Improving the accuracy of genomic prediction in dairy cattle using the biologically annotated neural networks framework

BACKGROUND

Biologically annotated neural networks (BANNs) are feedforward Bayesian neural network models that utilize partially connected architectures based on SNP-set annotations. As an interpretable neural network, BANNs model SNP and SNP-set effects in their input and hidden layers, respectively. Furthermore, the weights and connections of the network are regarded as random variables with prior distributions reflecting the manifestation of genetic effects at various genomic scales. However, its application in genomic prediction has yet to be explored.

RESULTS

This study extended the BANNs framework to the area of genomic selection and explored the optimal SNP-set partitioning strategies by using dairy cattle datasets. The SNP-sets were partitioned based on two strategies-gene annotations and 100 kb windows, denoted as BANN_gene and BANN_100kb, respectively. The BANNs model was compared with GBLUP, random forest (RF), BayesB and BayesCπ through five replicates of five-fold cross-validation using genotypic and phenotypic data on milk production traits, type traits, and one health trait of 6,558, 6,210 and 5,962 Chinese Holsteins, respectively. Results showed that the BANNs framework achieves higher genomic prediction accuracy compared to GBLUP, RF and Bayesian methods. Specifically, the BANN_100kb demonstrated superior accuracy and the BANN_gene exhibited generally suboptimal accuracy compared to GBLUP, RF, BayesB and BayesCπ across all traits. The average accuracy improvements of BANN_100kb over GBLUP, RF, BayesB and BayesCπ were 4.86%, 3.95%, 3.84% and 1.92%, and the accuracy of BANN_gene was improved by 3.75%, 2.86%, 2.73% and 0.85% compared to GBLUP, RF, BayesB and BayesCπ, respectively across all seven traits. Meanwhile, both BANN_100kb and BANN_gene yielded lower overall mean square error values than GBLUP, RF and Bayesian methods.

CONCLUSION

Our findings demonstrated that the BANNs framework performed better than traditional genomic prediction methods in our tested scenarios, and might serve as a promising alternative approach for genomic prediction in dairy cattle.

Assessing different cross-validation schemes for predicting novel traits using sensor data: an application to dry matter intake and residual feed intake using milk spectral data

Feed efficiency is important for economic profitability of dairy farms; however, recording daily dry matter intakes (DMI) is expensive. Our objective was to investigate the potential use of milk mid-infrared (MIR) spectral data to predict proxy phenotypes for DMI based on different cross-validation schemes. We were specifically interested in comparisons between a model that included only MIR data (Model M1), a model that incorporated different energy sink predictors, such as body weight, body weight change, and milk energy (Model M2), and an extended model that incorporated both energy sinks and MIR data (Model M3). Models M2 and M3 also included various cow level variables (stage of lactation, age at calving, parity) such that any improvement in model performance from M2 to M3, whether through a smaller root mean squared error (RMSE) or a greater squared predictive correlation (R2), could indicate a potential benefit of MIR to predict residual feed intake. The data used in our study originated from a multi-institutional project on the genetics of feed efficiency in US Holsteins. Analyses were conducted on 2 different trait definitions based on different period lengths: averaged across weeks vs. averaged across 28-d. Specifically, there were 19,942 weekly records on 1,812 cows across 46 experiments or cohorts and 3,724 28-d records on 1,700 cows across 43 different cohorts. The cross-validation analyses involved 3 different k-fold schemes. First, a 10-fold cow-independent cross-validation was conducted whereby all records from any one cow were kept together in either training or test sets. Similarly, a 10-fold experiment-independent cross-validation kept entire experiments together whereas a 4-fold herd-independent cross-validation kept entire herds together in either training or test sets. Based on cow-independent cross-validation for both weekly and 28-d DMI, adding MIR predictors to energy sinks (Models M3 vs M2) significantly (P < 10-10) reduced average RMSE to 1.59 kg and increased average R2 to 0.89. However, adding MIR to energy sinks (M3) to predict DMI either within an experiment-independent or herd-independent cross-validation scheme seemed to demonstrate no merit (P > 0.05) compared with an energy sink model (M2) for either R2 or RMSE (respectively, 0.68 and 2.55 kg for M2 in herd-independent scheme). We further noted that with broader cross-validation schemes, i.e., from cow-independent to experiment-independent to herd-independent schemes, the mean and slope bias increased. Given that proxy DMI phenotypes for cows would need to be almost entirely generated in herds having no DMI or training data of their own, herd-independent cross-validation assessments of predictive performance should be emphasized. Hence, more research on predictive algorithms suitable for broader cross-validation schemes and a more earnest effort on calibration of spectrophotometers against each other should be considered.

Standardization method, testing scenario, and accuracy of the infrared prediction model affect the standardization accuracy of milk mid-infrared spectra

The widespread use of milk mid-infrared (MIR) spectroscopy for phenotype prediction has urged the application of prediction models across regions and countries. Spectra standardization is the most effective way to reduce the variability in the spectral signal provided by different instruments and labs. This study aimed to develop different standardization models for MIR spectra collected by multiple instruments, across 2 provinces of China, and investigate whether the standardization method (piecewise direct standardization, PDS, and direct standardization, DS), testing scenario (standardization of spectra collected on the same day or after 7 mo), infrared prediction model accuracy (high or low), and instrument (6 instruments from 2 brands) affect the performance of the standardization model. The results showed that the determination coefficient (R2) between absorbance values at each wavenumber provided by the primary and the secondary instruments increased from less than 0.90 to nearly 1.00 after standardization. Both PDS and DS successfully reduced spectra variation among instruments, and performed significantly better than non-standardization (P < 0.05). However, DS was more prone to overfitting than PDS. Standardization accuracy was higher when tested using spectra collected on the same time compared with those collected 7 mo after (P < 0.05), but great improvement in model transferability was obtained for both scenarios compared with the non-standardized spectra. The less accurate infrared prediction model (for C8:0 and C10:0 content) benefited the most (P < 0.05) from spectra standardization compared with the more accurate model (for total fat and protein content). For spectra collected after 7 mo from standardization, after PDS the RMSE between predictions obtained by different machines decreased on average by 86 and 94% compared with the values before standardization, for C8:0 and C10:0 respectively. The secondary instrument had no significant effect on the R2 between predictions (P > 0.05). The variation in the spectral signal provided by different instruments was successfully reduced by standardization across 2 provinces in China. This study lays the foundations for developing a national MIR spectra database to provide consistent predictions across provinces to be used in dairy farm management and breeding programs in China. Besides, this provides opportunities for data exchange and cooperation at international levels.

Predicting methane emissions of individual grazing dairy cows from spectral analyses of their milk samples

Data on the enteric methane emissions of individual cows are useful not just in assisting management decisions and calculating herd inventories but also as inputs for animal genetic evaluations. Data generation for many animal characteristics, including enteric methane emissions, can be expensive and time consuming, so being able to extract as much information as possible from available samples or data sources is worthy of investigation. The objective of the present study was to attempt to predict individual cow methane emissions from the information contained within milk samples, specifically the spectrum of light transmittance across different wavelengths of the mid-infrared (MIR) region of the electromagnetic spectrum. A total of 93,888 individual spot measures of methane (i.e., individual samples of an animal's breath when using the GreenFeed technology) from 384 lactations on 277 grazing dairy cows were collapsed into weekly averages expressed as grams per day; each weekly average coincided with a MIR spectral analysis of a morning or evening individual cow milk sample. Associations between the spectra and enteric methane measures were performed separately using partial least squares regression or neural networks with different tuning parameters evaluated. Several alternative definitions of the enteric methane phenotype (i.e., average enteric methane in the 6 d preceding or 6 d following taking the milk sample or the average of the 6 d before and after the milk sample, all of which also included the enteric methane emitted on the day of milk sampling), the candidate model features (e.g., milk yield, milk composition, and milk MIR) as well as validation strategy (i.e., cross-validation or leave-one-experimental treatment-out) were evaluated. Irrespective of the validation method, the prediction accuracy was best when the average of the milk MIR from the morning and evening milk sample was used and the prediction model was developed using neural networks; concurrently including milk yield and days in milk in the prediction model generated superior predictions relative to just the spectral information alone. Furthermore, prediction accuracy was best when the enteric methane phenotype was the average of at least 20 methane spot measures across a 6-d period flanking each side of the milk sample with associated spectral data. Based on the strategy that achieved the best accuracy of prediction, the correlation between the actual and predicted daily methane emissions when based on 4-fold cross-validation varied per validation stratum from 0.68 to 0.75; the corresponding range when validated on each of the 8 different experimental treatments focusing on alternative pasture grazing systems represented in the dataset varied from 0.55 to 0.71. The root mean square error of prediction across the 4-folds of cross-validation was 37.46 g/d, whereas the root mean square error averaged across all folds of leave-one-treatment-out was 37.50 g/d. Results suggest that even with the likely measurement errors contained within the MIR spectrum and gold standard enteric methane phenotype, enteric methane can be reasonably well predicted from the infrared spectrum of milk samples. What is yet to be established, however, is whether (a) genetic variation exists in this predicted enteric methane phenotype and (b) selection on estimates of genetic merit for this phenotype translate to actual phenotypic differences in enteric methane emissions.

Selective breeding can contribute to bovine tuberculosis control and eradication

Bovine tuberculosis (bTB) persists in many countries having a significant impact on public health and livestock industry finances. The incidence and prevalence of new cases in parts of the UK and elsewhere over the past decades warrant intensified efforts towards achieving Officially Tuberculosis Free (OTF) status in the respective regions. Genetic selection aiming to identify and remove inherently susceptible animals from breeding has been proposed as an additional measure in ongoing programmes towards controlling the disease. The presence of genetic variation among individual animals in their capacity to respond to Mycobacterium bovis exposure has been documented and heritability estimates of 0.06-0.18 have been reported. Despite their moderate magnitude, these estimates suggest that host resistance to bTB is amenable to improvement with selective breeding. Although relatively slow, genetic progress can be constant, cumulative and permanent, thereby complementing ongoing disease control measures. Importantly, mostly no antagonistic genetic correlations have been found between bTB resistance and other animal traits suggesting that carefully incorporating the former in breeding decisions should not adversely affect bovine productivity. Simulation studies have demonstrated the potential impact of genetic selection on reducing the probability of a breakdown to occur or the duration and severity of a breakdown that has already been declared. Furthermore, research on the bovine genome has identified multiple genomic markers and genes associated with bTB resistance. Nevertheless, the combined outcomes of these studies suggest that host resistance to bTB is a complex, polygenic trait, with no single gene alone explaining the inherent differences between resistant and susceptible animals. Such results support the development of accurate genomic breeding values that duly capture the collective effect of multiple genes to underpin selective breeding programmes. In addition to improving host resistance to bTB, scientists and practitioners have considered the possibility of reducing host infectivity. Ongoing studies have suggested the presence of genetic variation for infectivity and confirmed that bTB eradication would be accelerated if selective breeding considered both host resistance and infectivity traits. In conclusion, research activity on bTB genetics has generated knowledge and insights to support selective breeding as an additional measure towards controlling and eradicating the disease.

Using machine learning methods to predict dry matter intake from milk mid-infrared spectroscopy data on Swedish dairy cattle

In this research communication we compare three different approaches for developing dry matter intake (DMI) prediction models based on milk mid-infrared spectra (MIRS), using data collected from a research herd over five years. In dairy production, knowledge of individual DMI could be important and useful, but DMI can be difficult and expensive to measure on most commercial farms as cows are commonly group-fed. Instead, this parameter is often estimated based on the age, body weight, stage of lactation and body condition score of the cow. Recently, milk MIRS have also been used as a tool to estimate DMI. There are different methods available to create prediction models from large datasets. The main data used were total DMI calculated as a 3-d average, coupled with milk MIRS data available fortnightly. Data on milk yield and lactation stage parameters were also available for each animal. We compared the performance of three prediction approaches: partial least-squares regression, support vector machine regression and random forest regression. The full milk MIRS alone gave low to moderate prediction accuracy (R2 = 0.07-0.40), regardless of prediction modelling approach. Adding more variables to the model improved R2 and decreased the prediction error. Overall, partial least-squares regression proved to be the best method for predicting DMI from milk MIRS data, while MIRS data together with milk yield and concentrate DMI at 3-30 d in milk provided good prediction accuracy (R2 = 0.52-0.65) regardless of the prediction tool used.

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.

Prediction of dry matter intake and gross feed efficiency using milk production and live weight in first-parity Holstein cows

Direct measurement of dry matter intake (DMI) presents a major challenge in estimating gross feed efficiency (GFE) in dairy cattle. This challenge can, however, be resolved through the prediction of DMI and GFE from easy-to-measure traits such as milk production (i.e. milk yield, energy-corrected milk (ECM), butterfat, protein, lactose) and live weight (LW). The main objective of this study was, therefore, to investigate the feasibility of predicting dry matter intake and gross feed efficiency for first-parity Holstein cows using milk production traits and LW. Data comprised of 30 daily measurements of DMI and milk production traits, and 25 daily LW records of a group of 100 first-parity Holstein cows, fed a total mixed ration. Gross feed efficiency was calculated as kg ECM divided by kg DMI. The initial step was to estimate correlations of milk production traits and LW with DMI and GFE, to identify the best potential predictors of DMI and GFE. Subsequently, a forward stepwise regression analysis was used to develop models to predict DMI and GFE from LW and milk production traits, followed by within-herd validations. Means for DMI, butterfat yield (BFY) and LW were 21.91 ± 2.77 kg/day, 0.95 ± 0.14 kg/day and 572 ± 15.58 kg/day, respectively. Mean GFE was 1.32 ± 0.22. Dry matter intake had positive correlations with milk yield (MY) (r = 0.32, p < 0.001) and LW (r = 0.76, p < 0.0001) and an antagonistic association with butterfat percent (BFP) (r =  - 0.55, p < 0.001). On the other hand, GFE was positively associated with MY (r = 0.36, p < 0.001), BFP (r = 0.53, p < 0.001) and BFY (r = 0.83, p < 0.0001), and negatively correlated with LW (r =  - 0.23, p > 0.05). Dry matter intake was predicted reliably by a model comprising of only LW and MY (R² = 0.79; root mean squared error (RMSE) = 1.05 kg/day). A model that included BFY, MY and LW had the highest ability to predict GFE (R² = 0.98; RMSE = 0.05). Live weight and BFY were the main predictor traits for DMI and GFE, respectively. The best models for predicting DMI and GFE were as follows: DMI (kg/day) =  - 54.21 - 0.192 × MY (kg/day) + 0.146 × LW (kg/day) and GFE (kg/day) = 4.120 + 0.024 × MY (kg/day) + 1.000 × BFY (kg/day) - 0.008 × LW (kg/day). Thus, daily DMI (kg/day) and GFE can be reliably predicted from LW and milk production traits using these developed models in first-parity Holstein cows. This presents a big promise to generate large quantities of data of individual cow DMI and GFE, which can be used to implement genetic improvement of feed efficiency.

Phenotypic and genetic characterization of the occurrence of noncoagulating milk in dairy sheep

Milk coagulation ability is of central importance for the sheep dairy industry because almost all sheep milk is destined for cheese processing. The occurrence of milk with impaired coagulation properties is an obstacle to cheese processing and, in turn, to the profitability of the dairy companies. In this work, we investigated the causes of noncoagulation of sheep milk; specifically, we studied the effect of milk physicochemical properties on milk coagulation status [coagulating and noncoagulating (NC) milk samples, which do or do not coagulate within 30 min, respectively], and whether mid-infrared spectroscopy (MIR) could be used to assess variability in coagulation status. We also investigated the genetic background of milk coagulation ability. Individual milk samples were collected from 996 Sarda ewes farmed in 47 flocks located in Sardinia (Italy). Considered traits were daily milk yield, milk composition traits, and milk coagulation properties (rennet coagulation time, curd firming time, and curd firmness), and MIR spectra were acquired. About 9% of samples did not coagulate within 30 min. A logistic regression approach was used to test the effect of milk-related traits on milk coagulation status. A principal component (PC) analysis was carried out on the milk MIR spectra, and PC scores were then used as covariates in a logistic regression model to assess their relationship with milk coagulation status. Results of the present work demonstrated that the probability of having NC samples increases as milk contents of proteins and chlorides and somatic cell score increase. The analysis of PC extracted from milk spectra that influenced coagulation status highlighted key regions associated with lactose and protein concentrations, and others not associated with routinely collected milk composition traits. These results suggest that the occurrence of NC is mostly related to damage of the epithelium secretory mammary cells, which occurs with the advancement of a lactation or due to unhealthy mammary gland status. Genetic analysis of milk coagulation status and of the extracted PC confirmed the genetic background of the milk coagulability of sheep milk.

Are subjectively scored linear type traits suitable predictors of the genetic merit for feed intake in grazing Holstein-Friesian dairy cows?

Measuring dry matter intake (DMI) in grazing dairy cows using currently available techniques is invasive, time consuming, and expensive. An alternative to directly measuring DMI for use in genetic evaluations is to identify a set of readily available animal features that can be used in a multitrait genetic evaluation for DMI. The objectives of the present study were thus to estimate the genetic correlations between readily available body-related linear type traits and DMI in grazing lactating Holstein-Friesian cows, but importantly also estimate the partial genetic correlations between these linear traits and DMI, after adjusting for differences in genetic merit for body weight. Also of interest was whether the predictive ability derived from the estimated genetic correlations materialized upon validation. After edits, a total of 8,055 test-day records of DMI, body weight, and milk yield from 1,331 Holstein-Friesian cows were available, as were chest width, body depth, and stature from 47,141 first lactation Holstein-Friesian cows. In addition to considering the routinely recorded linear type traits individually, novel composite traits were defined as the product of the linear type traits as an approximation of rumen volume. All linear type traits were moderately heritable, with heritability estimates ranging from 0.27 (standard error = 0.14) to 0.49 (standard error = 0.15); furthermore, all linear type traits were genetically correlated (0.29 to 0.63, standard error 0.14 to 0.12) with DMI. The genetic correlations between the individual linear type traits and DMI, when adjusted for genetic differences in body weight, varied from -0.51 (stature) to 0.48 (chest width). These genetic correlations between DMI and linear type traits suggest linear type traits may be useful predictors of DMI, even when body weight information is available. Nonetheless, estimated genetic merit of DMI derived from a multitrait genetic evaluation of linear type traits did not correlate strongly with actual DMI in a set of validation animals; the benefit was even less if body weight data were also available.

Multiple Country Approach to Improve the Test-Day Prediction of Dairy Cows' Dry Matter Intake

Simple Summary

Dry matter intake, related to the number of nutrients available to an animal to meet its production and health needs, is crucial for the economic, environmental, and welfare management of dairy herds. Because the equipment required to weigh the ingested food at an individual level is not broadly available, we propose some new ways to approach the actual dry matter consumed by a dairy cow for a given day. To do so, we used regression models using parity (number of lactations), week of lactation, milk yield, milk mid-infrared spectrum, and prediction of bodyweight, fat, protein, lactose, and fatty acids content in milk. We chose these elements to predict individual dry matter intake because they are either easily accessible or routinely provided by regional dairy organizations (often called “dairy herd improvement” associations). We succeeded in producing a model whose dry matter intake predictions were moderately related to the actual values.

Abstract

We predicted dry matter intake of dairy cows using parity, week of lactation, milk yield, milk mid-infrared (MIR) spectrum, and MIR-based predictions of bodyweight, fat, protein, lactose, and fatty acids content in milk. The dataset comprised 10,711 samples of 534 dairy cows with a geographical diversity (Australia, Canada, Denmark, and Ireland). We set up partial least square (PLS) regressions with different constructs and a one-hidden-layer artificial neural network (ANN) using the highest contribution variables. In the ANN, we replaced the spectra with their projections to the 25 first PLS factors explaining 99% of the spectral variability to reduce the model complexity. Cow-independent 10 × 10-fold cross-validation (CV) achieved the best performance with root mean square errors (RMSE_CV) of 3.27 ± 0.08 kg for the PLS regression and 3.25 ± 0.13 kg for ANN. Although the available data were significantly different, we also performed a country-independent validation (CIV) to measure the models’ performance fairly. We found RMSE_CIV varying from 3.73 to 6.03 kg for PLS and 3.69 to 5.08 kg for ANN. Ultimately, based on the country-independent validation, we discussed the developed models’ performance with those achieved by the National Research Council’s equation.

Comparison of methods to predict feed intake and residual feed intake using behavioral and metabolite data in addition to classical performance variables

Predicting dry matter intake (DMI) and feed efficiency by leveraging the use of data streams available on farm could aid efforts to improve the feed efficiency of dairy cattle. Residual feed intake (RFI) is the difference between predicted and observed feed intake after accounting for body size, body weight change, and milk production, making it a valuable metric for feed efficiency research. Our objective was to develop and evaluate DMI and RFI prediction models using multiple linear regression (MLR), partial least squares regression, artificial neural networks, and stacked ensembles using different combinations of cow descriptive, performance, sensor-derived behavioral (SMARTBOW; Zoetis), and blood metabolite data. Data were collected from mid-lactation Holstein cows (n = 124; 102 multiparous, 22 primiparous) split equally between 2 replicates of 45-d duration with ad libitum access to feed. Within each predictive approach, 4 data streams were added in sequence: dataset M (week of lactation, parity, milk yield, and milk components), dataset MB (dataset M plus body condition score and metabolic body weight), dataset MBS (dataset MB plus sensor-derived behavioral variables), and dataset MBSP (dataset MBS plus physiological blood metabolites). The combination of 4 datasets and 4 analytical approaches resulted in 16 analyses of DMI and RFI, using variables averaged within cow across the study period. Additional models using weekly averaged data within cow and study were built using all predictive approaches for datasets M, MB, and MBS. Model performance was assessed using the coefficient of determination, concordance correlation coefficient, and root mean square error of prediction. Predictive models of DMI performed similarly across all approaches, and models using dataset MBS had the greatest model performance. The best approach-dataset combination was MLR-dataset MBS, although several models performed similarly. Weekly DMI models had the greatest performance with MLR and partial least squares regression approaches. Dataset MBS models had incrementally better performance than datasets MB and M. Within each approach-dataset combination, models with DMI averaged over the study period had slightly greater model performance than DMI averaged weekly. Predictive performance of all RFI models was poor, but slight improvements when using MLR applied to dataset MBS suggest that rumination and activity behaviors may explain some of the variation in RFI. Overall, similar performance of MLR, compared with machine learning techniques, indicates MLR may be sufficient to predict DMI. The improvement in model performance with each additional data stream supports the idea of integrating data streams to improve model predictions and farm management decisions.

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

Fourier-transform infrared (FTIR) spectroscopy is a powerful high-throughput phenotyping tool for predicting traits that are expensive and difficult to measure in dairy cattle. Calibration equations are often developed using standard methods, such as partial least squares (PLS) regression. Methods that employ penalization, rank-reduction, and variable selection, as well as being able to model the nonlinear relations between phenotype and FTIR, might offer improvements in predictive ability and model robustness. This study aimed to compare the predictive ability of 2 machine learning methods, namely random forest (RF) and gradient boosting machine (GBM), and penalized regression against PLS regression for predicting 3 phenotypes differing in terms of biological meaning and relationships with milk composition (i.e., phenotypes measurable directly and not directly in milk, reflecting different biological processes which can be captured using milk spectra) in Holstein-Friesian cattle under 2 cross-validation scenarios. The data set comprised phenotypic information from 471 Holstein-Friesian cows, and 3 target phenotypes were evaluated: (1) body condition score (BCS), (2) blood β-hydroxybutyrate (BHB, mmol/L), and (3) κ-casein expressed as a percentage of nitrogen (κ-CN, % N). The data set was split considering 2 cross-validation scenarios: samples-out random in which the population was randomly split into 10-folds (8-folds for training and 1-fold for validation and testing); and herd/date-out in which the population was randomly assigned to training (70% herd), validation (10%), and testing (20% herd) based on the herd and date in which the samples were collected. The random grid search was performed using the training subset for the hyperparameter optimization and the validation set was used for the generalization of prediction error. The trained model was then used to assess the final prediction in the testing subset. The grid search for penalized regression evidenced that the elastic net (EN) was the best regularization with increase in predictive ability of 5%. The performance of PLS (standard model) was compared against 2 machine learning techniques and penalized regression using 2 cross-validation scenarios. Machine learning methods showed a greater predictive ability for BCS (0.63 for GBM and 0.61 for RF), BHB (0.80 for GBM and 0.79 for RF), and κ-CN (0.81 for GBM and 0.80 for RF) in samples-out cross-validation. Considering a herd/date-out cross-validation these values were 0.58 (GBM and RF) for BCS, 0.73 (GBM and RF) for BHB, and 0.77 (GBM and RF) for κ-CN. The GBM model tended to outperform other methods in predictive ability around 4%, 1%, and 7% for EN, RF, and PLS, respectively. The prediction accuracies of the GBM and RF models were similar, and differed statistically from the PLS model in samples-out random cross-validation. Although, machine learning techniques outperformed PLS in herd/date-out cross-validation, no significant differences were observed in terms of predictive ability due to the large standard deviation observed for predictions. Overall, GBM achieved the highest accuracy of FTIR-based prediction of the different phenotypic traits across the cross-validation scenarios. These results indicate that GBM is a promising method for obtaining more accurate FTIR-based predictions for different phenotypes in dairy cattle.

Development of a subacute ruminal acidosis risk score and its prediction using milk mid-infrared spectra in early-lactation cows

A routine monitoring for subacute ruminal acidosis (SARA) on the individual level could support the minimization of economic losses and the ensuring of animal welfare in dairy cows. The objectives of this study were (1) to develop a SARA risk score (SRS) by combining information from different data acquisition systems to generate an integrative indicator trait, (2) the investigation of associations of the SRS with feed analysis data, blood characteristics, performance data, and milk composition, including the fatty acid (FA) profile, (3) the development of a milk mid-infrared (MIR) spectra-based prediction equation for this novel reference trait SRS, and (4) its application to an external data set consisting of MIR data of test day records to investigate the association between the MIR-based predictions of the SRS and the milk FA profile. The primary data set, which was used for the objectives (1) to (3), consisted of data collected from 10 commercial farms with a total of 100 Holstein cows in early lactation. The data comprised barn climate parameters, pH and temperature logging from intrareticular measurement boluses, as well as jaw movement and locomotion behavior recordings of noseband-sensor halters and pedometers. Further sampling and data collection included feed samples, blood samples, milk performance, and milk samples, whereof the latter were used to get the milk MIR spectra and to estimate the main milk components, the milk FA profile, and the lactoferrin content. Because all measurements were characterized by different temporal resolutions, the data preparation consisted of an aggregation into values on a daily basis and merging it into one data set. For the development of the SRS, a total of 7 traits were selected, which were derived from measurements of pH and temperature in the reticulum, chewing behavior, and milk yield. After adjustment for fixed effects and standardization, these 7 traits were combined into the SRS using a linear combination and directional weights based on current knowledge derived from literature studies. The secondary data set was used for objective (4) and consisted of test day records of the entire herds, including performance data, milk MIR spectra and MIR-predicted FA. At farm level, it could be shown that diets with higher proportions of concentrated feed resulted in both lower daily mean pH and higher SRS values. On the individual level, an increased SRS could be associated with a modified FA profile (e.g., lower levels of short- and medium-chain FA, higher levels of C17:0, odd- and branched-chain FA). Furthermore, a milk MIR-based partial least squares regression model with a moderate predictability was established for the SRS. This work provides the basis for the development of routine SARA monitoring and demonstrates the high potential of milk composition-based assessment of the health status of lactating cows.

Genetic mechanisms underlying feed utilization and implementation of genomic selection for improved feed efficiency in dairy cattle

The economic importance of genetically improving feed efficiency has been recognized by cattle producers worldwide. It has the potential to considerably reduce costs, minimize environmental impact, optimize land and resource use efficiency, and improve the overall cattle industry’s profitability. Feed efficiency is a genetically complex trait that can be described as units of product output (e.g., milk yield) per unit of feed input. The main objective of this review paper is to present an overview of the main genetic and physiological mechanisms underlying feed utilization in ruminants and the process towards implementation of genomic selection for feed efficiency in dairy cattle. In summary, feed efficiency can be improved via numerous metabolic pathways and biological mechanisms through genetic selection. Various studies have indicated that feed efficiency is heritable, and genomic selection can be successfully implemented in dairy cattle with a large enough training population. In this context, some organizations have worked collaboratively to do research and develop training populations for successful implementation of joint international genomic evaluations. The integration of “-omics” technologies, further investments in high-throughput phenotyping, and identification of novel indicator traits will also be paramount in maximizing the rates of genetic progress for feed efficiency in dairy cattle worldwide.

Predicting bovine tuberculosis status of dairy cows from mid-infrared spectral data of milk using deep learning

Bovine tuberculosis (bTB) is a zoonotic disease in cattle that is transmissible to humans, distributed worldwide, and considered endemic throughout much of England and Wales. Mid-infrared (MIR) analysis of milk is used routinely to predict fat and protein concentration, and is also a robust predictor of several other economically important traits including individual fatty acids and body energy. This study predicted bTB status of UK dairy cows using their MIR spectral profiles collected as part of routine milk recording. Bovine tuberculosis data were collected as part of the national bTB testing program for Scotland, England, and Wales; these data provided information from over 40,500 bTB herd breakdowns. Corresponding individual cow life-history data were also available and provided information on births, movements, and deaths of all cows in the study. Data relating to single intradermal comparative cervical tuberculin (SICCT) skin-test results, culture, slaughter status, and presence of lesions were combined to create a binary bTB phenotype labeled 0 to represent nonresponders (i.e., healthy cows) and 1 to represent responders (i.e., bTB-affected cows). Contemporaneous individual milk MIR spectral data were collected as part of monthly routine milk recording and matched to bTB status of individual animals on the single intradermal comparative cervical tuberculin test date (±15 d). Deep learning, a sub-branch of machine learning, was used to train artificial neural networks and develop a prediction pipeline for subsequent use in national herds as part of routine milk recording. Spectra were first converted to 53 × 20-pixel PNG images, then used to train a deep convolutional neural network. Deep convolutional neural networks resulted in a bTB prediction accuracy (i.e., the number of correct predictions divided by the total number of predictions) of 71% after training for 278 epochs. This was accompanied by both a low validation loss (0.71) and moderate sensitivity and specificity (0.79 and 0.65, respectively). To balance data in each class, additional training data were synthesized using the synthetic minority over sampling technique. Accuracy was further increased to 95% (after 295 epochs), with corresponding validation loss minimized (0.26), when synthesized data were included during training of the network. Sensitivity and specificity also saw a 1.22- and 1.45-fold increase to 0.96 and 0.94, respectively, when synthesized data were included during training. We believe this study to be the first of its kind to predict bTB status from milk MIR spectral data. We also believe it to be the first study to use milk MIR spectral data to predict a disease phenotype, and posit that the automated prediction of bTB status at routine milk recording could provide farmers with a robust tool that enables them to make early management decisions on potential reactor cows, and thus help slow the spread of bTB.

Infrared Spectrometry as a High-Throughput Phenotyping Technology to Predict Complex Traits in Livestock Systems

High-throughput phenotyping technologies are growing in importance in livestock systems due to their ability to generate real-time, non-invasive, and accurate animal-level information. Collecting such individual-level information can generate novel traits and potentially improve animal selection and management decisions in livestock operations. One of the most relevant tools used in the dairy and beef industry to predict complex traits is infrared spectrometry, which is based on the analysis of the interaction between electromagnetic radiation and matter. The infrared electromagnetic radiation spans an enormous range of wavelengths and frequencies known as the electromagnetic spectrum. The spectrum is divided into different regions, with near- and mid-infrared regions being the main spectral regions used in livestock applications. The advantage of using infrared spectrometry includes speed, non-destructive measurement, and great potential for on-line analysis. This paper aims to review the use of mid- and near-infrared spectrometry techniques as tools to predict complex dairy and beef phenotypes, such as milk composition, feed efficiency, methane emission, fertility, energy balance, health status, and meat quality traits. Although several research studies have used these technologies to predict a wide range of phenotypes, most of them are based on Partial Least Squares (PLS) and did not considered other machine learning (ML) techniques to improve prediction quality. Therefore, we will discuss the role of analytical methods employed on spectral data to improve the predictive ability for complex traits in livestock operations. Furthermore, we will discuss different approaches to reduce data dimensionality and the impact of validation strategies on predictive quality.

Large-scale phenotyping in dairy sector using milk MIR spectra: Key factors affecting the quality of predictions

Methods and technologies enabling the estimation at large scale of important traits for the dairy sector are of great interest. Those phenotypes are necessary to improve herd management, animal genetic evaluation, and milk quality control. In the recent years, the research was very active to predict new phenotypes from the mid-infrared (MIR) analysis of milk. Models were developed to predict phenotypes such as fine milk composition, milk technological properties or traits related to cow health, fertility and environmental impact. Most of models were developed within research contexts and often not designed for routine use. The implementation of models at a large scale to predict new traits of interest brings new challenges as the factors influencing the robustness of models are poorly documented. The first objective of this work is to highlight the impact on prediction accuracy of factors such as the variability of the spectral and reference data, the spectral regions used and the complexity of models. The second objective is to emphasize methods and indicators to evaluate the quality of models and the quality of predictions generated under routine conditions. The last objective is to outline the issues and the solutions linked with the use and transfer of models on large number of instruments. Based on partial least square regression and 10 datasets including milk MIR spectra and reference quantitative values for 57 traits of interest, the impact of the different factors is illustrated by evaluating the influence on the validation root mean square error of prediction (RMSEP). In the displayed examples, all factors, when well set up, increase the quality of predictions, with an improvement of the RMSEP ranging from 12% to 43%. This work also aims to underline the need for and the complementarity between different validation procedures, statistical parameters and quality assurance methods. Finally, when using and transferring models, the impact of the spectral standardization on the prediction reproducibility is highlighted with an improvement up to 86% with the tested models, and the monitoring of individual spectrometer stability over time appears essential. This list inspired from our experience is of course not exhaustive. The displayed results are only examples and not general rules and other aspects play a role in the quality of final predictions. However, this work highlights good practices, methods and indicators to increase and evaluate quality of phenotypes predicted at a large scale. The results obtained argue for the development of guidelines at international levels, as well as international collaborations in order to constitute large and robust datasets and enable the use of models in routine conditions.

Opportunities and limitations of milk mid-infrared spectra-based estimation of acetone and β-hydroxybutyrate for the prediction of metabolic stress and ketosis in dairy cows

Subclinical (SCK) and clinical (CK) ketosis are metabolic disorders responsible for big losses in dairy production. Although Fourier-transform mid-infrared spectrometry (FTIR) to predict ketosis in cows exposed to great metabolic stress was studied extensively, little is known about its suitability in predicting hyperketonemia using individual samples, e.g. in small dairy herds or when only few animals are at risk of ketosis. The objective of the present research was to determine the applicability of milk metabolites predicted by FTIR spectrometry in the individual screening for ketosis. In experiment 1, blood and milk samples were taken every two weeks after calving from Holstein (n = 80), Brown Swiss (n = 72) and Swiss Fleckvieh (n = 58) cows. In experiment 2, cows diagnosed with CK (n = 474) and 420 samples with blood β-hydroxybutyrate [BHB] <1.0 mmol/l were used to investigate if CK could be detected by FTIR-predicted BHB and acetone from a preceding milk control. In experiment 3, correlations between data from an in farm automatic milk analyser and FTIR-predicted BHB and acetone from the monthly milk controls were evaluated. Hyperketonemia occurred in majority during the first eight weeks of lactation. Correlations between blood BHB and FTIR-predicted BHB and acetone were low (r = 0.37 and 0.12, respectively, P < 0.0001), as well as the percentage of true positive values (11.9 and 16.6%, respectively). No association of FTIR predicted ketone bodies with the interval of milk sampling relative to CK diagnosis was found. Data obtained from the automatic milk analyser were moderately correlated with the same day FTIR-predicted BHB analysis (r = 0.61). In conclusion, the low correlations with blood BHB and the small number of true positive samples discourage the use of milk mid-infrared spectrometry analyses as the only method to predict hyperketonemia at the individual cow level.

Predicting milk fatty acids and energy balance of dairy cows in Australia using milk mid-infrared spectroscopy

Mid-infrared spectroscopy (MIRS) is traditionally used for analysing milk fat, protein and lactose concentrations in dairy production, but there is growing interest in using it to predict difficult, or expensive-to-measure, phenotypes on a large scale. The resulting prediction equations can be applied to MIRS data from commercial herd-testing, to facilitate management and feeding decisions, or for genomic selection purposes. We investigated the ability of MIRS of milk samples to predict milk fatty acids (FAs) and energy balance (EB) of dairy cows in Australia. Data from 240 Holstein lactating cows that were part of two 32-day experiments, were used. Milk FAs were measured twice during the experimental period. Prediction models were developed using partial least-square regression with a 10-fold cross-validation. Measures of prediction accuracy included the coefficient of determination (R2cv) and root mean-square error. Milk FAs with a chain length of ≤16 were accurately predicted (0.89 ≤ R2cv ≤ 0.95), while prediction accuracy for FAs with a chain length of ≥17 was slightly lower (0.72 ≤ R2cv ≤ 0.82). The accuracy of the model prediction was moderate for EB, with the value of R2cv of 0.48. In conclusion, the ability of MIRS to predict milk FAs was high, while EB was moderately predicted. A larger dataset is needed to improve the accuracy and the robustness of the prediction models.

Comparing regression, naive Bayes, and random forest methods in the prediction of individual survival to second lactation in Holstein cattle

In this study, we compared multiple logistic regression, a linear method, to naive Bayes and random forest, 2 nonlinear machine-learning methods. We used all 3 methods to predict individual survival to second lactation in dairy heifers. The data set used for prediction contained 6,847 heifers born between January 2012 and June 2013, and had known survival outcomes. Each animal had 50 genomic estimated breeding values available at birth and up to 65 phenotypic variables that accumulated over time. Survival was predicted at 5 moments in life: at birth, at 18 mo, at first calving, at 6 wk after first calving, and at 200 d after first calving. The data sets were randomly split into 70% training and 30% testing sets to evaluate model performance for 20-fold validation. The methods were compared for accuracy, sensitivity, specificity, area under the curve (AUC) value, contrasts between groups for the prediction outcomes, and increase in surviving animals in a practical scenario. At birth and 18 mo, all methods had overlapping performance; no method significantly outperformed the other. At first calving, 6 wk after first calving, and 200 d after first calving, random forest and naive Bayes had overlapping performance, and both machine-learning methods outperformed multiple logistic regression. Overall, naive Bayes has the highest average AUC at all decision points up to 200 d after first calving. Random forest had the highest AUC at 200 d after first calving. All methods obtained similar increases in survival in the practical scenario. Despite this, the methods appeared to predict the survival of individual heifers differently. All methods improved over time, but the changes in mean model outcomes for surviving and non-surviving animals differed by method. Furthermore, the correlations of individual predictions between methods ranged from r = 0.417 to r = 0.700; the lowest correlations were at first calving for all methods. In short, all 3 methods were able to predict survival at a population level, because all methods improved survival in a practical scenario. However, depending on the method used, predictions for individual animals were quite different between methods.