolytic agent tranexamic acid (Padj=3.2e-6). A special case is Mg2+ (Padj=2.9e-6), which can be of both endogenous and exogenous origin as Mg2+ is often administered for the treatment of cardiac dissrhythmia. To analyze the impact of tranexamic acid and other exogenous compounds such as D-mannitol, paracetamol-sulfate, propofol-glucuronide, 4-hydroxy-propofol1-OH-D-glucuronide

and

4-hydroxy-propofol-4-OH-D-glucuronide

on

prognostication, all spectral areas corresponding to known exogenous compounds were excluded prior to data analysis. In subsequent analysis, which employed 26 features, an average prediction accuracy of 82.4% (vs. 80% including exogenous compounds) and an area under the ROC curve of 0.87 (vs. 0.87) were obtained. Overall sensitivity and specificity amounted to 75.8% and 86.5%, respectively, while the respective values before exclusion of exogenous compounds had amounted to 72.7% and 84.6%. In addition, we investigated whether improved prognostication performance could be achieved by combining the plasma data with the corresponding, previously published 24-h urine data,17 which had been scaled to creatinine and also log2 transformed. The final data matrix consisted of 1419 rows representing 718 plasma and 701 urine features and 85 columns representing the 33 and 52 AKI and non-AKI patients, respectively. One RF classification run with t-test 11 ACS Paragon Plus Environment

Journal of Proteome Research

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

based feature selection employing a leave-five-out cross-validation was performed. Results showed an averaged prognostication accuracy of 81.2% and an area under the ROC curve of 0.87, values similar to those obtained for plasma only. Further analysis showed that prognostication was based on 25 features. Ranking of these features according to their p-values revealed, that the first 24 features were identical to the first 24 plasma features listed in Supplemental Table S3. The most significant urinary feature was tranexamic acid at rank 25 (padj=1.6e-5), which explains why a combination of urinary and plasma fingerprints did not outperform prognostication on plasma fingerprints alone. Twelve out of 52 non-AKI patients (23.1%) and 21 out of 33 AKI patients (63.6%), respectively, suffered from chronic kidney disease (CKD), with a pvalue calculated by Fisher’s exact test of 0.0003 (compare to Supplemental Table S1). To test whether results obtained for AKI and non-AKI specimens had been dominated by CKD, we selected randomly 24 patients each from the AKI and non-AKI group, so that each group included 12 patients with and 12 patients without CKD. A Student’s t-test yielded after correction for multiple testing 73 significant features, 72 of which had been also part of the significant features obtained when all 85 samples were included (Supplemental Table S3). Random Forests based prognostication with leave-two-out cross-validation obtained an averaged prognostication accuracy of 72.9%, as well as an area under the ROC curve of 0.84. On average 17.5 features were employed by the algorithm. Sensitivity and specificity amounted to 70.8% and 75.0%, respectively. The corresponding permutation test was performed once, with an average total accuracy of 41.7%, an area under the ROC curve of 0.44, a sensitivity of 50.0% and a specificity of 66.7%. As both groups contained the same number of patients with and without CKD, these results clearly showed that CKD incidence did not exert a major effect on prognostication of AKI.

12 ACS Paragon Plus Environment

Page 12 of 32

Page 13 of 32

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Proteome Research

Quantification of metabolites In addition to the analysis of the NMR fingerprints, the plasma levels of 16 organic metabolites and the dications Ca2+ and Mg2+ were quantified from the 85 1D 1H NMR spectra. Mean plasma concentrations and standard deviations for the non-AKI and AKI group, respectively, as well as p-values based on twosided t-tests are given in Table 1. Note that due to the relatively small number of quantified metabolites no correction for multiple testing was applied. Metabolites that differed significantly (p