Predicting sperm production of young dairy bulls using collection history and management factors

Abstract

Selection of elite young dairy bulls by using genomic data shortened the generation interval and increased pressure to collect and market germplasm at an early age. The objectives of this study were (1) develop prediction models for daily, weekly, and monthly total sperm (TSp) production from collection history, health status, and management factors, and (2) assess the ability of these models to forecast future TSp production, as well as differences in prediction accuracy by seasonality or age of bull. Data consisted of 43,918 daily processing records from 1,037 Holstein and Jersey bulls between 10 and 28 mo of age at collection. Potential explanatory variables included year and season of collection, barn location, collection frequency, breed, scrotal circumference, TSp in previous months, health events, and age at arrival, first collection, and current collection. Linear regression, random forest (RF), Bayesian regularized neural network, model tree, multilayer perceptron neural network with multiple layers, and extreme learning machine were used to predict daily, weekly, and monthly TSp (R v3.5.1, https://www.r-project.org/). In the additive approach, all prior data were used for training; however, in the fixed-window approach, records from 3 previous months were used for age-based prediction, records from 4 previous months or 1 yr were used for the monthly date-based analyses, and records from 1 previous month or year were used for the weekly date-based analyses. Model performance was measured by root mean squared error (RMSE) and the correlation (r) between actual and predicted TSp in testing sets. In monthly analyses, RF with additive training performed best in age-based (RMSE = 13.6 billion cells, r = 0.93) and date-based (RMSE = 11.9, r = 0.94) prediction, compared with linear regression (age-based RMSE = 16.6, r = 0.89; date-based RMSE = 15.5, r = 0.90) and Bayesian regularized neural network (age-based RMSE = 14.1, r = 0.92). On average, RMSE was 0.93 or 0.14 billion cells greater with fixed 4-mo or 1-yr training windows, respectively, than in the additive analyses. The most important management variables affecting TSp were collection frequency, TSp in previous months, and age at collection. Results indicate RF models with additive training can predict TSp output of individual bulls with ≥85% accuracy up to 4 mo into the future. Spikes in accuracy were associated with sire summary times and company processing changes, and accuracy tended to stabilize when bulls reached 19 to 20 mo of age.