Actually, even if tested by a benchmark dataset within the scope that can be covered by TargetP [6] or Predotar [8], the success rate by the current Plant-mPLoc predictor is also much higher than those by the two predictors, as demonstrated below.
Meuwissen and Goddard [ 3] reported that a nonlinear predictor such as BayesB gave much higher accuracies than a linear predictor (RR-BLUP), whereas Druet et al. [ 5] did not use RR-BLUP.
FLP and FTP predictors that were based on combination of these confounders and that were found by our approach had much higher error rates than the predictors that were based only on gene expression (FLP error rate > 0.63, FTP error rate > 0.74; for all the predictors that were based on the above mentioned confounders).
In the process of model selection, it became evident that the geographic location of observations affected larva density to a much higher extent than the causal predictors CTH and DWT (Table 2).
Comparing with models with bioclimatic/topographic variables, the models with topographic and NDVI-derived predictors had much higher niche similarity with the best model (Schoener's D was 0.95, 0.95 and 0.94 for C. longicollis, E. macquarii and M. expansa, respectively).
Previous studies demonstrated that the elastic net performs better than least absolute shrinkage and selection operator (LASSO), especially when the number of predictors is much higher than the number of observations [ 18].
Besides, we found the specificities of the combined predictor are much higher than what were estimated from the previous independent testing (Table 1), indicating that the false-positive rate of the combined predictor might be over-estimated, and the previously selected threshold values might be too stringent for the combined predictor.
The reduced model that used only network size and rate of environmental change as predictors had a much higher AIC score (ΔAIC ∼ 300), but still explained 78% of the variance in average heritability (P<2.2e−16; Table 2).
The results of Table 3 alternatively show that MFP based predictors have a much higher success rate than semi-empirical force fields, and that taking into account the solvation free energy does not improve the reliability of the later.
Because the number of predictors (markers) is generally much higher than the sample size (P > > n), fixed regression methods using ordinary least squares cannot be used for developing prediction models.
