Evaluating the performance of machine learning methods and variable selection methods for predicting difficult-to-measure traits in Holstein dairy cattle using milk infrared spectral data

Infrared spectroscopy coupled with machine learning algorithms for predicting the detailed milk mineral profile in dairy cattle

Milk minerals are not only essential components for human health, but they can be informative for milk quality and cow's health. Herein, we investigated the feasibility of Fourier Transformed mid Infrared (FTIR) spectroscopy for the prediction of a detailed panel of 17 macro, trace, and environmental elements in bovine milk, using partial least squares regression (PLS) and machine learning approaches. The automatic machine learning significantly outperformed the PLS regression in terms of prediction performances of the mineral elements. For macrominerals, the R2 ranged from 0.59 to 0.78. Promising predictability was achieved for Cu and B (R2 = 0.66 and 0.74, respectively) and more moderate ones for Fe, Mn, Zn, and Al (R2 from 0.48 to 0.58). These results provide a reliable basis for a rapid and cost-effective quantification of these traits, serving as a resource for dairy farmers seeking to enhance the quality of milk production and optimize cheese properties.

Effectiveness of near-infrared spectroscopy to predict the chemical composition of feces and total-tract apparent nutrients digestibility estimated with undigestible neutral detergent fiber or acid-insoluble ash in lactating buffaloes' feces

Following a comparison of nutrient total-tract digestibility estimates in lactating buffaloes using single-point undigestible NDF (uNDF) or acid-insoluble ash (AIA) as internal markers, the potential of fecal near-infrared spectroscopy (NIRS) to provide calibration equations for the assessment of the chemical composition of feces and nutrient total-tract digestibility estimated with internal markers was explored. Chemical analyses were performed on 147 fecal samples from lactating buffaloes reared on 5 farms in central Italy (Naples). Each farm fed a silage-based TMR to the buffaloes, and the TMR was sampled in the 2 d before the fecal collection. The TMR and individual fecal samples were collected and analyzed for DM, OM, ash, AIA, ether extract (EE), starch, fiber fractions (amylase-treated NDF without residual ash [aNDFom], amylase-treated NDF inclusive of residual ash [aNDF], ADF without residual ash [ADFom], ADF, hemicellulose, cellulose, ADL, uNDF), N, CP and CP bound to aNDF (NDICP) and to ADF (ADICP). The uNDF content was determined through a 240-h in vitro fermentation and employed, together with AIA as markers, to estimate the total-tract apparent digestibility and total-tract digestibility of DM, OM, ash, N, CP, EE, aNDFom, aNDF, NDIP, ADFom, and ADF, ADIN, ADL, hemicellulose, cellulose, starch, NFC, and the B3 fraction of N (NB3). No correlation was found between DM and OM digestibility estimated with AIA and uNDF as internal markers. Weak correlations were detected for all the other nutients digestibilities, and strong correlations were observed for EE, ADFom, hemicellulose, NDIN, ADIN, NB3, NFC, and starch. The sample set (n = 147) was divided in a calibration set (n = 111) and a validation set (n = 36) to "train" and "validate" the fecal NIRS curve through an external validation process. An estimation usable for preliminary or initial evaluation was obtained for N, CP, and aNDF fecal content. An excellent prediction was obtained for total tract digestibility of ADIN (R2 = 0.90) when estimated with uNDF as the internal marker. The NIRS technology was not able to accurately predict all the other traits and the estimated nutrient digestibility of lactating buffalo diets from fecal spectra.

Benchmarking machine learning and parametric methods for genomic prediction of feed efficiency-related traits in Nellore cattle

Genomic selection (GS) offers a promising opportunity for selecting more efficient animals to use consumed energy for maintenance and growth functions, impacting profitability and environmental sustainability. Here, we compared the prediction accuracy of multi-layer neural network (MLNN) and support vector regression (SVR) against single-trait (STGBLUP), multi-trait genomic best linear unbiased prediction (MTGBLUP), and Bayesian regression (BayesA, BayesB, BayesC, BRR, and BLasso) for feed efficiency (FE) traits. FE-related traits were measured in 1156 Nellore cattle from an experimental breeding program genotyped for ~ 300 K markers after quality control. Prediction accuracy (Acc) was evaluated using a forward validation splitting the dataset based on birth year, considering the phenotypes adjusted for the fixed effects and covariates as pseudo-phenotypes. The MLNN and SVR approaches were trained by randomly splitting the training population into fivefold to select the best hyperparameters. The results show that the machine learning methods (MLNN and SVR) and MTGBLUP outperformed STGBLUP and the Bayesian regression approaches, increasing the Acc by approximately 8.9%, 14.6%, and 13.7% using MLNN, SVR, and MTGBLUP, respectively. Acc for SVR and MTGBLUP were slightly different, ranging from 0.62 to 0.69 and 0.62 to 0.68, respectively, with empirically unbiased for both models (0.97 and 1.09). Our results indicated that SVR and MTGBLUBP approaches were more accurate in predicting FE-related traits than Bayesian regression and STGBLUP and seemed competitive for GS of complex phenotypes with various degrees of inheritance.

SVR Chemometrics to Quantify β-Lactoglobulin and α-Lactalbumin in Milk Using MIR

Protein content variation in milk can impact the quality and consistency of dairy products, necessitating access to in-line real time monitoring. Here, we present a chemometric approach for the qualitative and quantitative monitoring of β-lactoglobulin and α-lactalbumin, using mid-infrared spectroscopy (MIR). In this study, we employed Hotelling T2 and Q-residual for outlier detection, automated preprocessing using nippy, conducted wavenumber selection with genetic algorithms, and evaluated four chemometric models, including partial least squares, support vector regression (SVR), ridge, and logistic regression to accurately predict the concentrations of β-lactoglobulin and α-lactalbumin in milk. For the quantitative analysis of these two whey proteins, SVR performed the best to interpret protein concentration from 197 MIR spectra originating from 42 Cornell University samples of preserved pasteurized modified milk. The R2 values obtained for β-lactoglobulin and α-lactalbumin using leave one out cross-validation (LOOCV) are 92.8% and 92.7%, respectively, which is the highest correlation reported to date. Our approach introduced a combination of preprocessing automation, genetic algorithm-based wavenumber selection, and used Optuna to optimize the framework for tuning hyperparameters of the chemometric models, resulting in the best chemometric analysis of MIR data to quantitate β-lactoglobulin and α-lactalbumin to date.

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 β-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 (CV_r) and herd-out (CV_h). 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 CV_r and 2% to 139% for CV_h, whereas the stacking ensemble increased the R² value from 4% to 70% for CV_r and 4% to 150% for CV_h. Considering the best model, with the CV_r 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.

Novel prediction models for hyperketonemia using bovine milk Fourier-transform infrared spectroscopy

Metabolic diseases driven by negative energy balance in dairy cattle contribute to reduced milk production, increased disease incidence, culling, and death. Cow side tests for negative energy balance markers are available but are labor-intensive. Milk sample analysis using Fourier transform infrared spectroscopy (FTIR) allows for sampling numerous cows simultaneously. FTIR prediction models have moderate accuracy for hyperketonemia diagnosis (beta-hydroxybutyrate (BHB) ≥ 1.2 mmol/L). Most research using FTIR has focused on homogenous datasets and conventional prediction models, including partial least squares, linear discriminant analysis, and ElasticNet. Our objective was to evaluate more diverse modeling options, such as deep learning, gradient boosting machine models, and model ensembles for hyperketonemia classification. We compiled a sizable, heterogeneous dataset including milk FTIR and concurrent blood samples. Blood samples were tested for blood BHB, and wavenumber data was obtained from milk FTIR analysis. Using this dataset, we trained conventional prediction models and other options listed above. We demonstrate prediction model performance is similar for convolutional neural networks and ensemble models to simpler algorithm options. Results obtained from this study indicate that deep learning and model ensembles are potential algorithm options for predicting hyperketonemia in dairy cattle. Additionally, our results indicate hyperketonemia prediction models can be developed using heterogeneous datasets.

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 [α_S1-casein (CN), α_S2-CN, β-CN, κ-CN, and glycosylated-κ-CN], and major whey proteins (β-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 (r_a) 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.

Phenotypic Analysis of Fourier-Transform Infrared Milk Spectra in Dairy Goats

The infrared spectrum of bovine milk is used to predict many interesting traits, whereas there have been few studies on goat milk in this regard. The objective of this study was to characterize the major sources of variation in the absorbance of the infrared spectrum in caprine milk samples. A total of 657 goats belonging to 6 breeds and reared on 20 farms under traditional and modern dairy systems were milk-sampled once. Fourier-transform infrared (FTIR) spectra were taken (2 replicates per sample, 1314 spectra), and each spectrum contained absorbance values at 1060 different wavenumbers (5000 to 930 × cm^-1), which were treated as a response variable and analyzed one at a time (i.e., 1060 runs). A mixed model, including the random effects of sample/goat, breed, flock, parity, stage of lactation, and the residual, was used. The pattern and variability of the FTIR spectrum of caprine milk was similar to those of bovine milk. The major sources of variation in the entire spectrum were as follows: sample/goat (33% of the total variance); flock (21%); breed (15%); lactation stage (11%); parity (9%); and the residual unexplained variation (10%). The entire spectrum was segmented into five relatively homogeneous regions. Two of them exhibited very large variations, especially the residual variation. These regions are known to be affected by the absorbance of water, although they also exhibited wide variations in the other sources of variation. The average repeatability of these two regions were 45% and 75%, whereas for the other three regions it was about 99%. The FTIR spectrum of caprine milk could probably be used to predict several traits and to authenticate the origin of goat milk.

In-line near-infrared analysis of milk coupled with machine learning methods for the daily prediction of blood metabolic profile in dairy cattle

Precision livestock farming technologies are used to monitor animal health and welfare parameters continuously and in real time in order to optimize nutrition and productivity and to detect health issues at an early stage. The possibility of predicting blood metabolites from milk samples obtained during routine milking by means of infrared spectroscopy has become increasingly attractive. We developed, for the first time, prediction equations for a set of blood metabolites using diverse machine learning methods and milk near-infrared spectra collected by the AfiLab instrument. Our dataset was obtained from 385 Holstein Friesian dairy cows. Stacking ensemble and multi-layer feedforward artificial neural network outperformed the other machine learning methods tested, with a reduction in the root mean square error of between 3 and 6% in most blood parameters. We obtained moderate correlations (r) between the observed and predicted phenotypes for γ-glutamyl transferase (r = 0.58), alkaline phosphatase (0.54), haptoglobin (0.66), globulins (0.61), total reactive oxygen metabolites (0.60) and thiol groups (0.57). The AfiLab instrument has strong potential but may not yet be ready to predict the metabolic stress of dairy cows in practice. Further research is needed to find out methods that allow an improvement in accuracy of prediction equations.

Challenging Sustainable and Innovative Technologies in Cheese Production: A Review

It is well known that cheese yield and quality are affected by animal genetics, milk quality (chemical, physical, and microbiological), production technology, and the type of rennet and dairy cultures used in production. Major differences in the same type of cheese (i.e., hard cheese) are caused by the rennet and dairy cultures, which affect the ripening process. This review aims to explore current technological advancements in animal genetics, methods for the isolation and production of rennet and dairy cultures, along with possible applications of microencapsulation in rennet and dairy culture production, as well as the challenge posed to current dairy technologies by the preservation of biodiversity. Based on the reviewed scientific literature, it can be concluded that innovative approaches and the described techniques can significantly improve cheese production.

Genetic parameters of differential somatic cell count, milk composition, and cheese-making traits measured and predicted using spectral data in Holstein cows

Mastitis is one of the most prevalent diseases in dairy cattle and is the cause of considerable economic losses. Alongside somatic cell count (SCC), differential somatic cell count (DSCC) has been recently introduced as a new indicator of intramammary infection. The DSCC is expressed as a count or a proportion (%) of polymorphonuclear neutrophils plus lymphocytes (PMN-LYM) in milk somatic cells. These numbers are complemented to total somatic cell count or to 100 by macrophages (MAC). The aim of this study was to investigate the genetic variation and heritability of DSCC, and its correlation with milk composition, udder health indicators, milk composition, and technological traits in Holstein cattle. Data used in the analysis consisted in single test-day records from 2,488 Holstein cows reared in 36 herds located in northern Italy. Fourier-transform infrared (FTIR) spectroscopy was used to predict missing information for some milk coagulation and cheese-making traits, to increase sample size and improve estimation of the genetic parameters. Bayesian animal models were implemented via Gibbs sampling. Marginal posterior means of the heritability estimates were 0.13 for somatic cell score (SCS); 0.11 for DSCC, MAC proportion, and MAC count; and 0.10 for PMN-LYM count. Posterior means of additive genetic correlations between SCS and milk composition and udder health were low to moderate and unfavorable. All the relevant genetic correlations between the SCC traits considered and the milk traits (composition, coagulation, cheese yield and nutrients recovery) were unfavorable. The SCS showed genetic correlations of -0.30 with the milk protein proportion, -0.56 with the lactose proportion and -0.52 with the casein index. In the case of milk technological traits, SCS showed genetic correlations of 0.38 with curd firming rate (k₂₀), 0.45 with rennet coagulation time estimated using the curd firming over time equation (RCT_eq), -0.39 with asymptotic potential curd firmness, -0.26 with maximum curd firmness (CF_max), and of -0.31 with protein recovery in the curd. Differential somatic cell count expressed as proportion was correlated with SCS (0.60) but had only 2 moderate genetic correlations with milk traits: with lactose (-0.32) and CF_max (-0.33). The SCS was highly correlated with the log PMN-LYM count (0.79) and with the log MAC count (0.69). The 2 latter traits were correlated with several milk traits: fat (-0.38 and -0.43 with PMN-LYM and MAC counts, respectively), lactose percentage (-0.40 and -0.46), RCT_eq (0.53 and 0.41), t_max (0.38 and 0.48). Log MAC count was correlated with k₂₀ (+0.34), and log PMN-LYM count was correlated with CF_max (-0.26) and weight of water curd as percentage of weight of milk processed (-0.26). The results obtained offer new insights into the relationships between the indicators of udder health and the milk technological traits in Holstein cows.

Comparison of Single-Breed and Multi-Breed Training Populations for Infrared Predictions of Novel Phenotypes in Holstein Cows

In general, Fourier-transform infrared (FTIR) predictions are developed using a single-breed population split into a training and a validation set. However, using populations formed of different breeds is an attractive way to design cross-validation scenarios aimed at increasing prediction for difficult-to-measure traits in the dairy industry. This study aimed to evaluate the potential of FTIR prediction using training set combining specialized and dual-purpose dairy breeds to predict different phenotypes divergent in terms of biological meaning, variability, and heritability, such as body condition score (BCS), serum β-hydroxybutyrate (BHB), and kappa casein (k-CN) in the major cattle breed, i.e., Holstein-Friesian. Data were obtained from specialized dairy breeds: Holstein (468 cows) and Brown Swiss (657 cows), and dual-purpose breeds: Simmental (157 cows), Alpine Grey (75 cows), and Rendena (104 cows), giving a total of 1461 cows from 41 multi-breed dairy herds. The FTIR prediction model was developed using a gradient boosting machine (GBM), and predictive ability for the target phenotype in Holstein cows was assessed using different cross-validation (CV) strategies: a within-breed scenario using 10-fold cross-validation, for which the Holstein population was randomly split into 10 folds, one for validation and the remaining nine for training (10-fold_HO); an across-breed scenario (BS_HO) where the Brown Swiss cows were used as the training set and the Holstein cows as the validation set; a specialized multi-breed scenario (BS+HO_10-fold), where the entire Brown Swiss and Holstein populations were combined then split into 10 folds, and a multi-breed scenario (Multi-breed), where the training set comprised specialized (Holstein and Brown Swiss) and dual-purpose (Simmental, Alpine Grey, and Rendena) dairy cows, combined with nine folds of the Holstein cows. Lastly a Multi-breed CV2 scenario was implemented, assuming the same number of records as the reference scenario and using the same proportions as the multi-breed. Within-Holstein, FTIR predictions had a predictive ability of 0.63 for BCS, 0.81 for BHB, and 0.80 for k-CN. Using a specific breed (Brown Swiss) as the training set for prediction in the Holstein population reduced the prediction accuracy by 10% for BCS, 7% for BHB, and 11% for k-CN. Notably, the combination of Holstein and Brown Swiss cows in the training set increased the predictive ability of the model by 6%, which was 0.66 for BCS, 0.85 for BHB, and 0.87 for k-CN. Using multiple specialized and dual-purpose animals in the training set outperforms the 10-fold_HO (standard) approach, with an increase in predictive ability of 8% for BCS, 7% for BHB, and 10% for k-CN. When the Multi-breed CV2 was implemented, no improvement was observed. Our findings suggest that FTIR prediction of different phenotypes in the Holstein breed can be improved by including different specialized and dual-purpose breeds in the training population. Our study also shows that predictive ability is enhanced when the size of the training population and the phenotypic variability are increased.