Near Infrared Reflectance Spectroscopy (NIRS) as a rapid tool to predict the total and poultry digestible amino acid content of animal meals. Theo van Kempen, Jean-Christophe Bodin, and David Jackson. Rhone-Poulenc Animal Nutrition.

12th Annual Carolina Swine Nutrition Conference, November 13th, 1996

NIR technology: Can we measure amino acid digestibility and energy values?

Theo van Kempen, Rhône-Poulenc Animal Nutrition, Antony, France

Bottom line:

The NIRS may be used as a tool to predict the digestible amino acid and likely the energy content of feedstuffs. Such information is useful from a quality control perspective to find outliers in incoming raw material and to rank batches of feedstuffs based on nutritionally relevant parameters; for a nutritionist, such information may be used, in combination with laboratory data, to reformulate diets.

Introduction

The feed industry requires rapid and accurate information concerning the nutritive value of feedstuffs. On one hand, such information is needed to negotiate the proper price for a feedstuff, and on the other hand, to correctly include this feedstuff in a complete diet such that the animal’s nutrition requirements are met with the lowest feed cost. Preferably, such information is obtained rapidly so that a routine evaluation of a feedstuff is possible.

Currently, many feedmills use NIRS technology to predict the protein, moisture, fat, and ash contents of feedstuffs to obtain the needed information on a feedstuff. However, these parameters are of limited value since they only correlate to a limited extend with the true nutritional value of a feedstuff. Better parameters to evaluate are the amino acid availability, and the net energy of a feedstuffs.

The objective of this paper is to review the basics of NIRS and the potential of this technique as a tool to routinely evaluate feedstuffs for nutritionally relevant parameters.

Part I: Understanding NIRS

Color

To understand the basic principle of NIRS, it is necessary to understand the basis of color. White light is composed of all the colors in the rainbow. When this white light falls onto an object, certain colors within this spectrum are absorbed by the object, while the rest is reflected (or transmitted). The light reflected can be observed by the eye, and its composite remaining color is interpreted by the eye as the color of the object.

As an example, when sunlight falls onto a clover leaf, both the red and the blue light are absorbed to aid in photosynthesis. The other colors are reflected, and are interpreted by the eye as green.

Physics behind light absorption

The physical basis of light absorption is linked to the nature of molecular bonds and to the nature of light. Light consists of photons traveling all at a fixed speed. While traveling, these photons exhibit a ‘wave’ motion around the path of travel. This wave motion is specific for each color of light, and is often expressed as the frequency (the number of waves traveled per second), or as the wavelength (the distance traveled while completing 1 ‘wave’).

Molecular bonds are the links between atoms within a molecule. These bonds are not static connections, since the atoms they connect vibrate relative to each other. This vibration causes the molecular bond to stretch and compress, resulting in a wave motion of the two atoms relative to each other with a frequency specific to the atoms involved.

When light hits such a vibrating bond, two situations may arise:

1) When the frequency (‘color’) of this light does not match the frequency of the vibration of the molecular bond, then this light is reflected

2) When the frequency (‘color’) of this light matches the frequency of the vibration of the molecular bond, then this light is absorbed, and its energy is used to intensify the vibration (the objects becomes warmer).

Light absorption by feedstuffs

Feedstuffs are made up of mainly organic matter. The molecular bonds which are most common in a feedstuff are thus bonds between hydrogen, carbon, oxygen, sulfur, phosphorus, and nitrogen. The frequency of the vibration between these molecules is such that these bonds generally absorb light in the near-infrared region, or the region which extends just beyond red in the rainbow (not visible for humans).

NIRS uses the principle that molecular bonds absorb specific frequencies of light to obtain information about the number and type of organic bonds present in a feedstuff; in other words, by ‘looking’ at the feedstuff’s color in the near-infrared region.

NIRS Equipment

The principle of a NIRS machine is that a feedstuff is illuminated with light of a specific and known frequency (or wavelength) in the near-infrared region. The absorption of light by the feedstuff is then measured as the difference between the amount of light emitted by the NIRS machine and the amount of light reflected by the sample. The machine layout is explained in Figure 1.

Fig. 1. Schematical layout of NIRS machine. Light is focused (1) onto a concave mirror (2), which separates the light into its composing wavelengths (3). One wave-length is selected (4), and falls onto the feedstuff (5). The amount of this light which is reflected by the feedstuff is measured (6) to obtain the absorption corresponding to one wavelength. By changing the orientation of the mirror (2), different wavelengths of light are selected to obtain absorption measures for all wavelengths of interest. The graph depicting the wavelength tested versus the absorption is called the spectrum, and an example spectrum for coconut meal is presented in the top right corner (7).

Instead, to predict the composition of feedstuffs it is necessary to make calibrations. A calibration is simply NIRS spectral information correlated with composition values (obtained by, e.g., wet chemistry). This concept is most easily explained by a fictitious example (see Fig. 4). Let us imagine the spectra obtained for 3 soybean samples which are identical except for their protein content. The protein content of these samples was either 44%, 46%, or 48%. In this example, the spectra from these samples deviates only in the region of 1900-2000 nanometer (nm; the wavelength of the light), and this difference must be related to the differences in protein content of the sample, since the samples were otherwise identical.

Plotting the absorption at 1950 nm for each spectrum against the protein level shows that a linear relationship exists between the absorption of light at 1950 nm and the protein level. It is easy to image that such a correlation could be used to predict the protein content of a fourth soybean sample.

In reality, spectra are much more complicated. Protein does not absorb at only one region of the spectra (as seen for methionine in Fig. 2). Additionally, when protein levels increase in a sample, other compounds, e.g., carbohydrates, must decrease, and thus other regions of the spectrum must change. The spectra of the three samples should thus be different in several regions, and it is not so easy to identify a region of the spectrum in which absorption correlates with the protein content of the samples. Nevertheless, this approach is basically the outlined procedure which is used to determine the correlations between spectral information and chemical composition, which are then subsequently used as calibrations to predict unknown samples.

Such calibrations have been successfully applied in the feed industry to predict parameters such as protein, moisture, NDF, ADF, etc. The main reason for this success is the rapidity of the measurement: values can be obtained within 5 minutes of receiving a sample.

Calibration validation

In the process of calculating correlations, the NIRS software has no notion of the functionality of the correlations it determines. When a spectral peak caused by the presence of carbohydrates happens to correlate (negatively) with the protein level in the reference samples, then a calibration might well be developed based on this carbohydrate peak.

Such a calibration is not very useful. Although a relationship exists between carbohydrates and protein in the reference population, this relationship does not hold in the general population. It is thus essential to prevent the use of such calibrations in real-life applications.

In order to improve the chance of developing valid calibrations, the reference database should contain the maximum amount of variation. In practical terms, at least 100 different samples should be used. These samples must be representative of the population to be predicted in the future, and should thus be obtained from different harvest years, the geographical regions from which this feedstuff is obtained, and cover all the genotypes of interest.

The actual functionality of a calibration should be tested in a process called validation. Ideally, a number of samples covering the range of samples which can be encountered in real life are used. All these samples are predicted using the calibration to be tested, and analyzed with the reference technique to obtain the true values. The difference between the predicted and true value is used to calculate the root mean square error of prediction (RMSEP), which is a measure for the accuracy of the calibration.

However, this exercise is tedious and costly, and because of this, NIRS software developers have included ingenious validation procedures in their software. It never hurts, though, to check the validity of a calibration for yourself.

Part II: NIRS to predict the energy content of feedstuffs

Several research groups have developed NIRS calibrations for energy values, however, the majority of this work has been focused on ruminants. Energy predictions for forages are common and are utilized commercially, which suggests that energy can be predicted with an adequate precision (DeBoever et al. 1993). However, the energy content of ruminant feeds is often determined by proximate analyses in combination with some factor of rumen or in vitro digestibility, rather than by in vivo assays.

For poultry, apparent metabolizable energy (AME) predictions were obtained with a prediction error of 80 kcal/kg feed by Valdez and Leeson (1992). However, in the sample population tested by Valdez the AME value of the feedstuffs was linearly dependent on the fat content of the test diet. The utility of NIRS to predict the fat content has been well documented, but unfortunately this does also imply that the calibration generated is not necessarily effective in predicting AME in cases where the fat content does not correlate with the AME content of a diet. Flinn et al. (1994) obtained a prediction error for AME ranging from 90 kcal/kg DM for cereals, to 275 kcal/kg for mixed feeds. This prediction error, however, appears close to the normal variability observed.

For pigs, digestible energy calibrations were developed by Van der Meer et al. (1988). This paper, though, only lists the standard error of the calibration (a measure of how well the calibration describes the population used to make the calibration), which is of little value for the true prediction error of a calibration. The success of NIRS in the ruminant area suggests, nevertheless, that the NIRS is capable of predicting energy values for feeds or feedstuffs for swine. In addition, the theoretical basis of NIRS suggests that energy predictions should be achievable. Until a serious attempt to do so is published, though, it remains to be seen if it is feasible in real life.

Part III: NIRS to predict the digestible amino acid content of feedstuffs

Calibration development

Since 1981, Rhône-Poulenc Animal Nutrition has tested feedstuffs in poultry for ileal amino acid digestibility according to the procedure of Green et. al (1986). Most samples used in these tests have been retained in cold storage. Currently, a large database exists containing samples from virtually every category of feedstuffs that would be of interest to the feedstuff industry and derived from many countries around the world.

The samples contained in this database were used for the development of NIRS calibrations. Course samples, such as grains, were ground using a Retch grinder (model ZM1000, Haan, Germany) with a 1 mm screen, while the other samples were scanned as-is. Prior to scanning, samples were acclimated to the NIRS room temperature of 23°C. Scans were made using a NIRSystems model 6500 NIRS (Perstorp Analytical, Höganäs, Sweden; a complete setup costs in the order of $75000) equipped with a natural product cell. These scans were subsequently correlated to the ileal digestibility data obtained in vivo.

The validation procedure used to test calibrations is as follows: one sample was deleted from the database, after which a calibration was developed using all other samples in the database. This calibration was subsequently used to predict the deleted sample. This procedure was repeated for all the samples in the database, and the difference between the true and predicted values was used to calculate the prediction error (RMSEP or Root mean square error of prediction).

Calibrations were developed and tested for either specific feedstuffs (e.g., only meat & bone meal), feedstuff categories (e.g., all animal protein products, including e.g., fish meal, meat & bone meal, etc.), or a global database (containing all feedstuffs). In principle, a calibration based on only one feedstuff yields the best results. However, this requires that sufficient samples are present in the database to cover all possible variation. Including samples from a different category increases the variation, and may improve the prediction accuracy of the calibration.

Table I. Root mean square errors of prediction (RMSEP; g/100g) and derived r²_rmsep values for meat & bone meal within the global (based on all feedstuffs), product group (based on feedstuffs of animal origin), product calibration (based on meat & bone meal samples). (The three-letter amino acid code followed by a ‘P’ indicates ‘poultry digestible’; nd=not determined)

	Lys	LysP	Met	MetP	Thr	ThrP
	Population statistics:

mean	2.85	2.21	0.85	0.70	1.87	1.42
stdev	0.74	0.62	0.26	0.21	0.42	0.35

	Global calibration:

RMSEP	0.23	0.30	0.14	0.13	0.12	0.14
r²_rmsep	0.90	0.77	0.72	0.63	0.92	0.80

	Animal meal calibration:

RMSEP	0.26	0.31	0.14	0.13	nd	0.16
r²_rmsep	0.87	0.74	0.69	0.62	nd	0.80

	Meat & bone meal calibration:

RPSEP	0.32	0.35	0.14	0.13	0.13	0.16
r²_rmsep	0.81	0.67	0.72	0.62	0.90	0.79

Table II. Population statistics and performance (g/100g) for a part of the calibration developed using the ‘global’ database. (The three-letter amino acid code followed by a ‘P’ indicates ‘poultry digestible’.)

	ISI statistics		validation
	SECV	1-VR	RMSEP	r²_rmsep
Lys	0.19	0.98	0.25	0.97
LysP	0.23	0.97	0.29	0.94
Met	0.09	0.96	0.11	0.94
MetP	0.09	0.95	0.10	0.94
SAA	0.14	0.96	0.19	0.94
SAAP	0.14	0.93	0.18	0.90
Thr	0.10	0.99	0.13	0.98
ThrP	0.12	0.97	0.15	0.96
Try	0.06	0.92	0.08	0.87
TryP	0.07	0.89	0.08	0.83

Table III. Calibration performance (g/100g) of the global calibration as applied to meat & bone meal and fish meal. (The three-letter amino acid code followed by a ‘P’ indicates ‘poultry digestible’.)

	Meat & bone meal
	mean	stdev	RMSEP	r²_rmsep
Lys	2.85	0.74	0.25	0.90
LysP	2.21	0.62	0.30	0.77
Met	0.85	0.26	0.14	0.72
MetP	0.70	0.21	0.13	0.63
Thr	1.87	0.42	0.12	0.92
ThrP	1.42	0.35	0.14	0.83
	Fish meal
	mean	stdev	RMSEP	r²_rmsep
Lys	4.88	0.66	0.22	0.89
LysP	4.47	0.75	0.26	0.87
Met	1.76	0.29	0.17	0.61
MetP	1.65	0.29	0.17	0.65
Thr	2.74	0.36	0.16	0.80
ThrP	2.49	0.44	0.18	0.83

For this reason, calibration and validation statistics within certain animal meals are provided in Table III, and based on this table one can really decide what the value of a calibration is. This table shows that for the amino acids listed, 60-90% of the variation within the populations was explained by the global calibration.

Advantages of a global calibration

The global calibration performed most accurately for high-protein feedstuffs. Although it works for other feedstuffs, the prediction error makes it unattractive for use for feedstuffs with low levels of amino acids.

Nevertheless, such predictions may still be interesting when no or limited data are available for a certain product. For example, samples of coconut meal, mung, and kapok beans from Asia and an extruded soybean-rapeseed mix product from Europe were received at the Rhône-Poulenc Laboratory for in vivo testing. For all four samples, the global calibration predicted the composition with an accuracy sufficient to get an idea of the nutritional value of the product.

Another application might be to use the global calibration for mixed feeds. Although no official tests have been carried out with mixed feeds, it appears that the current calibrations predict the total and digestible amino acid profiles of mixed feedstuffs with a reasonable accuracy. However, synthetic amino acids are not detected as amino acids (they are included in the protein estimate) since they were not included in the calibration samples.

NIRS in comparison to other techniques

To fully understand the value of the current global calibration it is important to look at alternative techniques for obtaining total and digestible amino acid values for feedstuffs. A multitude of techniques exist, but most of them have either one or both of the following disadvantages: they take too much time, or they are too expensive. Techniques which are expensive and/or cumbersome can not be applied as a routine evaluation system, and to compare the NIRS to such a technique is not very meaningful.

Instead, NIRS is to be used in situations where a answer is needed quickly, and in such situations a compromise with respect to accuracy is often unavoidable. As an example, currently many feedmills use book values (often personalized) in their formulations. Such a bookvalue, although accurate for a population mean, does not explain any of the variation within a population. The prediction error (RMSEP) in this case is thus equal to the standard deviation of the population, and the explained variation (r²) is 0.

A second alternative would be to measure the protein content of a batch of a feedstuff, and to use regression to predict the digestible amino acid content. The accuracies of such regressions vary. For maize and wheat, our calculations indicate that they are useless (r²_msep =0), while in meat & bone meal they explain a reasonable percentage of variation (r²_rmsep=0.59). Other products evaluated fall somewhere in between. This means that even in the best regression prediction (meat & bone meal), regression yields results which are still inferior to the NIRS predictions. So, as a rapid and practical technique to predict total and digestible amino acid contents of feedstuffs the NIRS scores quite well in comparison to some of the available alternatives (Table IV).

Table IV: Comparison (RMSEP in g/100g and derived r²_rmsep values) of NIRS digestible amino acid prediction accuracy with the accuracy of other rapid techniques to predict the digestible amino acid content of animal meal samples. The estimated in vivo accuracy (in g/100g) is provided for comparison only, since this technique may not be considered as ‘rapid’.

		Book value	regres-sion N	NIRS	In vivo

Digestible lysine:
meat & b	rmsep	0.62	0.37	0.30	0.20
	r²_rmsep	0.00	0.59	0.77
fish meal	rmsep	0.75	0.60	0.26	0.34
	r²_rmsep	0.00	0.41	0.87

Digestible methionine:
meat & b	rmsep	0.21	0.14	0.13	0.06
	r²_rmsep	0.00	0.57	0.63
fish meal	rmsep	0.29	0.16	0.17	0.12
	r²_rmsep	0.00	0.74	0.65

Digestible threonine:
meat & b	rmsep	0.35	0.16	0.14	0.13
	r²_rmsep	0.00	0.77	0.84
fish meal	rmsep	0.44	0.30	0.18	0.19
	r²_rmsep	0.00	0.52	0.83

Comparison with in vivo estimations

When evaluating a calibration, it is of importance to keep the quality of the reference data in mind. Since the calibration was calculated based on correlations between spectral and the (error containing) reference data.

The errors of the digestible amino acid assay are first, the amino acid analysis of the feedstuff, and second, the error associated with the in vivo part of the digestibility assay. Amino acids were analyzed by HPLC, and the error of this technique is generally considered to ±5%.

Subsequently, digestibility was evaluated in vivo, and due to animal variation and the analysis of ileal juice for amino acids this assay generally adds another 2% to the overall error. The total error of the digestible amino acid assay may thus be estimated at 7%, or for meat & bone meal, 0.20% of digestible lysine.

The prediction error obtained by NIRS ranged, for the data presented, from similar to up to 2 times worse than the in vivo assay error (see Table IV). This is in the range which is deemed acceptable for NIRS techniques. It should thus be concluded that these first-generation calibrations are acceptable.

Poultry as a model for swine digestibility

The calibration results discussed were for poultry digestible lysine. Little work has been performed on predicting swine ileal digestibility, the reason being that large databases of feedstuff samples with accompanying swine ileal digestibility data are not available. Work within our laboratory based on almost 100 feed samples tested in both pigs and poultry, however, suggests that a relationship exists between poultry digestible and swine digestible amino acid contents of feedstuffs. For lysine, this relationship, which explained 99.7% of the variation, is depicted in Fig. 6; for methionine, 99.7% of the variation was explained as well, and the prediction error was 0.02% digestible methionine. For other amino acids, relationships were obtained which explained at least 98% of the variability.

Given the almost perfect relationship between swine and poultry digestible amino acid contents of feedstuffs, calibrations based on swine ileal digestible data should be able to be generated. Alternatively, poultry predictions could be made, and these could subsequently be converted to swine predictions. The success of such a procedure, however, depends on the willingness of the industry to accept the correlations between poultry and swine digestibility.

Practical use of NIRS calibrations

In a commercial feedmill, it is practically and technically not possible to adjust ration formulations based on the nutritional profile of each batch of raw material received. Additionally, no technique exists which can determine the digestible amino acid profile of each batch of feedstuff with sufficient accuracy and within the cost and time constraints of a feedmill. NIRS, however, would fit within the time and cost constraints, and approaches the accuracy limit.

From our perspective, the presented NIRS calibrations are most interesting for monitoring the nutritional profile of a feedstuff upon reception. This would allow the feedmill to detect outliers which require additional analyses and which may require a re-evaluation of a supplier or the price being paid for the product. Over time (when several subsequent batches of a feedstuff have been predicted), in conjunction with laboratory analyses, the NIRS data may be used to adjust formulations.

Such a procedure also helps to evaluate NIRS. When laboratory analyses confirm the NIRS results, the user builds confidence in the NIRS predictions. Over time, the user will increase his reliance on the NIRS technique, resulting in less samples being selected for laboratory analysis.

In the situation where a feedmill has multiple bins for one type of a highly variable feedstuff in order to classify incoming material, NIRS predictions may also be used to determine in which quality category a certain batch belongs. The value of NIRS for feedstuff categorization is illustrated in Figure 7, which is based on the samples in the RPAN database. Meat & Bone meal was categorized based on its digestible lysine content in the classes 1-2%, 2-3%, and 3-4% digestible lysine. In the test, 30 of the 41 samples were categorized correctly by the NIRS, 6 samples were ‘borderline’ (categorization error acceptable since it falls within the limit of the reference technique), and 5 samples were categorized incorrectly (statistically, 74% should be categorized correctly when 80% of the variation in a population is explained; Shenk & Westerhouse, 1993).

Work in progress

As indicated above, our initial work to calibrate the NIRS for digestible amino acids was successful. The prediction errors obtained, although not perfect, were low enough so that the calibrations become attractive for practical use in evaluating high protein animal meals.

Other calibrations (based on product databases) have been developed for maize and soybean meals, and the validation results of these calibrations suggest that it is possible to calibrate the NIRS for these products as well. Digestibility trials to finalize these calibrations are already underway.

Selected references

DeBoever, J.L., Cottyn, B.G., Vanacker, J.M., and Boucqué, Ch.V., 1993. Recent developments in the use of NIRS for evaluating compound feeds and raw materials for ruminants. Paper 1; Proceedings Conference NIR Spectroscopy - Developments in Agriculture and Food. Birmingham, UK. ADAS Drayton Research Centre FEU.

Flinn, P.C., Windham, W.R., Johnson, R.J., 1994. Current status of near infrared (NIR) spectroscopy in Australia for predicting metabolisable energy of poultry feeds. Proceedings, 9th European Poultry Conference. Glasgow, UK. pp.106-109.

Green, S., 1986. Digestibility of amino acids in practical feed formulation. Poultry International, October: 140-143.

Green, S., S. Bertrand, M. Duron, and Maillard, R., 1987. Digestibilities of amino acids in maize, wheat, and barley meals, determined with intact and caecectomized cockerels. Br. Poultry Sci., 28: 631-641.

van der Meer, J.M., Vedder, H., and Wever, G., 1988. NIRS for the prediction of organic matter and energy digestibility in pig feeds. In vitro newsletter No. 4: 25-26.

Murray, I., and Hall, A.P., 1983. Animal feed evaluation by use of Near Infrared Reflectance (NIR) Spectrocomputer. Anal. Proc. 20: 75-70.

Low, A.G., 1990. In: J. Wiseman and D.J.A. Cole (Editors), Protein evaluation in pigs and poultry in feedstuff evaluation. Butterworths, London, pp. 91-114.

Osborne, B.G., Fearn, T., and Hindle, P.H., 1993. Practical NIR spectroscopy: with applications in food and beverage analysis (2nd edition). Longman Singapore Publishers, Singapore.

Rostagno, H.S., and Pupa, J.M.R., 1995. Diet formulation for broilers based on total versus digestible amino acids. J. Appl. Poultry Res. 4: 293-299.

Rhône-Poulenc, 1993. Rhodimet nutrition guide, 2^nd edition. Rhône-Poulenc Animal Nutrition, Antony, France.

Shenk, J.S., and Westerhaus, M.O., 1993. Monograph. Analysis of agriculture and food products by near infrared reflectance spectroscopy. Infrasoft International, Port Matilda, PA, USA.

Valdez, E.V., and Leeson, S., 1992. Near infrared reflectance analysis as a method to measure metabolizable energy in complete poultry feeds. Poultry Sci. 71: 1179-1187.

Commonly used abbreviations & terms:

1-VR Measure of the expected explained variation

Bias Systematic difference between true and predicted values

H Measure of spectral similarity

ISI Infrasoft International (major supplier of software for NIRS)

NIR Near infrared; 700-2500 nm.

NIRS Near infrared reflectance spectroscopy

r² Explained variation

RMSEP Root mean square error of prediction (true variability of an estimate)

SEC Standard error of calibration (variation within the reference population not explained by the calibration)

SECV Standard error of cross validation (expected variability of an estimate)

RMSEP Variability of prediction error

T Student T value (statistical similarity to reference population)

VIS Visible region; 400-700 nm