Comprehensive utilization of NMR spectra–evaluation of NMR-based quantification of amino acids for research and patient care

Subjects

Ninety EDTA-plasma samples were collected from left over material of out-patients of a specialized clinic for metabolic diseases at the University Medicine Greifswald in July 2020 (Ethic votum: MN: BB 100/15).

Methods

All samples were subject to two freeze-thaw-cycles since the frozen original samples had to be aliquoted into two sample sets. One sample set was analysed at the University Medicine of Greifswald (¹H NMR) and one at the University Medicine of Goettingen (HPLC). A panel of 15 AAs (alanine, asparagine, glutamic acid, glutamine, glycine, histidine, isoleucine, leucine, lysine, methionine, ornithine, phenylalanine, threonine, tyrosine and valine) was measured on both platforms.

Samples were stored at −80 °C until analysis. All measurements were performed in duplicates.

¹H NMR measurements

After thawing, 300 μL of plasma were mixed with 300 μL of phosphate buffer [prepared with D2O and contained sodium 3-trimethylsilyl-(2,2,3,3-D4)-1-propionate (TSP) as reference, (pH 7.4)]. Spectra were recorded on a Bruker Avance IVDr system (AVANCE-III 600 NMR spectrometer operated by TOPSPIN 3.6 software (both Bruker Biospin, Ettlingen, Germany) and a 5-mm BBI probe, equipped with a z-gradient and an automated tuning and matching (ATMA) unit (Bruker Biospin, Ettlingen, Germany)). Specimens were automatically delivered to the spectrometer via Sample Jet (Bruker Biospin, Ettlingen, Germany) in standard 5 mm NMR tubes. The settings used correspond to those from the Bruker’s bodyfluid NMR methods package (B.I.Methods2.0, Bruker Biospin, Ettlingen, Germany). The acquisition temperature was set to 310K. A standard one-dimensional ¹H NMR pulse sequence with suppression of the water peak (Bruker pulse sequence “noesygppr1d”; NOESY, Nuclear Overhauser and Exchange Spectroscopy) was used. The pulse sequence follows the scheme: RD–gz1–90°–t–90°–tm–gz2–90°–AQ, where RD is the relaxation delay (4 s), 90° denotes a hard 90° RF pulse, t is a short delay (4 µs), tm is the mixing time (10 m s), gz1 and gz2 are magnetic field gradients along the z-axis applied for 1 m s each, and AQ is the acquisition period (2.7 s), during which 98,304 data points are collected with a spectral width (SW) from −10 to 20 ppm.

A low power pulse is applied at the water resonance frequency to saturate the solvent signal during RD and tm. The receiver gain is set at 90.5 for all experiments. For pre-processing, a line broadening of 0.3 Hz, a zero filling to produce 128 k data points and zero order phase correction were applied. Automated quantification of small molecules was performed on the spectra using the Bruker IVDr Quantification in Plasma/Serum Analysis (B.I.Quant-PS2.0, Bruker Biospin, Ettlingen, Germany). The results obtained include, among other measurands, the substance group of AAs. Figure 1 visualizes an example for automated spectra evaluation.

Figure 1:

¹H NMR signals for alanine, lysine, ornithine and histidine. Black line, original signal; blue line, calculated fit signal; yellow line, baseline. Blue area, area used for quantification; red area, residue (area not used for quantification); e.g. from another signal overlap. Figures modified from Bruker IVDr quantification in plasma/serum B.I.Quant-PS2.0 analysis report.

Statistical analysis

Medians and interquartile ranges (IR: 1st–3rd quartile) were calculated per method and measurand. Duplicate measurements of each method were evaluated by calculation of coefficients of variation (CV in %) according to [13]. All other analyses were conducted with the first of the duplicate measurements each in order to reflect the clinical routine, where only one measurement is available. For comparison of ¹H NMR and HPLC measurements, Spearman correlation coefficients were calculated and Passing-Bablok regression was carried out to assess constant differences (regression line intercept) and proportional differences (regression line slope) between the methods. Since concentration levels vary considerably between the different measurands, intercepts were expressed in relation to the median HPLC-measured value in order to assure comparability across the measurands (normalized intercept=100*intercept/medianHPLC [%]). For illustration, scatterplots with the according regression lines are plotted for selected AAs along with reference intervals according to Blau et al. [14]. Statistical analysis was performed using R version 4.3.2 (R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.).

Imprecision

Imprecision was higher in ¹H NMR compared to HPLC measurements whereby the CVs of the HPLC method stayed well below 4 % for all investigated AAs. For NMR, CVs vary considerably among the investigated AA. While alanine, glycine and tyrosine showed low CVs, which would also be acceptable in clinical routine, histidine and threonine would not be usable in a clinical context due to their very high CVs. With respect to the latter, the high imprecision might be due to signal overlaps e.g. a lactate signal overlap with threonine [15] or of creatine as well as creatinine with lysine, see Figure 1. In case of ornithine not only signal overlap, but also a high number of rather small peaks complicates the quantification. Similar effects have been described by other authors [1], 12], 15]. Possible ways to improve imprecision and sensitivity may be application of additional NMR experiments such as 2D (JRES) or increasing magnetic field strengths and computational analysis after data collection [1], 2].

Wolak-Dinsmore et al. [16] reported CVs for inter-assay and intra-assay imprecision for valine, leucine and isoleucine, measuring three plasma pools by ¹H NMR and using a newly developed deconvolution algorithm for the quantification of branched chain AAs (BCAA, i.e. valine, leucine and isoleucine) in plasma or serum. Jimenez et al. [15] reported CVs obtained in a multi-laboratory trial using 37 serum and 37 plasma commercially sourced samples analyzed in 11 600 MHz spectrometers. In another study, AAs were detected in honey, using a 500 MHz NMR platform [17]. Imprecision was assessed by measuring a standard solution containing the principal compounds of honey. All studies revealed CVs partly comparable to our values, depending on the material and quantification methods used. The authors argue [16], that some CVs are fairly high due to high concentrations of large proteins and lipoproteins which limits sensitivity of small molecules in low concentrations.

Method comparison

The method comparison of AA concentrations between ¹H NMR and HPLC revealed highest correlation coefficients for phenylalanine (r=0.99), tyrosine (r=0.97) and alanine (r=0.90). Glycine (r=0.86) and valine (r=0.81) showed good correlation coefficients. Among these, phenylalanine and alanine also showed a low imprecision and only small bias to the HPLC measurements, suggesting their quantitative use in research and making them promising candidates for use as measurands in patient care. As an example, when applying a critical phenylalanine cut-off used in the newborn screening program to detect PKU, the concordance between the two analytical methods in classifying PKU was exceptionally high. Of the 177 samples for which phenylalanine concentrations from both methods were available, only two discordant classifications were observed when using a threshold of 115 μmol/L, which is the cut-off employed in our laboratory for routine newborn screening. It needs to be clearly pointed out, that for the use in patient care a CE-mark needs to be obtained, which is a costly and time-consuming process.

The remaining above-mentioned AAs, with a good correlation between ¹H NMR and HPLC results, showed systematic bias. A systematic bias often occurs in different methods and can be handled in both settings, patient care and research, if well described. Consequently, these AAs are also candidates for measurements in research and patient care.

Correlation coefficients were small (r ≤ 0.8) mostly for those AAs displaying a high imprecision in NMR, namely ornithine, histidine, isoleucine, leucine, lysine and threonine. High imprecision, in this case of the NMR based results, negatively impacts the concordance between NMR and HPLC.

Still, those methods with high imprecision may deliver additional information in terms of qualitative results when applied as add-on in the course of NMR measurements for other purposes. For example, in epidemiological approaches with large numbers of probands, qualitative results like “negative and positive” or “low and high” may be used to distinguish between patients and healthy controls.

We therefore suggest that current NMR quantification algorithms for AAs may be used for:

1. quantitative use in research and candidate for quantitative use in patient care (alanine and phenylalanine)

2. quantitative use in research and candidate for quantitative use in patient care if the systematic bias is taken into account (tyrosine, valine, glycine)

3. qualitative results in epidemiological studies (ornithine, threonine, histidine, isoleucine, leucine, lysine)

4. not evaluable (glutamine, glutamic acid, asparagine, methionine)

To our knowledge, very few studies on the comparison of NMR-measured AA levels with HPLC-measured levels exist, none of these used human plasmas. In Nord et al. (18), two quantification methods for NMR-measured AA levels in beer, based on signal integration and on partial least-squares (PLS) regression, respectively, are described. The PLS method was applied to 19 AAs and the signal integration method was applied for five AAs. NMR based AA levels were compared to HPLC based results in 58 samples. Depending on the quantification method and the considered measurand, the quantification models showed varying quality: all AAs quantified by signal integration and eight AAs quantified by the PLS model showed high comparability between NMR and HPLC results (squared correlation coefficient>0.95 and regression line slope close to 1), whereas quality of PLS prediction models were lower for the remaining AAs. Possible explanations for the lower quality were lower concentration levels and signal overlap for some of the measurands. Correlation coefficients were partly comparable to the values found in our study (phenylalanine and tyrosine). Deviations in the results might be explained with the differences in the quantification methods and the material (beer vs. plasma). Also, in case of the signal integration method, the considered measurands were chosen such that their signals were lying in regions with few other signals in the beer spectrum, which might explain better correlation for some of the measurands than in our study.

The authors [18] assume a limit for quantification of AAs with the described methods of 5–10 mg/L (=0.03–0.06 mmol/L for phenylalanine, e.g.), which is mostly comparable to our values, see Supplementary Table S1. Interestingly, a decrease of detection limit by a factor of 10 was presumed by enhancement of technology such as higher field strength, cryoprobe technology or larger number of scans. Further comparison studies in human plasma from both, healthy and diseased probands, are needed to improve the data base.

Multiple studies show the potential of AAs to serve as biomarkers in prognosis and diagnosis of various diseases, e.g. diabetes [6], 19], cancer [5], 20] or cardiovascular disease [21], [22], [23]. NMR-platforms are increasingly used in metabolic profiling studies due to their inherent advantages like robust and reproducible measurements [1], 12], also revealing insights in the role of AAs in the aforementioned clinical outcomes [16], [24], [25], [26], [27], [28]. Our results strongly support the use of NMR-based measurements of AAs, not only in research, but some of the investigated AAs may be regarded as candidates for measurements in patient care. Of course, they still need to undergo the CE-marking process prior to clinical use. Improvement of knowledge of NMR-based measurement and improvement of NMR-technology, e.g. by including more peaks per measurand, adding further experiments next to NOESY or increasing field strength, may prepare the ground for a wider use of AAs in disease diagnosis and prognosis. However, it is important to point out that results should be handled with care since the analytical quality varies considerably among the investigated AAs. The above stated categories should be taken into account when using NMR quantification for scientific purposes.

Strengths and limitations

To our knowledge this is the first independent study based on human plasma samples from patient care to compare NMR AA measurements to an established patient care HPLC method. For AA analysis recommended LC-MS/MS was not available for this study. The used number of 90 sample may not be sufficient to cover all important metabolomic states that may interfere with reliable quantification of AAs. Samples were limited to a specialized clinic covering only a small number of diseases. So far, the manufacturer does not provide automated reports for all desirable AA and therefore e.g. citrulline, and arginine are missing. The report for glutamine, glutamic acid, asparagine and methionine did not produce enough results above the LOD. Different from classical laboratory medicine measurement procedures, calibration relies on one substance only (TSP) and controls for whole spectra are not available. Regarding NMR experiments, only the NOESY experiment was performed. Further experiments allowing for longer acquisition time and more data points were not conducted. Independent quality controls for each specific measurand were not available and therefore could not be included. Furthermore, the study relies on one HPLC and NMR device only and should be repeated using multiple devices for each method. While NMR is already quite often used in research contexts, it is so far not suitable for routine and especially not in emergency settings. While the short sample preparation and measurement of approximately 20–30 min represent an advantage, high investment costs and the considerably low through-put represent draw-backs that still need to be overcome. In the analysis, we used Spearman correlation (as most of the data was not normally distributed), which may level out relevant differences in the concentrations.