A comparison of the prediction of apparent metabolisable energy content of starchy grains and cereal by-products for poultry from its chemical components , in vitro analysis or near-infrared reflectance spectroscopy

Regression models including chemical composition, in vitro digestibility and near infrared reflectance spectroscopy (NIRS) were compared in order to predict the energy value of several feed ingredients for poultry. The nitrogen-corrected apparent metabolisable energy content (AMEn) in cockerels and its proportion on total gross energy (AMEn/GE) were determined in 94 batches from six starchy grains and six cereal byproducts. Two preliminary trials were also designed to adapt in vitro methods for prediction of in vivo energy values for poultry. Mean concentrations of AMEn of the ingredient studied ranged from 2,464 to 3,595 kcal kg DM, and those of AMEn/GE from 53.7 to 80.0%. The most precise model of prediction of AMEn and AMEn/GE values was that based on NIRS equations (Rcv = 0.823 and 0.861, respectively). The best single chemical predictor of these energy values was the neutral detergent fibre concentration (R = 0.616 and 0.736, respectively). Further inclusion of ether extract and ash contents in the AMEn model and those of starch and ether extract in the AMEn/GE model allowed increasing coefficients of determination up to 0.791 and 0.839, respectively. A model including linear and quadratic effects of in vitro organic matter digestibility (IVOMd) provided a similar prediction of AMEn/GE values (R = 0.833). However the prediction of AMEn from IVOMd was worse (R = 0.62), as variations among batches of GE concentration (from 4,225 to 5,896 kcal kg DM) were little related to in vitro digestibility values. Additional key words: cockerels, energy content, energy utilization, NIRS, prediction models.


Introduction
Direct determination of energy values of feeds in in vivo trials is expensive and time-consuming; it also requires animal facilities and relatively large amounts of experimental diets.Chemical analyses, in vitro methods and near infrared reflectance spectroscopy (NIRS) techniques have been used in several animal species to estimate energy content of feeds and diets.These methods are rapid and economical, which make them more adequate to take into account the variability of the raw materials used by the poultry feed industry.However, its capability to estimate feed energy contents must be validated with in vivo determined values.
At present, several regression equations are available in poultry for the estimation of apparent metabolisable energy (AME) values from chemical components in compound feeds (e.g.Carpenter and Glegg, 1956;Sibbald etal, 1980;Fisher, 1982;Carree?al, 1984;EEC, 1986).However, prediction equations for feed ingredients are scarcer and its validity is limited to the conditions where they were obtained (Dolz and De Bias, 1992;Francesch, 2001).This approach is also limited by the time required for the chemical analyses and their accuracy.
The use of multiple-enzymatic in vitro methods has been proven to be a good alternative to chemical analyses to simulate the digestive processes and to predict energy values with a greater precision in non ruminant species as pigs (Boisen and Fernández, 1997;Noblet and Jacquelin-Peyraud, 2007) and rabbits (Ramos et al, 1992;Pascual et al., 2000).A two-step in vitro method using pepsin, pancreatin, bile acids and enterokinase has been tested in poultry complete diets by Valdes and Leeson (1992c).Its repeatability was similar to in vivo trials but the residual standard deviation of the prediction was high for some of the diets studied.
Previous studies have also shown that NIRS technique allows estimating succesfully the major chemical constituents and the digestion efficiency in several animal species (Roberts et al, 2004), including the energy values of complete poultry feeds (Valdes and Leeson (1992a).However, AME values of a limited number of single ingredients were not well predicted from NIRS, neither when using equations calculated for a limited number of ingredients or when using equations derived from complete diets (Pérez-Vendrell et al, 1992;Valdes andLeeson 1992d, 1994;Garnsworthy etal, 2000).
The aim of this research has been to establish a method of prediction of in vivo apparent metabolisable energy (AMEn) values in several poultry feed ingredients, using chemical analysis, in vitro digestion and NIRS techniques.

Ingredients
Sixty batches of six starchy grains: wheat (Triticum aestivum L., Triticum turgidum L.), barley (Hordeum vulgare L.), corn {Zea mays L.), sorghum {Sorghum vulgare L.), rye {Sécale cereale L.) and peas {Pisum sativum L.) and 34 batches of six cereal byproducts (corn gluten feed, rice bran, wheat bran and dry distillers grains and solubles (DDGS) from wheat, corn and sorghum), were sampled from the COREN SCL poultry feed manufacturing plant throughout a 3-yr period.The number of samples and the mean and range of chemical composition within each ingredient are shown in Table 1.

Apparent metabolisable energy determination
Energy values for feed ingredients were determined in vivo by using the difference method.Experimental diets were made by substituting with the ingredients studied a 40% of three basal diets formulated to avoid an excessive imbalance in dietary essential nutrients before and after substitution.The chemical and raw material composition of the basal diets is shown in Table 2.
Trials were conducted in 20 series, each series including the evaluation of four to five feed ingredients and the corresponding basal diet.Eight adult cockerels (Warren) were randomly assigned to each experimental Abbreviations used: ADF (acid detergent fibre), ADL (acid-detergent lignin), AME (apparent metabolisable energy), AMEn (nitrogen-corrected AME), CF (crude fibre), CP (crude protein), DDGS (dry distillers grains and solubles), DM (dry matter), GE (gross energy), IVDMd (in vitro digestibility of dry matter), IVOMd (in vitro digestibility of organic matter), NDF (neutral detergent fibre), NIRS (near infrared reflectance spectroscopy), S (sugars), SEC (root mean square error), SECV (standard error of cross-validation), SEP (standard error of prediction).ADF: acid detergent fibre.d ADL: acid-detergent lignin.e SD: standard diet.Animals were housed in individual metabolic cages (0.3x0.5x0.4 m high) with wire floors, and kept in an environmentally controlled room.Feed ingredients were ground (in a hammer mill, 6 mm of grill size), mixed with basal diets and given in mash form to birds.
Determination of AME of the experimental diets was made following the European reference method (Bourdillon et al., 1990).The droppings were dried in a forced-draught oven at 80°C to constant weight.After drying, the excreta samples were ground in a coffee mill and then stored in a sealed container at 4°C prior to chemical analysis.
The AME values were calculated using the following formula with appropriate corrections made for differences in DM content:

Feed intake
Nitrogen-corrected AME (AMEn) was calculated by correction to zero nitrogen retention by simple multiplication with 8.22 kcal g _1 of nitrogen retained in the body as described by Hill and Anderson (1958).

Chemical analyses
The proximate composition of feed ingredients, experimental diets and bird excreta were analyzed in duplicated samples using the procedures of AOAC ( 2000) for dry matter (DM) (930.15),ash (923.03),total sugars (974.06),ether extract (920.39) and crude fibre (978.10).Concentration of neutral detergent fibre (NDF), acid detergent fibre (ADF) and acid-detergent lignin (ADL) was determined according to the sequential method of Van Soest et al. (1991) using heat stable amylase (A3306, Sigma) and sodium sulfite, and expressed without residual ash.Starch content was measured following the alpha-amyloglucosidase method (996.11;AOAC, 2000).Nitrogen was measured by combustion (method 968.06;AOAC, 2000) using a VarioMax ELEMENTAR analyzer (Hanau, Germany).Gross energy was determined in an adiabatic bomb calorimeter (Parr Instruments, USA) standardized with BIPEA reference samples.

In vitro technique
Determination of in vitro digestibility of dry matter (IVDMd) and organic matter (IVOMd) of feed ingredient samples was based on the multi-enzymatic method proposed for pigs by Boisen and Fernández (1997).In each of the series, a reference sample (corn grain) was included.These samples were used as a blanck to correct IVDMd and IVOMd values for differences among the successive series.In the first digestion step, series of up to 30 duplicated samples were incubated with pepsin at pH 2 and 39°C during 2 h.In the second digestion step, samples were incubated with pancreatin (a mixture of protease, amylase and lipase) at pH 6.8 and 39°C during 7 h.
A preliminary in vitro trial was done to set the digestion duration at the second step (by comparing eight incubation times increasing from 4 to 19 h with wheat and corn grain samples).Another preliminary test was done to determine in the same samples the value of doing a third digestion step using microbial carbohydrases (Viscozyme 120 L, 120 FBG g _1 ) at pH=4.8 and 39°C, as described by Boisen and Fernández (1997).The in vitro digestibilities of dry matter and organic matter were calculated from the difference between concentrations in the sample and the indigested residue, after corrections for values obtained with reference samples.
NIRS analysis was performed on ground (0.7 mm) samples of the ingredients studied using a near-infrared reflectance spectrophotometer (model 6500; FOSS-NIR System, Silver Spring, MD) equipped with spinning sample cup module.Samples were scanned between 400 and 2,498 run and spectra were recorded with the ISI NIRS 3 software version 3.11 (Infrasoft International, Port Matilda, PA).Measurements were performed in duplicate with repacking of the cup, and spectra for the subsamples were averaged to provide one spectra per sample.

Statistics
Prediction equations of AMEn in vivo values from chemical and in vitro analysis were developed by stepwise regression analysis, using PROC REG of SAS (1990).The stepwise procedure introduced variables in the model only if they contributed to a significant improvement (P< 0.05) in the estimation of the dependent variable.
The NIRS calibrations were developed using the allsample set of full-scan mean spectra (n = 94).The population boundaries for calibration were set with a maximum standardized H (distance between a sample and the centroid of the group) value of 3.0 (Shenk and Westerhaus, 1991), and no samples were marked as outliers.The NIRS models were set up with a modified partial least squares regression, after scattering correction with the standard normal variate transformation combined with detrending.Additional mathematical pretreatments were performed by second derivative treatment.Cross-validation was used to select the optimal number of partial least squares factors and to avoid overfitting (Shenk and Westerhaus, 1995).No outlier elimination pass was accepted.Calibration equations were evaluated in terms of coefficient of determination (R 2 c) and root mean square error (SEC).Validation errors were combined in a standard error of cross-validation (SECV).Prediction error was measured by dividing the calibration samples into subsets (n = 4) with one subset reserved for validation and the remaining for calibration.Cross-validation was repeated until all subsets were used for validation once.Shenk and Westerhaus (1996) reported that the SECV is the best single estimate of the prediction capability of NIRS equations, and that this statistic is similar to the average standard error of prediction (SEP) from 10 randomlychosen prediction sets.The repeatability standard deviation and coefficient of variation of laboratory procedures and NIRS spectra was also determined from ten subsamples of two batches, one of corn grain and another one of peas.

Average and standard deviation of AMEn values in the feedstuffs studied
Values of AMEn of the ingredients studied determined by difference are shown in Table 3. Mean values varied from 2,464 (wheat bran) to 3,595 kcal kg 1 DM (corn grain).Standard deviation of AMEn was 490 kcal kg 1 DM for all the samples studied and 154 kcal kg 1 DM for average variation within ingredients.The degree of metabolizity of gross energy (GE), expressed as the proportion AMEn/GE (%) for each ingredient was also calculated, and the average values are shown in Table 3. Absolute AMEn concentrations determined in the current study for high starch cereal grains (corn, sorghum, wheat) were slightly below (by about 4%) than the average values assigned to these ingredients by several international Feed Tables (NRC, 1994;INRA, 2002;FEDNA, 2003;CVB, 2004).This difference might be explained by a higher proportional weight of endogenous energy losses in the birds used in the current study (fed near maintenance level), with respect to productive animals.In the same way, AMEn values of corn grain and corn DDGS were 3.75% higher in layer hens than in cockerels (Losada et ah, unpublished data).However, in vivo AMEn values measured for low starch grains (rye and barley) were close, whereas those obtained for cereal byproducts were clearly higher (between 8 and 29%) than those assigned as average in the Feed Tables.This result indicates a relative underestimation of this group of feed ingredients when using mean table values, which was directly proportional to its NDF content (r=0.71;P<0.01).The highest deviations were observed for DDGS, which might also reflect recent improvements of the method of production of these byproducts still not considered in Feed Tables.In the same way, recent work in poultry (Batal and Dale, 2006;Fastinger et ah, 2006) has reported AMEn values for corn DDGS even higher than those obtained in the current study.

Prediction of AMEn/GE and AMEn from chemical composition
A stepwise regression analysis was made to predict AMEn/GE and AMEn of all the ingredients studied from the determined chemical composition traits.The regression equations obtained are presented in Table 4.The NDF concentration was the first independent variable included in both models, explaining 73.8 and 61.6% of the variation of AMEn/GE and AMEn, respectively.This relationship indicates the strong negative effect  of dietary fibre on energy utilization in poultry, an effect that was consistent throughout the whole interval of NDF studied (see Fig. 1).Type of fibre, expressed either as proportion of ADL on NDF or ADF, or by the concentration of hemicelluloses and cellulose (calculated, respectively, from the differences NDF-ADF and ADF-ADL), had no significant influence beyond that of dietary level of fibre.This result could reflect the inability of birds to digest any of the constituents of the insoluble fibre from these ingredients.In the model of prediction of AMEn/GE, the stepwise procedure included in two further steps significant corrections (P< 0.05) to take into account the relatively high digestion efficiency of starch and ether extract fractions.In the case of the AMEn model, ether extract and ash content were included in steps 2 and 3, as they were able to explain part of the variation of GE concentration among batches (from 4,225 to 5,896 kcal kg-1 DM).The inclusion of these additional independent variables allowed to decrease the RSD of the models from 5.76 to 4.55% (AMEn/GE) and from 265 to 198 kcal AMEn kg 1 DM (see Table 4).

Prediction from in vitro digestibilities
In the first preliminary in vitro trial, organic matter digestibility (IVOMd) increased with time at the second incubation step from 4 to 7 h, especially in the case of corn grain, and reached a plateau after that (see Table 5).Accordingly, the relative IVOMd values of wheat and  corn grain samples decreased from 1.05 at 4 h to 0.967 at 7 h, with little variation at higher incubation times (see Table 5).Furthermore, relative values between wheat and corn after 7 h of incubation were close to the relative in vivo AMEn value obtained for the same samples (0.95).In the second preliminary trial, the addition of fibrolytic enzymes in a third incubation step led to IVOMd values much higher (92.2 and 90.9% for the wheat and corn samples, respectively) than those obtained with the two-steps technique and than those determined for the proportion AMEn/GE in the same samples.Moreover, the relative value wheat/corn obtained (1.01), led to a worse prediction of in vivo AMEn relative value than that determined with the two-steps technique.According to the results obtained in these preliminary trials, the duration of the digestion in the second in vitro step was set at 7 h, and third step was not done in the further in vitro trials of this experiment.The results of IVDMd and IVOMd obtained with this procedure for each of the ingredients studied are presented in Table 3.
A regression analysis showed a significant (P < 0.001) linear effect of IVDMd and IVOMd on in vivo AMEn/GE and AMEn values (see Table 6 and Fig. 2).A significant (P<0.01)quadratic effect was also observed, as the differences between in vivo AMEn/GE and IVd values were smaller in cereal byproducts than in starchy grains.In the case of AMEn, prediction was significantly improved (P < 0.001) when GE or chemical constituents related to GE concentration, as ether extract, ash or crude protein, were also included in the model (up to RSD values of 211 or 171 kcal kg -1 DM, respectively).Absolute values of IVDMd and IVOMd were similar for each of the ingredients studied (see Table 3), although the RSD of the regression equations to predict in vivo energy values were slightly improved when including IVOMd instead of IVDMd as independent    6).The comparison between regression equations based on chemical constituents (Table 4) or the combination of in vitro digestibilities and chemical constituents (see above), shows that both models led to a similar accuracy of prediction of AMEn/GE and AMEn values.Otherwise, the repeatability of IVDMd and IVOMd (CV R = 1.05%) was similar to that reported in complete diets for pigs (0.9%, Noblet and Jaguelin-Peyraud, 2007) and rabbits (1.09%, Carabaño et ah, 2008).This value was better than that obtained for NDF analyses (3.5%).

Prediction from NIRS analysis
Calibration and cross validation statistics of prediction of nutrient composition, in vitro digestibility and energy value of the ingredients tested from NIRS analysis is shown in Table 7.The repeatability of the NIRS method was estimated from the variability of the energy values predicted in homogeneous analytical conditions.The coefficient of variation obtained for AMEn/GE and AMEn were, respectively, 0.393 and 0.497% (corn grain) and 0.504 and 0.479% (peas subsamples).The coefficients of determination and values of SECV obtained confirm the utility of NIRS to predict chemical composition (Pérez-Marín et ah, 2004) and in vitro digestibility in poultry diets (Valdes and Leeson, 1992b).These relationships among NIRS and laboratory analyses could explain its accuracy to predict the energy value of feed ingredients.The ratio of SD to SECV values in Table 7 for AMEn and AMEn/GE was 2.72 and 2.70, which makes the prediction «good», according to Williams and Sobering (1996).This ratio should be ideally at least of three, unless variance of the reference data is low, as it is the case in the current study.Differences in the variance of data also explain the higher coefficients of determination observed in the current study with respect to those determined in a shorter range of ingredients variation, as samples of barley (R 2 = 0.61, Pérez-Vendrell etal, 1992) orwheat (R 2 = 0.45, Garnsworthy et al, 2000).On the other hand, prediction results in the current study were poorer than those obtained by Valdes and Leeson (1992a) for complete poultry diets, which might be explained by a higher error of the determination of AMEn by the difference with respect to the direct method.

Conclusions
The results of the current study indicate that AMEn/GE values for poultry of starchy grains and cereal byproducts can be predicted with a good precision using different regression models.The accuracy of the equations was slightly higher for NIRS than for regression models including in vitro digestibility or a combination of chemical constituents of the ingredients studied.All the techniques were less accurate for predicting AMEn than AMEn/GE.This was especially the case of in vitro digestibility, where additional inclusion of several chemical constituents was required to reach a comparable accuracy level than that obtained with the other methods.

Figure 1 .
Figure 1.Relationship between: a) AMEn/GE andNDF content, and b) AMEn andNDF content of the samples studied.

Figure 2 .
Figure 2. Relationship between: a) AMEn/GE and IVOMd, and b) AMEn and IVOMd of the samples studied.

Table 1 .
Chemical composition (%) of the feed ingredients studied

Table 1 (cont.). Chemical
composition (%) of the feed ingredients studied a n: number of samples.b NDF: neutral detergent fibre, deviation.

Table 2 .
Ingredient and chemical composition of the basal diets (% as fed basis)

Table 3 .
In vivo apparent metabolisable energy (AMEn, kcal kg -1 DM), degree of metabolizity of gross energy (AMEn/GE a %) and in vitro dry matter (IVDMd, %) and organic matter (IVOMd, %) digestibilities of the ingredients studied a GE: Gross energy.b SD: Standard deviation.c DDGS: dry distillers grains and solubles.

Table 5 .
Effect of the incubation time with pancreatin on the 2-steps in vitro digestibility of the organic matter (%) of wheat and corn samples

Table 7 .
Coefficients of determination and root mean square errors of calibration (R 2 c, SEC) and cross validation (R 2 cv, SECV) to predict the chemical composition (%) and the energy value of the ingredients studied (n= 94)