Isotope Ratio Outlier Analysis (IROA) for HPLC–TOFMS-Based Metabolomics of Human Urine

Abstract

Metabolic fingerprinting by mass spectrometry aims at the comprehensive, semiquantitative analysis of metabolites. Isotope dilution, if successfully implemented, may provide a more reliable, relative quantification. Therefore, the ¹³C labeled yeast extract of the IROA TruQuant kit was added as an internal standard (IS) to human urine samples measured in full-scan mode on a high-performance liquid chromatography-time-of-flight mass spectrometer (HPLC–TOFMS) system. The isotope ratio approach enabled the analysis of 112 metabolites. The correlation with reference data did not improve significantly using ¹²C/¹³C ratios compared to absolute ¹²C peak areas. Moreover, using an intricate ¹³C-labeled standard increased the complexity of the mass spectra, which made correct signal annotation more challenging. On the positive side, the ratio approach helps to reduce batch effects, but it does not perform better than computational methods such as the “removebatcheffect” function in the R package Limma.

1. Introduction

Metabolic fingerprinting, also called untargeted metabolomics, aims at analyzing all detectable metabolites in a given sample. It is a semiquantitative approach that focuses on spotting significant differences in metabolite abundance among different groups of a sample set [1,2,3,4,5]. To this end, liquid chromatography–mass spectrometry (LC–MS) is widely used.

Issues such as matrix effects (i.e., the variability of ionization efficiency); analytical variability, e.g., decline in instrument performance when analyzing large sample batches; and difficulties in annotating the detected features are hurdles of this approach [6,7,8]. Metabolic profiling, on the other hand, circumvents these obstacles by using known amounts of labeled internal standards (IS), and tailored sample preparation and LC–MS methods for the optimal detection and absolute quantification of a predefined but often relatively small set of target metabolites [9,10]. Often relative, as in metabolic fingerprinting, instead of absolute quantification is sufficient. In this context, stable isotope-labeled internal standards can improve analytical repeatability and allow for the relative quantification of a large number of metabolites in a single LC–MS run. The question remains, which internal standards are best to use when all the detectable metabolites are targeted? One approach is to use a stable isotope-labeled reagent to derivatize a sample aliquot or a pool sample to generate a sample specific IS [11,12,13,14]. Yet, this approach is only selective to certain compound classes that are covered by the chosen derivatization. Global labeling of the whole metabolome of an organism was used in targeted [15] and untargeted [16] analyses, but it was limited to the respective organism. Dethloff et al. [17] used partially labeled mouse plasma as an IS for amino acid analysis in human plasma. Furthermore, ¹³C-labeled yeast extracts have been used (ISOtopic solutions, Vienna, Austria) [9]. Employing a complex ¹³C labeled yeast extract is also the basis of the IROA TruQuant kit (IROA Technologies LLC^®, Bolton, MA, USA).

Isotopic ratio outlier analysis (IROA) is a mass-spectrometry-based technique that helps to distinguish features of biological origin from nonbiological artefacts and, secondly, aims at providing more reliable quantitative data by correcting for ion suppression and analytical variance [18]. The kit includes two types of standards. The long-term reference sample (LTRS) is a fully labeled S. cerevisiae yeast cell extract, labeled at 5% and 95% ¹³C and mixed 1:1. This translates to a unique isotopic pattern for all features of a biological origin, whereas artefacts would exhibit the isotopic pattern of natural abundance, allowing for a distinct filter criterion. Note that the LTRS is injected as is, without mixing with any specimens of interest. Thereby, the LTRS serves as a quality control (QC) for instrumental performance. It also allows for a validated annotation of the detected features that show the expected isotopic pattern, yielding a reference library. The second standard is the internal standard (IS), which only contains the 95% ¹³C labeled yeast extract. It is added to the samples to be analyzed, providing the advantages of a traditional internal standard by correcting for ion suppression and analytical variance. It also offers a more robust annotation process, benefiting from the reference library built using the LTRS measurements [6,18]. The key advantage compared to a traditional internal standard is that it comprises a very large number of labeled compounds, i.e., all metabolites present in yeast. However, as the exact quantities of the labeled compounds are not known, it allows for relative but not for absolute quantification.

Metabolomics data acquired in separate experimental batches may exhibit variation unrelated to biological differences among different sample groups. The sources of this additional analytical variance are, for example, varying experiment times, reagents, and instrument performance. Even when using the same reagents and instruments, the performance of the high-performance liquid chromatography (HPLC) system and the mass spectrometer response may deviate over time. These effects are aptly termed ‘batch effects’ [19]. Their impact can be mitigated using batch-effect correction algorithms (BECAs) such as “removebatcheffect” (RBE), which is a function in the Limma package in R [20,21], where the batch effect is modeled by an additional coefficient. However, they can only be applied in cases where data have been acquired in distinct batches and not in cases where samples were measured continuously over prolonged periods of time. Furthermore, BECA’s effectiveness is questionable in some cases as they could be misapplied, and the fitted model might introduce bias into the data [19]. Taking measures early in the workflow might produce better results than circumventing batch effects via meta-analysis. Having a reference sample such as the LTRS could help to account for changes in chromatographic conditions, instead of using more lenient thresholds in feature alignment and weak signal recovery [22]. The implemented IS can also correct for shifts in the detector response.

Several studies [6,23,24] have employed the isotopic ratio outlier analysis technique in model organisms. In such experiments, only one type of sample matrix is involved, and any features that do not exhibit the IROA isotopic pattern are indeed artefactual. Here, however, we used the labeled yeast extract, i.e., the IS of the TruQuant kit, as an internal standard to analyze human urine specimens. Thus, we are dealing with two different matrices. If a feature does not exhibit the IROA isotopic pattern, it could merely be a compound that is urine-specific or not produced in yeast in detectable quantities. With this in mind, we investigated the added value of incorporating the IROA TruQuant kit in our untargeted analysis of human urine.

We analyzed a set of 244 spot urine specimens from the German Chronic Kidney Disease (GCKD) project collected at the baseline time point [25]. The samples were analyzed by means of a high-performance liquid chromatography (HPLC) system coupled to a time-of-flight mass spectrometer (TOFMS). The resulting data are referred to as TOF. To generate reference datasets, a subset of 56 of the urine specimens was additionally analyzed with two different methods. The first was MS-based quantitative data of amino acids (AAs), referred to as “quant”. The second, referred to as “NMR”, stemmed from a nuclear magnetic resonance (NMR) quantitative analysis.

2. Materials and Methods

2.1. Chemicals

The IROA TruQuant IQQ Workflow Kit was purchased from Sigma–Aldrich (Taufkirchen, Germany). HPLC grade solvents for LC–MS analysis were purchased from VWR (Vienna, Austria). Purified water used to prepare the solutions was from a PURELAB Plus system (ELGA, LabWater, Celle, Germany)

2.2. Sample Preparation

The IROA Kit was applied to 244 urine samples from the German Chronic Kidney Disease (GCKD) project [25], from the baseline visit. The study was executed in accordance with the Declaration of Helsinki and registered in the national registry for clinical studies (DRKS 00003971). All study procedures and protocols were approved by the ethics committees of all participating institutions (Friedrich Alexander University Erlangen–Nuremberg, the Medical Faculty of the Rheinisch–Westfälische Technische Hochschule Aachen, Charité University Medicine Berlin, the Medical Center University of Freiburg, Medizinische Hochschule Hannover, the Medical Faculty of the University of Heidelberg, Friedrich Schiller University Jena, the Medical Faculty of the Ludwig Maximilians University Munich, and the Medical Faculty of the University of Würzburg). The study was carried out in accordance with relevant guidelines and regulations. Written declarations of informed consent were obtained from all study participants before inclusion. All specimens were stored at −80 °C until preparation of samples for mass spectrometry.

Various factors affect the abundance of urinary metabolites, such as liquid and food intake and many processes regulating the body’s solute and water content. Normalization is needed to balance out these differences among individuals and reveal actual biological variance. Creatinine is commonly used to achieve that as it is mainly excreted by glomerular filtration at a constant rate without reabsorption in the renal tubule [26]. Urine was diluted with pure water to a similar creatinine concentration to correct for urinary output. Since the individual dilution of each sample is impractical, samples were grouped based on the creatinine concentration, and a fixed dilution factor was used for the respective group, see

Table 1. Dilution of urine samples based on creatinine concentration.

Creatinine (mM)	Dilution Factor
1–3	undiluted
3–6	1:2
6–12	1:4
12–24	1:8

. Creatinine concentrations determined in the course of a corresponding NMR analysis of the samples were utilized.

Samples were diluted to a volume of 20 µL. For the internal standard, five vials of IROA–IS were used. To each vial, 600 µL of pure water was added, instead of the recommended 1200 µL, to keep a proper concentration after dilution with urine. All five resulting IS solutions were mixed to form a homogenous 3 mL IROA–IS solution. Then, 10 µL of IROA–IS was added to each diluted urine sample, resulting in a final volume of 30 µL. Two in-house QC urine samples were prepared similarly. QC1 and QC2 have creatinine concentrations of 19.71 mM and 8.92 mM, respectively, and were diluted 1:8 and 1:4 with water correspondingly, to reach a volume of 120 µL. Sixty microliters of IS was then added to each QC, which allowed one to maintain the same sample:IS ratio used to prepare the samples. The final volume of 180 µL of each QC was necessary to allow multiple injections throughout the experiment.

For the LTRS, 80 µL of water was added to reach a similar concentration of the IS in the samples. IS blank was prepared by adding 80 µL water to a 40 µL IS solution. A water blank was included.

For NMR analysis, samples were prepared by mixing 400 μL human urine with 200 μL of 0.1 M phosphate buffer, pH 7.4, which contained 3.9 mM boric acid to impair bacterial growth. Furthermore, 50 μL of 0.75% (wt) trimethylsilylpropanoic acid (TSP) in deuterium oxide (D₂O) and 10 μL of 81.97 mM formic acid (FA) were added as internal reference standards. NMR spectra were acquired on a Bruker 600 MHz Avance III HD NMR spectrometer (Bruker BioSpin GmbH, Rheinstetten, Germany) equipped with a helium cooled cryo–probe and an automated sample changer. For each sample, a 1D ¹H spectrum was acquired employing a Carr–Purcell_Meiboom–Gill (CPMG) pulse sequence to facilitate the suppression of macromolecular signals. From the obtained spectra, metabolites were quantified by the fitting of reference signals employing the Chenomx NMR suite v. 8.6 (Chenomx Inc., Edmonton, Canada).

To investigate whether the urine specimens used in the experiment contained a sufficient concentration of metabolites, we performed the recommended calibration experiment according to the protocol in the Kit’s product information sheet [27]. The content of the internal standard (IS) vial was solubilized in 1.2 mL pure water and that of the LTRS in 40 μL pure water as suggested in the protocol. A urine specimen with a high creatinine level (20 mM) was diluted with pure water to creatinine concentrations of 0.25, 0.5, 1.25, 1.5, 2, 2.5, 3.75, 5, 7.5, 10, and 20 mM, respectively. Forty-microliter aliquots of each dilution were dried in triplicate employing an infrared vortex vacuum evaporator (CombiDancer, Hettich AG, Baech, Switzerland) and reconstituted in 40 μL IS solution.

As a result, a graph (see Figure S5) has been obtained that indicates the creatinine concentration to which urine specimens shall be adjusted. According to the product information sheet [27], this concentration “yields an overall mass spectral signal that is equal to the overall mass spectral signal of the IS. This is the amount of sample that will most accurately be measured using the IS in the future, i.e., well balanced by the standard 40 μL of IS.” In Figure S5, this is the intersection between the lines of the normalized IS MSTUS (blue squares), the normalized ¹²C MSTUS (green crosses), and the line of the ¹²C values corrected for ion suppression (red circles). In our case, this corresponds to about 6.5 mM creatinine. We then repeated the original experiment using a subset of 26 urine specimens with original creatinine concentrations equal to or greater than 6.5 mM. Aliquots of 40 μL of urine with a creatinine concentration of 6.5 mM (either pure or prediluted to 6.5 mM) were dried and reconstituted in 40 μL IS (dissolved according to protocol in 1.2 mL) so that each sample contained a final creatinine concentration of 6.5 mM. The LTRS and IS blanks were also prepared according to the protocol. Additionally, aliquots from the same 26 urine specimens were diluted to 2 mM creatinine to obtain concentrations comparable to our original data set, and aliquots of 40 μL were also dried and reconstituted in 40 μL IS.

2.3. Instrumentation

A Thermo Scientific Dionex Ultimate 3000 UHPLC system (Idstein, Germany) was coupled through an ESI source to a Maxis Impact QTOFMS (Bruker Daltonics, Bremen, Germany). Chromatographic separation was accomplished with a Waters™ Atlantis™ PREMIER BEH C18 AX (1.7 µm 150 × 2.1 mm id) column kept at 35 °C. The aqueous eluent A is formic acid 0.1% (v/v) and 1 mM ammonium formate in water. The organic eluent B is formic acid 0.1% (v/v) and 10 mM ammonium formate in water: ACN (10:90, v/v). A flow rate of 0.3 mL/min was used. Elution was accomplished by the following eluent B gradient: 0–30% in 10 min, 30–100% in 5 min, 100% for 5 min, back to 0% in 2 min, followed by equilibration (0% B) for 8 min. Sample volumes of 5 µL were injected.

Electrospray ionization was performed in positive ionization mode using the following settings to operate the source and the mass spectrometer: a drying gas (nitrogen) temperature of 220 °C and a flow rate of 10 L/min; nebulizer gas pressure (nitrogen): 2.6 bar; end plate offset: 500 V; capillary voltage: 4500 V; mass range: 50–1000 m/z; and acquisition rate: 2 spectra/s. An external calibration of the mass spectrometer was employed before the measurements, using a sodium formate cluster solution (10 mM sodium formate in 50:50 v/v water/isopropanol). For internal recalibration, each run involved an injection of the sodium formate cluster solution through a six-port valve. A mass spectral resolution of R = 21,000 (for m/z 90.977) was obtained. The samples were injected in random order. On average, QCs were injected after every 20 samples, whereas LTRS, IS blank, and water blank were injected after every 40 samples. In total, 331 injections were performed.

2.4. Analysis of Amino Acids

Absolute concentrations of amino acids were determined by HPLC–MS/MS, using Agilent 1200 series HPLC system (Agilent, Waldbronn, Germany) with API 4000 QTRAP (AB SCIEX, Darmstadt, Germany) after derivatization with propyl chloroformate/1–propanol using 10 μL urine. Deviating from the established protocol [28], the derivatizing reagent 2 was composed of 17.4% propyl chloroformate and 82.6% isooctane, replacing chloroform entirely with isooctane.

2.5. Software

DataAnalysis version 4.1 (Bruker Daltonics, Bremen, Germany) was used for manual inspection and the processing of the HPLC–TOFMS mass spectra and chromatograms, internal recalibration of the spectra, and picking accurate masses. MSConvert Version: 3.0.19 (a tool of ProteoWizard; Palo Alto, CA, USA) [29] was used to convert the files to .mzxml format, using the parameter “vendor” of the peak picking filter. The IROA kit comes with dedicated software, ClusterFinder™ V3 (IROA Technologies LLC^®, Bolton, MA, USA). It was used to establish a library based on the measurements of the LTRS runs. The untargeted analysis was done using MZmine 2 [30], where the chromatographic peaks were detected, automatically integrated, deconvoluted, aligned, and filtered. The integrated peak areas were extracted for further statistical evaluation.

We were provided with an upgraded version of the proprietary software ClusterFinder™ V4.2.11A (CF4). This was used for the analysis of the calibration experiment, the creatinine concentration investigation, and the analysis of the original samples that were employed for the investigation of batch effects. The parameters used are listed in Table S6. As the kit is based on 95% ¹³C for the IS, a noticeable m-1 peak (and m-2 depending on the C-number, etc.) is formed for the IS signal that increases with carbon number. For a semi-quantitative comparison of the endogenous metabolite amount to the IS amount in the sample, the peak areas of the isotopologue envelops are summed up in CF4. CF4 calculates the complete ratio between the sum of natural abundance isotopologues and the sum of 95% ¹³C isotopologues, in contrast to considering only the monoisotopic peak in each envelope of isotopologues when using MZmine. Since the absolute amount of substance of the metabolites in the IS is not known, we cannot compare the quantities of different metabolites in a sample to each other. In our work, we only compared the relative levels of the same metabolites between samples. As the isotope pattern does not change for a given metabolite in the IS or the urine specimen between samples, using the monoisotopic peak ratio is, in our opinion, a sufficient substitute for the summed-up isotopologue envelops. For comparison, the new datasets for the 26 urine specimens diluted either to 2.0 or 6.5 mM creatinine were analyzed with both CF4 and with MZmine.

2.6. Data Processing

Internal recalibration using the sodium formate cluster allowed for a mass tolerance of only 3 mDa. The LTRS data files were evaluated in ClusterFinder, and a reference library was created. This library consists of the peak pairs of the main ¹²C and ¹³C peaks, i.e., the unlabeled and the fully labeled peaks of each compound, respectively. ClusterFinder did not manage to handle all 331 data files at once. Therefore, further data evaluation had to be performed in MZmine 2 using the reference library to identify peak pairs. Peak pairs were matched manually, and orphan entries were excluded. Further data curation was done based on quality control. For instance, the ¹²C peaks should be minimal in the IS blank, so if a considerable ¹²C peak (¹²C/¹³C > 20%) was detected there, the pair was excluded. Additionally, the ratio of the pairs in the LTRS samples should be very close to one; if it deviated by more than 30%, the pair was excluded too. To confirm the annotations of the metabolites, a cross-identification was performed using our in-house library, generated using the Mass Spectrometry Metabolite Library (MSMLS™) (IROA Technologies LLC^®, Bolton, MA, USA).

2.7. Data Evaluation

To assess the added value of using the kit, the data was evaluated using either the absolute peak areas of the ¹²C signal, referred to as “TOF absolute”, or the ¹²C/¹³C ratios, referred to as “TOF ratios”. ¹²C peak areas were corrected for the dilution factor. The data were compared to data from two other sources. Quantitative data (targeted analysis) acquired for the same sample set, measured in-house on the 4000 QTrap in MRM mode, referred to as “quant”. The other data set for the same sample set was also produced via an in-house NMR analysis, referred to as “NMR”.

The IS would have an added value if the TOF ratios data correlated better to the other two datasets than the TOF absolute values. Spearman correlation coefficients were calculated. Spearman correlations were plotted, and a regression line was fitted. To analyze whether there are significant differences in the agreement of the TOF absolute or the TOF ratio method with either the quant or NMR method, we computed for both cases in each data set individual differences (“diff” values) from the respective regression lines (see Figure 1 for a typical example). These sets of differences were then analyzed by means of statistical tests. First, a Shapiro–Wilk normality test was performed to see if values in each set were normally distributed. If they were, a paired t-test was performed; otherwise, a paired Wilcoxon test was performed.

3. Results

Across all samples, 423 potential IROA features were found out of 12,938 unknown features. Quality control based on accurate masses, retention time, and feature intensity (see Section 2.6 Data processing) was applied to filter out and match the IROA ¹²C and fully labeled ¹³C signals. This left 224 features, resulting in 112 IROA pairs. Using our in-house retention time library, the identity of 27 out of these 112 were confirmed, see Table S1.

To assess the added value of using the kit, the data were evaluated once using the absolute peak areas of the ¹²C signal, referred to as “TOF absolute”, and then again using the ratios to the respective internal standards (¹²C/¹³C), referred to as “TOF ratios”. Subsequently, AAs detected with the two other methods, ‘‘quant’’ and “NMR”, see Table S1, were compared to both “TOF absolute” and “TOF ratios”. Leucine and isoleucine were chromatographically improperly resolved in the TOF data; hence, they were excluded, leaving 11 AAs covered both by quant and TOF absolute/TOF ratios. Figure 2 shows the Spearman correlation plots of four exemplary AAs ranging from high correlation (top) to low correlation (bottom), where the TOF data is represented by either TOF absolute (left) or TOF ratios (right) and compared to quant. Similar comparisons against quant were performed for all 11 mutual AAs. The corresponding Spearman correlation coefficients are given in

Table 2. Spearman correlation coefficients for TOF absolute vs. quant and TOF ratios vs. quant.

Metabolite	TOF Absolute	TOF Ratios
arginine	0.694	0.8224
aspartate	0.505	0.407
glutamate	0.917	0.9118
glutamine	0.919	0.879
lysine	0.761	0.843
phenylalanine	0.978	0.9894
proline	0.559	0.5985
serine	0.858	0.8811
threonine	0.890	0.8044
tryptophan	0.922	0.9613
tyrosine	0.980	0.9782

. Neither substantial differences using TOF absolute versus TOF ratios nor a clear trend can be seen. For a better visual comparison, the coefficients of both comparisons for all 11 AAs were plotted in a scatter plot, see Figure 3.

To analyze whether there are significant differences in the agreement of the TOF absolute or the TOF ratio method with either the quant or NMR method, we computed, for both cases in each data, set individual differences (“diff” values) from the respective regression lines (see Section 2.7 Data evaluation). These sets of diff values were then analyzed by means of statistical tests. In case of normally distributed diff values (Table S2) a t-test was used; otherwise, a paired Wilcoxon test [31] was performed. The results shown in

Table 3. Paired Wilcoxon or t-test results comparing “diff” values between TOF absolute and TOF ratios against quant.

Metabolite	p-Value	Paired Test Type
arginine	0.735	Wilcoxon
aspartate	0.999	t-test
glutamate	1.000	t-test
glutamine	0.971	Wilcoxon
lysine	0.999	t-test
phenylalanine	0.997	Wilcoxon
proline	0.999	t-test
serine	1.000	t-test
threonine	0.998	t-test
tryptophan	0.977	Wilcoxon
tyrosine	0.810	Wilcoxon

suggest no significant difference between the correlation of TOF absolute and TOF ratios against quant.

Similar comparisons were performed for six AAs that could be quantified by NMR (Table S1). Figure 4 shows the Spearman correlation plots for the comparisons of TOF absolute and TOF ratios to NMR for four exemplary AAs. The Spearman correlation coefficients of the six AAs are shown in

Table 4. Spearman correlation coefficients for “TOF absolute” vs. “NMR” and “TOF ratios” vs. “NMR”.

Metabolite	TOF Absolute	TOF Ratios
aspartate	0.456	0.333
lysine	0.669	0.625
phenylalanine	0.720	0.767
threonine	0.789	0.673
tyrosine	0.849	0.849
tryptophan	0.733	0.775

. The “diff” values were also calculated for each dataset. The results of paired Wilcoxon or t-tests are shown in

Table 5. Paired Wilcoxon or t-test results comparing “diff” values between “TOF absolute” and “TOF ratios” against “NMR”.

Metabolite	p-Value	Paired Test Type
aspartate	0.999	t-test
lysine	0.999	t-test
phenylalanine	0.977	Wilcoxon
threonine	0.861	Wilcoxon
tyrosine	0.996	t-test
tryptophan	0.998	t-test

. Similar to quant, no significant difference between “TOF absolute” and “TOF ratios” in the correlation is seen. Furthermore, Figure S1 shows Spearman correlation plots between “NMR” and “quant” for the six overlapping AAs.

Overall, a poor correlation was observed for aspartate and proline in all comparisons due to the low aspartate and proline concentrations in urine, see Figure S2. The correlation with the NMR data was overall lower, which can be attributed to difficulties in NMR data analysis due to overlapping signals.

To further investigate the differences between TOF absolute and TOF ratios, relative standard deviation (RSD) values of the QCs were considered in both datasets. Note that two urine samples served here as QCs and were analyzed repeatedly throughout the batch, resulting in 14 measurements each. The RSDs of the 112 features were calculated as TOF absolute and TOF ratios for QC1 and QC2 separately; then, in addition, the average RSD of both was calculated, resulting in a total of six lists of RSD values. Figure S3 shows histograms of the average RSD values for TOF absolute and TOF ratios. TOF absolute exhibits more features in the lower RSD range. Statistical analysis was performed to check whether the difference is significant. Shapiro–Wilk normality tests [32] for all six lists of RSD values show significant p-values (Table S3). Thus, paired Wilcoxon tests were conducted (Table S4). The resulting p-values together with the data presented in Figure S3 suggest that the RSD in TOF absolute is significantly lower than in TOF ratios. This can be caused by the erroneous automatic peak integration of the internal standard peaks, which will in turn lead to erroneous ratios. In general, abundant analytes with peaks way above noise level are less challenging for automatic peak integration algorithms, and vice versa. To further investigate that, the 112 features were divided into four quantiles (Q1–Q4) based on the peak area of the respective internal standards. Q1 contains the features with the largest IS peak areas, and Q4 has the lowest. Shapiro and paired Wilcoxon tests between the average RSD values in each quantile were conducted, see Table S5. Q1 shows a nonsignificant Wilcoxon p-value, suggesting that features with the abundant IS signal show no significant difference between absolute and ratios. Although Q2 already shows a significant p-value, it is higher than those of Q3 and Q4. Individual features were checked, and the three features with the highest RSD in TOF ratios were inspected. The first one shows integration errors despite having an abundant IS; this is justified by the isomer coeluting close to the IS peak, Figure S4. The second has a mismatch between the endogenous feature and its IS. The third has again a closely eluting compound interfering with the integration.

The ratio approach was also tested for its usability for batch-effect correction. For this, a subset of 56 samples randomly drawn from the original set of 244 measurements were denoted Batch I. The corresponding 56 samples were measured again, six months later, using the same methods (Batch 2). As a comparison, the batch effect–correction algorithm “removebatcheffect” (RBE) of the Limma package was tested [21].

Two feature lists were created for the comparison containing the data from both batches each. The first is the feature list yielded by the typical fingerprinting analysis workflow, containing all detectable features meeting the analysis parameters. It comprised 3291 unidentified features and is referred to as the “T–list”. To account for RT shifts between the two batches, a higher than usual RT tolerance (0.3 min instead of 0.2 min) was used for peak alignment in the T-List. The second list was a feature list of metabolites that were identified using the IROA kit. It comprised 115 metabolites showing the IROA isotopic pattern, which were detected in both batches, and their respective internal standards. The LTRS measurement in both batches was used to identify and align the metabolites. From the second list, two further lists were generated. The list comprising the absolute peak areas of the metabolite without considering their ISs is referred to as the IROA–Abs list. The list in which the peak areas of the metabolites were divided by those of their corresponding ISs is called IROA–ratio list. The PCA score plots of T–list and IROA–Abs list showed a pronounced shift between the two batches (Figure 5a,b). This can be clearly observed by considering the clustering of the respective QC samples. The closer the clusters of the same QC, the less pronounced are the batch effects, and vice versa. Leek et al. demonstrated that normalization alone does often not remove batch effects [33]. We applied the probabilistic quotient normalization (PQN) [34] followed by a Z–transformation for each batch separately in the T-list. This reduced the shift in the QC samples (Figure 5c). Using the IROA–ratio list also provided a considerable improvement over the IROA–Abs list, Figure 5d. However, applying PQN followed by RBE produced the best results, whether on the T-list or the IROA–Abs list, with data points from the two batches pairing well and QCs clustering more closely (Figure 5e,f).

The original and new LTRS measurements (from the creatinine concentration test) were both analyzed separately by CF4 with the same parameters. The new LTRS measurements yielded 417 complete bins (uncured), compared to 232 complete bins (uncured) for the original dataset (higher dilution of the LTRS).

In the results file from CF4, three types of relevant outputs are considered: Raw, i.e., the absolute ¹²C peak area, corresponds to “TOF absolute”; ¹²C/¹³C Ratios that correspond to “TOF ratios”; and Suppression Corrected, i.e., peak areas that are corrected for ion suppression, as the name states. To this end, in CF4, the ¹²C/¹³C ratio is multiplied by the corresponding least suppressed ¹³C area. Spearman correlation coefficients with results obtained by a targeted MS analysis (see Section 2 for details) were calculated for both 2 mM and 6.5 mM creatinine, employing all three outputs. The results are shown in Table S7. For comparison, the corresponding coefficients from analyzing the same samples in MZmine are given.

The ability to correct for batch effects was considered again by reanalyzing the same original two batch effect datasets using CF4. Four relevant outputs were considered, i.e., Raw, Ratios, Suppression Corrected, and Normalized (MSTUS). The resulting PCAs are shown in Figure S6.

4. Discussion

The data shows no significant improvement if TOF ratios are used instead of the TOF absolute when comparing the results to reference measurements. This poses the pivotal question of whether the use of an IS in untargeted LC–MS metabolomics carries any benefit. The answer lies within the nature of the metabolic fingerprinting approach itself. Trying to be as universal as possible makes one miss out on the perks of tailored methods that are designed to take full advantage of having an IS. This includes manual peak integration, or the manual inspection of automatically integrated peaks, which cannot be practically applied to metabolic fingerprinting.

A complex matrix such as urine that is further mixed with a complex labeled yeast extract poses a challenge for automated data analysis. One example is the case of aspartate. In both Figure 2 and Figure 4, aspartate shows a better correlation to quant without the IS. When inspecting the chromatographic peaks, we notice two main problems: there are samples where the ¹²C aspartate peak is so low in abundance that it falls within the noise, see Figure 6a. This explains the generally lower correlation even in “TOF absolute”. Furthermore, a peak with an m/z of 138.0549 elutes close to aspartate. The mass difference to ¹³C aspartate (m/z 138.0578) is below the 3 mDa tolerance used in data processing, hence resulting in an interference. The tentative identification of the neighboring peak suggests the ¹²C peak of trigonelline, a product of niacin (vitamin B3) metabolism, which is excreted in urine. The abundance of trigonelline varies across the samples. Figure 6b–d show that the larger the peak of trigonelline, the higher the interference with the ¹³C aspartate coming from the IS. While using a stricter m/z tolerance might solve this problem, it could cause problems in data processing for other compounds. This is in agreement with the work by Qiu et al. [35], who showed the advantage of higher resolution MS instrumentation for peak detection and identification.

Furthermore, employing a labeled cell extract as an internal standard will cause low abundant metabolites to have low signal intensities of the respective labeled compounds, resulting in a higher variance or difficulties in data integration.

In general, using a yeast extract for urine analysis is not optimal. We, however, did not address the other aspects of using the kit here, such as the range of covered metabolites, which would be readily affected by the matrix type. We, instead, focus on a few identified metabolites and investigate the use of a complex IS and its ability to improve the data correlation to a reference dataset. The lack of improvement when using ratios instead of absolute peak areas can also be explained by an analysis of stability. We did not observe any shifts or the deterioration of the HPLC–TOFMS performance when analyzing the 244 samples. This is exemplified by the PCA scores plot shown in Figure 7a. The QC samples, two urine samples analyzed 14 times each, cluster closely together.

Regarding the batch effects, the application of the kit helped to reduce them. Nevertheless, the IROA–ratio list might show even better results when comparing several batches measured over a prolonged time or even measured on different instruments. Additionally, since RBE works when there are distinct batches, IROA ratios would be more beneficial if samples are measured in a single batch over a long period of time and a gradual decline in instrument performance occurs throughout the batch. This, however, is beyond the scope of this study.

5. Conclusions

Using a complex ¹³C labeled yeast extract enabled the analysis of 112 metabolites using the ratio approach. While the correlation with reference data did not improve significantly, the ratio approach helped to reduce the batch effects. However, most features in the data set are not covered by the approach due to a missing internal standard signal, which is also due to the mismatch between the sample (urine) and internal standard matrix (yeast). Moreover, complications arising from simultaneously handling a vast number of features in the untargeted approach undermine the benefits of incorporating this IS into our workflow. Matrix interference; co-eluting isobaric compounds; and imprecisions in peak integration, -alignment, and -gap–filling are all potential factors. Their impact can be mitigated by means of improved chromatography, cutting-edge MS instrumentation, upgraded processing software, and tailored post processing algorithms.