Symposium Review: Development of genomic evaluation for methane efficiency in Canadian Holsteins

: Reducing methane (CH 4 ) emissions from agriculture, among other sectors, is a key step to reduce global warming. There are many strategies to reduce CH 4 emissions in ruminant animals, including genetic selection, which yields cumulative and permanent genetic gains over generations. A single-step genomic evaluation for Methane Efficiency (ME) was officially implemented in April 2023 for the Canadian Holstein breed, aiming to reduce CH 4 emissions without impacting production levels. This evaluation was achieved by using milk mid-infrared (MIR) spectral data to predict individual cow CH 4 production. The genetic evaluation model included milk MIR predicted CH 4 (CH4 MIR ), along with milk yield (MY), fat yield (FY), and protein yield (PY), as correlated traits. Traits were expressed in kg/day (MY, FY, and PY) or g/day (CH4 MIR ). The MiX99 software was used to fit the single-step, 4-trait animal model. Genomic breeding values for CH4


Symposium Review Genetics
Abstract: Reducing methane (CH 4 ) emissions from agriculture, among other sectors, is a key step to reduce global warming.There are many strategies to reduce CH 4 emissions in ruminant animals, including genetic selection, which yields cumulative and permanent genetic gains over generations.A single-step genomic evaluation for Methane Efficiency (ME) was officially implemented in April 2023 for the Canadian Holstein breed, aiming to reduce CH 4 emissions without impacting production levels.This evaluation was achieved by using milk mid-infrared (MIR) spectral data to predict individual cow CH 4 production.The genetic evaluation model included milk MIR predicted CH 4 (CH4 MIR ), along with milk yield (MY), fat yield (FY), and protein yield (PY), as correlated traits.Traits were expressed in kg/day (MY, FY, and PY) or g/day (CH4 MIR ).The MiX99 software was used to fit the single-step, 4-trait animal model.Genomic breeding values for CH4 MIR were then obtained by re-parameterization, using recursive genetic linear regression coefficients on MY, FY, and PY, giving a measure of ME that is genetically independent of the production traits.The estimated breeding values were expressed as Relative Breeding Values (RBV) with a mean of 100 and standard deviation of 5 for the genetic base population, where a higher value indicates the animal produces lower predicted CH 4 .This national genomic evaluation is another tool that will lower the dairy industry's carbon footprint by reducing CH 4 emissions without impacting production traits.
C oncerns about the effects of climate change on environmental sustainability are growing.Increasingly, communities, organizations, and individuals are actively thinking about how they can be more sustainable and create a balanced ecosystem for future generations.This collective effort is reflected in the commitments made by numerous global dairy industry stakeholders, including Dairy Farmers of Canada, to achieve net-zero greenhouse gas (GHG) emissions by 2050.Methane (CH 4 ), a potent GHG, which remains in the atmosphere for about 12 years and makes up 14% of Canada's GHG emissions, has been under the spotlight as it is responsible for nearly half the net global temperature change due to human activities in the last decade (Environment and Climate Change Canada, 2022).The goal is to reduce the global CH 4 production and limit global warming to 1.5°C by 2050 (IPCC, 2021;The Global Methane Pledge, 2021).
Even though the agriculture industry is not the sole source of increasing global CH 4 emissions, it has the potential to mitigate this increase and contribute to climate cooling by reducing its rate of CH 4 emissions.Looking at the dairy industry specifically and accounting for all on-farm and off-farm GHGs involved in the production of 1 kg of milk on Canadian dairy farms, CH 4 appears as the largest contributor to the milk footprint at close to 50% (Groupe AGEGO, 2018).It mostly comes from digestion (enteric fermenta-tion) and, to a lesser extent manure management.Previous studies have identified that an average Holstein cow produces roughly 426 to 463 g of CH 4 per day (Denninger et al., 2019;Kamalanathan et al., 2023).Moreover, CH 4 production is heritable ranging from 0.12 to 0.45 (Lassen and Løvendahl, 2016;Breider et al., 2019;Kamalanathan et al., 2023), which presents an opportunity to decrease CH 4 emissions by using genetic selection.
Using genetics to select cows with reduced CH 4 emissions is a permanent and cumulative solution for reducing the dairy sector's GHG emissions.However, measuring CH 4 is expensive using standard methods (i.e., GreenFeed and sniffers) and difficult, resulting in few animals with recorded CH 4 (Shetty et al., 2017).Alternative methods, such as the use of mid-infrared (MIR) technology, offer cost-effective solutions to predict CH 4 emissions on a larger scale (Dehareng et al., 2012;Shadpour et al., 2022).The objective of this paper was to present the development and implementation of a routine genomic evaluation system for Methane Efficiency (ME), launched officially in Canada in April 2023 for the Holstein breed.200 cows) between 2016 and 2022 as part of the Efficient Dairy Genome Project (https: / / genomedairy .ualberta.ca/ ) and the Resilient Dairy Genome Project (RDGP, http: / / www .resilientdairy.ca/).Animal Care Committee approval was obtained from the University of Guelph (animal utilization protocol number: 3503) and from the University of Alberta (animal utilization protocol number: AUP00000170).The data recording at these 2 stations (one GreenFeed per station) is described in Kamalanathan et al. (2023) and Liu et al. (2022).Briefly, first lactation cows between 120 and 150 DIM at the Ontario Dairy Research Station were moved into a tie-stall area.For CH 4 emission testing, the GreenFeed system (C-Lock Inc., Rapid City, SD, USA) was moved in front of the animal and CH 4 and carbon dioxide concentrations were measured for roughly 10 min, 4 times a day (8am, 12pm, 4pm, and 8pm, from 2016 to 2019) or 3 times a day (8am, 12pm, and 4pm, from 2019 to 2023) for 5 consecutive days.This assured the sampling of numerous short-term records at different times of the day across multiple days as suggested as the best practice for the GreenFeed system.Methane production data for each day, in g/day, was the average of the estimated CH 4 production during the 3 or 4 testing sessions.
At the Dairy Research and Technology Center CH 4 emissions were recorded on mixed parity cows between 3 and 240 DIM housed in a tie-stall barn, starting in 2016.Methane measurements were recorded via the GreenFeed system for 12 consecutive days twice per day at 12 h intervals that shifted by an hour each day to cover all 24 h.In 2019, the collection was changed to 3 times per day (8am, 12pm, and 4pm) for 5 consecutive days to better mirror the sampling at the Ontario Dairy Research Station.

MIR DATA
Individual milk sample MIR spectra collected routinely in Lactanet labs in the provinces of Quebec, Ontario, Alberta (samples from both Alberta and Saskatchewan), and British Columbia were included in the analysis.Roughly 13 million spectra records on 1.6 million Holstein dairy cows from 7,171 herds were collected from 2018 to 2022 using the CombiFoss TM 7 instruments (Foss Electric A/S, Hillerød, Denmark).Uninformative spectral regions and those related to the high absorption of water were removed, resulting in 241 spectral data points.Those regions remaining for the analysis were roughly: 1,000 to 1,550 cm −1 (FOSS MIR pins 260 to 402); 1,705 to 1,820 cm −1 (FOSS MIR pins 442 to 472); and 2,700 to 2,955 cm −1 (FOSS MIR pins 701 to 767).Milk MIR spectra were standardized between laboratories and across time using the approach described by Bonfatti et al. (2017).This methodology uses principal component analysis (PCA) to inspect shifts in PCA scores over time to define subsets with homogenous spectra for the creation of standardization matrices.Principal components for each lab were obtained using the prcomp function in R, using data scaled by the scale2 function.All major principal components explaining more than 1% of variance were further investigated to define the changes in patterns over time.For quality control of the spectra within each subset, outliers were detected using Mahalanobis distance and removed based on the Chi-squared test statistic (P < 0.001).Following the standardization, spectral smoothing pretreatment was applied using a Savitzky-Golay filter with a third-order polynomial and filter width 11 (Savitzky and Golay, 1964).The standardized spectra from all labs were combined with test-day production records and only those recorded from first lac-tation animals with a calving age of at least 16 mo and between 5 and 305 DIM were retained for further analyses.

MIR PREDICTION OF CH 4 EMISSIONS
Previous analysis showed that the multilayer perceptron artificial neural network based on Bayesian regularization had better prediction performance compared with linear regression models such as PLS regression model, due to the ability to model complex patterns (Shadpour et al., 2022).Therefore, the impact of different input variables was tested using solely this approach.Only those variables available in a complete test day production record were tested for inclusion as potential input variables along with milk MIR (i.e., milk, fat, and protein yields, season, lactation, age, and DIM).Ultimately, milk MIR data was used as a sole input variable to have a more flexible prediction to apply to historical data.Following the approach by Shadpour et al. (2022), predictions were performed using the "brnn" function available in the R software (Perez Rodriguez and Gianola, 2022), fitting 2-layers neural networks containing 2 neurons.The number of epochs was set to 100 and all other model parameters were set as the default values.
After determining the optimum Bayesian regularized artificial neural network prediction model and the milk MIR input using information from preliminary analysis and Shadpour et al. (2022), predicting daily CH 4 and the weekly average CH 4 production were compared.Weekly averages were calculated based on the week of measurement of each animal.Methane records were matched with the closest test day milk MIR data.The weekly averages yielded higher accuracies compared with the daily averages and were therefore used for the final model.Only records from first lactation cows within 5 and 305 DIM with at least 2 daily CH 4 measurements in the weekly average were included in the analysis and outliers were removed based on the standard deviation (>|3.5|SD).Additionally, a record was required to have a corresponding milk MIR spectrum within 11 d from the middle day of the week measurement.After editing, the data set used for the prediction included 496 animals, where 96% of cows had their milk MIR data within the same week.Descriptive statistics of the data are shown in Table 1.
Prediction accuracy (r) was measured as phenotypic Pearson correlation between recorded and predicted CH 4 .Methane predictions using weekly CH 4 averages had a prediction accuracy of 0.70.This result was similar to the values previously reported by Shadpour et al. (2022), when using milk MIR data and no milk production data for prediction (i.e., r = 0.57-0.72).Moreover, RMSE for the prediction equation was 58.62 g/d.The ability of milk MIR information to accurately predict CH 4 can be partially explained by the fact that milk MIR spectra include detailed information on milk components including milk fatty acid profiles, which is related to enteric fermentation and CH 4 emissions (Delfosse et al., 2010;Dijkstra et al., 2011;Dehareng et al., 2012).

ESTIMATION OF VARIANCE COMPONENTS
where y is a vector of observations for CH4 MIR , MY, PY, and FY, b is a vector of all fixed effects (age at calving, DIM and year-season of calving), htd is a vector of the random HTD effect, a is a vector of random animal additive genetic effects, p is a vector of random permanent environmental (PE) effects, e is a vector of random residuals, and X, Z 1 , Z 2 , and Z 3 are the respective incidence matrices.Random effects were assumed to be normally distributed, with means equal to zero.Covariance structure of the 4-trait analysis was as follows: where HTD is the (co)variance (4 × 4) matrix between traits for HTD effects; G is the genetic covariance (4 × 4) matrix between traits for animal additive genetic effects; P is the (co)variance (4 × 4) matrix between traits due to PE effects; R is the residual covariance (4 × 4) matrix between traits; ⊗ is the Kronecker product; A is the additive genetic relationship matrix, and I is an identity matrix.Average estimates of genetic parameters from the multi-trait analyses are reported in Table 2. Average heritability for CH4 MIR was 0.23 (0.01), which was in line with previously reported heritability estimates for CH4 MIR (Kandel et al., 2017) and other CH 4 traits (Lassen and Løvendahl, 2016;Kamalanathan et al., 2023;van Breukelen et al., 2023).Average heritability estimates for MY, FY, and PY were 0.38 (0.01), 0.27(0.01),and 0.28(0.01),respectively.These values were in line with those estimated for the official genetic evaluation for these traits (Lactanet, 2021).The CH4 MIR had a moderately positive genetic correlation of 0.38 with FY.Kandel et al. (2017) also found a positive genetic correlation with FY that increased from approximately 0.10 to 0.20 across 120 to 180 DIM when using a DIM and milk MIR data in the model to predict daily CH 4 emissions.Similarly, a positive genetic correlation of 0.21 between FY and CH 4 production was reported by Pszczola et al. (2019).This means that cows with high FY emit more CH 4 , which can be explained by the positive relationship between milk fatty acids (MFA) and CH 4 production (Dijkstra et al., 2011).Odd-and branched-chain MFA are related to volatile fatty acids in the rumen (Vlaeminck and Fievez, 2005), which play a role in methanogenesis (Ellis et al., 2008).Moreover, MIR was reported to be a good predictor of MFA (Soyeurt et al., 2011;Fleming et al., 2017), and accordingly, CH4 MIR has a positive moderate correlation with FY.Nonetheless, a slight negative correlation of −0.13 and −0.11 was estimated for MY and PY with CH4 MIR , respectively.Kandel et al. (2017) also found a similar negative genetic correlation with MY around −0.20 and with PY that ranged from approximately −0.18 to −0.08 across 120 to 180 DIM.

GENOMIC EVALUATION OF METHANE EFFICIENCY
A multiple-trait single-step genomic evaluation (ssGBLUP) was implemented at Lactanet Canada using MiX99 and related software (MiX99 Development Team, 2017).Results presented here refer to data extracted for April 2023 official evaluations, which included 773,743 CH4 MIR , MY, FY, and PY records from 541,565 primiparous Holsteins between 120 and 185 DIM (daughters of 10,765 sires) from 6,128 herds.Average MY, FY and PY were 32.5 (SD = 6.2), 1.3 (SD = 0.3), and 1.1 (SD = 0.2) kg/d, respectively.On average, cows produced 492 g of CH4 MIR per day (average body weight and DMI; 634.9 (SD = 61.9)kg and 20.1 (SD = 2.2) kg/d, respectively), which was slightly higher than previous studies using GreenFeed (Denninger et al., 2019;Kamalanathan et al., 2023).However, this can be explained by the range in DIM used in the current study.
There were 134,963 genotyped animals in the 5 generations of pedigree, including 68,138 cows with data (i.e., CH4 MIR , MY, FY, and PY records) and genotypes and 7,921 sires of cows with data.Animals were genotyped either with a 50K SNP panel or imputed to 50K using F-Impute (Sargolzaei et al., 2014).The linear animal model for genomic prediction was the same for each of the 4 traits (model specified above).However, in ssGBLUP matrix A was replaced by H matrix, which combines pedigree and genomic information.Reliability of GEBV was approximated by a weighted (80:20) average of Direct Genomic Value (DGV) and animal model reliabilities (Sullivan et al., 2005;Sullivan, 2010).Direct Genomic Value reliabilities were calculated using SNP prediction error co-variances with the SNP-BLUP-REL software (Luke, Finland).Animal model reliabilities were calculated based on Effective Daughter Contributions (Sullivan, 2005).Methane Efficiency was defined as genetic residual CH 4 production, or CH 4 genetically independent of milk, fat, and protein yield, and derived using a recursive model operational tool (Jamrozik et al., 2017(Jamrozik et al., , 2021)), and the scale of the final ME is then reversed.As expected, ME had no genetic correlation with MY, PY, nor FY, meaning selection for ME will have no impact on milk production.Reliabilities of GEBV for ME, being a linear function of 4 traits, were approximated by a selection index method (Sullivan et al., 2005).

RELATIVE BREEDING VALUES
To maintain consistency with other Canadian genetic evaluations, ME was expressed as a Relative Breeding Value (RBV), with a mean of 100 and a standard deviation of 5 for base bulls (born 2008-2017, with an official ME evaluation).The higher values indicate a higher (more desirable) ME for an animal.Therefore, the higher a sire's RBV for ME, the less CH 4 their daughters are expected to produce.Moreover, predicted and recorded CH4 emissions were compared for low (<-1SD), medium (−1 ≥ SD ≤ +1), and high (>+1SD) RBV classes for ME (Figure 1).On average, cows in the lowest RBV class (<95; 72 cows) had the highest recorded CH 4 emissions at 518.The average reliability of RBVs for ME of the 68,138 genotyped cows was 86.7% (SD = 1.5%; ranging from 68% to 95%), whereas the average reliability for the 473,427 non-genotyped cows was 56.3% (SD = 3.6%).There are 2,142 Holstein sires with an official evaluation status for ME, with an average reliability of 95.9% (SD = 2.5%) ranging from 72% to 99%.To have an official evaluation, a sire is required to have 20 daughters from 5 herds and a minimum reliability of 70%.Daughter averages for CH4 MIR were regressed on sires' RBV for ME to translate RBV to an equivalent expected reduction in CH 4 for their daughters.Daughters of bulls with a breed average 100 RBV for ME produced on average 486.5 g/d of CH4 MIR .A 5-point RBV increase for ME (1 SD) has the expected effect of decreasing CH4 MIR in daughters by 7.55 g/d, or 3 kg per year, which is approximately a 1.5% reduction in CH 4 emissions per cow per year.Overall, herd owners consistently selecting animals with RBVs for ME of 105 or greater, can achieve a 20 to 30% reduction in methane emissions from their herd by 2050.

RELATIONSHIPS WITH OTHER TRAITS
To estimate proof correlations between ME and the 2 selection indexes (LPI and Pro$) as well as other routinely evaluated traits in Canada, 2,142 genotyped Holstein bulls with official evaluations for ME were selected.Methane Efficiency proofs were uncorrelated with LPI, Pro$, and the production traits (Milk, Fat, and Protein).Slightly positive and favorable correlations of 0.17 and 0.23 were identified with the Health and Fertility component of LPI and Metabolic Disease Resistance, respectively.In addition, ME did not have a significant correlation with Feed Efficiency (proof correlation of −0.14).This means that selection for Methane Efficiency will not affect Feed Efficiency, which is a trait calculated to be genetically independent of energy corrected milk and metabolic body weight.

LIMITATIONS
Some limitations in this study should be noted.The prediction of methane using milk MIR data can be further refined, as the current model is based on a limited sample size from 2 research stations.The plan is to enlarge the quantity of data used for the prediction as new records become available and to validate the prediction formally using an external data source, independent of the data set used for model training.Nevertheless, the prediction is accurate enough to start genomic selection for this important new trait in Canada and move the dairy cattle population in the right direction by reducing CH4 emissions.The strict protocol of CH 4 collection, the focus on using only mid-lactation first parity Holsteins, the robust and homogeneous milk MIR data recording and storage across Lactanet labs in Canada are all elements that provide a sound technical decision for the launch of genomic evaluation for Methane Efficiency.

CONCLUSIONS
Predicting average daily CH 4 production using milk MIR was proven to be a key and rapid alternative to direct CH 4 measurements.Predicted CH 4 using milk MIR data led to the development of routine genomic evaluations for ME that were officially implemented in April 2023 for the Holstein breed in Canada, focusing on selection for reduced CH 4 emissions without affecting milk, fat, and protein production levels.Methane Efficiency evaluations are a new tool to help reduce the dairy industry's environmental footprint and contribute to the goal of reaching net zero GHG emissions by 2050 without impacting milk production.
Variance components for CH4 MIR , milk yield (MY), fat yield (FY), and protein yield (PY) were estimated using data from the August 2022 data extraction.The final data contained 659,701 records from 462,120 first lactation cows (120-185 DIM) in 5,804 herds.Due to computational demand, 5 different subsets, each representing 10% of the herds from the final data set were used for the Oliveira et al. | Development of genomic… variance components estimation.On average, subsets contained 64,803 records from 45,137 first lactation cows in 580 herds.Variance components were estimated for each trait in AIREMLF90 using the AI-REML method (Misztal et al., 2014), with the following 4-trait linear animal model: 8 g/d and CH4 MIR at 521.0 g/d.Cows with RBVs between 95 and 105 (355 cows) had recorded CH 4 and CH4 MIR emissions of 492.8 g/d and 489.7 g/d, respectively.Lastly, cows with the highest RBVs for ME (>105; 44 cows) had average recorded CH 4 and CH4 MIR emissions of 464.0 g/d and 465.8 g/d, indicating that selecting animals with high ME evaluations will lead to a reduction in CH 4 emissions.
Figure 1.MIR predicted CH 4 and recorded CH 4 (g/d) for low, medium, and high classes of Methane Efficiency Relative Breeding Values (n = 496).