THE RESILIENT COW: NEXT-GENERATION SELECTION USING HIGH-FREQUENCY PHENOTYPES TO ACHIEVE PREDICTABLE PERFORMANCE IN UNPREDICTABLE CONDITIONS

<b>Project Methods</b><br> In Aim 1a, we will use data from commercial dairy farms served by Dairy Records Management Systems (DRMS; Raleigh, NC) that provide daily off-site backups. A data use agreement is in place, and our database contains more than 150 million daily MY records of 420,000 cows on 300 large commercial farms, including about 30 herds with AMS data that also include milking behavior. These data are linked to Dairy Herd Improvement records containing pedigree, calving, health, and fertility data. A relational database has already been created, and data cleaning and validation pipelines have been implemented.In Aim 1b, we will use data from research farms participating in the Resilient Dairy Genome Project (RDGP) led by the University of Guelph (Guelph, ON; http://www.resilientdairy.ca). UW-Madison is a full partner in Activity 3: "Feed Efficiency and Methane Reduction", where we are the largest contributor of daily feed intake phenotypes. Our data include 1.3 million daily DMI records of 11,700 Holstein cows on 42 research farms, along with milk, fat, protein, lactose, somatic cell count, milk urea nitrogen, BW, BCS, and (for a subset) CH4, CO2, β-hydroxybutyrate, and mid-infrared milk spectrum. Data from Blaine, the Ontario Dairy Research Center at the Elora Research Station (Elora; Ariss, ON), and the USDA-ARS Beltsville Agricultural Research Center (BARC; Beltsville, MD) also include detailed feeding behaviors, including the time, duration, intake, and feeding rate for every bunk visit throughout the day.In Aim 1c, we will use data from multiple sources that have precision livestock farming technologies offered by commercial vendors. The first consists of data from 1,050 cows with high-frequency phenotypes recorded by our SMARTBOW monitoring system at Blaine from May 2019 to present. These data include physical location (stall, feed bunk, alley, milking parlor), activity type (lying or standing), activity level (inactive, active, highly active), and rumination time of individual cows, recorded at 4-second intervals and summarized hourly. The second consists of guided AMS data of Fadul-Pacheco et al. (2021b), which are from February 2019 to present for a farm with 215 cows dispersed across four pens. The third consists of image data from lactating cows at Blaine and growing heifers at the UW-Madison Marshfield Agricultural Research Station (MARS; Stratford, WI). Image and video data from Blaine are collected by 40 RGB cameras positioned in the pens for behavior evaluation and four depth cameras (Intel RealSense D455) located at the exits of the milking parlor for BW and BCS evaluation. Image data at MARS are captured by a deployed edge-computing system composed of 30 depth cameras (Intel RealSense D455) and 30 NVIDIA Jetson Nano computers. Depth cameras are positioned for a top-down view above the water troughs. Animals are identified by image analyses using coat color patterns and alpha-numeric collars. Next, we will compute deviations from these expected curves using methods that include, but are not limited to, the LnVar, rauto, and skewness of daily deviations from predicted curves. Therefore, we will initially consider the polynomial quantile regression approach of Poppe et al. (2020).In Aim 2a, we will develop cohort-based methods that can be used to detect management and environmental disturbances in the absence of knowledge regarding the underlying causes of these events. Information about the specific pen locations of individual animals on specific dates is available in all data sets described previously, and the interaction of pen location and calendar date will define a cohort. Dates on which a supermajority of cows deviate significantly in a negative direction will be flagged as putative disturbances, while the proportions of animals that constitute a supermajority and the appropriate thresholds for declaring declines in performance as significant will be evaluated using various statistical measures based on the distribution of deviations for specific pens and calendar dates. The precision and recall of competing methods and their parameters will be evaluated by cross-validation.Following the identification of putative management and environmental disturbances, in Aim 2b, we will extract more information about herd management activities and local environmental conditions from subsets of herds with these detailed data, to determine the extent to which we are detecting real shocks and noises and not simply rediscovering routine herd management interventions. This step is important, because (for example) labeling cows as resilient if they fail to exhibit behavioral changes when in estrus would be detrimental to the overall breeding objective. Weather station data, which are readily available as THI from the (National Oceanic and Atmospheric Administration (NOAA), Asheville, NC) have been used to identify periods of heightened thermal stress in dairy cattle and other species (e.g., Nguyen et al., 2016; Misztal, 2017). These data will be matched with individual resilience phenotypes and cohort-based assessments of putative disturbances from Aim 2a to validate the impact of nutritional variation on daily phenotypes and resilience.In Aim 3a, we will carry out the core genetics and genomics components of the proposal, focusing on estimation of genetic parameters for the traits described in Aims 1a, 1b, and 1c, as well as their relationships with existing traits in the national breeding objective. The phenotypic measures of resilience will be as described previously, and the genetic data will represent a combination of pedigree and single nucleotide polymorphism (SNP) genomic data. Most of the cows contributing phenotypes to Aim 1c also have genomic data, because these herds include Blaine and several nearby commercial herds that cooperate with our group regularly and routinely genotype all heifer calves for the 79,294 (actual and imputed) SNPs used in Council on Dairy Cattle Breeding (CDCB; Bowie, MD) genetic evaluations. Most of the cows that contributed daily MY phenotypes to Aim 1a have only pedigree data, obtained from DRMS, but nearly all their sires have genomic data. The specific form of the statistical model for parameter estimation and breeding value prediction will likely vary by trait, but it will be a single- or multiple-trait version of the typical repeatability animal model, implemented using single-step genomic best linear unbiased prediction (ssGBLUP; Aguilar et al., 2011).Lastly, in Aim 3b we will carry out the critical steps of estimating the economic value of resilience and forecasting the gains that can be achieved in lifetime net profit by including one or more resilience traits in the breeding objective. Aim 3b of our proposal is inextricably linked to Aim 2, because we cannot calculate the economic value of resilience without knowing the frequency, severity, and duration of environmental and management disturbances that may compromise health and performance. We will use a bioeconomic Markov Chain model following Hietela et al. (2014) and calculate the economic value of resilience as the numerical approximation of the partial derivative of the profit function of resilience with respect to the population mean when the herd structure reaches steady state. Results from Aims 2a and 2b will inform the transition matrices of the Markov chain structure and the outcomes associated with the genetic expression of the resilience traits.