Abstract: The purpose of this study was to establish a rapid batch determination method for free arginine, histidine and isoleucine in milk by mid-infrared spectroscopy, and to carry out a large number of external verifications. A total of 217 Chinese Holstein milk samples from 4 provinces in North China, Central China and Northwest China were taken as the research objects, using 4 spectral preprocessing algorithms (SG smoothing, difference, multivariate scattering correction, standard normal transformation), 4 feature selection algorithms (known information region, adaptive heavy weighting algorithm, genetic algorithm and minimum angle regression algorithm) and 2 modeling algorithms (partial least squares regression and ridge regression), the MIR spectral quantitative prediction models of free arginine, histidine and isoleucine contents in milk were established. The optimal model was applied to the MIR spectra of 32 559 milk samples collected from 4 690 cows in 9 different dairy farms to explore the effects of lactation stage, pasture, parity and season on the predicted arginine, histidine and isoleucine contents by MIR. The results show that:1) The prediction model of arginine content based on CARS feature selection algorithm, non-spectral pretreatment algorithm and PLSR modeling algorithm was the best, R_P²=0.58, RMSEp=6.89 nmol·mL^-1; The prediction model of histidine content based on CARS feature selection algorithm, SG smoothing (window length is 11, 2-order polynomial) pretreatment and PLSR modeling algorithm was the best, R_P²=0.56, RMSEp=0.88 nmol·mL^-1; Based on 274 characteristic information wave points, SG smoothing (window length is 29, 3-order polynomial) pretreatment and PLSR modeling algorithm, the prediction model of isoleucine content was the best, R_P²=0.49, RMSEp=1.75 nmol·mL^-1; 2) When the optimal model was verified externally across regions, the prediction accuracy was reduced; 3) Applying the established model to the large-scale spectral database of E province (not participating in the establishment of the model), the contents of free arginine, histidine and isoleucine in milk was predicted, it was found that lactation stage, pasture and season had significant effects on the contents of free arginine, histidine and isoleucine in milk (P<0.001), while parity had no significant effect on arginine content, but had significant effect on histidine and isoleucine (P<0.001). The results show that it is feasible to predict the content of free amino acids in milk by MIR, especially, it has certain predictive ability in the trend analysis of milk amino acid content, and the prediction model needs more representative samples to optimize, so as to improve the accuracy and universality of the model.

Key words: mid-infrared spectroscopy(MIR), milk amino acid, prediction model, milk, machine learning