Heritability figures must be read with considerable caution: among other things, they reflect only the amount of variation within the studied population that can be explained by genes, roughly speaking.
Data from the Mackay lab for a dozen Drosophila melanogaster traits [16] [18] suggests that a relationship between network characteristics and heritability exists (Figure 1).
Fruit cracking also had a very low heritability (0.03, Figure 3), lower than the value of 0.22 observed by Durel et al., an intermediate estimated value for π in the BayesCπ model (0.1, Figure 2), and a short range of relative SNP effects, extending to less than 0.0034, all contributing to very low accuracies.
The finding that all ECAM phenotypes showed a continuous distribution and high broad-sense heritability (74 95%; Figure S4 and Figure S5) in both crosses indicates that several genes are likely to contribute to the melanization phenotype in Z. tritici.
However, this seems unlikely to be responsible for the strong heritability evident in figure 2. As an additional check, we also compared the intron expression statistic between parents of the trios and found that these were significantly correlated (r = 0.44; p = 0.0007), suggesting that a batch effect may be involved in the apparent heritability of the intron expression summary statistic.
Despite these efforts, it seems unlikely that common variants with very small effects will completely solve the missing heritability problem. Figure 1 frames the issue by plotting allele frequency (from small to large) against effect size (from rare to common).
DGV accuracies estimated using breed-specific trait heritabilities are in Figure 5A for WBSF, Figure 5B for REA and in Additional file 2: Figures S5, S6, S7, and S8 for all other traits [See Additional file 2].
In contrast, nonparametric methods, particularly the NWE, the RKHS, and SVM, provided predictions that are reasonably accurate, especially for traits with higher heritabilities (shown in Figure 6 and Figure 8).
