Random forest estimation of genomic breeding values for disease susceptibility over different disease incidences and genomic architectures in simulated cow calibration groups

A simulation study was conducted to investigate the performance of random forest (RF) and genomic BLUP (GBLUP) for genomic predictions of binary disease traits based on cow calibration groups. Training and testing sets were modified in different scenarios according to disease incidence, the quantitative-genetic background of the trait (h(2) = 0.30 and h(2) = 0.10), and the genomic architecture [725 quantitative trait loci (QTL) and 290 QTL, populations with high and low levels of linkage disequilibrium (LD)]. For all scenarios, 10,005 SNP (depicting a low-density 10K SNP chip) and 50,025 SNP (depicting a 50K SNP chip) were evenly spaced along 29 chromosomes. Training and testing sets included 20,000 cows (4,000 sick, 16,000 healthy, disease incidence 20%) from the last 2 generations. Initially, 4,000 sick cows were assigned to the testing set, and the remaining 16,000 healthy cows represented the training set. In the ongoing allocation schemes, the number of sick cows in the training set increased stepwise by moving 10% of the sick animals from the testing set to the training set, and vice versa. The size of the training and testing sets was kept constant. Evaluation criteria for both GBLUP and RF were the correlations between genomic breeding values and true breeding values (prediction accuracy), and the area under the receiving operating characteristic curve (AUROC). Prediction accuracy and AUROC increased for both methods and all scenarios as increasing percentages of sick cows were allocated to the training set. Highest prediction accuracies were observed for disease incidences in training sets that reflected the population disease incidence of 0.20. For this allocation scheme, the largest prediction accuracies of 0.53 for RF and of 0.51 for GBLUP, and the largest AUROC of 0.66 for RF and of 0.64 for GBLUP, were achieved using 50,025 SNP, a heritability of 0.30, and 725 QTL. Heritability decreases from 0.30 to 0.10 and QTL reduction from 725 to 290 were associated with decreasing prediction accuracy and decreasing AUROC for all scenarios. This decrease was more pronounced for RF. Also, the increase of LD had stronger effect on RF results than on GBLUP results. The highest prediction accuracy from the low LD scenario was 0.30 from RF and 0.36 from GBLUP, and increased to 0.39 for both methods in the high LD population. Random forest successfully identified important SNP in close map distance to QTL explaining a high proportion of the phenotypic trait variations.