Guelph, Guelph, Ontario, Canada c Department of Animal Science, Faculty of Agricultural Science, University of Guilan,

This article was downloaded by: [University of Guelph] On: 01 October 2014, At: 11:40 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK Journal of Applied Animal Research Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/taar20 Evaluation of broiler chicks responses to protein, methionine and tryptophan using neural network models A. Faridi a, A. Golian a, J. France b, A. Heravi Mousavi a & M. Mottaghitalab c a Department of Animal Sciences, Centre of Excellence, Ferdowsi University of Mashhad, Mashhad, Iran b Department of Animal and Poultry Science, Centre for Nutrition Modelling, University of Guelph, Guelph, Ontario, Canada c Department of Animal Science, Faculty of Agricultural Science, University of Guilan, Rasht, Iran Published online: 05 Feb 2014. To cite this article: A. Faridi, A. Golian, J. France, A. Heravi Mousavi & M. Mottaghitalab (2014) Evaluation of broiler chicks responses to protein, methionine and tryptophan using neural network models, Journal of Applied Animal Research, 42:3, 327-332, DOI: 10.1080/09712119.2013.867860 To link to this article: http://dx.doi.org/10.1080/09712119.2013.867860 PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the Content ) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http:// www.tandfonline.com/page/terms-and-conditions

Journal of Applied Animal Research, 2014 Vol. 42, No. 3, 327 332, http://dx.doi.org/10.1080/09712119.2013.867860 Evaluation of broiler chicks responses to protein, methionine and tryptophan using neural network models A. Faridi a *, A. Golian a, J. France b, A. Heravi Mousavi a and M. Mottaghitalab c a Department of Animal Sciences, Centre of Excellence, Ferdowsi University of Mashhad, Mashhad, Iran; b Department of Animal and Poultry Science, Centre for Nutrition Modelling, University of Guelph, Guelph, Ontario, Canada; c Department of Animal Science, Faculty of Agricultural Science, University of Guilan, Rasht, Iran (Received 30 July 2013; accepted 7 October 2013) In the present study, neural networks (NN) were developed to investigate the responses (average daily gain [ADG] and feed efficiency [FE]) of broiler chicks to protein and two amino acids (AA), namely methionine (Met) and tryptophan (Trp). Separate NN models were constructed for responses to protein and Met and protein and Trp. Comparisons between NN and response surface models revealed higher accuracy of prediction with the NN models. The relative importance of the input variables (protein and AA) on model output (ADG and FE) was assessed using a sensitivity analysis technique. Results indicated that dietary protein is a more important variable than AA (Met and Trp). Optimal values of the input variables (protein and AA) required to maximize ADG and FE in were obtained by subjecting all constructed NN to an optimization algorithm. The optimization algorithm for the protein and Met response models revealed that diets containing 216 g/kg of protein and 5.45 g/kg of Met lead to maximum ADG, whereas maximum FE is achieved with diets containing 222.6 and 5.85 g/kg of protein and Met, respectively. The optimization algorithm for protein and Trp responses showed that 234 g/kg of protein and 2.6 g/kg of Trp in the diet lead to maximum ADG, while maximum FE is achieved with diets containing 240 g/kg of protein and 3.1 g/kg Trp. The optimization results therefore suggest that protein and AA requirements for maximum FE are higher than for maximum ADG. Keywords: methionine; tryptophan; protein; neural networks; sensitivity analysis; optimization 1. Introduction Methionine (Met) is an amino acid (AA) of crucial importance in commercial poultry diets, as it is typically the first-limiting AA. Tryptophan (Trp) is the third or fourth limiting AA, after Met and lysine, in most diets composed of corn and soybean or peanut meal for broiler chickens (Rosa et al. 2001). Determining the AA requirement for growth is affected by many complex responses other than dietary AA balance for protein synthesis. However, the level of dietary protein is one of the most important factors affecting the requirements of chickens for AA (Abebe & Morris 1990; Garcia Neto et al. 2000). Combination of mathematical modeling and the continuous stream of information on AA responses in broiler chickens can lead to advances in understanding and more precise determination of the requirements of these birds. In this regard, the use of nonlinear data mining tools such as neural networks (NN), due to their practicality and successful application in several areas of poultry nutrition (Faridi & Golian 2011; Faridi et al. 2011), has much to offer. NN present a relatively new class of computing systems inspired by the biological structure of the human brain. On the basis of NN models ability to learn from examples, very complex and analytically unknown dependencies between input and output patterns can be estimated (Gabrijel & Dobnikar 2003). The objectives of this study were to: (1) develop separate NN models to predict the responses (average daily gain [ADG] and feed efficiency [FE]) of broiler chicks to protein and Met and protein and Trp; (2) determine the relative importance of protein and these two limiting AA on ADG and FE using sensitivity analysis; (3) find the optimum level of dietary protein and AA required to maximize ADG and FE of chicks; and (4) compare the predictive ability of NN models with response surface (RS) models. 2. Material and methods 2.1. Description of the data-sets The necessary information for developing the models was obtained from published papers, so Animal Care and Use Committee approval was not necessary for this study. Two data-sets, containing 40 dose-response treatments apiece for Met (Morris et al. 1992) and Trp (Abebe & Morris 1990) were extracted from the literature, and subsequently used for training and testing the *Corresponding author. Email: ako.faridi@stu-mail.um.ac.ir 2014 Taylor & Francis

328 A. Faridi et al. NN. Description of the data used to develop NN models for ADG and FE of broiler chicks in response to protein and AA is summarized in Table 1. 2.2. Neural network model development NN architectures have been discussed in detail in the literature (Basheer & Hajmeer 2000). Determining an appropriate network topology is one of the most critical tasks in NN model development. The NN topology is determined by size, synaptic connection weights, and the hidden-units activation function (Andrews et al. 2008). In our study, separate feed forward multilayer perceptrons, with two input variables each, were employed to predict ADG and FE in chicks during starting period. Variables of interest in the ADG and FE prediction models were the same and comprised protein and Met and protein and Trp. In all models developed, the hyperbolic tangent was used as an activation function and quasi-newton was used as a training algorithm. The data-set for each AA was divided into a training set (n = 28 data lines) and a testing set (n = 12). The training set was used to adjust the connection weights, whereas the test set was used to assess performance. The range of data used to develop the NN models is summarized in Table 2. The Statistica Neural Networks software version 8.0 was used to construct and train the NN (StatSoft 2009). Quantitative examination of the predictive ability of both models was made by R 2, mean square (MS) error, and bias. Table 1. Description of the data used to develop neural network and response surface models. 2.3. Sensitivity analysis A sensitivity analysis technique indicates the input variables considered the most important in model developed. In our study, the missing value problem method was applied to determine the relative importance of the input variables (protein and AA) to model output (ADG and FE). In this method, each input variable is replaced in turn with missing values and the effect upon the output error, named the variable sensitivity error (VSE), is assessed. By the same token, the variable sensitivity ratio (VSR) is a relative indication of the ratio between the VSE and the error in the model developed when all the variables are available (Hunter et al. 2000). The more important variable is matched with the higher VSR (Lou & Nakai 2001; Faridi & Golian 2011). 2.4. Model optimization Optimization is defined as finding a set of values for the input variables for which the predictive model yield the desired response. Therefore, optimized ADG and FE models describe the levels of dietary protein and AA required for maximum ADG and FE. The random search method, a common optimization method provided in Statistica (StatSoft 2009), was applied to the models developed. In this method, iterative random samples of the input variables are taken and the model predictions computed and compared with the best values found from previous iterations. If the newly found values are better than the previous ones, the new results are stored. This process is repeated until the end of the iterations is Amount (g/kg of diet) ADG (g/bird per day) FE (g gain/g feed intake) Model a Protein Amino acid b Min Max Mean ± SD Min Max Mean ± SD Methionine 140 2.95, 3.16, 3.37, 3.58, 3.79 15.2 19.6 17.8 ± 1.75 0.43 0.51 0.48 ± 0.033 160 3.38, 3.62, 3.86, 4.1, 4.34 18 21.4 20.1 ± 1.43 0.45 0.52 0.5 ± 0.028 180 3.8, 4.07, 4.34, 4.61, 4.88 18.9 21.5 20.74 ± 1 0.5 0.57 0.55 ± 0.026 200 4.22, 4.52, 4.82, 5.12, 5.42 21.5 23.3 22.8 ± 0.73 0.58 0.6 0.59 ± 0.01 220 4.31, 4.64, 4.97, 5.3, 5.63 20.7 24.1 22.8 ± 1.42 0.53 0.62 0.58 ± 0.04 240 4.7, 5.06, 5.42, 5.78, 6.14 20.9 23.4 22.54 ± 1.05 0.57 0.61 0.59 ± 0.015 260 5.1, 5.49, 5.88, 6.27, 6.66 20.3 23.5 22.5 ± 1.3 0.56 0.62 0.6 ± 0.023 280 5.5, 5.9, 6.33, 6.75, 7.17 21.4 23.6 23 ± 0.93 0.58 0.63 0.6 ± 0.018 Tryptophan 160 1.2, 1.44, 1.68, 1.92, 2.16 12 16.2 14.78 ± 1.76 0.47 0.55 0.52 ± 0.032 180 1.35, 1.62, 1.9, 2.16, 2.43 13.2 18.4 16.28 ± 2.12 0.48 0.58 0.55 ± 0.04 200 1.5, 1.8, 2.1, 2.4, 2.7 14 18.9 17.38 ± 2.04 0.51 0.6 0.57 ± 0.034 220 1.65, 1.98, 2.31, 2.64, 2.97 15.7 19 18.22 ± 1.4 0.55 0.63 0.6 ± 0.035 240 1.8, 2.16, 2.52, 2.9, 3.24 15.6 20.3 18.82 ± 1.9 0.55 0.63 0.61 ± 0.036 260 1.95, 2.34, 2.73, 3.12, 3.51 17.6 20.2 19.58 ± 1.1 0.59 0.65 0.63 ± 0.028 280 2.1, 2.52, 2.94, 3.36, 3.78 19 20.4 19.94 ± 0.56 0.61 0.66 0.64 ± 0.021 300 2.25, 2.7, 3.15, 3.6, 4.05 18.7 20 19.64 ± 0.55 0.6 0.65 0.63 ± 0.02 Note: ADG, average daily gain; FE, feed efficiency. a Methionine model (data from Morris et al. 1992; covers days 4 18 of age) and tryptophan model (data from Abebe and Morris 1990; covers days 4 18 of age). b The amino acid is the same in the model.

Journal of Applied Animal Research 329 Table 2. Ranges for the data used to develop the neural network and response surface models for broiler chicks responses to protein and amino acids (Met and Trp). Methionine Tryptophan Entity Min Max Mean ± SD Min Max Mean ± SD Protein (g/kg of diet) 140 280 210 ± 46.4 160 300 230 ± 46.4 Amino acid (g/kg of diet) 2.95 7.17 4.87 ± 1.05 1.2 4.05 2.41 ± 0.7 ADG (g/bird per day) 15.2 24.1 21.5 ± 2.1 12 20.4 18.8 ± 2.22 FE (g gain/g feed intake) 0.436 0.631 0.57 ± 0.052 0.473 0.661 0.6 ± 0.05 ADG, average daily gain; FE, feed efficiency. Table 3. Response surface models to predict average daily gain (ADG) and feed efficiency (FE) in response to protein and amino acids (Met and Trp). Methionine YADG ¼ 8.365 þ 0.06x 1 þ 8.27x 2 þ 0.08x 1 x 2 0.0011x 2 1 2.43x2 2 reached. The random search should be confined to the data range within which the models were developed; otherwise it may lead to unreliable results. 2.5. RS model Prediction ability of the NN models was compared to that of RS models. The RS methodology is a collection of statistical techniques for designing experiments, building models, and evaluating the effects of factors. ADG and FE were the investigated response functions, which were described using an empirical equation. The equation is as follows (Lou & Nakai 2001): y ¼ a 0 þ Xk a i x i þ X X aij x i x j þ Xk a ii x 2 i þ e i¼1 i<j i¼1 where x i and x j are the input variables (protein and AA) which influence the response of interest (y; ADG and YFE ¼ 0.0464 þ 0.0022x 1 þ 0.12x 2 þ 0.0012x 1 x 2 0.000018x 2 1 0.035x2 2 Tryptophan YADG ¼ 6.342 þ 0.0961x 1 þ 8.39x 3 þ 0.033x 1 x 3 0.000353x 2 1 2.77x2 3 Note: x 1, x 2, and x 3 denote protein, Met, and Trp, respectively. YFE ¼ 0.013 þ 0.0024x 1 þ 0.18x 3 þ 0.00068x 1 x 3 0.0000078x 2 1 0.059x2 3 FE), k is the number of input variables (k = 2), the coefficients a 0, a i, a ij, and a ii are the model parameters estimated by a multiple linear regression analysis using the least-squares method in the JMP statistical software (JMP 2007), and ε is the residual associated with the experiment. The RS models were developed on the same training set (n = 28) used to develop the NN models, and the testing set (n = 12) was used to evaluate their performance. 3. Results The RS equations developed for ADG and FE are shown in Table 3, while the goodness of fit statistics for the NN and RS models developed are shown in Tables 4 and 5, respectively. The NN models could accurately (R 2 > 0.90) predict ADG and FE in the testing data-sets, which were not used during the training processes. This is evidence Table 4. Statistics and information on neural network models developed for response of broiler chicks to protein and amino acids (Met and Trp). Methionine Tryptophan ADG FE ADG FE Entity Training Testing Training Testing Training Testing Training Testing Statistics R 2 0.94 0.987 0.97 0.981 0.98 0.98 0.97 0.99 MS error 0.234 0.181 0.00007 0.0001 0.09 0.133 7 10 5 0 Bias 0.008 0.15 0.002 0.001 0.009 0.037 7 10 5 0 Information Type of network Three-layer perceptron Training algorithm Quasi-Newton No. of hidden neuron 5 5 5 5 No. of data lines 28 12 28 12 28 12 28 12 ADG, average daily gain; FE, feed efficiency.

330 A. Faridi et al. Table 5. Statistics and information on response surface models developed for response of broiler chicks to protein and amino acids (Met and Trp). Methionine Tryptophan ADG FE ADG FE Entity Training Testing Training Testing Training Testing Training Testing Statistics R 2 0.87 0.94 0.83 0.90 0.92 0.93 0.93 0.97 MS error 2.9 2.5 0.0004 0.0004 0.29 0.46 0.0001 0.00006 Bias 1.5 1.16 0.009 0.0044 0.09 0.14 0.0057 0 Information Training algorithm Least-squares method No. of data lines 28 12 28 12 28 12 28 12 ADG, average daily gain; FE, feed efficiency. that overlearning did not occur with these models. Although performance of the RS models was acceptable, comparisons showed higher accuracy of prediction for the NN. Comparisons between experimental broiler chick response (ADG and FE) and NN-predicted response to Met and Trp are shown in Figures 1 and 2, respectively. The NN models developed were subjected to a sensitivity analysis test to determine the relative significance of the input variables and their rank in order of (a) ADG (g/ bird per day) 26 24 22 20 18 16 0.55 0.5 0.45 14 14 16 18 20 22 24 26 ADG (g/ bird per day) (b) 0.65 0.6 FE (g gain/ g intake) 0.4 0.4 0.45 0.5 0.55 0.6 0.65 FE (g gain/ g intake) Figure 1. Comparison between predicted and observed average daily gain (ADG; a) and feed efficiency (FE; b) in response to protein and Met for the neural network models developed. importance. Relative importance of the input variables was determined using the VSE and VSR values. In all the models developed, the level of dietary protein was the most influential input variable affecting ADG and FE of broiler chicks. The VSR obtained for the model output (ADG and FE), with respect to dietary levels of protein and AA, are shown in Table 6. Optimization of NN helps to address the question of what levels of dietary protein and AA lead to maximum (a) ADG (g/ bird per day) (b) FE (g gain/ g intake) 21 19 17 15 13 11 11 13 15 17 19 21 ADG (g/ bird per day) 0.7 0.65 0.6 0.55 0.5 0.45 0.45 0.5 0.55 0.6 0.65 0.7 FE (g gain/ g intake) Figure 2. Comparison between predicted and observed average daily gain (ADG; a) and feed efficiency (FE; b) in response to protein and Trp for the neural network models developed.

Journal of Applied Animal Research 331 Table 6. Sensitivity analysis for input variables in the neural network models for average daily gain (ADG) and feed efficiency (FE) in response to protein and amino acids (Met and Trp). Input variables Model Protein (g/kg of diet) Amino acid (g/kg of diet) Protein and methionine ADG VSR 40.14 8.18 FE VSR 34 20.18 Protein and tryptophan ADG VSR 29.1 7 FE VSR 16.2 9.4 VSR, variable sensitivity ratio. ADG and FE. The optimization results for all the NN models developed for response to protein and AA are summarized in Table 7. The optimization results for protein and Met show that maximum ADG (23.44 g/bird per day) and FE (0.61 g gain/g feed intake) were achieved with diets contained 216 and 222.6 g/kg of protein and 5.45 and 5.85 g/kg of Met, respectively. Optimization for protein and Trp revealed that diets containing 234 g/kg of protein and 2.6 g/kg of Trp lead to maximum ADG (19.7 g/bird per day), whereas maximum FE (0.64 g gain/g feed intake) was achieved with diets containing 240 g/kg of protein and 3.1 g/kg of Trp. 4. Discussion The present study was designed to investigate the responses (ADG and FE) of broiler chicks to protein and limiting AA through NN. Separate NN models were developed and compared with RS models in terms of predictive accuracy. High values of R 2 indicated that the relationships between dietary nutrients and broiler chick responses were well represented by the NN and RS models developed, though accuracy of the NN predictions was higher in both the training and the testing sets (Tables 4 and 5). Nonlinearity of response of broiler chicks to dietary nutrients made NN models, as nonlinear approximators, more efficient and accurate. In building NN, a key design decision is the number of layers and the corresponding neurons in each layer. It has been shown that NN models with one hidden layer will predict any continuous function (Cybenko 1989), therefore all NN developed here had one hidden layer, while the number of hidden neurons was selected based on the predictive performance of the models. In this study, sensitivity analysis was used to determine how different values of an independent variable impact a particular dependent variable. Moreover, variables that can be ignored in the analysis (VSR 1) can be identified as well as variables that must be retained (VSR 1). The greater the VSR value, the more important is the input variable. Studies designed to evaluate the effect of protein and limiting AA on broiler performance (Garcia Neto et al. 2000; Si et al. 2004) are in agreement with results obtained here (VSR 1; Table 6). The overall values of VSR indicated that in all NN the responses of broiler chicks were more sensitive to dietary protein than to dietary AA. After analyzing 55 experiments, Pesti (2009) concluded that dietary protein is a very significant contributor to variation in both broiler ADG and FE, even when purified AA are added to lower the protein level of the diet. Optimization is another goal in developing NN. The purpose of optimizing is to maximize or minimize some of the output variables by finding a set of design parameters that the aforementioned variables (model output) are dependent upon. In this paper, the objectives of optimization were to find the optimal levels of dietary protein and AA that maximize the ADG and FE responses. The optimization algorithm showed that Met required for optimum ADG and FE is a fixed percentage of dietary protein (2.5%). In agreement with these results, Morris et al. (1992) concluded that Met concentration in chick Table 7. Optimization of the neural network models to achieve maximum average daily gain (ADG) and feed efficiency (FE) in response to protein and amino acid (Met and Trp). Optimal value of input variable Model Protein (g/kg of diet) Amino acid (g/kg of diet) output variable at optimal point Protein and methionine ADG 216 5.45 Maximum ADG = 23.44 g/bird per day FE 222.6 5.85 Maximum FE = 0.61 g gain/g feed intake Protein and tryptophan ADG 234 2.6 Maximum ADG = 19.8 g/bird per day FE 240 3.1 Maximum FE = 0.64 g gain/g feed intake

332 A. Faridi et al. diets should not be less than 0.025 times dietary protein concentration. The optimization results for Trp showed that Trp required for optimum ADG and FE was 1.1% and 1.3% of dietary protein, respectively. As for Met, Morris and co-workers advocated a fixed ratio of dietary Trp to protein as requirement and stated that the amounts of Trp required for maximum growth and FE were each linear functions (12 g/kg) of dietary protein (Abebe & Morris 1990). Reported Trp requirements for broiler chicks during the starter period have varied widely. Levels have been suggested from as low as 1.3 1.6 to as high as 2.5 g per kg of diet (Hewitt & Lewis 1972; Smith & Waldroup 1988; Rogers & Pesti 1989). A likely cause of some of the variation in Trp requirement estimates is the difficulty in determining the amount of Trp in basal diets (Rosa & Pesti 2001). However, based on the current analysis, the National Research Council (NRC, 1994) Trp requirement estimate of 2 g/kg diet is probably too low for broilers chicks and at least 2.6 g Trp in the diet is necessary to maximize ADG during the starter period. The same findings were reported by Rosa and Pesti (2001), where the data on Trp requirement from the literature were pooled for nonlinear regression fitting to achieve the best estimate of the starter chick s requirement. Our results indicated that in all the NN developed the protein and AA requirements for maximum FE were higher than those for ADG. Similar results for Trp have been obtained in previous studies (Rosa & Pesti 2001). However, it is worth pointing out that the optimization approach conducted here, like others (Rosa & Pesti 2001; Pesti 2009), is flawed from the economic perspective as it disregards prices. The most economical levels of dietary protein and AA in feed may not necessarily be the levels that are required for maximum growth but the levels in diets providing the largest difference between costs and returns (Costa et al. 2001). 5. Conclusions In conclusion, NN models appear to provide a more accurate option than RS models in predicting the response of broiler chick to protein and AA. Sensitivity analysis revealed that in all the models developed, the responses of broiler chicks were more sensitive to protein than to AA (Met and Trp). The optimization results indicated that protein and AA requirements for maximum FE were higher than those for maximum ADG. References Abebe S, Morris TR. 1990. Effects of protein concentration on responses to dietary tryptophan by chicks. Br Poult Sci. 31:267 272. Andrews EAM, Morris Q, Bonner AJ. 2008. Neural networks approaches for discovering the learnable correlation between gene function and gene expression in mouse. Neurocomputing. 71:3168 3175. Basheer IA, Hajmeer M. 2000. Artificial neural networks: fundamentals, computing, design, and application. J Microbiol Methods. 43:3 31. Costa EF, Miller BR, Houston JE, Pesti GM. 2001. Production and profitability responses to alternative protein sources and levels in broiler rations. J Agric Appl Econ. 33:567 581. Cybenko G. 1989. Approximation by superpositions of a sigmoidal function. Math Cont Sign Syst. 2:303 314. Faridi A, Golian A. 2011. The use of neural network models to estimate early egg production in broiler breeder hens through dietary nutrient intake. Poult Sci. 90:2897 2903. Faridi A, Mottaghitalab M, Darmani-Kuhi H, France J, Ahmadi H. 2011. Predicting carcass energy content and composition in broilers using the group method of data handling-type neural networks. J Agric Sci Cam. 149:249 254. Gabrijel I, Dobnikar A. 2003. On-line identification and reconstruction of finite automata with generalized recurrent neural networks. Neural Netw. 16:101 120. Garcia Neto M, Pesti GM, Bakalli RI. 2000. Influence of dietary protein level on the broiler chicken s response to methionine and betaine supplements. Poult. Sci. 79:1478 1484. Hewitt D, Lewis D. 1972. The amino acid requirements of the growing chick. 1. Determination of amino acid requirements. Br Poult Sci. 13:449 463. Hunter A, Kennedy L, Henry J, Ferguson I. 2000. Application of neural networks and sensitivity analysis to improved prediction of trauma survival. Comput Meth Prog Biomed. 62:11 19. JMP. 2007. JMP user s guide. Version 7.0. Cary (NC): SAS Institute. Lou W, Nakai S. 2001. Artificial neural network-based predictive model for bacterial growth in a simulated medium of modified-atmosphere-packed cooked meat products. J Agric Food Chem. 49:1799 1804. Morris TR, Gous RM, Abebe S. 1992. Effects of dietary protein concentration on the response of growing chicks to methionine. Br Poult Sci. 33:795 803. [NRC] National Research Council. 1994. Nutrient requirements of poultry. 9th Revised ed. Washington (DC): National Academy of Sciences. Pesti GM. 2009. Impact of dietary amino acid and crude protein levels in broiler feeds on biological performance. J Appl Poult Res. 18:477 486. Rogers SR, Pesti GM. 1989. Effect of protein on the tryptophan requirement of the growing chick and subsequent effects on lipid metabolism. Poult Sci. 68:200(Abstr.). Rosa A, Pesti GM. 2001. Estimation of the tryptophan requirement of chickens for maximum body weight gain and feed efficiency. J Appl Poult Res. 10:135 140. Rosa AP, Pesti GM, Edwards Jr HM, Bakalli RI. 2001. Tryptophan requirement of different broiler genotypes. Poult Sci. 80:1718 1722. Si J, Fritts CA, Waldroup PW, Burnham DJ. 2004. Effects of tryptophan to large neutral amino acid ratios and overall amino acid levels on utilization of diets low in crude protein by broilers. J Appl Poult Res. 13:570 578. Smith Jr NK, Waldroup PW. 1988. Estimation of the tryptophan requirement of male broiler chickens. Poult Sci. 67:1174 1177. StatSoft. 2009. Statistica data analysis software system. Version 7.1. Tulsa, OK: StatSoft Inc.