Prediction of methane emissions in Austrian dairy cows using milk mid-infrared spectra and milk fatty acids: A nested cross-validation approach.

Milk mid-infrared (MIR) spectroscopy offers a rapid and cost-effective method for quantifying milk components and has shown potential for predicting physiological and environmental traits in dairy cattle, including methane (CH4) emissions. However, milk MIR-based CH4 predictions remain a black-box, as the biological mechanisms linking milk MIR spectra to CH4 emissions are not understood. This study aimed to predict CH4 emissions in dairy cows, measured using the GreenFeed system, based on milk MIR spectra and milk fatty acids (MFA). By comparing milk MIR-based models with those using MFA as predictors, we intended to disentangle the underlying biological signals associated with CH4 emissions. Data from 16 commercial Austrian dairy farms were available. All farms had a similar feeding system based on grass and corn silage, supplemented by varying amounts of concentrated feed and without grazing. Data included 807 records from 604 cows. Six partial least squares regression models were developed to predict CH4 emissions by using various combinations of predictor variables, including milk MIR spectra, MFA, major milk components, milk yield, days in milk and parity. A nested cross-validation framework was applied using 2 strategies: 5-fold cow-independent and leave-one-farm-out cross-validation. Mean CH4 emissions were 436 g/d. Positive correlations were observed between CH4 and de novo MFA throughout lactation, whereas preformed MFA showed negative associations, particularly in early lactation. These associations reflect signals associated with energy status. In 5-fold cow-independent cross-validation, the model based on milk MIR spectra showed moderate predictive accuracy (r = 0.52, R2 = 0.27, RMSE = 60 g/d), while the model based on milk MFA performed lower (r = 0.42, R2 = 0.18, RMSE = 63 g/d). However, combining MFA with fat%, protein% and lactose% slightly increased accuracy (r = 0.44, R2 = 0.20, RMSE = 62 g/d). The additional consideration of milk yield, days in milk and parity improved accuracy of MIR-based (r = 0.62, R2 = 0.39, RMSE = 55 g/d) and MFA-based models (r = 0.60, R2 = 0.36, RMSE = 56 g/d). The model including milk yield, days in milk and parity showed lower predictive performance (r = 0.49, R2 = 0.24, RMSE = 60 g/d), indicating that both MIR spectra and MFA together with fat%, protein% and lactose% provide additional information. Leave-one-farm-out cross-validation resulted in slightly lower predictive accuracies (MIR-based model r = 0.43, R2 = 0.22, RMSE = 63 g/d and MFA-based model: r = 0.41, R2 = 0.21, RMSE = 66 g/d). Models using all predictor traits performed best (MIR-based model r = 0.56, R2 = 0.33, RMSE = 58 g/d and MFA-based model: r = 0.58, R2 = 0.35, RMSE = 57 g/d). Correlations between predictions from the MIR- and MFA-based models including all predictors were high (r = 0.86), indicating a high agreement between the 2 modeling approaches. This study represents one of the first large-scale evaluations of milk MIR-based CH4 prediction models under commercial farm conditions. Milk MIR spectra and MFA explained only part of the variability in CH4 emissions, largely reflecting metabolic signals related to energy status. Correlations between CH4 emissions and MFA suggest that these models will favor cows with a negative energy balance during early lactation and with consistently lower de novo MFA synthesis throughout lactation. Further research is needed to gain more insight into the use of milk mid-infrared spectra for genetic selection to reduce CH4 emissions.