To assess the applicability of our

To assess the applicability of our method, we selected three binary-fate stem cell differentiation systems for which high-quality single-cell gene expression data are available. These examples include the differentiation of inner cell mass (ICM) into either primitive endoderm (PE) or epiblast (EPI) (Guo et al., 2010), the differentiation of different progenitor ghrelin receptor in the hematopoietic system (hematopoietic stem cell (HSC) into either multipotent progenitor (MPP) or megakaryocyte–erythroid progenitor (MEP), MPP into common myeloid progenitor (CMP) or common lymphoid progenitor (CLP), and CMP into either MEP or granulocyte–macrophage progenitor (GMP)) (Guo et al., 2013), and the differentiation of lung alveolar bipotential progenitor (BP) into either alveolar type 1 (AT1) or alveolar type 2 (AT2) (Treutlein et al., 2014). In the first example Gata6 for PE and Klf2 for EPI were predicted, which is in full agreement with previously reported experimental observations (Fujikura et al., 2002; Yeo et al., 2014; Gillich et al., 2012). In addition, many well-known lineage specifiers in the hematopoietic system, such as Cebpa (Radomska et al., 1998), Gata1 (Pevny et al., 1991), Gfi1 (Li et al., 2010) and Spi1 (PU.1) (Voso et al., 1994) were correctly predicted for appropriate subpopulations, demonstrating the validity of our approach. Finally, our predictions in the relatively under-studied lung BP developmental system provided novel candidate lineage specifiers with prior associations with lung development, including Hes1 (Ito et al., 2000) and Pou6f1 (Sandbo et al., 2009). To our knowledge, this is the first computational method that systematically predicts cell lineage specifiers based on cell subpopulation-specific TRNs. Our method does not require pre-selection of candidate genes, and can be applied to any binary-fate differentiation event for which single-cell gene expression data are available. Finally, this method is compatible with both single-cell RT-PCR and single-cell RNA-seq data. Given the increasing importance of single-cell gene expression data in stem cell biology, we believe that approaches like ours would be useful for the identification of lineage specifiers. This should aid in understanding stem cell lineage specification and the development of strategies for regenerative medicine (Li and Kirschner, 2014).
Materials and methods
Results Based on the proposed model above, we implemented a computational method for predicting opposing lineage specifier pairs of stem cell differentiation. The schematic view of the method is shown in (Fig. 1). Briefly, we first generated a raw TRN for each parental cell subpopulation by combining literature-based interactions, predicted TF-DNA binding interactions, and single-cell gene co-expression-based interactions (Fig. S1). We then performed network contextualization of raw TRNs using an improved version of the method developed in our group (Crespo et al., 2013), which removes interactions that are inconsistent with Booleanized gene expression profiles. Once contextualized TRNs were reconstructed, candidate opposing lineage specifier pairs were predicted based on their expression values, and their presence in the SCC of the parental cell subpopulation. We propose that the change in expression ratio between parental and daughter cell subpopulations is biologically more relevant than the expression ratio itself within each cell subpopulation, since the basal/effective level of expression differs among TFs due to several factors, including the binding strength and number of target genes (see Materials and methods section for complete explanation).
Discussion Understanding lineage specification has been partly hampered by the co-existence of different cell subpopulations within a heterogeneous stem cell population. In the current study we have proposed a model of binary-fate stem cell differentiation, in which each parental stem cell subpopulation is in a stable state maintained by its specific TRN stability core. Furthermore, this stability core is maintained by the balanced expression pattern of opposing lineage specifiers for different daughter cell subpopulations. Dysregulation of this balanced expression pattern induces differentiation. Based on this model, we have developed a computational method for predicting opposing lineage specifier pairs for a binary-fate differentiation event. Single-cell gene expression data enabled us to reconstruct TRNs and to identify their stability cores specific for different parental cell subpopulations. Indeed, subpopulation-specific TRNs exhibited significant network rewiring, as was previously reported in Moignard et al. (2013). Using these subpopulation-specific TRNs, our method was, albeit a few false negatives such as Gata4 and Nanog for PE and EPI lineage specifications, able to predict many known lineage specifiers in the two well-studied examples (Guo et al., 2010, 2013). This method was further applied to a less-studied example, the lung BP differentiation system (Treutlein et al., 2014), and predicted novel candidate lineage specifiers, several of which have been previously shown to have some association with lung development, and could be experimentally validated in future.