Functional validation

We apply the following two methods to validate whether the predicted regulons are functional related to their associated cell types: (1) SEEK analysis ( Zhu et al., 2015 ), and (2) CoCiter analysis ( Qiao et al., 2013 ). First, SEEK ( http://seek.princeton.edu/modSeek/mouse/ ) is a tool that provides the gene co-expression search function for over ~2000 mouse datasets from the Gene Expression Omnibus (GEO). We used the mouse version of SEEK to evaluate whether the genes in a regulon are co-expressed, and if so whether the datasets supporting the co-expression are associated with an interested cell type. If genes are significantly co-expressed in many datasets related to a certain cell type, it could be inferred that the function of this regulon is highly related to this cell type. Taking the erythroblast for example, we input gene list of regulon Lmo2 to SEEK web server and search for erythroblast related keywords (such as ‘erythroblast’, ‘hematopoietic’, etc.) from the complete dataset-list ranked by query-coexpression score. Then we choose a p < 0.01 cutoff to select significant datasets and finally use Fisher’s exact test to evaluate whether the selected datasets are significantly enriched in the top ranks. Second, the CoCiter ( Qiao et al., 2013 ) is a text mining approach against the up-to-date Medical Literature Analysis and Retrieval System Online (MEDLINE) literature database to evaluate the co-citation impact (CI, log-transformed paper count) between a gene list and a term. To assess significance of co-citation, a Monte Carlo approach is used to evaluate random expectations by randomly selecting 1000 gene sets with the same size as input gene list and then a permutation p value is calculated as the number of times that CI _random > CI _true divided by 1000. Here we used the function “gene-term” in CoCiter (use default parameters but set organism as mouse, http://www.picb.ac.cn/hanlab/cociter ) to check whether the genes in a regulon are significantly co-cited with a certain cell type in literatures.

Regulon module analysis

Regulon modules were identified based on the Connection Specificity Index (CSI) ( Fuxman Bass et al., 2013 ), which is a context-dependent measure for identifying specific associating partners. The evaluation of CSI involves two steps. First, the Pearson correlation coefficient (PCC) of activity scores is evaluated for each pair of regulons. Next, for a fixed pair of regulons, A and B, the corresponding CSI is defined as the fraction of regulons whose PCC with A and B is lower than the PCC between A and B.

Hierarchical clustering with Euclidean distance was performed based on CSI matrix to identify different regulon modules. We also used CSI > 0.7 as a cutoff to build the regulon association network to investigate the relationship of different regulons. The result was visualized by Cytoscape ( Shannon et al., 2003 ). We used the same strategy to identify submodules within M7. For each regulon module, its activity score associated with a cell type is defined as the average of the activity scores of its regulon members in all cells within this cell type. Then the top ranked cell types are identified for each module.

Quantifying cell type relationship

Using the gene regulatory network analysis as a guide, we quantified the relationship between different cell-types based on the similarity of the overall regulon activities, which is quantified by the Spearman correlation coefficient. The results were represented as a network, where a pair of cell types were connected if the Spearman correlation coefficient is greater than 0.8. Again, the result was visualized by using Cytoscape. Groups of related cell-types were identified by using the Markov Clustering Algorithm (MCL) ( van Dongen and Abreu-Goodger, 2012 ), as implemented in the ClusterMaker application in Cytoscape. We used the default setting except setting the inflation parameter as 2.

QUANTIFICATION AND STATISTICAL ANALYSIS

Details of the statistical tests used in this study are described briefly in the main text and more in-depth in the subsections above. They are also summarized below:

• (1)
  To evaluate the consistency of identified regulons in three Avg20 replicates in each of bladder, kidney and bone marrow tissue, we counted the number of overlapped regulon TFs between different replicates and applied the one-sided Fisher’s exact test to evaluate statistical significance.
• (2)
  t test was used evaluate whether the performance of Avg20 approach is better (p < 0.05) than that of using all single cells in each of bladder, kidney and bone marrow tissue based on silhouette value.
• (3)
  When applying SEEK to test whether the genes in regulons are co-expressed in certain cell type, we chose correlation significant datasets (p < 0.01) and then used one-sided Fisher’s exact test to evaluate whether datasets related to interested cell type are significantly enriched (p < 0.01) in the top ranks.
• (4)
  In CoCiter analysis, a permutation p value was introduced. It randomly selected 1000 gene sets with the same size of tested regulon and the p value was calculated as the number of times that co-citation impact of “random” larger than “true” divided by 1000.