Fig. 2

Consensus set-based predictor construction and evaluation. In panels (a) and (b), row (i) represents subtype predictor models and row (ii) corresponds to sets of predictions made by these models. Each column corresponds to a different predictor/subtype assignment set pair. a Consensus set construction (see Methods). For a given training set T, five initial sets of subtype assignments are obtained. First, the original PAM50 predictor is applied to T, resulting in subtype assignment set ST PAM. Next, three SCMs models are estimated on T, based on the MGLs D, W and HK (Additional file 1: Section 2). Here SCM X,T denotes an SCM estimated on T, based on MGL X. The resulting SCMs are subsequently applied to T resulting in three additional sets of subtype assignments, i.e. ST D, ST W, and ST HK. A final set of subtype assignments ST STG is obtained by the application of the STG predictor on the over/underexpression profile of ER, HER2, PGR and proliferation phenotypes, estimated on T. From the five subtype assignment sets in row (ii) a consensus set (CS) is derived consisting of those samples in T for which all five subtype assignments are concordant. b Construction and evaluation of the consensus set-based subtype predictors SSP.cs (left), SCM.cs (middle) and STG.cs (right), see Methods. For a given CS, three SSPs are constructed that differ only by their associated IGL S, H or P (Additional file 1: Section 2). Here SSP X,CS represents an SSP with associated IGL X, of which the centroids are estimated on CS. The SSPs are subsequently applied to a validation set V, leading to subtype assignment sets SV S, SV H and SV P, respectively. On the same CS also three SCMs are constructed, based on the MGLs D, W and HK. The resulting SCMs are subsequently applied to validation set V, yielding subtype assignment sets SV D, SV W and SV HK. Similar to SCMs, also three STG.cs predictors are constructed based on MGLs D, W and HK and applied to validation set V