CA2278870A1 - Multianalyte serum assays from mid-ir spectra of dry films on glass slides - Google Patents

Description

Multianalyte Serum Assays from mid-IR Spectra of Dry Films on Glass Slides R. Anthony Shaw* and Henry H. Mantsch Institute for Biodiagnostics, National Research Council of Canada, 435 Ellice Avenue, Winnipeg, Manitoba, Canada R3B 1Y6 Corresponding author: (e-mail) Anthony.Shaw@nrc.ca (fax) 204-984-5472.
(phone) 204-984-4626.
Running title: Serum analysis using glass slides

2 An analytical method based upon mid-infrared spectroscopy is proposed, and the advantages of this approach are discussed. The method involves drying a liquid specimen to a film, and deriving analyte levels from the infrared spectrum of that film. The specific aim of this study was to determine whether glass might serve as a suitable substrate for the simultaneous determination of several analytes in complex mixtures. Using human serum as a 'proof-of-concept' example, we show here that six commonly measured analytes may be determined from spectra originally measured employing barium fluoride substrates, but restricting the analytical models to absorptions within the region 2000-4000 cm's - i.e.
making use of only those absorptions that are accessible with glass substrates. Using partial least squares calibration models, it is shown that albumin, cholesterol, glucose, total protein, triglycerides, and urea, may be determined with standard errors that approach or meet the criteria required for routine clinical analysis. The practical advantages of such an approach are discussed.
Index headings: Analysis, Partial least squares regression, Dried films, Calibration, Spectral normalization, Automation, Serum

3 Serum and blood analyses are by far the most common tests carried out in the clinical laboratory. Not surprisingly, all of the major dissolved species are routinely assayed. For example, serum or blood glucose testing is essential to the diagnosis and monitoring of diabetes; serum cholesterol and triglycerides, together with lipoprotein assays, provide sensitive indicators of the risk of coronary disease; urea and creatinine are commonly tested to monitor kidney function, and depressed serum protein or albumin levels are associated with liver and renal conditions.'~2 Taken together, analyses for albumin, cholesterol, glucose, total protein, triglycerides, and urea account for thousands of individual tests in a typical clinical chemistry laboratory, with comparable numbers carried out as point-of-care tests, primarily for glucose, in the hospital environment. We have demonstrated previously that all of these species can be assayed on the basis of the infrared (IR) spectra of dried serum films.3 The drying step eliminates the single most difficult obstacle to routine quantitative assays using mid-IR
spectroscopy of aqueous solutions, namely the overwhelming absorptions of water itself. This general approach combines the following potential advantages:
no reagents are required, all six analyses are derived simultaneously from a single IR spectrum, very little sample is required (the previous exploratory study used 7 ~L serum aliquots), no dilution is necessary for any of these analyses, and the method is suited for automation. Not surprisingly, near-IR
spectroscopy has also been employed as the basis for serum analysis,4-6 although these investigations and other mid- and near-IR studies of whole blood'-9 have been

4 motivated largely by the prospect of a near-IR method for the in-vivo measurement of blood glucose levels.'o."
The original study basing serum analysis upon mid-infrared transmission spectroscopy of serum films made use of infrared-transparent barium fluoride windows.3 Although this general approach to analysis proved sound, the requirement of expensive barium fluoride windows diminishes the prospects for widespread adoption of the method in a clinical setting. For that reason, we have recently begun to explore the possibility of using alternative substrates that are generally not considered to be suitable for mid-IR transmission spectroscopy.
One candidate is ordinary glass.
Glass is generally considered to be among the worst possible choices as a window material for mid-IR transmission spectroscopy. Since glass is opaque at wavenumbers below 2000 crri', any spectral information in the fingerprint region is inaccessible. The spectral signatures above 2000 crri' that remain available - i.e. the C-H, N-H, and O-H stretching transitions - must therefore serve as the basis for analysis.
The aim of the present study was to assess the potential for using transmission spectroscopy of serum films, dried onto glass substrates, as the basis for quantifying serum albumin, cholesterol, glucose, total protein, triglycerides, and urea. Because the IR spectrum of the film itself is largely unaffected by the substrate (over spectral regions where the substrate is transparent!), we have explored the suitability of glass in this application by reanalyzing spectra acquired previously using BaF2 windows. In particular, we have derived new quantitation models based upon the mid-infrared spectra recorded earlier,3 but restricting the spectral range for analytical calibration models to the region above 2000 crri ~.
MATERIALS AND METHODS
IR spectroscopy. The experimental details have been reported previously.3 Serum specimens were diluted to a 1:1 mixture with 4 g/L aqueous potassium thiocyanate, which provided an absorption at 2060 crri ~ that served as the basis to normalize the spectra. Seven ~L of this mixture was spread evenly on a circular BaF2 window (13 mm diameter x 2 mm thickness) and dried under moderate vacuum to leave a glassy film. Single beam spectra of these films (256 scans, 4 cm-' resolution, MCT detector) were acquired using a Bio-Rad FTS-40A
Fourier transform IR spectrometer (Bio-Rad Laboratories, Cambridge, MA, USA
02319) and were ratioed against the single beam spectrum of a clean BaF2 window and converted to an absorbance scale. Spectra were acquired for 300 serum specimens, each of which was also analyzed for albumin, cholesterol, glucose, total protein, triglycerides, and urea using standard clinical chemistry methods. Illustrative spectra of albumin, glucose, and urea were measured for powder samples using a Split-Pea ATR accessory (Harrick Scientific, Ossining, N.Y.), while spectra of cholesterol and tripalmitin were recorded for dilute solutions in carbon tetrachloride.

Calibration model development. Partial least squares (PLS) calibration models'2 were developed separately for each of the six analytes, employing only spectral regions lying in the transmission window of glass (i.e. above 2000 crri').
The final models were based upon PLS trials using the average of the two replicate spectra for each specimen. Spectra were acquired for over 300 specimens; approximately 2/3 of these were employed to calibrate the PLS
model (training set) and the remaining 1/3 made up the test set to assess the predictive accuracy of the models (the precise number of spectra in each set varies, since not all reference assays were available for all specimens).
Preprocessing included normalization to the thiocyanate absorption at 2060 crri', and derivation (Savitzky-Golay), with trials using unnormalized spectra included here for comparison. All spectral manipulations and PLS trials were carried out using GRAMS/32 software (Galactic Industries, Salem, NH, USA 03079), which includes tests to identify statistical and spectral outliers. Typically four to eight spectra (of 300) were removed in the development of calibration models, most commonly due to large spectral residual residuals in the PLS model.
RESULTS
The IR absorption spectra of the five target species in the range 2700-3700 crri' are plotted in Figure 1. These spectra illustrate that the absorption patterns arising from the X-H stretching vibrations alone are clearly sufficient to distinguish among the analytes of interest. For example, tripalmitin (a serum triglyceride) yields prominent absorptions at 2850 and 2920 crri' arising from the symmetric and asymmetric stretching vibrations of the 15 acyl chain groups. The spectrum of cholesterol shows a rich array of CH stretching absorptions that reflect the diversity of CH oscillators. For albumin, the spectrum is dominated by strong, broad absorption centered at 3200 cm' arising from the amide A (NH stretching) vibration, while the absorption profile for urea comprises two major features due to the symmetric and asymmetric NH2 stretching modes.
Finally, the glucose spectrum is dominated by the absorption bands arising from OH stretching vibrations, with weaker absorptions below 3000 crri ~ from the methyne CH stretching modes.
Table I lists the normal concentration ranges for albumin, cholesterol, glucose, total protein, triglycerides, and urea in human serum. Protein is by far the most abundant dissolved species, and the IR spectrum of serum (Figure 2) is dominated by protein absorptions, with other species contributing bands that are superimposed upon the protein absorption profile. The same Figure illustrates a finding that is not unexpected; the IR spectra measured for films dried on glass are equivalent to those measured on BaF2 substrates over the common transmission window.
Although the IR spectra of serum specimens are superficially very similar to one another as a result of the high protein content, it emerged that the six target assays could be carried out on the basis of the serum IR spectra. To illustrate this, Figure 3 compares the IR-predicted (test set) analyte levels to the reference assays for albumin, cholesterol, glucose, total protein, triglycerides, and urea. Linear regression of the IR-based vs. reference assays yielded regression lines with slopes typically 0.9-1.0, and intercepts that demonstrated little or no bias in the IR-based analytical results (Table II). The number of PLS
factors was set at the point where the inclusion of additional factors produced no substantial improvement in the accuracy for the test set. This gauge is illustrated by Figure 4, which shows the trends in the RMS errors for the training and test sets as the number of PLS factors was increased in the triglyceride calibration model. The model was limited to 5 factors, since there was no substantial improvement in the predictive accuracy with inclusion of additional factors.
The PLS models (spectral range employed and number of PLS factors) are summarized in Table III.
Because the samples were prepared manually, some imprecision in sample preparation was inevitable. To partly correct for that influence, the spectra were originally acquired for specimens that were diluted with aqueous KSCN. The spectra were then normalized to yield a common intensity for the 2060 crri' absorption of the SCN- ion. To demonstrate the practical benefit of this procedure, we have derived a second set of PLS analytical models based upon the unnormalized spectra; a summary of these models is included in Tables II and III, and scatterplots comparing IR-predicted assays to reference analyses appear in Figure 5. In all but one case, the normalization procedure yielded PLS
models at least 25% more accurate (as gauged by the SEP) than the corresponding models based upon unnormalized spectra.

DISCUSSION
Infrared spectroscopy has emerged as the method of choice in a wide variety of analytical applications, and both near- and mid-IR spectroscopy are finding use in a rapidly expanding range of both clinical/analytical and diagnostic applications.'3,'a Although mid-IR spectroscopy is generally recognized to provide a better molecular 'fingerprint' than near-IR, the practical advantages of near-IR spectroscopy have been perceived as overwhelmingly favorable for analytical work. The two primary advantages of near-IR over mid-IR
spectroscopy are that large samples are easily handled (optical pathlengths are generally measured in millimeters or centimeters, compared to microns in the mid-IR) and that remote sampling is possible using fiber optics.
Mid-IR spectroscopy has potential advantages over near-IR spectroscopy in certain analytical situations. One clear advantage is that analyses are feasible for very small sample sizes (microliters). Another is that the relatively narrow bandwidths and comparatively rich mid-IR absorption spectrum may provide for more accurate analysis than is feasible using near-IR. One hindrance to the widespread adoption of mid-IR spectroscopy as an analytical method has been the perception that specialized, expensive window materials are required. The present study illustrates that the transmission window of ordinary glass is wide enough that it may be generally useful as a substrate in analytical mid-IR
applications; in particular, the absorption bands that are observable (CH, NH, and OH stretching vibrations) form patterns that are rich enough to discriminate among at least five major serum constituents, and hence to form the basis for their quantitation. As a general observation, it is worth noting that the majority of analytical methods that are based upon infrared spectroscopy make use of only limited spectral windows, whether those applications are based upon near-IR or mid-IR spectroscopy. As a practical matter, what is required of the substrate supporting the sample is not that it be completely transparent throughout the mid-(or near-) IR, but rather that the spectral window encompasses the analytically relevant absorptions.
The case of serum highlights the fact that the XH stretching region can serve as the basis to simultaneously quantitate several analytes. This is probably the most surprising finding, since the 'fingerprint' region 800-1800 crri' is employed so commonly as the primary basis to distinguish organic species.
Routine, multianalyte analysis of complex liquids may therefore be contemplated using this general approach.
In addition to providing a means to analyze very small liquid volumes, and capitalizing upon the relatively narrow bandwidths, analysis based upon dried films offers certain other practical benefits. One potential advantage is that the dried films, following analysis, may be archived for subsequent reanalysis if necessary. The fact that the analysis is based upon dried films also raises the possibility of preparing the films at remote locations and transporting the films rather than the entire specimen to the laboratory for analysis. The effective optical pathlength may also be varied by adjusting the amount of material that is dried to form the film, permitting convenient analytical methods for both concentrated and dilute liquid specimens. For example, we have developed IR-based assays using films dried from 7-15 pL of (diluted) serum,3 and employed approximately the same sample volume (12 pL) for the analysis of urine urea, creatinine, and total protein.'S Amniotic fluid, in contrast, requires dried aliquots of approximately 35 p,L in order to provide spectra with a suitable absorbance range.'6." Once the optimal sample volume has been determined for a particular application, an automated liquid-handling system might readily be programmed to sample and dispense the appropriate volume, depending on the analysis at hand.
Finally, the proof-of-concept analytical methods presented here are almost certainly not optimized. Adjustment of the sample volume may also lead to improved performance of the method. Development of a practical clinical analytical method would include automation of the sample preparation - a practical necessity that will undoubtedly improve the precision (and accuracy) of the assays described here.
CONCLUSION
We have demonstrated that several serum analytes may be determined simultaneously from IR spectra of dried films, based only upon those absorptions in the spectral region above 2000 crn'. This finding indicates that glass can be used as a replacement for specialized, expensive substrates in most analytical methods based upon IR spectroscopy of dried films. More generally, we note that many other potential substrates exist that may prove useful despite their being opaque over much of the IR spectral region. The main consideration for analytical work is that the limited transmission window encompasses analytically useful absorptions of the target species.

References (1 ) Wallach, J. Interpretation of diagnostic tests. A synopsis of laboratory medicine, 5th ed.; Little, Brown and Company: London, 1992.
(2) Burtis C.A.; Ashwood, E.R. Eds., Tietz. Fundamentals of clinical chemistry, 4th ed.; W.B. Saunders: Philadelphia, 1996.
(3) Shaw, R.A.; Kotowich, S.; Leroux, M.; Mantsch, H.H. Ann. Clin. Biochem, 1998, 35, 624-632.
(4) Hall, J.W.; Pollard A. Clin. Chem. 1992, 38, 1623-1631

(5) Hall, J.W.; Pollard A. Clin. Biochem. 1993, 26, 483-490

(6) Hazen K.H.; Arnold, M.A.; Small, G.W. Analyt. Chim. Acta 1998, 371, 255-67.

(7) Bhandare, P.; Mendelson, Y.; Peura, R.A.; Janatsch, G.; Kruse-Jarres, J.;
Marbach, R.; Heise, H.M. Appl. Spectrosc. 1993, 8, 1214-1221.

(8) Ward, K.J.; Haaland, D.M.; Robinson, M.R.; Eaton, R.P. Appl. Spectrosc.
1992, 46, 959-965.

(9) Haaland, D.M.; Robinson, M.R.; Koepp, G.W.; Thomas, E.V.; Eaton, R.P.
Appl. Spectrosc. 1992, 46, 1575-1578.

(10) Arnold, M.A. Current Opinion Biotech. 1996, 7, 46-49.
(11 ) Khalil, O.S. Clin. Chem. 1999, 45, 165-177.
(12) Martens, H.; Naes, T. Multivariate Calibration; John Wiley & Sons; New York, 1989.

(13) Mantsch, H.H.; Jackson M., Eds. Progress in Biomedical Optics. Infrared Spectroscopy: New Tool in Medicine; Proc. SPIE 3257; SPIE: Bellingham, WA, 1998.
(14) Ng, M.N.; Simmons, R. Anal. Chem. 1999, 71, 3438-3508.
(15) Shaw, R.A.; Low Ying S.; Leroux M.; Mantsch, H.H. Manuscript in preparation.
(16) Liu, K.Z.; Mantsch, H.H. Am. J. Obstet. Gynecol. 1999, 980, 696-702.
(17) Liu, K.Z.; Shaw, R.A.; Dembinski, T.C.; Reid, G.J.; Low Ying, S.;
Mantsch, H.H. Submitted to Am. J. Obstet. GynecoL

TABLE I. Reference intervals (normal physiological ranges) for serum analytes Analyte Reference intervals Total protein 60-83 g/L (adult) Albumin 32-48 g/L (adult) Urea 2.5-6.4 mmoI/L

Glucose 3.6-5.8 mmol/L

Cholesterol 3.9-6.1 mmoI/Lb Triglycerides 0.5-1.8 mmol/L

References 1,2 Desirable range (5t" percentile to 75t" percentile) for 40 year old males.
Desirable upper limit for women is ~5.5 mmol/L.
Desirable range for adult males. Desirable upper limit for women is 1.5 mmol/L.

TABLE II. Summary of regression lines relating IR-based assays (Y) to reference assays (X) as Y = AX + B.
Spectraa N A B r SEP~

Albumin Normalized 100 0.90 (0.03)3.4 (1.0) 0.95 2.0 g/L

Unnnormalized 0.88 (0.04)3.6 (1.3) 0.92 2.6 "

Cholesterol Normalized 87 0.95 (0.02)0.15 (0.10)0.98 0.29 mmoI/L

Unnnormalized 1.02 (0.03)-0.20 (0.14)0.97 0.36 "

Glucose Normalized 99 0.85 (0.05)1.1 (0.4) 0.86 1.5 mmoI/L

Unnormalized 0.84 (0.04)1.2 (0.3) 0.91 1.5 mmol/L

Protein Normalized 103 0.87 (0.03)7.6 (2.2) 0.93 3.1 g/L

Unnnormalized 0.93 (0.04)4.8 (2.4) 0.93 3.1 "

TriglyceridesNormalized 82 0.94 (0.03)0.06 (0.06)0.97 0.27 mmol/L

Unnnormalized 0.94 (0.03)0.08 (0.08)0.95 0.32 "

Urea Normalized 100 0.98 (0.02)0.42 (0.21 0.98 1.2 mmoI/L
) Unnnormalized 0.90 (0.03)1.2 (0.4) 0.96 1.6 a Models based upon spectra normalized to KSCN internal calibrant ('Normalized') and on the same spectra but without normalization ('Unnormalized').
b Standard errors are in parentheses.
~ The standard error of prediction (SEP) is the root-mean-square difference between the IR-based and reference assays: SEP = [E; (C;~R - C~ref)2~1/2/N

TABLE III. PLS models.
# PLS factors Spectral region Normalized Unnormalized (cm-') s ectraa s ectrab Albumin 2800-3500 9 9 Cholesterol 2800-3000 5 5 Glucose 2800-3500 12 12 Protein 2800-3500 10 13 Tri I cerides 2800-3000 9 8 U rea 2800-3500 11 10 a Models based upon spectra normalized to the absorption intensity of the KSCN
internal calibrant.
b Models based upon unnormalized spectra.

Figure Captions Figure 1. Mid-IR absorption spectra of the target analytes in the XH
stretching region.
Figure 2. Mid-IR absorption spectra for a single serum specimen, measured for films dried on BaF2 (lower trace) and on a glass microscope slide (upper trace).
Figure 3. Comparison of analyte levels predicted using the IR-based quantitation models (validation, or test set), based upon the normalized spectra, to the reference assays. The best-fitting regression lines (see Table II) and 95%
confidence intervals are also plotted.
Figure 4. Trends in the standard errors of calibration (training set) and validation (test set) with increasing factors in the PLS model for cholesterol.
Figure 5. Comparison of analyte levels predicted using the IR-based quantitation models (validation, or test set), based upon the unnormalized spectra, to the reference assays. The best-fitting regression lines (see Table II) and 95%
confidence intervals are also plotted.

Claims