Figure 3: Phenotypic clustering.

Using spectral clustering with Mahalanobis distance (Supplementary Figs 5 and 6), candidate proteins were grouped into nine separate clusters. (a) Distribution of clusters in three-dimensional space according to Z-scores for the parameters E-cad, Jun-A or Cyt-A. The number of proteins classified in each cluster is shown in parenthesis. (b) Silhouette plot shows a silhouette value (−1 to +1) for each point in the data set measuring how similar it is to points within its own cluster; positive values mean correct classification in its cluster. (c) Adequacy of experimental clusters based on their protein interaction landscape. Around 10,000 random clusters were plotted according to their Commuter time Kernel maximal score (CK-max) to produce the background curve of random clusters (green line). Red arrow points to where experimental clusters are found. (d) Heat map of the optimized clusters based on their corresponding Z-scores for E-cad, Jun-A and Cyt-A. Name of protein depleted is shown on the right of each panel. Data in a–d represent median Z-score values of three independent replicates for each RNAi oligo. (e) Functional enrichment. Following manual curation of the literature, each of the candidate proteins was attributed one or more functions on actin remodelling (functional group; x axis). Enrichment of the 13 functions in each cluster is shown in two ways: the distribution of actin functions relative to all functions found in the candidate protein data set (grey bars, that is, all proteins that cap, sever and so on.) or in the subset of functions found among proteins in the cluster (yellow bars). N=156.