Imputation of missing milk Fourier transform mid-infrared spectra using existing milk spectral databases: A strategy to improve the reliability of breeding values and predictive models.

The use of milk Fourier transform mid-infrared (FT-MIR) spectrometry to develop management and breeding tools for dairy farmers and industry is growing and supported by the availability of numerous new predicted phenotypes to assess the nutritional quality of milk and its technological properties, but also the animal health and welfare status and its environmental fingerprint. For genetic evaluations, having a long-term and representative spectral dairy herd improvement (DHI) database improves the reliabilities of estimated breeding values (EBV) from these phenotypes. Unfortunately, most of the time, the raw spectral data used to generate these estimations are not stored. Moreover, many reference measurements of those phenotypes, needed during the FT-MIR calibration step, are available from past research activities but lack spectra records. So, it is impossible to use them to improve the FT-MIR models. Consequently, there is a strong interest in imputing those missing spectra. The innovative objective of this study was to use the existing large spectral DHI database to estimate missing spectra by selecting probable spectra using, as the match criteria, common dairy traits recorded for a long time by DHI organizations. We tested 4 match criteria combinations. Combination 1 required to have equal fat and protein contents between the sample for which a spectrum was to be estimated and the reference samples in the DHI database. Combination 2 also required an equal urea content. Combination 3 requested equal fat, protein, and lactose contents. Finally, combination 4 included all criteria. When more than one spectrum was found during the search, their average was the estimated spectrum for the query sample. Concretely, this study estimated missing spectra for 1,700 samples using 2,000,000 spectral DHI records. For assessing the effect of this spectral estimation on the prediction quality, FT-MIR equations were used to predict 11 phenotypes, selected as their quantification used different FT-MIR regions. They were related to the milk fat and mineral composition, lactoferrin content, quantity of eructed methane, body weight (BW), and dry matter intake. The accuracy between predictions obtained from actual and estimated spectra was evaluated by calculating the mean absolute error (MAE). The criteria in the fourth and second combinations were too strict to estimate a spectrum for most samples. Indeed, for many samples, no spectra with the same values for those matching criteria was found. The third match criteria combination had a poorer prediction performance for all studied traits and spectral absorptions than the first combination due to fewer matched samples available to compute the missing spectrum. By allowing a range for matching lactose content (±0.1 g/dL milk), we showed that this new combination increased the number of selected samples to compute missing spectra and predict better the infrared absorption at different wavenumbers, especially those related to the lactose quantification. The prediction performance was further improved by performing queries on the entire Walloon DHI spectral database (6,625,570 spectra), and it varied among the studied phenotypes. Without considering the traits used for the matching, the best predictions were obtained for the content of saturated fatty acids (MAE = 0.15 g/dL milk) and BW (MAE = 12.80 kg). Yet, the predictions for the unsaturated fatty acids were less accurate (MAE = 0.13 and 0.018 g/dL milk for monounsaturated and polyunsaturated fatty acids), likely because of the poorer predictions of spectral regions related to long-chain fatty acids. Similarly, poorer predictions were observed for the amount of methane eructed by dairy cows (MAE = 47.02 g/d), likely because it is not directly related to fat content or composition. Prediction accuracies for the remaining traits were also low. In conclusion, we observed that increasing the number of relevant matching criteria helps improve the quality of FT-MIR predicted phenotypes and the number of spectra used during the search. So, it would be of great interest to test in the future the suitability of the developed methodology with large-scale international spectral databases to improve the reliability of EBV from these FT-MIR-based phenotypes and the robustness of FT-MIR predictive models.

Usefulness of mid-infrared spectroscopy as a tool to estimate body condition score change from milk samples in intensively fed dairy cows.

Directly measuring individual cow energy balance is not trivial. Other traits such as body condition score (BCS) and BCS change (ΔBCS) can, however, be used as an indicator of cow energy status. Body condition score is a metric used worldwide to estimate cow body reserves, but the estimation of ΔBCS was, until now, conditional on the availability of multiple BCS assessments. The aim of the present study was to estimate ΔBCS from milk mid-infrared (MIR) spectra and days in milk (DIM) in intensively fed dairy cows using statistical prediction methods. Daily BCS was interpolated from cubic splines fitted through the BCS records and daily ΔBCS was calculated from these splines. The ΔBCS records were merged with milk MIR spectra recorded on the same week. The dataset comprised 37,077 ΔBCS phenotypes across 9,403 lactations from 6,988 cows in 151 herds based in Quebec, Canada. Partial least squares regression (PLSR) and a neural network (NN) were then used to estimate ΔBCS from (1) MIR spectra only, (2) DIM only, or (3) MIR spectra and DIM together. The ΔBCS data in both the first 120 and 305 DIM of lactation were used to develop the estimates. Daily ΔBCS had a standard deviation of 4.40 × 10-3 BCS units in the 120-d dataset and of 3.63 × 10-3 BCS units in the 305-d dataset. A 4-fold cross-validation was used to calibrate and test the prediction equations. External validation was also conducted using more recent years of data. Irrespective of whether based on the first 120 or 305 DIM, or when MIR spectra only, DIM only or MIR spectra and DIM were jointly used as prediction variables, NN produced the lowest root mean square error (RMSE) of cross-validation (1.81 × 10-3 BCS units and 1.51 × 10-3 BCS units, respectively, using the 120-d and 305-d dataset). Relative to predictions for the entire 305 DIM, the RMSE of cross-validation was 15.4% and 1.5% lower in the first 120 DIM when using PLSR and NN, respectively. Predictions from DIM only were more accurate than those using just MIR spectra data but, irrespective of the dataset and of the prediction model used, combining DIM information with MIR spectral data as prediction variables reduced the RMSE compared with the inclusion of DIM alone, albeit the benefit was small (the RMSE from cross-validation reduced by up to 5.5% when DIM and spectral data were jointly used as model features instead of DIM only). However, when predicting extreme ΔBCS records, the MIR spectral data were more informative than DIM. Model performance when predicting ΔBCS records in future years was similar to that from cross-validation demonstrating the ability of MIR spectra of milk and DIM combined to estimate ΔBCS, particularly in early lactation. This can be used to routinely generate estimates of ΔBCS to aid in day-to-day individual cow management.

Mid-infrared spectroscopy for large-scale phenotyping of bovine colostrum gross composition and immunoglobulin concentration.

Immunoglobulin G is the fundamental antibody for acquisition of passive transfer of immunity in ruminant newborns. Colostrum, in fact, must be administered as soon as possible after birth to ensure a successful transfer of IgG from the dam to the calf. Assessment of colostrum Ig concentration and gross composition via gold standards is expensive, time consuming, and hardly implementable for large-scale investigations. Therefore, in the present study we evaluated the predictive ability of mid-infrared spectroscopy (MIRS) as an indirect determination method. A total of 714 colostrum samples collected within 6 h from parturition from Italian Holstein cows, 30% primiparous and 70% pluriparous, were scanned using a benchtop spectrometer after dilution in pure water. The prediction models were developed by correlating spectral information with the reference measurements: IgG concentration (93.54 ± 33.87 g/L), total Ig concentrations (102.82 ± 35.04 g/L), and content of protein (14.71 ± 3.51%), fat (4.61 ± 3.04%), and lactose (2.36 ± 0.51 mg/100 mg). We found a good to excellent performance in prediction of colostrum IgG concentration and traditional composition traits in cross-validation (R2CV ≥ 0.92) and a promising and good predictive ability in external validation with R2V equal to 0.84, 0.89, and 0.74 for IgG, protein, and fat, respectively. In the case of IgG and protein content, for example, the coefficient of determination in external validation was greater than 0.84. The other Ig fractions, A and M, presented insufficient prediction accuracy likely due to their extremely low concentration compared with IgG (4.56 and 5.06 g/L vs. 93.54 g/L). The discriminant ability of MIRS-predicted IgG and protein content was outstanding when trying to classify samples according to the quality level (i.e., low vs. high concentration of IgG). In particular, the cut-off that better discriminate low- from high-quality colostrum was 75.40 g/L in the case of the MIRS-predicted IgG and 13.32% for the MIRS-predicted protein content. Therefore, MIRS is proposed as a rapid and cheap tool for large-scale punctual IgG, protein, and lactose quantification and for the screening of low-quality samples. From a practical perspective, there is the possibility to install colostrum models in the MIRS benchtop machineries already present in laboratories in charge of official milk testing. Colostrum phenotypes collected on an individual basis will be useful to breeders for the definition of specific selection strategies and to farmers for management scopes. Finally, our findings may be relevant for other stakeholders, given the fact that colostrum is an emerging ingredient for the animal and human food and pharmaceutical industry.

Effectiveness of mid-infrared spectroscopy for the prediction of cow milk metabolites.

Proton nuclear magnetic resonance (1H NMR) spectroscopy is acknowledged as one of the most powerful analytical methods with cross-cutting applications in dairy foods. To date, the use of 1H NMR spectroscopy for the collection of milk metabolic profile is hindered by costly and time-consuming sample preparation and analysis. The present study aimed at evaluating the accuracy of mid-infrared spectroscopy (MIRS) as a rapid method for the prediction of cow milk metabolites determined through 1H NMR spectroscopy. Bulk milk (n = 72) and individual milk samples (n = 482) were analyzed through one-dimensional 1H NMR spectroscopy and MIRS. Nuclear magnetic resonance spectroscopy identified 35 milk metabolites, which were quantified in terms of relative abundance, and MIRS prediction models were developed on the same 35 milk metabolites, using partial least squares regression analysis. The best MIRS prediction models were developed for galactose-1-phosphate, glycerophosphocholine, orotate, choline, galactose, lecithin, glutamate, and lactose, with coefficient of determination in external validation from 0.58 to 0.85, and ratio of performance to deviation in external validation from 1.50 to 2.64. The remaining 27 metabolites were poorly predicted. This study represents a first attempt to predict milk metabolome. Further research is needed to specifically address whether developed prediction models may find practical application in the dairy sector, with particular regard to the screening of dairy cows' metabolic status, the quality control of dairy foods, and the identification of processed milk or incorrectly stored milk.

Correlation between breeding values for milk fatty acids and Nordic Total Merit index traits for Danish Holstein and Danish Jersey.

Milk fatty acid composition is gaining interest in the Danish dairy industry both to develop new dairy products and as a management tool. To be able to implement milk fatty acid (FA) composition in the breeding program, it is important to know the correlations with the traits in the breeding goal. To estimate these correlations, we measured milk fat composition in Danish Holstein (DH) and Danish Jersey (DJ) cattle breeds using mid-infrared spectroscopy. Breeding values were estimated for specific FA and for groups of FA. Correlations with the estimated breeding values (EBV) underlying the Nordic Total Merit index (NTM) were calculated within breed. For both DH and DJ, we showed that FA EBV had moderate correlations with the NTM and production traits. For both DH and DJ, the correlation of FA EBV and NTM were in the same direction, except for C16:0 (0 in DH, 0.23 in DJ). A few correlations differed between DH and DJ. The correlation between claw health index and C18:0 was negative in DH (-0.09) but positive in DJ (0.12). In addition, some correlations were not significant in DH but were significant in DJ. The correlations between udder health index and long-chain FA, trans FA, C16:0, and C18:0 were not significant in DH (-0.05 to 0.02), but were significant in DJ (-0.17, -0.15, 0.14, and -0.16, respectively). For both DH and DJ, the correlations between FA EBV and nonproduction traits were low. This implies that it is possible to breed for a different fat composition in the milk without affecting the nonproduction traits in the breeding goal.

Prediction of detailed blood metabolic profile using milk infrared spectra and machine learning methods in dairy cattle.

The adoption of preventive management decisions is crucial to dealing with metabolic impairments in dairy cattle. Various serum metabolites are known to be useful indicators of the health status of cows. In this study, we used milk Fourier-transform mid-infrared (FTIR) spectra and various machine learning (ML) algorithms to develop prediction equations for a panel of 29 blood metabolites, including those related to energy metabolism, liver function/hepatic damage, oxidative stress, inflammation/innate immunity, and minerals. For most traits, the data set comprised observations from 1,204 Holstein-Friesian dairy cows belonging to 5 herds. An exception was represented by ß-hydroxybutyrate prediction, which contained observations from 2,701 multibreed cows pertaining to 33 herds. The best predictive model was developed using an automatic ML algorithm that tested various methods, including elastic net, distributed random forest, gradient boosting machine, artificial neural network, and stacking ensemble. These ML predictions were compared with partial least squares regression, the most commonly used method for FTIR prediction of blood traits. Performance of each model was evaluated using 2 cross-validation (CV) scenarios: 5-fold random (CVr) and herd-out (CVh). We also tested the best model's ability to classify values precisely in the 2 extreme tails, namely, the 25th (Q25) and 75th (Q75) percentiles (true-positive prediction scenario). Compared with partial least squares regression, ML algorithms achieved more accurate performance. Specifically, elastic net increased the R2 value from 5% to 75% for CVr and 2% to 139% for CVh, whereas the stacking ensemble increased the R2 value from 4% to 70% for CVr and 4% to 150% for CVh. Considering the best model, with the CVr scenario, good prediction accuracies were obtained for glucose (R2 = 0.81), urea (R2 = 0.73), albumin (R2 = 0.75), total reactive oxygen metabolites (R2 = 0.79), total thiol groups (R2 = 0.76), ceruloplasmin (R2 = 0.74), total proteins (R2 = 0.81), globulins (R2 = 0.87), and Na (R2 = 0.72). Good prediction accuracy in classifying extreme values was achieved for glucose (Q25 = 70.8%, Q75 = 69.9%), albumin (Q25 = 72.3%), total reactive oxygen metabolites (Q25 = 75.1%, Q75 = 74%), thiol groups (Q75 = 70.4%), total proteins (Q25 = 72.4%, Q75 = 77.2.%), globulins (Q25 = 74.8%, Q75 = 81.5%), and haptoglobin (Q75 = 74.4%). In conclusion, our study shows that FTIR spectra can be used to predict blood metabolites with relatively good accuracy, depending on trait, and are a promising tool for large-scale monitoring.

Predicting milk protein fractions using infrared spectroscopy and a gradient boosting machine for breeding purposes in Holstein cattle.

In recent years, increasing attention has been focused on the genetic evaluation of protein fractions in cow milk with the aim of improving milk quality and technological characteristics. In this context, advances in high-throughput phenotyping by Fourier transform infrared (FTIR) spectroscopy offer the opportunity for large-scale, efficient measurement of novel traits that can be exploited in breeding programs as indicator traits. We took milk samples from 2,558 Holstein cows belonging to 38 herds in northern Italy, operating under different production systems. Fourier transform infrared spectra were collected on the same day as milk sampling and stored for subsequent analysis. Two sets of data (i.e., phenotypes and FTIR spectra) collected in 2 different years (2013 and 2019-2020) were compiled. The following traits were assessed using HPLC: true protein, major casein fractions [αS1-casein (CN), αS2-CN, ß-CN, κ-CN, and glycosylated-κ-CN], and major whey proteins (ß-lactoglobulin and α-lactalbumin), all of which were measured both in grams per liter (g/L) and proportion of total nitrogen (% N). The FTIR predictions were calculated using the gradient boosting machine technique and tested by 3 different cross-validation (CRV) methods. We used the following CRV scenarios: (1) random 10-fold, which randomly split the whole into 10-folds of equal size (9-folds for training and 1-fold for validation); (2) herd/date-out CRV, which assigned 80% of herd/date as the training set with independence of 20% of herd/date assigned as the validation set; (3) forward/backward CRV, which split the data set in training and validation set according with the year of milk sampling (FTIR and gold standard data assessed in 2013 or 2019-2020) using the "old" and "new" databases for training and validation, and vice-versa with independence among them; (4) the CRV for genetic parameters (CRV-gen), where animals without pedigree as assigned as a fixed training population and animals with pedigree information was split in 5-folds, in which 1-fold was assigned to the fixed training population, and 4-folds were assigned to the validation set (independent from the training set). The results (i.e., measures and predictions) of CRV-gen were used to infer the genetic parameters for gold standard laboratory measurements (i.e., proteins assessed with HPLC) and FTIR-based predictions considering the CRV-gen scenario from a bi-trait animal model using single-step genomic BLUP. We found that the prediction accuracies of the gradient boosting machine equations differed according to the way in which the proteins were expressed, achieving higher accuracy when expressed in g/L than when expressed as % N in all CRV scenarios. Concerning the reproducibility of the equations over the different years, the results showed no relevant differences in predictive ability between using "old" data as the training set and "new" data as the validation set and vice-versa. Comparing the additive genetic variance estimates for milk protein fractions between the FTIR predicted and HPLC measures, we found reductions of -19.7% for milk protein fractions expressed in g/L, and -21.19% expressed as % N. Although we found reductions in the heritability estimates, they were small, with values ranging from -1.9 to -7.25% for g/L, and -1.6 to -7.9% for % N. The posterior distributions of the additive genetic correlations (ra) between the FTIR predictions and the laboratory measurements were generally high (>0.8), even when the milk protein fractions were expressed as % N. Our results show the potential of using FTIR predictions in breeding programs as indicator traits for the selection of animals to enhance milk protein fraction contents. We expect acceptable responses to selection due to the high genetic correlations between HPLC measurements and FTIR predictions.

Methodological guidelines: Cow milk mid-infrared spectra to predict reference enteric methane data collected by an automated head-chamber system.

Various methodological protocols were tested on milk samples from cows fed diets affecting both methanogenesis and milk synthesis to identify the best approach for the prediction of GreenFeed system (GF) measured methane (CH4) emissions by milk mid-infrared (MIR) spectroscopy. The models developed were also tested on a data set from cows fed chemical inhibitors of CH4 emission [3-nitrooxypropanol (3NOP)] that just marginally affect milk composition. A total of 129 primiparous and multiparous Holstein cows fed diets with different methanogenic potential were considered. Individual milk yield (MY) and dry matter intake were recorded daily, whereas fat- and protein-corrected milk (FPCM) was recorded twice a week. The MIR spectra from 2 consecutive milkings were collected twice a week. Twenty CH4 spot measurements with GF were taken as the basic measurement unit (BMU) of CH4. The equations were built using partial least squares regression by splitting the database into calibration and validation data sets (excluding 3NOP samples). Models were developed for milk MIR spectra by milking and on day spectra obtained by averaging spectra from 2 consecutive milkings. Models based on day spectra were calibrated by using CH4 reference data for a measurement duration of 1, 2, 3, or 4 BMU. Models built from the average of the day spectra collected during the corresponding CH4 measurement periods were developed. Corrections of spectra by days in milk (DIM) and the inclusion of parity, MY, and FPCM as explanatory variables were tested as tools to improve model performance. Models built on day milk MIR spectra gave slightly better performances that those developed using spectra from a single milking. Long duration of CH4 measurement by GF performed better than short duration: the coefficient of determination of validation (R2V) for CH4 emissions expressed in grams per day were 0.60 vs. 0.52 for 4 and 1 BMU, respectively. When CH4 emissions were expressed as grams per kilogram of dry of matter intake, grams per kilogram of MY, or grams per kilogram of FPCM, performance with a long duration also improved. Coupling GF reference data with the average of milk MIR spectra collected throughout the corresponding CH4 measurement period gave better predictions than using day spectra (R2V = 0.70 vs. 0.60 for CH4 as g/d on 4 BMU). Correcting the day spectra by DIM improved R2V compared with the equivalent DIM-uncorrected models (R2V = 0.67 vs. 0.60 for CH4 as g/d on 4 BMU). Adding other phenotypic information as explanatory variables did not further improve the performance of models built on single day DIM-corrected spectra, whereas including MY (or FPCM) improved the performance of models built on the average of spectra (uncorrected by DIM) recorded during the CH4 measurement period (R2V = 0.73 vs. 0.70 for CH4 as g/d on 4 BMU). When validating the models on the 3NOP data set, predictions were poor without (R2V = 0.13 for CH4 as g/d on 1 BMU) or with (R2V = 0.31 for CH4 as g/d on 1 BMU) integration of 3NOP data in the models. Thus, specific models would be required for CH4 prediction when cows receive chemical inhibitors of CH4 emissions not affecting milk composition.

Predicting dry matter intake in Canadian Holstein dairy cattle using milk mid-infrared reflectance spectroscopy and other commonly available predictors via artificial neural networks.

Dry matter intake (DMI) is a fundamental component of the animal's feed efficiency, but measuring DMI of individual cows is expensive. Mid-infrared reflectance spectroscopy (MIRS) on milk samples could be an inexpensive alternative to predict DMI. The objectives of this study were (1) to assess if milk MIRS data could improve DMI predictions of Canadian Holstein cows using artificial neural networks (ANN); (2) to investigate the ability of different ANN architectures to predict unobserved DMI; and (3) to validate the robustness of developed prediction models. A total of 7,398 milk samples from 509 dairy cows distributed over Canada, Denmark, and the United States were analyzed. Data from Denmark and the United States were used to increase the training data size and variability to improve the generalization of the prediction models over the lactation. For each milk spectra record, the corresponding weekly average DMI (kg/d), test-day milk yield (MY, kg/d), fat yield (FY, g/d), and protein yield (PY, g/d), metabolic body weight (MBW), age at calving, year of calving, season of calving, days in milk, lactation number, country, and herd were available. The weekly average DMI was predicted with various ANN architectures using 7 predictor sets, which were created by different combinations MY, FY, PY, MBW, and MIRS data. All predictor sets also included age of calving and days in milk. In addition, the classification effects of season of calving, country, and lactation number were included in all models. The explored ANN architectures consisted of 3 training algorithms (Bayesian regularization, Levenberg-Marquardt, and scaled conjugate gradient), 2 types of activation functions (hyperbolic tangent and linear), and from 1 to 10 neurons in hidden layers). In addition, partial least squares regression was also applied to predict the DMI. Models were compared using cross-validation based on leaving out 10% of records (validation A) and leaving out 10% of cows (validation B). Superior fitting statistics of models comprising MIRS information compared with the models fitting milk, fat and protein yields suggest that other unknown milk components may help explain variation in weekly average DMI. For instance, using MY, FY, PY, and MBW as predictor variables produced a predictive accuracy (r) ranging from 0.510 to 0.652 across ANN models and validation sets. Using MIRS together with MY, FY, PY, and MBW as predictors resulted in improved fitting (r = 0.679-0.777). Including MIRS data improved the weekly average DMI prediction of Canadian Holstein cows, but it seems that MIRS predicts DMI mostly through its association with milk production traits and its utility to estimate a measure of feed efficiency that accounts for the level of production, such as residual feed intake, might be limited and needs further investigation. The better predictive ability of nonlinear ANN compared with linear ANN and partial least squares regression indicated possible nonlinear relationships between weekly average DMI and the predictor variables. In general, ANN using Bayesian regularization and scaled conjugate gradient training algorithms yielded slightly better weekly average DMI predictions compared with ANN using the Levenberg-Marquardt training algorithm.

Predicting methane emission in Canadian Holstein dairy cattle using milk mid-infrared reflectance spectroscopy and other commonly available predictors via artificial neural networks.

Interest in reducing eructed CH4 is growing, but measuring CH4 emissions is expensive and difficult in large populations. In this study, we investigated the effectiveness of milk mid-infrared spectroscopy (MIRS) data to predict CH4 emission in lactating Canadian Holstein cows. A total of 181 weekly average CH4 records from 158 Canadian cows and 217 records from 44 Danish cows were used. For each milk spectra record, the corresponding weekly average CH4 emission (g/d), test-day milk yield (MY, kg/d), fat yield (FY, g/d), and protein yield (PY, g/d) were available. The weekly average CH4 emission was predicted using various artificial neural networks (ANN), partial least squares regression, and different sets of predictors. The ANN architectures consisted of 3 training algorithms, 1 to 10 neurons with hyperbolic tangent activation function in the hidden layer, and 1 neuron with linear (purine) activation function in the hidden layer. Random cross-validation was used to compared the predictor sets: MY (set 1); FY (set 2); PY (set 3); MY and FY (set 4); MY and PY (set 5); MY, FY, and PY (set 6); MIRS (set 7); and MY, FY, PY, and MIRS (set 8). All predictor sets also included age at calving and days in milk, in addition to country, season of calving, and lactation number as categorical effects. Using only MY (set 1), the predictive accuracy (r) ranged from 0.245 to 0.457 and the root mean square error (RMSE) ranged from 87.28 to 99.39 across all prediction models and validation sets. Replacing MY with FY (set 2; r = 0.288-0.491; RMSE = 85.94-98.04) improved the predictive accuracy, but using PY (set 3; r = 0.260-0.468; RMSE = 86.95-98.47) did not. Adding FY to MY (set 4; r = 0.272-0.469; RMSE = 87.21-100.76) led to a negligible improvement compared with sets 1 and 3, but it slightly decreased accuracy compared with set 2. Adding PY to MY (set 5; r = 0.250-0.451; RMSE = 87.66-100.94) did not improve prediction ability. Combining MY, FY, and PY (set 6; r = 0.252-0.455; RMSE = 87.74-101.93) yielded accuracy slightly lower than sets 2 and 3. Using only MIRS data (set 7; r = 0.586-0.717; RMSE = 69.09-96.20) resulted in superior accuracy compared with all previous sets. Finally, the combination of MIRS data with MY, FY, and PY (set 8; r = 0.590-0.727; RMSE = 68.02-87.78) yielded similar accuracy to set 7. Overall, sets including the MIRS data yielded significantly better predictions than the other sets. To assess the predictive ability in a new unseen herd, a limited block cross-validation was performed using 20 cows in the same Canadian herd, which yielded r = 0.229 and RMSE = 154.44, which were clearly much worse than the average r = 0.704 and RMSE = 70.83 when predictions were made by random cross-validation. These results warrant further investigation when more data become available to allow for a more comprehensive block cross-validation before applying the calibrated models for large-scale prediction of CH4 emissions.

Correlations of feed intake predicted with milk infrared spectra and breeding values in the Dutch Holstein population.

Feed is a major cost in dairy production, and substantial genetic variation in feed efficiency exists between cows. Therefore, breeders aim to improve feed efficiency of dairy cattle. However, phenotypic data on individual feed intake on commercial farms is scarce, and accurate measurements are very costly. Several studies have shown that information from Fourier-transformed infrared spectra of milk samples (milk infrared, milk IR) can be used to predict phenotypes such as energy balance and energy intake, but this is usually based on small data sets obtained under experimental circumstances. The added value of information from milk IR spectra for estimation of breeding values is unknown. The objectives of this study were (1) to develop prediction equations for dry matter intake (DMI) and residual DMI (rDMI) from milk IR spectra; (2) to apply these for a data set of milk IR spectra from commercial Dutch dairy farms; (3) to estimate genetic parameters for these traits; and (4) to estimate correlations between these predictions and other traits in the breeding goal. We used data from feeding trials where individual feed intake was recorded daily and for which milk IR spectra were determined weekly to develop prediction equations for DMI and rDMI with partial least squares regression. This data set contained over 7,600 weekly averaged DMI records linked with milk IR spectra from 271 cows. The equations were applied for a data set with test day information from 676 Dutch dairy herds with 621,567 records of 78,488 cows. Both milk IR-predicted DMI and rDMI were analyzed with an animal model to obtain genetic parameters and sire effect estimates that could be correlated with breeding values. A partial least squares regression model with 10 components from the milk IR spectra explained around 25% of DMI variation and less than 10% of rDMI variation in the validation set. Nearly all variation in the milk IR spectra was captured by 7 components; additional components contributed marginally to the spectral variation but decreased prediction errors for both traits. Accuracies of predictions of DMI and rDMI from milk IR spectra for a large feeding experiment were 0.47 and 0.26 on average, respectively, with small differences between ration treatments (ranging from 0.43 to 0.55 and from 0.21 to 0.34, respectively) and among lactation stages (ranging from 0.24 to 0.59 and from 0.13 to 0.36, respectively), with the highest prediction accuracies in early lactation. The estimated heritabilities for predicted DMI and rDMI were 0.3 and 0.4, respectively, which suggests genetic potential for both predicted traits. The correlations of sire estimates for milk IR-predicted DMI with official Dutch breeding values were strongest with milk production (0.33), longevity (0.26), and fertility (-0.27), indicating that cows that eat more produce more, live longer, and have poorer fertility. The correlations of sire estimates for predicted DMI and rDMI with the official breeding values for DMI were low (0.14 and 0.03, respectively). This implies that the added value of including milk IR-predicted DMI information in the estimation procedure of breeding values for DMI would be considered insufficient for practical application.

The use of milk Fourier-transform mid-infrared spectroscopy to diagnose pregnancy and determine spectral regional associations with pregnancy in US dairy cows.

Accurate early diagnosis of pregnancy is important for timely reproductive management of dairy farms. Fourier-transform mid-infrared (FT-MIR) milk spectral data are routinely used for determining milk components such as fat and protein, whereas milk composition is known to change with advancing stages of pregnancy. The objectives of this study were to compare partial least squares discriminant analysis (PLS-DA) and a Bayesian variable selection regression model (BayesC) for the diagnosis of pregnancy status (PS) from milk FT-MIR data and to infer any spectral regions that might be highly associated with PS at various stages of pregnancy. Conception dates on confirmed pregnant cows were obtained from Holstein cows within 123 herds in Michigan, Ohio, and Indiana during 2018 and 2019. Milk samples from these pregnant cows at 7 different stages of pregnancy were case-control matched to open contemporary herd mates to be within the same stage (±10 d for days in milk) of lactation for the same milk sample test date. The FT-MIR data were obtained for all of these milk samples. Ten-fold herd-independent cross-validation was used to compare PLS-DA versus BayesC using the area under the receiver operating characteristic curve (AUC). The BayesC model demonstrated higher mean AUC compared with PLS-DA at all stages exceeding 60 d of pregnancy. The mean BayesC AUC at stage 1 (1-30 d) was 0.58 ± 0.02, which was superior to a random guess (AUC = 0.50) yet too low to be of practical use. The mean BayesC AUC at stage 7 (≥180 d) was 0.13 greater compared with that of stage 1 (1-30 d) and 0.07 to 0.10 greater compared with stages 2, 3, 4, 5, and 6 (31-180 d in 30-d increments). The mean AUC of stages 2 to 6 were 0.03 to 0.06 greater compared with stage 1 yet again too low to be of practical use. Because of high multicollinearity between many adjacent wavenumbers, a spatially constrained clustering algorithm was used to adaptively partition wavenumbers into 68 windows before inferring associations of spectral regions with pregnancy. Pregnancy status was highly associated with wavenumber windows 1,063 to 1,134 cm-1, 1,201 to 1,257 cm-1, and 1,260 to 1,432 cm-1 based on an estimated BayesC posterior probability of association (PPA) approaching 100% for each of these windows at all pregnancy stages. Other windows ranging from 1,730 to 1,764 cm-1, 1,775 to 1,992 cm-1, 1,995 to 2,163 cm-1, and 2,167 to 2,316 cm-1 had varying medium to high PPA (30% to 100%) across stages. The estimated PPA in wavenumber regions from 1,477 to 1,507 cm-1, and 1,510 to 1,574 cm-1 was weaker in stages 1 and 2 compared with later stages, whereas for the regions 2,984 to 3,077 cm-1 and 3,081 to 3,133 cm-1 the effect of pregnancy was greater for stage 1 compared with other stages. Despite our conclusion that milk FT-MIR data poorly diagnose PS, our study provides new insights into spectral regions that are strongly associated with PS and warrant greater attention.

Pregnancy status predicted using milk mid-infrared spectra from dairy cattle.

Accurate and timely pregnancy diagnosis is an important component of effective herd management in dairy cattle. Predicting pregnancy from Fourier-transform mid-infrared (FT-MIR) spectroscopy data is of particular interest because the data are often already available from routine milk testing. The purpose of this study was to evaluate how well pregnancy status could be predicted in a large data set of 1,161,436 FT-MIR milk spectra records from 863,982 mixed-breed pasture-based New Zealand dairy cattle managed within seasonal calving systems. Three strategies were assessed for defining the nonpregnant cows when partitioning the records according to pregnancy status in the training population. Two of these used records for cows with a subsequent calving only, whereas the third also included records for cows without a subsequent calving. For each partitioning strategy, partial least squares discriminant analysis models were developed, whereby spectra from all the cows in 80% of herds were used to train the models, and predictions on cows in the remaining herds were used for validation. A separate data set was also used as a secondary validation, whereby pregnancy diagnosis had been assigned according to the presence of pregnancy-associated glycoproteins (PAG) in the milk samples. We examined different ways of accounting for stage of lactation in the prediction models, either by including it as an effect in the prediction model, or by pre-adjusting spectra before fitting the model. For a subset of strategies, we also assessed prediction accuracies from deep learning approaches, utilizing either the raw spectra or images of spectra. Across all strategies, prediction accuracies were highest for models using the unadjusted spectra as model predictors. Strategies for cows with a subsequent calving performed well in herd-independent validation with sensitivities above 0.79, specificities above 0.91 and area under the receiver operating characteristic curve (AUC) values over 0.91. However, for these strategies, the specificity to predict nonpregnant cows in the external PAG data set was poor (0.002-0.04). The best performing models were those that included records for cows without a subsequent calving, and used unadjusted spectra and days in milk as predictors, with consistent results observed across the training, herd-independent validation and PAG data sets. For the partial least squares discriminant analysis model, sensitivity was 0.71, specificity was 0.54 and AUC values were 0.68 in the PAG data set; and for an image-based deep learning model, the sensitivity was 0.74, specificity was 0.52 and the AUC value was 0.69. Our results demonstrate that in pasture-based seasonal calving herds, confounding between pregnancy status and spectral changes associated with stage of lactation can inflate prediction accuracies. When the effect of this confounding was reduced, prediction accuracies were not sufficiently high enough to use as a sole indicator of pregnancy status.

Application of machine-learning methods to milk mid-infrared spectra for discrimination of cow milk from pasture or total mixed ration diets.

The prevalence of "grass-fed" labeled food products on the market has increased in recent years, often commanding a premium price. To date, the majority of methods used for the authentication of grass-fed source products are driven by auditing and inspection of farm records. As such, the ability to verify grass-fed source claims to ensure consumer confidence will be important in the future. Mid-infrared (MIR) spectroscopy is widely used in the dairy industry as a rapid method for the routine monitoring of individual herd milk composition and quality. Further harnessing the data from individual spectra offers a promising and readily implementable strategy to authenticate the milk source at both farm and processor levels. Herein, a comprehensive comparison of the robustness, specificity, and accuracy of 11 machine-learning statistical analysis methods were tested for the discrimination of grass-fed versus non-grass-fed milks based on the MIR spectra of 4,320 milk samples collected from cows on pasture or indoor total mixed ration-based feeding systems over a 3-yr period. Linear discriminant analysis and partial least squares discriminant analysis (PLS-DA) were demonstrated to offer the greatest level of accuracy for the prediction of cow diet from MIR spectra. Parsimonious strategies for the selection of the most discriminating wavelengths within the spectra are also highlighted.

Validation of milk mid-infrared spectroscopy for predicting the metabolic status of lactating dairy cows in Australia.

Increased concentrations of some serum biomarkers are known to be associated with impaired health of dairy cows. Therefore, being able to predict these biomarkers, especially in the early stage of lactation, would enable preventive management decision. Some health biomarkers may also be used as phenotypes for genetic improvement for improved animal health. In this study, we validated the accuracy and robustness of models for predicting serum concentrations of ß-hydroxybutyrate (BHB), fatty acids, and urea nitrogen, using milk mid-infrared (MIR) spectroscopy. The data included 3,262 blood samples of 3,027 lactating Holstein-Friesian cows from 19 dairy herds in Southeastern Australia, collected in the period from July 2017 to April 2020. The models were developed using partial least squares regression and were validated using 10-fold random cross-validation, herd-year by herd-year external validation, and year by year validation. The coefficients of determination (R2) for prediction of serum BHB, fatty acids, and urea obtained through random cross-validation were 0.60, 0.42, and 0.87, respectively. For the herd-year by herd-year external validation, the prediction accuracies held up comparatively well, with R2 values of 0.49, 0.33, and 0.67 for of serum BHB, fatty acids, and urea, respectively. When the models were developed using data from a single year to predict data collected in future years, the R2 remained comparable, however, the root mean squared errors increased substantially (4-10 times larger than compared with that of herd-year by herd-year external validation) which could be due to machine differences in spectral response, the change in spectral response of individual machines over time, or other differences associated with farm management between seasons. In conclusion, the mid-infrared equations for predicting serum BHB, fatty acids, and urea have been validated. The prediction equations could be used to help farmers detect cows with metabolic disorders in early lactation in addition to generating novel phenotypes for genetic improvement purposes.

Predicting pregnancy status from mid-infrared spectroscopy in dairy cow milk using deep learning.

Accurately identifying pregnancy status is imperative for a profitable dairy enterprise. Mid-infrared (MIR) spectroscopy is routinely used to determine fat and protein concentrations in milk samples. Mid-infrared spectra have successfully been used to predict other economically important traits, including fatty acid content, mineral content, body energy status, lactoferrin, feed intake, and methane emissions. Machine learning has been used in a variety of fields to find patterns in vast quantities of data. This study aims to use deep learning, a sub-branch of machine learning, to establish pregnancy status from routinely collected milk MIR spectral data. Milk spectral data were obtained from National Milk Records (Chippenham, UK), who collect large volumes of data continuously on a monthly basis. Two approaches were followed: using genetic algorithms for feature selection and network design (model 1), and transfer learning with a pretrained DenseNet model (model 2). Feature selection in model 1 showed that the number of wave points in MIR data could be reduced from 1,060 to 196 wave points. The trained model converged after 162 epochs with validation accuracy and loss of 0.89 and 0.18, respectively. Although the accuracy was sufficiently high, the loss (in terms of predicting only 2 labels) was considered too high and suggested that the model would not be robust enough to apply to industry. Model 2 was trained in 2 stages of 100 epochs each with spectral data converted to gray-scale images and resulted in accuracy and loss of 0.97 and 0.08, respectively. Inspection on inference data showed prediction sensitivity of 0.89, specificity of 0.86, and prediction accuracy of 0.88. Results indicate that milk MIR data contains features relating to pregnancy status and the underlying metabolic changes in dairy cows, and such features can be identified by means of deep learning. Prediction equations from trained models can be used to alert farmers of nonviable pregnancies as well as to verify conception dates.

The use of milk mid-infrared spectroscopy to improve genomic prediction accuracy of serum biomarkers.

Breeding objectives in the dairy industry have shifted from being solely focused on production to including fertility, animal health, and environmental impact. Increased serum concentrations of candidate biomarkers of health and fertility, such as ß-hydroxybutyric acid (BHB), fatty acids, and urea are difficult and costly to measure, and thus limit the number of records. Accurate genomic prediction requires a large reference population. The inclusion of milk mid-infrared (MIR) spectroscopic predictions of biomarkers may increase genomic prediction accuracy of these traits. Our objectives were to (1) estimate the heritability of, and genetic correlations between, selected serum biomarkers and their respective MIR predictions, and (2) evaluate genomic prediction accuracies of either only measured serum traits, or serum traits plus MIR-predicted traits. The MIR-predicted traits were either fitted in a single trait model, assuming the measured trait and predicted trait were the same trait, or in a multitrait model, where measured and predicted trait were assumed to be correlated traits. We performed all analyses using relationship matrices constructed from pedigree (A matrix), genotypes (G matrix), or both pedigree and genotypes (H matrix). Our data set comprised up to 2,198 and 9,657 Holstein cows with records for serum biomarkers and MIR-predicted traits, respectively. Heritabilities of measured serum traits ranged from 0.04 to 0.07 for BHB, from 0.13 to 0.21 for fatty acids, and from 0.10 to 0.12 for urea. Heritabilities for MIR-predicted traits were not significantly different from those for the measured traits. Genetic correlations between measured traits and MIR-predicted traits were close to 1 for urea. For BHB and fatty acids, genetic correlations were lower and had large standard errors. The inclusion of MIR predicted urea substantially increased prediction accuracy for urea. For BHB, including MIR-predicted BHB reduced the genomic prediction accuracy, whereas for fatty acids, prediction accuracies were similar with either measured fatty acids, MIR-predicted fatty acids, or both. The high genetic correlation between urea and MIR-predicted urea, in combination with the increased prediction accuracy, demonstrated the potential of using MIR-predicted urea for genomic prediction of urea. For BHB and fatty acids, further studies with larger data sets are required to obtain more accurate estimates of genetic correlations.

Prediction of indicators of cow diet composition and authentication of feeding specifications of Protected Designation of Origin cheese using mid-infrared spectroscopy on milk.

The ability of mid-infrared spectroscopy (MIR) to predict indicators (1) of diet composition in dairy herds and (2) for the authentication of the cow feeding restrictions included in the specification of 2 Protected Designation of Origin (PDO) cheeses (Cantal and Laguiole) was tested on 7,607 bulk milk spectra from 1,355 farms located in the Massif Central area of France. For each milk sample, the corresponding cow diet composition data were obtained through on-farm surveys. The cow diet compositions varied largely (i.e., from full grazing for extensive farming systems to corn silage-based diets, which are typical of more intensive farming systems). Partial least square regression and discriminant analysis were used to predict the proportion of different feedstuffs in the cows' diets and to authenticate the cow feeding restrictions for the PDO cheese specifications, respectively. The groups for the discriminant analysis were created by dividing the data set according to the threshold of a specific feedstuff. They were issued based on the specifications of the restriction of the PDO cheese. The pasture proportion in the cows' diets was predicted by MIR with an coefficient of determination in external validation (R2V) = 0.81 and a standard error of prediction of 11.7% dry matter. Pasture + hay, corn silage, conserved herbage, fermented forage, and total herbage proportion in the cows' diets were predicted with a R2V >0.61 and a standard error of prediction <14.8. The discrimination models for pasture presence, pasture ≥50%, and pasture ≥57% in the cows' diets achieved an accuracy and specificity ≥90%. A sensitivity and precision ≥85% were also observed for the pasture proportion discrimination models, but both of these indexes decreased at increasing thresholds from 0 to 50, and 57% pasture in the cows' diets. An accuracy ≥80% was also observed for pasture + hay ≥72%, herbage ≥50%, pasture + hay ≥25%, absence of fermented herbage, absence of corn silage, and corn silage ≤30% in the cows' diets, but for several models, either the sensitivity or precision was lower than the accuracy. Models built on the simultaneous respect of all the criteria of the feeding restrictions of PDO cheese specifications achieved an accuracy, specificity, sensitivity, and precision >90%. Both the regression and discriminant MIR models for bulk milk can provide useful indicators of cow diet composition and PDO cheese specifications to producers and consumers (farmers, dairy plants).

The relationship of excessive energy deficit with milk somatic cell score and clinical mastitis.

Periparturient cows go through a period of immune suppression often marked by immune cell dysfunction. Further exacerbation of this dysfunction through early-lactation excessive energy deficit (EED) has been associated with increased susceptibility to infectious conditions such as mastitis. Our objective was to explore the association of milk somatic cell score (SCS) and clinical mastitis (CM) diagnosis in cows identified with EED, diagnosed using each of the following: blood and milk ß-hydroxybutyrate (BHB), milk predicted blood nonesterified fatty acid (mpbNEFA) concentrations, or milk de novo fatty acid (FA) relative percentages (rel %). We analyzed data collected from 396 multiparous Holstein cows from 2 New York farms in a prospective cohort study. Coccygeal vessel blood samples and composite milk samples were collected twice weekly from 3 to 18 days in milk (DIM) for a total of 4 time points per cow (T1, T2, T3, T4). Blood was analyzed using a hand-held meter, and milk was analyzed using Fourier-transform mid-infrared spectrometry for milk BHB and mpbNEFA concentrations, milk de novo FA rel %, and somatic cell count. Excessive energy deficit was diagnosed as blood BHB ≥ 1.2 mmol/L, milk BHB ≥ 0.14 mmol/L, mpbNEFA ≥ 0.55 mmol/L, or de novo FA ≤ 22.7 rel %, depending on the model. Clinical mastitis cultures were collected from 4 to 60 DIM by on-farm personnel. Incidence of hyperketonemia as determined by blood BHB was 13.4%, and incidence of CM was 23.9%. Separate repeated-measures ANOVA models were developed for each EED diagnostic analyte for parity groups 2, 3, and ≥4 to assess differences in SCS; t-test analyses were similarly used to assess the association of each diagnostic analyte with CM at each time point. For all diagnostic analytes, apart from milk BHB, cows diagnosed with EED tended to have lower SCS than their non-EED counterparts. This was especially apparent at T1 for all parity groups, and at T2, T3, and T4 for blood BHB and mpbNEFA. For EED diagnosis via mpbNEFA, mean SCS were lower in parity ≥4, with a difference in mean SCS between EED and non-EED animals of 0.7 SCS units, equating to a somatic cell count in EED animals approaching half that of non-EED (EED = 67,000 cells/mL, non-EED = 107,000 cell/mL). No important relationships were observed between CM diagnosis and blood BHB, milk BHB, or mpbNEFA. For de novo FA rel %, reductions in this analyte were noted before CM diagnosis at all time points. Although the relationship between EED and CM is still unclear, our findings suggest that cows in EED, diagnosed using blood BHB or mpbNEFA during the first 18 DIM, have a tendency toward lower SCS compared with their non-EED counterparts.

A comparison of 4 different machine learning algorithms to predict lactoferrin content in bovine milk from mid-infrared spectra.

Lactoferrin (LF) is a glycoprotein naturally present in milk. Its content varies throughout lactation, but also with mastitis; therefore it is a potential additional indicator of udder health beyond somatic cell count. Condequently, there is an interest in quantifying this biomolecule routinely. First prediction equations proposed in the literature to predict the content in milk using milk mid-infrared spectrometry were built using partial least square regression (PLSR) due to the limited size of the data set. Thanks to a large data set, the current study aimed to test 4 different machine learning algorithms using a large data set comprising 6,619 records collected across different herds, breeds, and countries. The first algorithm was a PLSR, as used in past investigations. The second and third algorithms used partial least square (PLS) factors combined with a linear and polynomial support vector regression (PLS + SVR). The fourth algorithm also used PLS factors, but included in an artificial neural network with 1 hidden layer (PLS + ANN). The training and validation sets comprised 5,541 and 836 records, respectively. Even if the calibration prediction performances were the best for PLS + polynomial SVR, their validation prediction performances were the worst. The 3 other algorithms had similar validation performances. Indeed, the validation root mean squared error (RMSE) ranged between 162.17 and 166.75 mg/L of milk. However, the lower standard deviation of cross-validation RMSE and the better normality of the residual distribution observed for PLS + ANN suggest that this modeling was more suitable to predict the LF content in milk from milk mid-infrared spectra (R2v = 0.60 and validation RMSE = 162.17 mg/L of milk). This PLS +ANN model was then applied to almost 6 million spectral records. The predicted LF showed the expected relationships with milk yield, somatic cell score, somatic cell count, and stage of lactation. The model tended to underestimate high LF values (higher than 600 mg/L of milk). However, if the prediction threshold was set to 500 mg/L, 82% of samples from the validation having a content of LF higher than 600 mg/L were detected. Future research should aim to increase the number of those extremely high LF records in the calibration set.