PSX-2 Data Analytics of a Global Nirs Database of Soybean Meal Nutrient Content

Abstract

Abstract The nutrient and energy content of soybean meal (SBM) is variable due to genetics, country of origin, agronomic conditions, and to a lesser extent, the solvent-extraction process. The analytical variability among and within-labs can increase the difficulty of detecting differences among sources and establish nutrient interrelationships to estimate some nutrients based on others that are less variable. Near-infrared reflectance spectroscopy (NIRS) can reduce laboratory analytical inconsistencies and increase the sample size for feedstuff evaluation enhancing feed formulation accuracy. The PNE system (Adisseo) has unique NIRS curves based on direct calibrations generated with in vivo data of standard chicken metabolizable energy (AME and AMEn) and amino acid (AA) digestibility methods. This study evaluated the relationship between AA concentrations and crude protein (CP) and among proximate parameters (CP, crude fiber, ether extract, nitrogen-free extract, and ash), AA concentrations, and digestibilities with AME AMEn. The PNE database used contained 77,573 samples of SBM from Argentina (9,684), Brazil (62,322), and the U.S. (5,567) scanned in diverse countries worldwide between 2018 and 2021. Correlation and regression analyses between AA and CP were conducted. The pairwise correlations (P < 0.001) between AA and CP ranged between 0.34 and 0.82. Concentrations of all AAs in SBM can be predicted (P < 0.001) from the positive linear relationships with CP content. However, all coefficients of determination (R2) were low (0.11 to 0.68). The SBM country of origin was a significant factor and interacted with CP (P < 0.001), but R2 improved only slightly when included in these linear models. Multiple linear regression (MLR) and neural networks (NN) analyses were conducted to estimate AME and AMEn based on proximate values, AA concentrations and digestibility, and country of origin. All factors were significant (P < 0.001) in the MLR models for AME (R2 = 0.78, AICc = 643,940.6) and AMEn (R2 = 0.77, AICc = 642,785). Additionally, to have R2 below 0.80, the MLR models had collinearity for digestibility data and branched-chain AAs concentrations. The NN models had a better fit for training and validation datasets for AME (R2 = 0.88, RASE = 39.96) and AMEn (R2 = 0.87, RASE = 37.21). The NN models allowed us to visualize the greater importance of AA digestibility coefficients (His, Val, Trp, Cys, Arg, Leu, Thr) than the concentration of AAs or proximate values to estimate SBM energy values. In conclusion, AA concentrations of SBM cannot be accurately predicted from its CP content. The AA content of SBM varies by country of origin and may vary independently of CP content. The energy value of SBM may depend more on AA digestibility than proximate parameters. Data analytics of NIRS databases with direct calibration of in vivo parameters of energy utilization and AA digestibility elucidates relevant nutritional factors.