a, Pipeline for the definition of interest point clusters using a subset of the data. Images (936, corresponding to 5% of the entire dataset) were randomly selected from the dataset to construct a pool of interest points. Each interest point was numerically described with a 40 dimensional feature vector encoding the intensity distribution, localization and contrasts to the interest point neighbourhood. Combining k-d tree-like and thresholding-based clustering with density-based clustering, the interest points were grouped into 100 clusters. b, The remaining interest points of the dataset were then assigned to the identified clusters. Thus each image was represented as the distribution of intensity in each of the 100 interest point clusters. c, Non-negative factorization of the data tensor of proteins × features × mitotic stages (left panel) produced a non-negative tensor of reduced dimension (middle) for which entries can be interpreted as the fraction of protein belonging to each cluster over time (right, each cluster is represented by a different colour and the height of a coloured bar at a given mitotic stage represents the fraction of the protein in the corresponding cluster at this stage).