Genomic and molecular control of cell type and cell type conversions.

Organisms are made of a limited number of cell types that combine to form higher order tissues and organs. Cell types have traditionally been defined by their morphologies or biological activity, yet the underlying molecular controls of cell type remain unclear. The onset of single cell technologies, and more recently genomics (particularly single cell genomics), has substantially increased the understanding of the concept of cell type, but has also increased the complexity of this understanding. These new technologies have added a new genome wide molecular dimension to the description of cell type, with genome-wide expression and epigenetic data acting as a cell type 'fingerprint' to describe the cell state. Using these genomic fingerprints cell types are being increasingly defined based on specific genomic and molecular criteria, without necessarily a distinct biological function. In this review, we will discuss the molecular definitions of cell types and cell type control, and particularly how endogenous and exogenous transcription factors can control cell types and cell type conversions.

Defining cell type

The cells of an organism are made up of a limited number of ‘cell types’ that are reused in different tissues and combine to form organs and systems. For example macrophages, phagocytic immune cells, are found throughout the body, as are connective fibroblast cells. However, defining a cell type, especially now that single cell technologies are revealing ever more heterogeneity between cells, is challenging, and it remains unclear how many cell types there are, or exactly how fine the differences are that demarcate two cell types. There have been several estimates for the total number of cell types in an organism, with numbers ranging from between hundreds to thousands of distinct human cell or sub-cell types. Classical taxonomic approaches estimated the number of cell types in a selection of chordata as between 99 and 122, and around 200 cell types in humans. Systematic attempts to count cell types, using a variety of techniques, particularly newer gene expression data, generally come to a much higher number of cell types. CELLPEDIA is a human annotated database of cell type, based mainly on taxonomy, gene expression data and text mining of publications, it suggests 2260 taxonomic categories for cell types. CELLPEDIA also uses tissue location to define cell type, which may inflate the total. However, the same ‘cell types’, isolated from different tissue locations, can show radically different gene expression patterns,1, 2 hence tissue location can also be an important determinant of cell type. CellFinder takes a different approach, using a mixture of database amalgamation, text mining, and human annotation, it comes to a total of 1058 human cell types, and readily concedes there are many more cell types to discover. Cell Ontology (CL) describes 2200 ‘classes’ of cell or sub-cell type, and, like the related CellFinder and LifeMap databases uses cell type definitions to map the cell types into a hierarchical model of development.8, 9 These newer studies put the total number of cell types considerably higher than previous estimates, and the true number of cell types seems to be increasing as researchers develop new tools to more accurately map gene expression and the epigenetic status of cells.

Identification of different cell types in the immune system

An illustrative example of how improvements in technology can drive the discovery of cell types is the proliferation of new immune cell types along the T cell lineage (Fig. 1). Initially defined by morphology alone, T cells were indistinguishable from B cells, and were labelled simply as ‘lymphocytes’, i.e. cells that occupy lymphoid tissue, but they had no known function. Later, they were discriminated based on their tissue of origin; bursa of Fabricius-derived lymphocytes (bone marrow-derived in mammals) became B cells, and thymus-derived lymphocytes became T cells.11, 12 However, B and T cells only became recognized as distinct cell types in the 1960s as B cells were definitively identified as the source of the humoral (i.e. antibody) immune response, whilst T cells were initially recognized as ‘B cell helpers’ a few years later. The widespread adoption of monoclonal antibody technology led to a burst of activity in defining further T cell types. The cluster of differentiation (CD) antibodies, are a set of defined monoclonal antibodies against a variety of cell surface targets. Two CD antibodies can separate T cells into two distinct cell types: CD4+ T helper cells and CD8+ T cytotoxic cells. T helper cells play a supporting role in immune responses, whilst T cytotoxic cells perform cytotoxic killing of virus-infected cells, importantly, their cell morphology is basically identical and they can only be discriminated by their biological activity and cell surface markers. Further application of monoclonal antibodies and careful flow cytometry experiments divided T helper cells into a wide range of other T helper cell types. For example, naïve T helper cells, that have not encountered their antigen are defined by the absence of CD25, whilst experienced (those that have encountered their antigen) T helper (Th) cells differentiate into four major types, namely, Th1, Th2, Th17, and Tregs (regulatory T cells), along with many more less well characterized T helper cell types.15, 17 Importantly, these cell types are not just finer definitions of subpopulations, but each T helper cell type has a distinct biological function. The four best characterized T helper cell types are Th1, Th2, Th17 and Treg cells, which are important in responding to intracellular pathogens, helminth infection, extracellular pathogens, and maintaining self-tolerance, respectively. However, many more T helper cell types have been discovered (e.g. Th9, Th3, TR1, Th22, Tfh, Thab, nTreg, etc.),15, 19 these new T helper cell types have less clear biological roles, but take part in a range of specific activities, including airway inflammation, allergic reactions, B cell responses and immune-related diseases, amongst other roles.

Fig. 1

Gradual refinement of the definition of CD4+ T helper cell types. Schematic of the refinement of the CD4+ T helper cells, from the original cell type ‘leukocyte’, through to a plethora of distinct Th (T helper) cell types. The technology/technique used to separate the cell types is indicated in grey at the branch point.

T and B cells were originally defined based on the organ they were first purified from, and the tissue of origin can have a strong influence on cell type. For example, gene expression microarrays of macrophages purified from different tissues showed greater overall variation in gene expression patterns, when compared to other immune cells,1, 21 or compared to just other lymphoid cells. Dendritic cells (DCs), antigen-presenting cells of the immune system, highlight the opposite problem of separating cell types. DCs and macrophages are challenging to experimentally separate accurately, as they share many of the same cell surface markers. Consequently, there is argument about the difference between macrophages and DCs, and a model has been put forward that suggests DCs and macrophages are a ‘spectrum’ cell type, with phagocytic cells (macrophages) on one end and antigen-presenting cells (DCs) on the other, with several cell types sitting in the middle of the spectrum, each possessing more or less macrophage or DC character.24, 25 Molecular characterization suggests that, from the perspective of gene expression at least, macrophages and DCs can be distinguished based on a unique gene expression signature,21, 26 and macrophages and DCs respond differently to inflammatory stimuli. Yet, arguments over the differences between these cell types remains.24, 27, 28, 29

Heterogeneity in embryonic stem cells; defining cell type by biological function

One of the better studied cell types are mouse embryonic stem cells (mESCs), which are derived from early embryos, and maintain the ability to regenerate a full mouse. Although mESCs have many similarities with the inner cell mass (ICM) of the early blastocyst, particularly in the activity of key transcription factors such as OCT4, SOX2, KLF4 and NANOG, there remains debate about their exact origin and cell type, as when the ICM converts to mESCs the cells undergo many gene expression changes.26, 33 mESCs as a cell culture were thought to be relatively homogenous, yet careful study of mESCs revealed small numbers of cells in a typical cell culture with altered gene expression profiles.34, 35 In mESCs, the expression level of the essential pluripotency gene Nanog36, 37 naturally fluctuates, and about 5–20% of mESCs express very low levels.37, 38, 39 In culture, mESCs cycle Nanog on and off, which helps prime mESCs to differentiate, and so these cells have a distinct phenotype and arguably cell type. Nanog is by no means the only example of heterogeneity in mESCs. STELLA, a marker of primordial germ cells, is expressed in 20–30% of mESCs, and those cells with STELLA more closely resemble the ICM, whilst those without STELLA express developmentally later epiblast-specific genes. Indeed, there are multiple cell types contained within a typical mESC culture, including small numbers of cells with radically different biological function. Normally, mESCs very rarely contribute to extraembryonic tissues, such as the trophectoderm (placenta) or primitive endoderm.30, 41 However, mESC cultures contain about 15% of cells that are Hhex+ (a homeobox protein that specifically marks endoderm), and these cells can contribute to extraembryonic tissues in mouse chimeras. Although the Hhex+ and Hhex-mESC's gene expression signature is nearly identical, they have different biological potential, and so can be considered a distinct cell type. One caveat is that these Hhex+ cells still contribute to the epiblast and embryo proper, so it is not a pure population of cells. A rarer subset of cells within mESC cultures express the endogenous retrovirus MERVL. MERVL is specifically expressed at the 2 cell stage of embryonic development,42, 43 and using a MERVL-Tomato reporter, the ∼2% of mESCs that express MERVL can contribute to extraembryonic tissues, although again, the MERVL+ cells can also contribute to the embryo proper, and the cells can interconvert between MERVL+ and MERVL- cells, suggesting instability in their cell type. It was initially thought that these MERVL expressing cells closely resemble the 2 cell (2C) stage of the embryo, where MERVLs are also specifically expressed, however, recent single cell RNA-seq data suggests these 2C-like cells may more closely resemble the blastocyst, so their ultimate identity remains unclear. Ultimately, the relationship between all of these heterogeneous cell types or sub-cell types within mESC cultures remains unclear. For example, despite their capability of both 2C-like and Hhex+ cells to contribute to extraembryonic tissues, it is unclear how they are related to each other, along with other potential cell types revealed by single cell genomics.

Ultimately mESCs are an in vitro artifact, a ‘trapped’ version of the blastocyst ICM that can grow indefinitely, but still maintain pluripotency. It is possible to capture many additional embryonic cell types, of which some appear to represent earlier timepoints in the developmental process. One such cell type are ‘Extended pluripotent stem cells’ (EPCs), that can contribute to extraembryonic tissues, and have distinct gene expression compared to mESCs. Other embryonic cell types appear to be developmentally later than mESCs, such as Epiblast stem cells (EpiSCs), that more closely resemble the developing epiblast and have a primitive endoderm-like gene expression signature,45, 46 and lack Esrrb activity. The similar but distinct EpiLCs (epiblast-like cells), lack the primitive endoderm gene expression signature found in EpiSCs, and are instead biased towards a primordial germ cell fate.48, 49 Finally, region-selective EpiSCs (rsEpiSCs) are biased to colonize just the posterior part of the developing embryo, suggesting an even later developmental phenotype than EpiSCs. These and other embryonic cell types indicate that at specific stages, with the right conditions, transient cell types can be captured and maintained in vitro.

A continuum of cell states in the transitions between cell types

It is challenging to define at what point two cell types are distinct, and where two cells are simply at one end of a continuum. Single cell data suggests that cells can transit through stages where the cell type-signatures of both origin and destination cells are simultaneously present. For example, in developing lung, some cells express markers for both alveolar type 1 and 2 cells simultaneously, and early in the embryo some cells simultaneously express genes for the primitive endoderm and epiblast. Consequently, identifying cell types in developmental processes is challenging. Potentially there is a continuum of expression as cells pass through developmental stages, and at any one point along that process the cell is not stable and may collapse into a more stable and distinct cell type (Fig. 2). Single cell mass spectrometry of cell surface markers in developing human B cells revealed a continuous spectrum of B cell development stages, rather than specific barriers, something similar was seen for in vitro differentiation of cells to neurons, and in in vitro transdifferentiation of cells to myoblasts. This calls into question the existence of cell types during development and, instead of development proceeding in jumps across energy barriers to local energy minima (or distinct cell types), cells develop in a continuous manner with intermediate stages where cells can continue to choose their developmental outcome (Fig. 2). Crucially, as cells differentiate to alternate cell types they lose developmental potential, and consequently most, if not all, adult cells cannot transdifferentiate. There appear to be many epigenetic blocks that lock cells into a specific cell type and limit the cells capability to dedifferentiate and transdifferentiate. A major candidate for the control of cell type is transcriptional control, which may act to lock cells into a cell type.

Fig. 2

Cells traverse pathways from origin cell types to destination cell types. A hypothetical map of cell fate conversion between an origin cell type and a destination cell type. Each node in the network is a new cellular state, and each edge is a transition between a cell state. Only parts of the network can form stable cell types, and many branching pathways exist. As the cells differentiate they move through intermediate stages, each step with a slightly different gene regulatory network underlying the cell state. When the cell reaches its destination, it becomes locked into that cell type, and can no longer traverse the intermediate states. Figures were drawn using glbase.

Transcriptional control of cell type

Cell type is thought to be controlled through the activity of transcription factors (TFs), that respond to either internal or external cellular cues. TFs bind to DNA and regulate gene expression, and interact with local chromatin to control cell type. Although a comprehensive model describing exactly how TFs perform these feats remains frustratingly elusive.59, 60

TFs can be expressed in both a cell type-specific and cell type-independent manner. Many, about 60%, of TFs are cell type-specific. Cell type-specific TFs can function as ‘master regulators’, a class of TF that can specify cell type in the absence of any other activity. The prototypical example is MyoD (Myod1), which when overexpressed converts fibroblasts to myoblasts, activating an entire gene expression program in the absence of specific external cues (Fig. 3A).

Fig. 3

Mechanisms of transcription factor mediated cell type determination. (A) A single ‘master’ transcription factor can activate an entire gene expression program to alter cell type. In this example, the transcription factor MyoD can convert fibroblasts into myoblasts. (B) Different combinations of transcription factor pairs can specify alternate cell lineages. In this example one OCT (OCT4/BRN2) factor pairs with a SOX (SOX2/SOX17) factor to maintain/specify one of three cell fates: embryonic stem cells (pluripotency), a primitive endoderm (yolk sac), or neural progenitor cell fate. (C) The same transcription factor can have multiple biological roles in different cell types. The example shown here is the transcription factor STAT3, which is expressed in most tissues but has widely divergent biological functions in different cell types.

However, a single master regulator for each cell type seems to be a relatively rare phenomenon, and often the same TF can act in multiple cell types. For example, knocking down Gata3 in mouse embryos leads to a failure to establish mature blastocysts, likely due to a trophectoderm defect, as when Gata3 is overexpressed in mESCs it drives them to a trophectoderm cell fate. Yet, despite its importance in the early embryo, Gata3 is also a critical factor in the specification CD4+ Th2 cells. A further difficulty with the idea of master regulators is tremendous degeneracy in the DNA sequences that individual TFs use to bind to DNA. For example, the homeodomain TFs all bind to a similar version of the same sequence of DNA, despite the involvement of homeodomain proteins in a wide range of developmental processes. This is not restricted to just one family of TFs, as almost all TF families bind to very similar DNA motifs, leading to the vexing issue of finding cell type-specific activity between different TFs that bind to the very similar sequences of DNA. One solution is for pairs (or more) of TFs to combine together to specify a developmental process. For example, the combination of OCT4-SOX2 is critical for pluripotency, but OCT4-SOX17, binding to a slightly different DNA motif acts to specify primitive endoderm, whilst another OCT/POU-family containing complex, BRN2-SOX2, specifies neural progenitors (Fig. 3B). Complex cell type specific assembly of TF complexes is not limited to OCT/POU-SOX factor pairs, as GATA1, GATA2 and PU.1 can assemble on a variety of specific DNA motifs to direct erythroid and neutrophil cell fates. TF–TF pairing appears to be widespread; a systematic analysis of genome-wide TF binding discovered 603 potential constrained TF–TF pairs,72, 73, 74 suggesting a combinatorial code that adds complexity to regulate the diversity of cell types and biological processes.

TFs that have a cell type-independent pattern of expression might not seem a promising area to explore for cell type-specific control, but these TFs can also exert specificity in the correct setting. Around 30% of TFs are cell type-independent at both the RNA, and protein level. It might seem that these TFs are involved in basal cell activities, and indeed many are, but cryptically, many cell type-independent TFs can have highly cell type-specific function. For example, STAT3, despite being expressed almost uniformly in cells and tissues, specifies pluripotency in mESCs, an anti-inflammatory response in macrophages, a pro-inflammatory response in dendritic cells, and has a critical role in T helper17 cells type differentiation, amongst many other cell type-specific roles76, 77, 78 (Fig. 3C). Ultimately, many TFs have widely overlapping functions in multiple cell types, and as yet, no comprehensive model of TF cell type control exists.

Exogenous expression of transcription factors can drive conversion of cell type

TFs have been instrumental in the forced conversion of one cell type to another. The earliest use of a TF to drive transdifferentiation was the transfection of Myod1 (MyoD), to convert cells to myoblasts, and Cebpa and Cebpb to convert B cells into macrophages. However, the most dramatic demonstration of the power of TFs was the conversion of fibroblasts to mESCs using just four TFs: OCT4, SOX2, KLF4 and c-MYC. Since this breakthrough many other transdifferentiation protocols have been discovered, along with the use of small molecules to convert cell type, for example the conversion of fibroblasts to neurons or fibroblasts to mESCs. Intriguingly, many of the small molecules used in these protocols directly interfere with epigenetic control, such as DZNep (methylation inhibitor), VPA (histone deacetylase inhibitor), or Tranylcypromine (histone demethylase and monoamine oxidase inhibitor), indicating that epigenetic control is a major factor in the determination of cell type.

Transdifferentiation protocols mediated by TFs nonetheless remain relatively few, and there are many target cell types we would like to in vitro differentiate, but cannot. Consequently, there has been a lot of activity in designing systematic computational approaches to predict candidates and improve existing approaches. Many approaches attempt to identify cell type-specific genes, as these are relatively easy to identify, and their specific presence in a cell type is often (although by no means always), indicative of function. Computational efforts to identify transdifferentiation factors, has included modelling development onto patterns of gene expression, and approaches to discover ‘core’ TFs that are both cell type-specific and expressed at high levels. Mogrify used a cell ontology tree to map cell type-specific genes against their developmental pattern and so identify TFs specific to a developmental lineage. Mogrify also includes nearest neighbour protein–protein interactions to overcome limitations in discovering cell type-independent TFs. This technique was very successful in discovering previously known transdifferentiation TFs, and was used to predict and then validate TFs that transdifferentiated keratinocytes into endothelial cells. CellNet describes another approach using vast amounts of microarray data to build cell type-specific gene regulatory networks, and then to apply these networks to predict cell type-specific regulatory modules, and so candidate TFs for transdifferentiation. CellNet set out to solve a common problem in transdifferentiation and differentiation experiments where the differentiated cells fail to completely silence the gene expression program of the originating cell type, and remain immature. CellNet was successfully used to improve the transdifferentiation of B cells to macrophages, and also identified an alternate colon cell fate for cells that were transdifferentiating to hepatocytes. Pairs of TFs often antagonize each other's function, hence pairs of TFs with opposing gene expression in two cell types could be used to predict master regulators of lineages.91, 92 Methods that combine gene expression data with epigenetic data have also been successful in predicting transdifferentiation TFs. Another approach extended the discovery of cell type-regulatory modules by looking at gene expression in other mammalian species, and discovered many primate-specific long non-coding RNAs (lncRNAs) with putative cell type-specific functions. Indeed, lncRNAs are also expressed in a cell type-specific pattern,26, 95 and are good candidates for cell type-specific control.96, 97 However, a comprehensive explanation of how TFs, lincRNAs, and other non-coding RNAs can control cell type, and why transdifferentiation is rare in the adult organism remains unclear. One possible solution to this problem are developmental landscapes and cellular pathways that describe the routes cells can traverse to alter cell type.

Developmental landscapes

The concept of cell type is a powerful and attractive model to explain how a limited supply of information can encode a wide array of complex developmental patterns. Development is a highly-ordered process marked by major stages of cell differentiation during gastrulation that establish the three somatic germ lineages of the mesoderm (blood, muscle), endoderm (lung, digestive tract) and ectoderm (skin, brain). As cells take part in development they differentiate to specific cell types and lose developmental competency. What mechanistically underlies these cell type conversions in development is less well understood. What is clear is that combinations of interconnected genes form ‘gene regulatory networks’ (Fig. 4A), and the genes and proteins regulate each other at multiple levels to maintain a semi-stable cell type. But what ultimately builds Waddington's epigenetic landscapes, or the ocean expanses of Cook's islands is unknown. These two models suggest cell types are like depressions in a topological landscape (Waddington), or islands of cell type stability separated by expanses of unstable sea (Cook's islands) (Fig. 4B). Gene regulatory networks underlying cell types may function as ‘attractors’ for cell types to cluster around, similar to probabilistic cell state maps that can construct landscapes for cells, or cellular network entropy, or landscapes constructed based on molecular similarity. Ultimately, conceptual ideas such as Waddington's epigenetic landscape, and Cook's islands, still lack robust biological mechanisms that can explain all aspects of cell type control. Specifically, why there is a limited number of cell types at all, why there are so many unstable intermediary states between cell types, why some cell types are stable and some are not, and why there are strict limits placed on cell type transdifferentiation. Nonetheless, it is becoming possible to model the organization of cell type on a large scale,26, 103 even if we cannot as yet understand the process in detail. New technologies will continue to be applied to this problem; its deeper understanding will have important implications for understanding cellular regeneration and ultimately tissue regeneration and how these techniques can be applied to the understanding of development and human disease.

Fig. 4

Gene regulatory control of cell type. (A) Two gene regulatory networks are shown for two hypothetical cell states (red and blue). Each node in the network is a gene or protein, and each edge is a regulatory interaction, such as a TF binding to regulate a gene, or a kinase activating or repressing a protein, etc. Sets of genes organize into self-stabilizing networks that maintain cell type, and impede transdifferentiation. (B) One view of cell type is the concept of a ‘landscape’ for all cell type possibilities. In this idea, cell types exist in a probability space, and cell types are then defined as the local energy minima in which a cell type can stably exist (purple depressions), surrounded by barriers that prevent conversion of cell type (orange hills).