Fig. 2
Impact of kinship derived from different types of markers on prediction accuracy and model fit. Prediction accuracy and model fit were evaluated on simulated mouse traits controlled by 10 QTNs at three levels of heritability. The heritability (h2) was set to 0.25, 0.5, and 0.75 (a–c, respectively). QTNs were randomly sampled from a total of 12,227 available markers on 1940 mice individuals from the WTCHG dataset. Four types of markers were used to derive kinship: (1) all markers (method of gBLUP), (2) 10 true QTNs plus 1000 randomly selected Non-QTN Markers (NQM), (3) estimated QTNs from SUPER (method of sBLUP), and (4) 10 true QTNs. We compared the impact of the different kinship derivations on prediction accuracy as the Pearson correlation coefficient between predicted and observed phenotypes, and on model fit as twice the negative log likelihood (−2LL). The accuracy (blue bars) and −2LL (red bars) are displayed as the means of 40 replicates. The standard errors are indicated by the whiskers on the bars. Each replicate used a five-fold cross-validation. Inferences were used to evaluate prediction accuracy and references were used to evaluate −2LL. The phenotypes of inferences were not used for estimating QTNs in sBLUP to ensure independent tests on inferences. As expected, the set of estimated QTNs (i.e., sBLUP) performed a bit worse than the ideal scenario (true QTNs). However, sBLUP performed better than QTN + NQM and much better than all markers (gBLUP)
