Fig. 8

Fertility in the main cohort can be predicted from DMCs. A–C Random Forest models were built on three different training sets from the DNA methylation values of 107 fertility-related DMCs without missing values. The models were then run on three different testing sets, and predicted fertility status was compared against the actual fertility status of bulls included in each type of testing set. Receiver operating characteristics (ROC) curves and four performance indicators are displayed: accuracy (correct prediction rate), area under the ROC curve (AUC), sensitivity (true positive rate) and specificity (true negative rate). D–F For each model, principal component analysis was run on the 107 fertility-related DMCs and an individual factor map is displayed, with bulls colored according to the consistency between their predicted and actual fertility status (red: fertile predicted as fertile, FF; blue: subfertile predicted as subfertile, SS; orange: fertile predicted as subfertile, FS; light blue: subfertile predicted as fertile, SF). Confidence ellipses are indicated. A, D the main cohort was split into a training set of 67 bulls and a testing set of 33 bulls, the proportions of fertile and subfertile bulls being the same in each set. The average performance indicators were calculated after 50 iterations with a random resampling of the training and testing sets. B, E The training set included bulls originating from semen collection center 1 while the testing set included bulls originating from semen collection center 2. C, F The training set included bulls from center 2 while the testing set included bulls from center 1. Whatever the model, the performance indicators were within the same range; misclassified bulls were found in both fertility classes and exhibited a DNA methylation pattern typical of the opposite class