Fig 1.
Distribution of chill coma recovery time among 176 DGRP lines.
The histograms depict the distribution of line means for (A) females and (B) males for chill coma recovery time.
Fig 2.
Prediction accuracy using common or all variants with raw line means or Box-Cox transformed line means.
Prediction accuracy of GBLUP for 100 replicates of 5-fold cross-validation (CV) are plotted as box plots, for females and males separately. We performed the analysis using either raw line means (A) or line means transformed by Box-Cox transformation (B).
Fig 3.
Genetic architecture affects estimated additive genetic variance.
Different genetic architectures were simulated using DGRP genotypes for an additive model (A), a major gene model (B) or an epistatic model (C). Broad sense heritability was assumed to be 37%, the same as the observed chill coma recovery time. A total of 100 simulations were performed and the additive genetic variance (expressed as the proportion of additive genetic variance of total variance) was estimated and summarized in histograms.
Fig 4.
We performed LOOCV in females (A) and males (B) separately. In each of the 176 folds, the top GWAS associations and/or epistatic interactions in the training set were used to build the genomic relationship matrix and make prediction of the validation line. Accuracy of prediction (left y-axis, correlation between predicted and observed phenotypes) is plotted against the P-value threshold for the additive model and additive + epistatic model. For the additive + epistatic model, the epistatic pairs (on average the top 30 pairs in females and the top 3,232 pairs in males) that achieved the highest prediction accuracy in an epistatic only model was used, while the threshold for GWAS association was varied. The black line indicates the number of variants (right y-axis) significant in GWAS at varying thresholds.
Fig 5.
Accounting for genetic architecture improves genomic prediction.
Scatter plots showing the predicted phenotypes and observed phenotypes for females (A-D) and males (E-H) under different GBLUP models. Each panel represents a model indicated by the text above the plot with the prediction accuracy in the parenthesis. Each point in the scatter plots represents one fold of LOOCV under the indicated model.