Cost-effective and rapid quantitative detection of amino acids in milk based on milk Fourier-transform mid-infrared spectroscopy and machine learning techniques

Rapid assessment of milk amino acids (AA) is essential for milk quality monitoring and AA-enriched product development. Herein, we used Fourier-transform mid-infrared (MIR) spectroscopy combined with machine learning (ML) to develop prediction models for milk AA contents. Dataset comprised 789 milk samples from dairy cows, buffaloes, goats, sheep, camels, and donkeys. First, dataset was divided into calibration, farm-independent external validation (FEV), and species-independent external validation (SEV) sets. Second, ten ML methods were used to develop models. Finally, model performance was assessed using cow-independent cross-validation, FEV, and SEV. Most methods achieved similar predictive accuracy in cross-validation. In a limited FEV set, spike and slab regression, elastic net, and least absolute shrinkage and selection operator tended to show higher predictive accuracy. The ratio of performance to deviation (RPD) of the “best” model ranged from 2.07 to 3.39 in cross-validation, 1.18 to 3.22 in FEV, and 0.15 to 8.59 in SEV. Satisfactory accuracy was observed for Ile, Leu, Phe, Val, Lys, Asp, Glu, Ala, Tyr, Arg, total essential AA, and total AA models (RPD ≥ 2.0). In SEV, the “best” model exhibited the highest performance for goat milk. These results provide a basis for the accurate, rapid, and cost-effective quantification of milk AA.