Fig. 6

Compositional signatures discriminate between normal and cancer tissues. We used the 53 complex members that were identified as variable in both the 11 cell lines or reprogramming dataset to query an independent proteomic dataset obtained from human colon tissue samples [30]. a A nearest centroid method [58] was used to classify 14 cancer and seven normal tissues. Variable complex members have a better discriminative power than random protein features. Average accuracy was measured for 100 feature sets randomly sampled from variable complex members (n = 53) versus all quantified proteins (n = 6148). The size of the feature set was increased from 4 to 28 in steps of 4. The average accuracy for the variable complex members (red) were significantly higher in comparison to randomized features (black) (e.g. for n = 20, Wilcoxon rank sum test, p value = 9.9E-24). Error bars represent standard error. b For variable complex members and random proteins, two set of features (n = 20) were selected as representing the average accuracy. Cancer and normal proteomic profiles with the representative features were grouped by hierarchical clustering (average-linkage, using Euclidean distance) and presented as dendrograms for variable complex members and random protein features. Gray arrowheads indicate wrongly classified samples