Assessing the predictive potential of the Holstein bovine sperm methylome for bull fertility using machine learning

Bull fertility is a key parameter in the dairy cattle industry, with significant economic implications. DNA methylation is an epigenetic mark essential for fertility, and recent works have highlighted sperm DNA methylation as a promising source of biomarkers for male fertility, but no operational tool has been developed for predicting bull fertility based on DNA methylation. Here, we leveraged existing methylation data to evaluate the potential of DNA methylation for predicting bull fertility with machine learning models.

Our study was conducted on a cohort of 351 Holstein commercial bulls, classified as subfertile, neutral and fertile based on a field fertility index (non-return rate of the inseminated cows after 56 days post-insemination). This dataset includes the largest cohort assembled for bull fertility prediction from epigenetic markers. We developed classification models to discriminate subfertile and fertile individuals using a subset of predictors derived from reduced representation bisulfite sequencing data obtained on frozen semen. Four machine learning algorithms were compared: Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest, Gradient Boosting, and Neural Networks. To optimize performances of each algorithm and conduct an unbiased assessment of classification accuracy, we used a three-fold cross-validation before final testing on an independent test set. To increase the robustness of the results, model performances were averaged over 20 partitions of the data into training and test sets, to account for variability in bull distribution across sets. All models performed better than random classification. The best model was LASSO according to classification accuracy (0.62) and Random Forest according to Area Under Curve (0.69) computed with the independent test set.

Even with this moderate performance, our models would lead to a 12.4% point increase of the probability of identifying subfertile individuals. These results suggest that sperm DNA methylation holds potential as a relevant source of predictive markers for bull fertility.