Fig. 4
Interaction between prediction methods and genetic architecture. Prediction accuracies were evaluated on mouse phenotypes simulated with different numbers of Quantitative Trait Nucleotides (QTNs) at different heritabilities. The QTNs were sampled from 12,227 markers genotyped on 1940 mice individuals from the WTCHG dataset. The heritabilities (h2) were set to 0.75, 0.5, 0.25, and 0.1 (a–d, respectively). Five-fold cross-validation was conducted to evaluate prediction accuracies by comparing four methods: sBLUP, Bayesian LASSO, cBLUP, and gBLUP. The cross-validations were replicated 40 times when gBLUP, cBLUP, and sBLUP were used and 20 times when Bayesian LASSO was used. The first 80,000 iterations were used as “burn in” and the next 60,000 iterations were used to derive Bayesian estimations. The advantage of the new developed method are under condition of less QTNs (sBLUP) and low heritability (cBLUP)
