| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
ANIMAL NUTRITION |
Department of Animal Science, North Carolina State University, Raleigh 27695
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
Key Words: Amino Acids • Animal Meal • Infrared • Quality Control
|
Introduction |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
ise de Zootechnie et al., 2000
). Amino acids are typically determined by HPLC. Although a well-established technique, HPLC is time-consuming, expensive, and prone to errors within and between laboratories (van der Meer, 1990). Thus, it is not suitable for routine use by the animal feed industry for evaluating the nutritional quality of incoming ingredients. Instead, the industry typically relies on book values or values extrapolated from CP for their ingredient matrices. Using these values generates errors in feed formulation and manufacturing, which increase costs and result in augmented nutrient excretion (van Kempen and Simmins, 1997).
Near infrared reflectance spectroscopy (NIRS) is a common quality control tool in the feed industry. Typically, NIRS is used for predicting the CP content of feed samples, but work by van Leeuwen et al. (1991) and van Kempen and Bodin (1998) have shown that it can also be used to predict total and digestible AA. Other infrared spectroscopic techniques, such as mid-infrared spectroscopy (Fourier transform infrared or FTIR) and Raman spectroscopy, are not common in the feed industry, but given their potential for higher-quality spectral data, they may be of interest for the determination of nutrient contents in feed ingredients (van Kempen, 2001). Fourier transform infrared spectroscopy, however, has limited flexibility for analyzing solids. Raman instruments suitable for routine use are only recently becoming available, and their practical utility requires study. The objective of this study was to compare these three techniques for their ability to predict AA in animal meals.
|
Materials and Methods |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
Meat and bone meal samples (n = 141) were collected from major North American rendering companies over a period of 2 yr. These samples included typical materials as well as over- and underprocessed samples so that the variability in the database was maximized. Samples were stored in Nalgene bottles (Fisher Scientific, Pittsburgh, PA) at 4°C.
Because AA analyses are prohibitively expensive for developing calibrations using large sample numbers, and to ensure that each sample included in the calibration was unique, a subset of samples was selected for calibration development. For this, samples were scanned "as is" using a NIRS spectrometer (NIRSystems model 6500, Foss Inc., Silver Spring, MD) in the reflectance mode. Spectra (log 1/R) were obtained using a full-size transport cell (model IH-0331) and were recorded between 1,100 and 2,500 nm with 64 scans co-added over a period of approximately 2 min. Scanning resolution was 2 nm. Samples (n = 54) were selected using the "sample select" routine of Infrasoft International (Port Matilda, PA), such that these 54 samples uniformly covered the spectral variation in the original dataset. These samples were assayed for AA using HPLC at the University of Missouri Chemistry Laboratory (Columbia). Assay data were not corrected for AA losses during hydrolysis. A summary of the AA composition data for these database samples is listed in Table 1. Because spectral variation in this dataset was maximized through the sample selection procedure, variation in AA content should not be used as an indicator for variation in amino acid content of animal meals in the market place.
|
Samples were scanned as received with a near infrared reflectance spectrometer as described above.
FTIR
Samples were first ground to pass a 0.5-mm screen with a grinder (model ZM100, F. Kurt Retsch GmbH & Co., Haan, Germany). Samples were then scanned with an FTIR spectrometer (model Magna 760, Nicolet, Madison, WI) equipped with an extended KBr beam-splitter and a mercury cadmium telluride-A detector. Scanning was conducted while the sample was pressed on a multi-bounce total reflection (ATR) accessory with a 40° zinc selenide through crystal (Spectratech, Shelton, CT). Absorbance spectra (log 1/R) were recorded from 4,000 to 650 cm–1 (2,500 to 15,485 nm), with 64 scans co-added over approximately 1 min. The scanning resolution was 2 cm–1.
Raman Spectroscopy
Samples were scanned with a Fourier transform-Raman spectrometer (model Nexus 670, Nicolet) equipped with a KBr beam-splitter, an InGaAs detector, and a laser operating at 9,393 cm–1. Samples were scanned as is in clear scintillation vials. Raman scattering intensity spectra were recorded from 3,800 to 400 cm–1 (2,632 to 25,000 nm), with 512 scans co-added over approximately 10 min. The scanning resolution was 4 cm–1.
Statistical Analyses
Spectral data from NIRS were smoothed using a quadratic function based on five data points, and second-order derivatives were calculated before developing calibrations. Spectral data from FTIR were reduced to a data spacing of 8 cm–1, and then first-order derivatives were calculated with a quadratic function based on five data points. Various modifications of Raman spectra, such as reduction, smoothing, baseline offset, and derivatives of different order, were calculated to find the appropriate data pretreatment for optimal calibrations. Smoothing and derivatives were calculated with the Savitzky-Golay routine (Esbensen et al., 1996).
Calibrations for nutritionally important amino acid content were developed with partial least squares regression (PLS) using full cross validation (Esbensen et al., 1996). The dependent variable was a specific AA, and the independent variables were the spectral data. Outliers were removed in two cycles to improve the prediction error of the calibrations developed. In the first cycle, all samples with a residual y variance over one were deleted (with y being the value of the AA evaluated). In the second cycle, all samples with a residual y variance over 0.5 were deleted. This procedure was chosen because it is an objective means of outlier removal, although it may not result in optimal calibrations. Spectral modifications, PLS regressions, and outlier removal were performed with The Unscrambler (Version 7.5, Camo Inc., Trondheim, Norway).
|
Results and Discussion |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
A typical NIRS spectrum of an animal meal sample is presented in Figure 1. This spectrum exhibits typical characteristics of NIRS spectra of solids. The baseline gradually increases as a result of scattering of light from the solid particles and peaks are indistinct and are difficult to assign to specific components in the animal meals. This spectrum also shows one of the strengths of NIRS—very low noise.
|
; Aufrere et al., 1996). The NIRS calibrations explained, on average, 70% of the variation in AA (ranging from 56% for valine and 82% for leucine, Table 2). On average, six samples were removed as outliers. For methionine, valine, and phenylalanine, no samples were removed. The average number of principal components used was five, with a range of three to nine.
|
, who used a database of similar size but with more random samples. Better calibrations are certainly feasible with a larger number of samples, as shown by Bodin et al. (1999) and Fontaine et al. (2001). FTIR
A typical FTIR spectrum of an animal meal sample is presented in Figure 1
. The spectrum shows several distinct peaks that can be assigned to specific functional groups present in the sample. Given the sharper nature of these peaks, they may be better suited for quantitative purposes than those observed in the NIRS spectra. However, this spectrum also contains noise, which may adversely affect the quality (prediction error) of calibrations. Also visible are water vapor peaks from water vapor in the air contained within the ATR accessory.
Because FTIR prediction of AA content in feedstuffs has not been reported in the literature, various modifications of the spectral data were evaluated to best calibrate AA content. Partial least squares regressions based on original, first-order, and second-order derivatives were performed. The first-order derivatives gave the best prediction as judged by the r2 (data not shown).
FTIR calibrations explained an average of 68 ± 3% of the variation in AA (ranging from 41% variation in histidine and 86% variation in leucine; Table 2). On average, 13 samples were removed as outliers, and calibrations were based on six principal components.
Because noise was apparent in the spectra, especially at lower wave numbers (4,000 to 2,000 cm–1), calibrations were recalculated using spectral data between 2,000 and 650 cm–1 (Table 2). Calibrations obtained using spectral data from 2,000 to 650 cm–1 explained more variation (r2 = 0.77 vs. 0.68 ± 0.03; P = 0.09) than did calibrations obtained using the entire FTIR spectrum. For individual AA, the variation explained ranged from 69% for lysine to 89% for cysteine. Again, a substantial number of samples was removed as outliers (an average of 15) and calibrations were based on five or six principal components.
Raman Spectroscopy
A typical Raman spectrum of an animal meal sample is shown in Figure 1. One peak is distinguishable in the same region where the FTIR spectrum showed two well-defined peaks. In general, the Raman spectra were noisy, even with a sample acquisition time of 10 min (co-adding 512 spectra). Treatments of the spectral data were evaluated similar to FTIR, but no successful calibrations were developed due to the poor quality of the data. These scans were not obtained in-house; therefore, no sample manipulations were evaluated to improve the spectra.
Raman spectroscopy is able to recognize most chemical bonds in proteins (Ozaki, 1999). However, our data show a signal so noisy that even with a 10-min data acquisition time per sample, spectra were unacceptable. The reason for this is unknown but was recently confirmed with unground meat samples (data not shown). Thus, for Raman spectroscopy, additional work is needed on sample presentation before it can be considered as an alternative to NIRS for predicting amino acids in animal meals.
Comparison of NIRS and FTIR
The spectral data obtained confirm that, using solid samples, NIRS excels in obtaining low-noise spectra, whereas FTIR obtains more information on the sample, but noise is a problem. Practically speaking, these advantages and disadvantages cancelled each other out, as for four of the nine calibrations, NIRS had some advantage over FTIR, whereas for five calibrations FTIR had some advantage as judged by variation explained (Figure 2). Across all calibrations, variation explained by NIRS was equal to variation explained by FTIR (70 vs. 68 ± 3%, respectively). However, NIRS calibrations were based on a larger number of samples (48 vs. 41 ± 1, NIRS vs. FTIR, respectively; P < 0.01).
|
Compared with NIRS, which has over two decades of history of application in determining nutrients for animals (Norris et al., 1976), FTIR has not been documented extensively as a tool for animal feed analysis. The data obtained based on entire spectra showed that FTIR and NIRS performed equally well, despite problems with noise in the FTIR spectra. Eliminating a noisy portion of the FTIR spectrum resulted in better calibrations with FTIR likely because better quality information can be obtained from the samples in mid infrared, although these calibrations were based on fewer samples (and thus likely less robust) because more samples were considered as outliers.
The reason for more noise at lower wave numbers is that sample penetration is a function of wave number when using an ATR, with deeper sample penetration at higher wave numbers. For example, at 1,000 cm–1, sample penetration depth is four times that at 4,000 cm–1 (Sedman et al., 1999). Removal of spectral data at lower wave numbers thus resulted in removal of spectral data obtained at shallower sample penetration, which is prone to more noise. Based on these shortened wave number ranges, calibrations were significantly improved. This suggests that the ATR technology used (based on a 40° ZnSe crystal) limits the quality of calibrations. However, currently no ATR crystals are available that result in deeper sample penetration without sacrificing practicality (hardness and chemical resistance). Other means to improve calibrations may be to grind samples even finer (increases assay time), press the samples firmer onto the ATR, or use a lower scan resolution, but the first two of these methods do not improve the practicality of FTIR for routine quality control of animal meals.
|
Implications |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
|
Footnotes |
|---|
2 Our appreciation goes out to S. Vaughan for scanning samples using Raman spectroscopy and to E. Harris for proofreading the manuscript.
3 Current address: Ajinomoto (China) Co. Ltd., 2201 Tower A, Full Link Plaza, No. 18 Chao Yang Men Wai Da Jie, Beijing 100020, China.
4 Correspondence: Provimi R&T Centre, Lenneke Marelaan 2, B-1932 Sint Stevens Woluwe, Belgium (e-mail: theovankempen{at}yahoo.com).
Received for publication February 9, 2004. Accepted for publication May 13, 2004.
|
Literature Cited |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
