To identify finer substructure amongst these classes, classes with more than 200 cells were selected for subclustering. The largest 50% of the cells from each of these clusters was subjected to a variable gene selection, scaling, and independent component analysis. The independent component space is highly dependent on the number of components K that are selected for computation. To automatically nominate a value for K, we took advantage of the fact that the fastICA algorithm begins with a whitening step, in which a singular value decomposition is used to select the top K eigenvectors (i.e. principal components) for maximization of non-Gaussianity(Hyvärinen, 1999). We therefore calculated the number of statistically meaningful principal components using the Jackstraw method (Chung and Storey, 2015), to obtain a suitable value for K. In almost all instances of subclustering, this value was used, with a few exceptions where K was increased slightly. These values of K were then used to compute ICs for each subclustered class.
