As Leek et al. reported [ 35], large batch effects are associated with DNA preparation group and processing date.
As summarized Table 1 none of the PCs were associated with the batch effects anymore for the batch corrected data; and the CpGs that were associated with the batch effects were reduced to a few (12, 25, 23 in Dataset 2 and 2, 6, 8 in Dataset 3 for QNβ, lumi and ABnorm normalized data, respectively).
In spite of the significant reductions, the numbers of CpGs remaining associated with batch effects were still much higher than expected.
After QNβ, lumi or ABnorm, the number of CpGs associated with batch effects were reduced to 24, 32, and 26percentt for Dataset 2; and 37, 46, and 35percentt for Dataset 3, respectively.
Batch effects were removed by aligning the within-batch medians for all measurements.
Batch effects were then adjusted by the COMBAT method [ 30].
Experimental batch effects were adjusted for by including experimental batch as a covariate in the statistical model [ 56].
Unsupervised hierarchical clustering (Pearson correlation and Ward linkage) and principal components analysis were used to search for batch effects, but no significant batch effects were observed.
Batch effects are the technical artifacts that are not associated with the underlying biology, but rather to unrelated factors, such as laboratory conditions, experiment time, reagent lots and/or laboratory personnel differences.
In analyzing the 225 muscle files, we discovered a batch effect which was associated with distinctly different expression values by clusters of sample processing dates for some genes (Additional file 1).
