Validation of microarray experiments by qrPCR. Concordance between microarray and qrPCR FCs (averaged across within-experiment replicates) for top-ranked (n = 15) and random-stratified sampling methods (n = 15) for each microarray experiment is shown in the six scatter plots on the right (B, D, F, H, J, L); corresponding MvA plots are shown on the left (A, C, E, G, I, K). The log2 fold-change (FC) values for qrPCR were determined from the calibration curve performed for each selected gene. Deleting the red data points in panels B, J, and L, which had large influence on the regression slope (standardized slope dfBeta values > 1) had little effect on the CCC values (data not shown). Results of a robust regression approach for detecting curvature in H, J, and L yielded inconsistent results; p > 0.10 (H and J) and p < 0.10 (L) (see Methods); least squares regression tests for quadratic effects were likewise inconsistent; p values for the quadratic term for H, J, and L were 0.28, 0.08, and 0.004, respectively. The regression lines are: the identity line (red), least-squares (blue) and loess (yellow). Random-stratified sampling consistently yields better results than top-ranked sampling.
Miron et al. BMC Bioinformatics 2006 7:333 doi:10.1186/1471-2105-7-333