M. Kathleen Alam

Chemometric Analysis of NMR Spectroscopy Data: A Review

Abstract The application of chemometric or multivariate analysis techniques to nuclear magnetic resonance (NMR) spectroscopic data is reviewed. Descriptions of the different processing and data manipulation procedures being utilized to produce reproducible input data sets for chemometric analysis are discussed. A brief review of some of the standard supervised and unsupervised chemometric classification methods as applied to NMR data is presented. The application of spectral resolution algorithms in the decomposition of NMR data to obtain pure component spectra and concentrations is also described. Specific examples of the use of chemometrics in NMR for a wide range of different fields are presented. The limitations, advantages and future directions of chemometric analysis in NMR are also discussed.

1H NMR-based metabonomic investigation of tributyl phosphate exposure in rats

Tributyl phosphate (TBP) is a toxic organophosphorous compound widely used in many industrial applications, including significant usage in nuclear processing. The industrial application of this chemical is responsible for occupational exposure and environmental pollution. In this study, (1)H NMR-based metabonomics has been applied to investigate the metabolic response to TBP exposure. Male Sprague-Dawley rats were given a TBP-dose of 15 mg/kg body weight, followed by 24h urine collection, as was previously demonstrated for finding most of the intermediates of TBP. High-resolution (1)H NMR spectroscopy of urine samples in conjunction with statistical pattern recognition and compound identification allowed for the metabolic changes associated with TBP treatment to be identified. Discerning NMR spectral regions corresponding to three TBP metabolites, dibutyl phosphate (DBP), N-acetyl-(S-3-hydroxybutyl)-L-cysteine and N-acetyl-(S-3-oxobutyl)-L-cysteine, were identified in TBP-treated rats. In addition, the (1)H NMR spectra revealed TBP-induced variations of endogenous urinary metabolites including benzoate, urea, and trigonelline along with metabolites involved in the Krebs cycle including citrate, cis-aconitate, trans-aconitate, 2-oxoglutarate, succinate, and fumarate. These findings indicate that TBP induces a disturbance to the Krebs cycle energy metabolism and provides a biomarker signature of TBP exposure. We show that three metabolites of TBP, dibutylphosphate, N-acetyl-(S-3-hydroxybutyl)-L-cysteine and N-acetyl-(S-3-oxobutyl)-L-cysteine, which are not present in the control groups, are the most important factors in separating the TBP and control groups (p<0.0023), while the endogenous compounds 2-oxoglutarate, benzoate, fumarate, trigonelline, and cis-aconetate were also important (p<0.01).

Characterization of pH variation in lysed blood by near-infrared spectroscopy

Near-infrared spectra (1300–2500 nm) collected from lysed blood solutions were shown to correlate with the pH of the solutions measured potentiometrically. Cross-validated partial least-squares (PLS) models were developed from these spectral data, which provided standard error of prediction (SEP) values below 0.05 pH units for a pH range of 1.0 (6.8–7.8). Experiments were designed to eliminate possible correlation between pH and other components in the blood in order to ensure that variations in the spectral data correlated to pH were due to hydrogen ion changes only. Further work was performed to discern the primary source of pH information in the lysed blood spectra by using spectra collected from plasma and histidine solutions. The blood, plasma, and histidine data sets were compared with the use of loading vectors from principal component analysis (PCA). These loading vectors show that variations in the spectra of the titrated amino acid histidine mimic those seen in lysed blood, but not those seen in plasma. These results suggest that histidine residues of hemoglobin are providing the spectral variation necessary for pH modeling in the lysed blood solutions. It is further shown that the observed pH-sensitive histidine bands do not arise from the exchangeable proton on the imidazole ring of histidine; rather they arise from the variation in the C–H bonds of the C2 and/or the C4 carbons of the imidazole ring as they are influenced by the titration of the nitrogen-bound proton of the imidazole ring.

Measurement of pH in Whole Blood by Near-Infrared Spectroscopy

Whole blood pH has been determined in vitro by using near-infrared spectroscopy over the wavelength range of 1500 to 1785 nm with multivariate calibration modeling of the spectral data obtained from two different sample sets. In the first sample set, the pH of whole blood was varied without controlling cell size and oxygen saturation (O2 Sat) variation. The result was that the red blood cell (RBC) size and O2 Sat correlated with pH. Although the partial least-squares (PLS) multivariate calibration of these data produced a good pH prediction cross-validation standard error of prediction (CVSEP) = 0.046, R2 = 0.982, the spectral data were dominated by scattering changes due to changing RBC size that correlated with the pH changes. A second experiment was carried out where the RBC size and O2 Sat were varied orthogonally to the pH variation. A PLS calibration of the spectral data obtained from these samples produced a pH prediction with an R2 of 0.954 and a cross-validated standard error of prediction of 0.064 pH units. The robustness of the PLS calibration models was tested by predicting the data obtained from the other sets. The predicted pH values obtained from both data sets yielded R2 values greater than 0.9 once the data were corrected for differences in hemoglobin concentration. For example, with the use of the calibration produced from the second sample set, the pH values from the first sample set were predicted with an R2 of 0.92 after the predictions were corrected for bias and slope. It is shown that spectral information specific to pH-induced chemical changes in the hemoglobin molecule is contained within the PLS loading vectors developed for both the first and second data sets. It is this pH specific information that allows the spectra dominated by pH-correlated scattering changes to provide robust pH predictive ability in the uncorrelated data, and visa versa.

Accurate wavelength measurements of a putative standard for near-infrared diffuse reflection spectrometry.

The diffuse reflection (DR) spectrum of a sample consisting of a mixture of rare earth oxides and talc was measured at 2 cm−1 resolution, using five different accessories installed on five different Fourier transform near-infrared (FT-NIR) spectrometers from four manufacturers. Peak positions for 37 peaks were determined using two peak-picking algorithms: center-of-mass and polynomial fitting. The wavenumber of the band center reported by either of these techniques was sensitive to the slope of the baseline, and so the baseline of the spectra was corrected using either a polynomial fit or conversion to the second derivative. Significantly different results were obtained with one combination of spectrometer and accessory than the others. Apparently, the beam path through the interferometer and DR accessory was different for this accessory than for any of the other measurements, causing a severe degradation of the resolution. Spectra measured on this instrument were removed as outliers. For measurements made on FT-NIR spectrometers, it is shown that it is important to check the resolution at which the spectrum has been measured using lines in the vibration-rotation spectrum of atmospheric water vapor and to specify the peak-picking and baseline-correction algorithms that are used to process the measured spectra. The variance between the results given by the four different methods of peak-picking and baseline correction was substantially larger than the variance between the remaining five measurements. Certain bands were found to be more suitable than others for use as wavelength standards. A band at 5943.13 cm−1 (1682.62 nm) was found to be the most stable band between the four methods and the six measurements. A band at 5177.04 cm−1 (1931.61 nm) has the highest precision between different measurements when polynomial baseline correction and polynomial peak-picking algorithms are used.

1H Nuclear Magnetic Resonance (NMR) Metabolomic Study of Chronic Organophosphate Exposure in Rats

1H NMR spectroscopy and chemometric analysis were used to characterize rat urine obtained after chronic exposure to either tributyl phosphate (TBP) or triphenyl phosphate (TPP). In this study, the daily dose exposure was 1.5 mg/kg body weight for TBP, or 2.0 mg/kg body weight for TPP, administered over a 15-week period. Orthogonal signal correction (OSC) -filtered partial least square discriminant analysis (OSC-PLSDA) was used to predict and classify exposure to these organophosphates. During the development of the model, the classification error was evaluated as a function of the number of latent variables. NMR spectral regions and corresponding metabolites important for determination of exposure type were identified using variable importance in projection (VIP) coefficients obtained from the OSC-PLSDA analysis. As expected, the model for classification of chronic (1.5–2.0 mg/kg body weight daily) TBP or TPP exposure was not as strong as the previously reported model developed for identifying acute (15–20 mg/kg body weight) exposure. The set of majorly impacted metabolites identified for chronic TBP or TPP exposure was slightly different than those metabolites previously identified for acute exposure. These metabolites were then mapped to different metabolite pathways and ranked, allowing the metabolic response to chronic organophosphate exposure to be addressed.

Multivariate analysis and quantitation of 17O-nuclear magnetic resonance in primary alcohol mixtures

Multivariate techniques were used to address the quantification of (17)O-nuclear magnetic resonance (NMR) spectra for a series of primary alcohol mixtures. Due to highly overlapping resonances, quantitative spectral evaluation using standard integration and deconvolution techniques proved difficult. Multivariate evaluation of the (17)O-NMR spectral data obtained for 26 mixtures of five primary alcohols demonstrated that obtaining information about spectral overlap and interferences allowed the development of more accurate models. Initial partial least squares (PLS) models developed for the (17)O-NMR data collected from the primary alcohol mixtures resulted in very poor precision, with signal overlap between the different chemical species suspected of being the primary contributor to the error. To directly evaluate the question of spectral overlap in these alcohol mixtures, net analyte signal (NAS) analyses were performed. The NAS results indicate that alcohols with similar chain lengths produced severely overlapping (17)O-NMR resonances. Grouping the alcohols based on chain length allowed more accurate and robust calibration models to be developed.

Near-infrared spectroscopy of lysed blood: pH effects

Recent investigations by our group have demonstrated that near-infrared spectra collected from lysed blood solutions can be used to create clinically useful partial least squares (PLS) models for pH with standard errors of prediction below 0.05 pH units for a pH range of 1 (6.8 to 7.8). Further work was performed in order to discern the primary source of pH information in the spectra. Results from these experiments are presented using spectral data acquired over the spectral range of 1300 nm to 2500 nm from plasma, lysed blood and amino acids solutions. Data were analyzed by principal component analysis (PCA) and loading vectors were compared. Experiments were designed to eliminate possible correlation between pH and other components in the system in order to ensure variations in the spectral data were due to hydrogen ion changes only. Results indicate that variations in the spectral characteristics of histidine mimic those seen in lysed blood, but not those seen in plasma, suggesting that histidine residues from hemoglobin are providing the necessary variation for pH modeling in the lysed blood solutions.

Crystallization behavior of vapor-deposited hexanitroazobenzene (HNAB) films

Vapor-deposited hexanitroazobenzene (HNAB) has been shown to form an amorphous structure as-deposited that crystallizes over a period ranging from several hours to several weeks, depending on the ambient temperature. Raman spectroscopy and x-ray diffraction were used to identify three distinct phases during the crystallization process: the as-deposited amorphous structure, the HNAB-II crystal structure, and an as-yet unidentified crystal structure. Significant qualitative differences in the nucleation and growth of the crystalline phases were observed between 65°C and 75°C. While the same two polymorphs form in all cases, significant variation in the quantities of each phase was observed as a function of temperature.

Hemoglobin Correction for Near-Infrared pH Determination in Lysed Blood Solutions

The near-infrared (NIR) measurement of blood pH relies on the spectral signature of histidine residing on the hemoglobin molecule. If the amount of hemoglobin in solution varies, the size of the histidine signal can vary depending on changes in either the pH or hemoglobin concentration. Multivariate calibration models developed using the NIR spectra collected from blood at a single hemoglobin concentration are shown to predict data from different hemoglobin levels with a bias and slope. A simple, scalar path length correction of the spectral data does not correct this problem. However, global partial least-square (PLS) models built with data encompassing a range of hemoglobin concentration have a cross-validated standard error of prediction (CVSEP) similar to the CVSEP of data obtained from a single hemoglobin level. It will be shown that the prediction of pH of an unknown sample using a global PLS model requires that the unknown have a hemoglobin concentration falling within the range encompassed by the global model. An alternative method for correcting the predicted pH for hemoglobin levels is also presented. The alternative method updates the single-hemoglobin-level models with slope and intercept estimates from the pH predictions of data collected at alternate hemoglobin levels. The slope and intercept correction method gave SEP values averaging to 0.034 pH units. Since both methods require some knowledge of the hemoglobin concentration in order for a pH prediction to be made, a model for hemoglobin concentration is developed using spectral data and is used for pH correction.