Biological validation of milk mid-infrared-predicted dry matter intake, nitrogen use efficiency, and methane production across lactation in individual dairy cows.

Dry matter intake (DMI), nitrogen use efficiency (NUE) for milk production, and methane (CH4) production are key sustainability traits in dairy cows. Milk mid-infrared (MIR) spectroscopy offers a scalable, low-cost solution to quantify these traits. The objectives of this study were to 1) develop prediction models for each trait using milk MIR spectral data, 2) evaluate the predictive performance of these models using a year-independent validation strategy, whereby data from one or more years were excluded from model training and used exclusively for testing; 3) assess the biological plausibility of predictions within the same temporal and environmental context as the calibration data, by examining lactation-stage and parity patterns; and 4) assess the biological plausibility of predictions on an independent data set collected outside the calibration context. Data on milk MIR spectra, DMI, NUE for milk production, and CH4 production were collected at the Agroscope research farm in Posieux, Switzerland, between 2015 and 2024. The data sets used for model development included 2,069, 2,067, and 1,744 records from 264, 265, and 185 cows for DMI, NUE, and CH4 production, respectively. Prediction models included linear regression, partial least squares regression (PLSR), and neural networks (NN), evaluated under animal-independent 4-fold cross-validation and a year-independent validation, where data up to 2022 were used for calibration and data from 2023 for validation. Under animal-independent cross-validation, NN consistently outperformed PLSR, with best prediction accuracies of R2 = 0.64 (ratio of performance to deviation, RPD = 1.65) for DMI, R2 = 0.56 (RPD = 1.51) for NUE for milk production, and R2 = 0.52 (RPD = 1.43) for CH4 production. In the year-independent validation, prediction accuracy decreased substantially for all traits, with the linear regression model outperforming spectral-based models for NUE and CH4 production, suggesting that, for this specific validation scenario, spectral information added little to no predictive value beyond milk yield and stage of lactation. When the best models were applied to a historical database of 9,975 milk MIR spectra from 300 cows collected in the Agroscope experimental farm in Posieux between 2015 and 2024, predicted values reflected known biological patterns: NUE decreased over the course of lactation, DMI peaked at mid-lactation, and CH4 production rose rapidly in early lactation. Multiparous cows showed consistently higher DMI and CH4 production than primiparous cows, although their NUE only exceeded that of primiparous cows in early to mid-lactation. When applied to an independent data set of 1,740 records from 1,706 cows on 39 new farms, the recovered biological patterns were inconsistent across traits and models. These results confirm the potential of MIR spectroscopy as a phenotyping tool when applied within the same environmental and temporal context as model development, while highlighting the need for caution when models are used on data from different farms and time periods.