Blast searches against the shotgun cDNA assemblies from SC and NBH embryos using the significant array probes resulted in further extension of some of the probe sequences.
We performed a hierarchical cluster analyses of the microarrays using the significant (FDR < 0.05) probes for either breed or sex with Cluster 3.0 [ 17].
Including these probes also increases the sensitivity of the analysis, however: the significant probe in cax7 had a FDR = 0.011, and would have also been excluded had we only used a threshold FDR of 0.01.
Next, we constructed a centroid-based gene expression signature using significant probes from the SAM analysis for validation in external gene expression studies (in Additional file 1: Table S1).
However, these genes were included in the subsequent ontological analysis as their modulation is highly significant in dataset 1 and 3. To directly compare the expression data from datasets 1 and 3, a scatter plot of the log2 fold-change was generated using the 58 probes with a significant fold-change in both comparisons (Fig. 2).
In order to have the same power for all tests, the GO test of serpentine vs. granitic soils was re-computed using only the 751 significant probes, and compared to the probes with the lowest 751 t-test p-vales from the other two combinations of the data.
To this end, we ranked all probe sets according to their absolute significance score and used the most significant 200 probe sets of each of these pairwise comparisons to generate the Venn diagram.
Using the background corrected probe intensities allows one to eliminate a significant portion of sequence-dependent probe behaviors, including the increased mean and variance commonly observed in GC-rich probes.
Using hierarchical clustering, the most significant probes for DNA methylation discriminated prednisolone response while the gene expression probes did not, reflecting the more variable nature of gene expression compared to DNA methylation [ 21, 22].
For CNA data, we used the 10%% probes having the most significant p-values using the univariate log-rank test from a Cox proportional hazard model splitting the linear predictor at the median.
