Experimental set-up. (a) Cells were grown at steady state in continuous chemostat cultures, with the specific growth rate controlled by the flow rate and the volume of medium in the reactor. Cells were harvested and used for transcription analysis and subsequent clustering of the transcription data. A simulated dataset was generated to illustrate the principles of consensus clustering. The dataset contained 80 members derived from four clusters (*, x, + and · in blue) in two experiments. The consensus clustering method consisted of three steps (panels b-d). (b) An ensemble of clusterings was obtained by multiple runs of mixture of Gaussians . Each run gave very different results (red ellipses), depending upon the initialization. (c) The results from multiple runs was used to form the transcript co-occurrence matrix (C), which was calculated as the empirical probability (over all runs) of observing each pair of transcripts (n,n') in the same cluster. (d) Based on the co-occurrence of transcripts a consensus clustering was generated. The co-occurrence matrix was also converted into a transcript-transcript distance matrix as D
= 1 - C
, which was used as input to a hierarchical clustering. The resulting consensus dendrogram showed the relationship between the clusters and was thereby a valuable tool in the biologic validation of the data.