Fig. 3

Intra-predictor concordance of SSPs and SCMs. Comparisons between predictors of the same type, e.g. the SSP of Hu vs. the SSP of Parker. The five panels show box and whisker (BW) plots for kappa statistics calculated over all subtypes and for each subtype separately, as indicated on the left hand side. Results for individual datasets are superimposed as dots. Each panel contains ten BW plots. From top to bottom these respectively indicate concordance for pairs of: (i) classic SSPs initially reported by Weigelt et al. [5], i.e. based on uncentered data (‘weigelt uncentered’), (ii) classic SSPs by Weigelt et al. [11], based on centered data (‘weigelt centered’). Estimates based on subtype assignments from the literature are superimposed as gray symbols with letters (see running text), (iii) classic SSPs without a normal-like subtype (SSP.classic.4s), (iv) classic SCMs (SCM.classic), (v) SSP.cs predictors, different CS and IGL, (vi) SSP.cs, same CS and different IGL, (vii) SSP.cs, same IGL and different CS, (viii) SCM.cs, different CS and MGL, (ix) SCM.cs, same CS and different MGL, and (x) SCM.cs, same MGL and different CS. Results for BWs (iii)-(x) are based on the hgu133plus2 compendium consisting of 11 datasets (2,019 samples after QC, Table 1). Vertical gray lines indicate kappa estimates that were pooled over all three groups of comparisons per predictor type. Top legend: type of concordance assessment indicated by the color of BW median values (indicated by a bar) (GL: gene list, IGL or MGL). Bottom legend: predictor type indicated by the color of a BW box. Numerical details of the BW plots and, highly similar, results for the analyses on all 3,908 arrays (including the hgu133a samples) are presented in Additional file 2: Tables S1 and S2