The development of CellNet — network-biology software that determines how cell types generated in vitro relate to their naturally occurring counterparts — could improve our ability to produce desirable cells in culture.
Over the past few years, websites such as Facebook and Google have attained an uncanny ability to understand us and to predict our behaviour, even before we have consciously decided what to do. This predictive power is achieved through the systematic application of statistical 'inference algorithms' to the vast numbers of connections and links that users establish when browsing the Internet — making up a 'social graph' that can be exploited to characterize distinct groups of Internet users. It would be wonderful to have such a graph to characterize distinct groups of cells. This could then be used in regenerative medicine to overcome the challenge of coercing stem cells to become the cell type needed for a particular therapy. Writing in Cell, Cahan et al.1 and Morris et al.2 describe a network-biology platform, CellNet, that takes a first step in this direction.
The most popular representation of the differentiation of cells from immature precursors to mature cell types was, for many years, the 'epigenetic landscape' diagram conceived by the biologist Conrad Hal Waddington3,4. This diagram evokes a set of one-way paths down which immature cells roll along defined routes to more-differentiated cellular states. But over the past decade, this simple model has morphed into the concept of a multidirectional cell-identity transfer hub.
In 2007, Yamanaka and colleagues5 reprogrammed ordinary human skin cells called dermal fibroblasts into induced pluripotent stem (iPS) cells using transcription factors that are highly expressed in embryonic stem (ES) cells, an equivalent cell type that is derived from early embryos. Both iPS cells and ES cells are pluripotent — they can, given the correct molecular cues, differentiate into almost any cell in the body, forming any one of hundreds of different cell types. Each of these mature cell types is characterized by distinct networks of highly expressed transcription factors, which regulate the expression of large sets of genes. Researchers have used transcription-factor cocktails specific to cell types of interest to try to directly convert one cell type, such as a fibroblast, into another, such as a neuron6 or a liver cell7 (known as a hepatocyte). The abundance of reports suggests that there are almost no limits to the number of possible cellular transformations.
But are these engineered cells genuine copies of cells that exist in the body? There are well-established gene-expression tools for determining whether a stem cell is pluripotent8,9. But remarkably, given the flurry of research into directing stem cells to take on a particular identity, or fate, there is no commonly accepted way to determine whether a differentiating cell is moving towards the right developmental destination. Cahan et al. designed the CellNet software to give researchers an idea of how closely matched a cultured cell type is to its presumed counterpart in the body. The program applies a sequence of statistical inference algorithms to create a global transcriptional graph, which resembles the social graphs used by websites such as Facebook (Fig. 1a).
To construct this transcriptional graph, the authors used publicly available gene-expression data from tissues and cells, and information from genome-wide transcription-factor binding studies performed by the ENCODE consortium. The CellNet program identifies gene-regulatory networks (GRNs) for specific cell types in the body, such as neurons and hepatocytes. GRNs are groups of genes that are coordinately regulated in distinct cell types and that are more highly interconnected with one another than with other genes. The program then compares these cell-type-specific GRNs with those from experimentally derived cultured cells to determine how accurately the derived cells mimic the 'real' cell type (Fig. 1b). In addition, CellNet suggests transcription factors that could be modulated to shift an in vitro cell type closer to its in vivo correlate.
Cahan and colleagues compared two strategies for producing mature cell types: differentiation from pluripotent stem cells and direct conversion from another cell type. Testing engineered neurons and hepatocytes, CellNet analysis revealed differences between the two approaches. Cell types derived from pluripotent stem cells were similar to the naturally occurring cell types, but directly converted cells could be abnormal. For example, neurons directly converted from fibroblasts retained substantial fibroblastic identity and expressed GRNs that were characteristic of cells from the heart and pancreas.
How can CellNet improve the quality of engineered cells? An inkling of its future utility comes from Morris and colleagues' study of directly converted hepatocytes, called induced hepatocytes (iHeps). A CellNet comparison of iHeps and actual hepatocytes revealed that iHeps did express GRNs that were characteristic of hepatocytes, but they also activated illicit, developmentally immature transcriptional programs. Following up on this observation, Morris et al. found that the iHeps would be better described as induced endoderm progenitors (iEPs). These endodermal precursors can give rise to many cell types that arise from the endoderm (an embryonic cell layer), including cells of the colon, liver and pancreas. Indeed, when the authors used CellNet as a guide to modify the transcription-factor cocktail used for direct conversion, they generated iEPs that could differentiate into mature colon cells when transplanted into mice, and could repair damaged colons.
The major limitation of predictive programs such as CellNet is a shortage of data. In the modern world of 'big data', more data always lead to better predictions. Although Google can profile hundreds of millions of search users and follow their behaviour over several years, the CellNet team were limited to published data generated from a few thousand genome-wide analyses of gene expression. Because of the lack of large data sets from human cells and tissues, the current version of CellNet is practical only for experimental studies of mouse cells.
The lack of large, high-quality genome-wide transcriptional profiles for normal human cell types is a major bottleneck in the development of stem-cell-based therapies and drug screens. We need to learn how to robustly define and mechanistically understand the molecular coordinates for differentiated cell types before we can give human stem cells the directions they need to arrive at the right fate. As the baseball player Yogi Berra once said, “If you don't know where you're going, you'll wind up someplace else”.
Cahan, P. et al. Cell 158, 903–915 (2014).
Morris, S. A. et al. Cell 158, 889–902 (2014).
Slack, J. M. Nature Rev. Genet. 3, 889–895 (2002).
Waddington, C. H. The Strategy of the Genes (George Allen & Unwin, 1957).
Takahashi, K. et al. Cell 131, 861–872 (2007).
Vierbuchen, T. et al. Nature 463, 1035–1041 (2010).
Sekiya, S. & Suzuki, A. Nature 475, 390–393 (2011).
Müller, F. J. et al. Nature 455, 401–405 (2008).
Müller, F. J. et al. Nature Methods 8, 315–317 (2011).
About this article
Cite this article
Müller, FJ., Loring, J. A compass for stem-cell differentiation. Nature 513, 498–499 (2014). https://doi.org/10.1038/513498a