Table 1

Comparison of the regression analysis in linear and log space.

Regression slope after log-transformation

Regression slope in linear space

not ideal

ideal (slope = 1)


not ideal

167

53

ideal (slope = 1)

14

6


Measured and theoretical concentrations from NBT Study I (with 8 sites, 10 peptides and 3 transitions per peptide) are natural log transformed (other common bases used for the log transform are 2 and 10). Weighted, robust regression lines are fitted to the linear data while robust regression is used for log-transformed data. The slope of each regression is assumed to be 1 (ideal) if the 95% confidence interval calculated as slope ± t(1-Δ/2),df * s.e. includes 1 (where, t(1-Δ/2),df is the 2-tailed t-distribution critical value for α = 0.05, df = (# replicates - 1), and s.e. is the standard error for the regression slope). As is evident from the table, log-transforming data results in more ideal slopes (closer to 1). Visual inspection of the regression lines and associated data indicate that this overestimation of the number of ideal peptides by log-transformed regression is unwarranted and misleading, as illustrated by examples in Figure 1d.

Mani et al. BMC Bioinformatics 2012 13(Suppl 16):S9   doi:10.1186/1471-2105-13-S16-S9

Open Data