Computational Stem Cell Biology: open questions and guiding principles

Abstract

Computational Biology is enabling an explosive growth in our understanding of stem cells and our ability to use them for disease modeling, regenerative medicine, and drug discovery. We discuss four topics that exemplify applications of computation to stem cell biology cell typing, lineage tracing, trajectory inference, and regulatory networks. We use these examples to articulate principles that have guided Computational Biology broadly and call for renewed attention to these principles as computation becomes increasingly important in Stem Cell Biology. We also discuss important challenges for this field with the hope that it will inspire more to join this exciting area.

Cahan and colleagues analyze how computational principles are best applied to stem cell biology. They focus on cell typing, lineage tracing, trajectory inference, and regulatory networks as examples articulating principles that have guided Computational Biology broadly and call for renewed attention to these principles as applied in stem cell research.

The intersection of Computational Biology and Stem Cell Biology

Computational tools have aided our understanding of stem cells and development at least since 1952, when Alan Turing used computer simulations to explore reaction diffusion as an explanation of embryonic patterning (Turing, 1952). In the intervening decades, computational tools have become inextricably linked to stem and developmental biology with the advent of high throughput technologies such as nucleic acid sequencing, and with the increased prevalence of modeling and simulation. In this Perspective, we discuss four current topics that epitomize the intersection of Stem Cell Biology and Computational Biology: cell typing, lineage tracing, trajectory inference, and regulatory networks. We use these examples to articulate general principles of how Computational Biology can be leveraged to meet the future needs of Stem Cell Biology. In the final section we highlight other topics at the interface between Stem Cell Biology and Computational Biology (an intersection that we refer to as Computational Stem Cell Biology (CSCB)) but that we do not discuss in depth due to space constraints.

Current challenges and future opportunities

Cell typing and assessing fidelity of cell fate engineering

Three specific aims in CSCB relate to the concept of cell identity and are linked by the fact that they can be achieved with related computational methods that operate on molecular profiles (Figure 1). These are concerned with i) establishing whether a cell population is pluripotent, ii) whether an individual cell is multipotent or pluripotent, and iii) the fidelity of engineered cell types when compared to their in vivo counterparts.

Figure 1.

Computational determination of population and single cell state, identity, and function from single cell and bulk molecular profiles. (A) First generation CSCB methods to predict whether a population of cells was pluripotent or not. (B) Second generation methods extended these approaches from bulk population to single cell level of resolution and allowed for prediction of multipotency. (C) Methods to quantitatively measure the similarity of engineer cells (i.e. those derived by directed differentiation or direct conversion) to their in vivo counterparts. (D) A pressing current challenge is to determine from a molecular profile the extent to which an engineered cell or cell population will behave as intended.

Prospective experimentation, such as a blastocyst complementation (Bradley et al., 1984), may be the gold standard for ascertaining whether a cell population is pluripotent, but it is not always possible nor desirable for all species (Müller et al., 2010). There are many cases where molecular surrogates of pluripotency are required, and traditionally these surrogates have been limited to a handful of markers (e.g. Pou5f1/Oct4 in pluripotent stem cells (Osorno et al., 2012), or Lgr5 in intestinal stem cells (Barker et al., 2007)). However, relying on a small number of markers is unlikely to yield a robust outcome, for example in cases of partial reprogramming (Knaupp et al., 2017) or parthenote-derived hESCs (Sagi et al., 2019). Therefore, computational techniques that use genome-wide data to address this question have emerged. One of the first was Pluritest, which compared microarray-generated gene expression profiles of cell lines to a compendium of ESCs and differentiated cell types (Müller et al., 2011). The Pluritest approach leveraged non-negative matrix factorization (Table 1) (Lee and Seung, 1999) to determine cell type conditional models of expression, which were then used as a basis to compute pluripotency scores on external samples. Subsequently, CellNet addressed the question of pluripotency by using a random forest classifier trained on gene regulatory networks extracted from a curated set of microarray expression data (Cahan et al., 2014; Morris et al., 2014). These and other ‘first generation’ methods including ScoreCard, TeratoScore and KeyGenes were trained on bulk data, which has well-known limitations relative to single cell data (Avior et al., 2015; Bock et al., 2011; Roost et al., 2015). Profiling groups of cells may yield a low pluripotency score if some cells remained pluripotent but others had differentiated, and this would be indistinguishable from cultures where all cells had differentiated. Moreover, bulk data obscures the detection of distinct sub-states of pluripotency more easily discernible at the single cell level (Neagu et al., 2020). Therefore, the second generation of computational pluripotency predicting methods relies on single cell genome-wide data, mostly scRNA-seq. Currently, there are two approaches to do so.

Table 1:

Glossary of terms

Term	Definition
Auto-regulatory loop	network motif in which nodes positively regulate themselves and each other
Boolean network	a network in which each node can take on only binary values (ON/OFF), which is determined by the set of edges into that node and a defined regulatory function (e.g. OR)
Cell fate engineering	deriving a cell population of desired identity through directed differentiation of stem cells or through direct conversion
Entropy	in information theory, entropy is the uncertainty associated with the outcome of random variable
Maximum likelihood	a technique of estimating the parameters of a model such that the probability of the observed data is at the maximum
Maximum parsimony	a technique to reconstruct a phylogenetic tree by minimizing the number of nucleotide changes required to build the tree
Mutual information	in information theory, MI can be considered the amount of information about one random variable is gained by determining the value of another random variable
Network motif	sub-graphs that occur frequently in a network (e.g. positive feedback)
Non-negative matrix factorization	a group of matrix factorization approximation algorithms
Random forest	a machine learning classification technique that is based on sets of decision trees generated by repeated sampling from the training data
Regulon	the set of genes that are transcriptionally regulated by a transcription factor
Simpson’s Paradox	the appearance of trends in grouped data that disappear or change when examining the groups individually

One approach to identify stem cells from scRNA-seq data is to search for higher level features of stem cells in the query data. For example, stem cells tend to sporadically express lineage regulators; a phenomenon called priming in hematopoietic stem cells (Hu et al., 1997; Månsson et al., 2007; Miyamoto et al., 2002). Expression heterogeneity can be quantified by borrowing a core notion from Information Theory - that of entropy, which quantifies the degree of uncertainty associated with a random variable. Several methods have used this concept, including SLICE, which infers potency based on entropy of GO-defined gene sets, and SCENT, which infers potency based on the entropy of signaling networks, and this concept has been extended to quantify the maturation of pluripotent stem cell (PSC)-derived cardiomyocytes (Guo et al., 2017; Kannan et al., 2020; Teschendorff and Enver, 2017). A central property of stem cells is that they differentiate into distinct lineages. Therefore, if lineage relationships can be gleaned from single cell data, then this information can be factored into a metric of fate potential. In a subsequent section, we describe Trajectory inference (TI) methods that perform precisely this task. StemID and FateID are examples of potency prediction methods that count the number of lineages predicted to emerge from each cluster of single cells (Grün et al., 2016; Herman et al., 2018). They then incorporate this information with transcriptional entropy to produce a potency score. These methods used a priori determined features of stem cells (e.g. transcriptional heterogeneity, number of predicted descent lineages), whereas a more recent approach, CytoTRACE, took a hypothesis-free approach to find features that tracked with fate potency across a range of well-annotated scRNAseq data sets (Gulati et al., 2020). Somewhat surprisingly, the number of genes expressed in a cell was discovered as the top predictor. Based on the rapid pace of innovation in this area, we expect that there will continue to be innovative approaches to derive and combine genome-wide features to predict cell fate potency.

The second class of approaches to predict cell fate potency is analogous to the first generation based on bulk data in that they compare gene expression from a query data set to a reference data set in which the stem and differentiated cells have been annotated. However, these comparisons are challenging due to differences in experimental conditions that may introduce batch effects that obscure meaningful comparisons. Although more standardized protocols for sample preparations may ameliorate these issues, the need for specialized computational approaches that go beyond standard clustering methods will remain (Ding et al., 2020; Ziegenhain et al., 2017). Already there are several methods, for example scmap, scID, singleCellNet, scPred, scClassify, Cell-BLAST, Moana, Garnett and CellAssign (Alquicira-Hernandez et al., 2019; Boufea et al., 2020; Cao et al., 2020; Kiselev et al., 2018; Lin et al., 2019; Pliner et al., 2019; Tan and Cahan, 2019; Wagner and Yanai, 2018; Zhang et al., 2019) that perform ‘cell typing’, recently reviewed by Abdelaal et al (Abdelaal et al., 2019). These methods are also emerging as a way to address the third aim in this section, which is to determine the extent to which the products of cell fate engineering (CFE) resemble their ‘natural’ counterparts.

Most methods dealing with predicting cell type and cell potency have adhered to best practices of scientific computational tools (e.g. availability of code, data and documentation to ensure reproducibility and to enable improvements and extensions). Two areas that still warrant improvement are guidance on how to optimize parameters and how to assess the confidence of the outputs. For example, CellNet returns a classification score which represents the fraction of decision trees that predicted that the sample was a given cell type, however the associated sensitivity and precision of this score are not easily extractable. Pluritest and KeyGenes return the distribution of scores for training data so that the user can visualize the similarity of their samples to true ESCs and the predicted in vivo human organ/tissue counterparts, respectively. While ideally these outputs would be coupled to performance metrics such as sensitivity and precision, they are nonetheless sufficient to generate reliable hypotheses that have been experimentally tested in several systems (Takasato et al., 2015).

There are several issues in predicting cell type and cell potency that warrant special attention and focused effort. First, most current cell-typing methods have been developed and benchmarked primarily for terminally differentiated cell types whereas, by definition, stem and progenitor cells are uncommitted. Therefore, a first step will be to generate and use appropriate training data, such as scRNAseq of embryonic and fetal sources. We note that because cells often exhibit a continuum of expression states as they differentiate (Sharma et al., 2020; Stumpf et al., 2017), supervised machine learning approaches to predict developmental staging and cell-maturation will require very large training datasets.

Second, these methods need to incorporate other sources of molecular data to produce better predictors of cell identity and cell potency. Just as bulk profiling approaches have obscured the heterogeneity of cell phenotypes, current approaches to pathway or cell ontologies lack sufficient molecular resolution to derive cell-type specific annotations. Therefore single-cell integration of multiple molecular layers is needed to describe and understand cell state transitions. In addition to scATAC-seq, emerging technologies include the dual profiling of RNA and open chromatin, RNA and protein surface markers, and RNA and DNA methylation, as reviewed in Hu et al (Hu et al., 2018). The development and support of well-curated databases that link high dimensional genetic, epigenetic, transcriptomic and proteomic data to specific phenotypes, such as tumorigenicity, immunogenicity, genome stability, editability, pluripotency and differentiation potential (e.g. EBI-EMBL Expression atlas, scExpression atlas, Stemformatics, as well as hPSCreg) will enable new prediction algorithms (Choi et al., 2019; Papatheodorou et al., 2020). Principled methods that tie these molecular readouts together will be useful beyond more finely resolved predictions of cell identity and cell potency. For example, methods that can take into account the underlying genetic variability and the role of specific variants in determining cell fate can elucidate cellular mechanisms and guide future experiments (van der Wijst et al., 2020).

Third, we need to develop methods and a corresponding nomenclature to handle cases where CFE yields cell types that do not exist or have not been detected in a developmentally anchored reference atlas (Kime et al., 2019; Tonge et al., 2014; Yang et al., 2019a). Imperfectly differentiated or reprogrammed cells exist in hybrid and often transient states and the extent of artificial cell types has not been systematically explored (Morris et al., 2014), but it is a phenomenon that will negatively impact attempts to deconvolute population composition from bulk-derived data (Burke et al., 2020). Therefore, more expansive cell-typing methods that recognize hybrid identities and that find the specific attributes shared between the engineered cells and in vivo cell types are needed. Along these lines, we should develop methods to predict the behavior or function of cells that have been produced with synthetic biology -- in other words, cells that are not intended to mimic an in vivo identity, but rather are engineered to perform specific functions (Figure 1) (Del Vecchio et al., 2017; Qian et al., 2017; Toda et al., 2018, 2019).

Finally, we will need to capitalize on in situ technologies, which can include information such as cell morphology, structural characteristics and the behavior of neighboring cells. These data can be used to perform more refined cell typing, and to distill information about the specific contributors of niche elements to stem cell maintenance and behavior. Ultimately, we predict that in situ data will be used to characterize the fidelity of engineered tissues and organs on the combined basis of cell identity, cell type proportions, and localization.

Lineage tracing

Whereas cell typing is concerned with cell identity, the central goal of lineage tracing is to identify all progeny that arise from an individual cell. Understanding lineage history is powerful in directing cell fate decisions to the desired outcome in the context of stem cell biology. For example, analyzing lineage tracing of alveolar epithelial type 2 (iAEC2) differentiation with a Continuous State Hidden Markov Model (Lin and Bar-Joseph, 2019) led to predictions that ultimately improved directed differentiation of iAEC2 cells (Figure 2A) (Hurley et al., 2020). Although many lineage tracing approaches can achieve high spatial resolution, they typically do not capture many molecular features of each cell under observation, and thus provide limited insight into the regulators and mechanisms of cell fate decisions. To overcome this limitation, single-cell transcriptomics has recently taken a core role, fueling a new wave of lineage tracing tools. These methods are based on uniquely labeling individual cells by leveraging naturally occurring somatic mutations (Leung et al., 2017; Lodato et al., 2015) or experimentally introducing heritable genetic makers (Kester and van Oudenaarden, 2018). While genetic labeling or barcoding offers theoretically unlimited diversity to track cells and their progeny within a defined population (Lu et al., 2011; McKenna et al., 2016; Porter et al., 2014; Sun et al., 2014) these methods were originally incompatible with scRNA-seq because they relied on DNA sequencing. In response, a suite of tools enabling barcode expression as transcripts has allowed the capture of lineage information in parallel with single-cell transcriptomes (Alemany et al., 2018; Biddy et al., 2018; Bowling et al., 2020; Frieda et al., 2017; Raj et al., 2018; Spanjaard et al., 2018; Tusi et al., 2018; Wagner et al., 2018; Yao et al., 2017).

Figure 2.

(A) Understanding how early state relates to eventual fate is valuable to design efficient cell engineering strategies. Enriching for or promoting desirable early cell states can enhance the yield of target cell types. (B) Via the simultaneous capture of lineage and cell identity across a differentiation/reprogramming process, early cell state can be linked to eventual cell fate. These lineage tracing strategies indicate the existence of heritable properties that guide fate determination. However, eventual cell fate cannot be predicted based only on gene expression. Additional information, from chromatin accessibility assays such as single-cell ATAC-seq, and other ‘hidden variables’ may serve a valuable role in uncovering these heritable properties. (C) Schematic of cell labeling to enable phylogeny construction. Heritable cell labels are introduced via a variety of experimental methods, or naturally occurring somatic mutations can be exploited. Accumulation and inheritance of labels is used to reconstruct phylogenetic trees.

Viral barcoding for clonal analysis and lineage tracing has been deployed across several stem cell differentiation and reprogramming paradigms. The accessibility and malleability of cells within these systems renders them amenable to multiple rounds of viral transduction, representing a more tractable approach for cell barcoding, relative to genome editing. Furthermore, cells can be sampled throughout differentiation or reprogramming, enabling progenitor states to be linked to eventual fate. This strategy, coupled with two computational methods (logistic regression and a multilayer perceptron neural network), was used to determine how well gene expression state in progenitors accounts for eventual cell fate in hematopoiesis (Weinreb et al., 2020). Later-stage sister cells were used to predict the dominant differentiation outcome, where gene expression at the later stage held greater predictive power than the expression state of the progenitors. These observations indicate the existence of heritable properties that guide fate determination that were not detected by scRNA-seq alone. Similar observations were made in transcription factor-mediated direct reprogramming of mouse embryonic fibroblasts (MEFs) to induced endoderm progenitors (iEPs) (Biddy et al., 2018). In contrast to these observations, an in vivo state-fate system uncovered distinct transcriptional states in HSCs that predicted differentiation capacity (Pei et al., 2020). One possible explanation for this apparent discrepancy is that the molecular states of ex vivo and in vitro progenitors are less well defined than their in vivo counterparts on a transcriptional versus epigenomic basis. Additional information from chromatin accessibility assays such as ATAC-seq and single-cell ATAC-seq will be invaluable to uncovering these heritable properties and in resolving these apparent discrepancies (Figure 2B). For example, an application of machine-learning and gene regulatory network analysis of gene expression and chromatin accessibility information showed that lineages that failed to convert to iEPs did so because the reprogramming TFs were unable to properly regulate genes essential to iEP function and identity (Kamimoto et al., 2020). These observations were consistent with prior work suggesting that reprogramming to pluripotency is inefficient due to the required target genes being ‘locked’ in heterochromatin and thus unavailable for targeting by reprogramming factors (Soufi et al., 2012). Collectively, this combination of lineage tracing, scRNAseq and computational sleuthing has implicated factors determining chromatin accessibility as the likely heritable properties that influence cell differentiation and reprogramming outcome.

The computational analysis of single-cell lineage tracing is in its infancy and is currently facing several unmet needs. First, particularly for CRISPR/Cas9 single-cell lineage tracing, broadly applicable methods for phylogenetic tree reconstruction based on maximum-likelihood and maximum parsimony approaches are emerging (Fig 2C) (Feng et al., 2019; Jones et al., 2020) but will need continuous development to accommodate the diversifying experimental lineage tracing toolbox. Indeed, novel computational methods are overcoming challenges presented by these complicated experimental platforms. For example, one computational method exploits the stochasticity of cell fate choice in development to overcome the requirement for multiple rounds of labeling to infer phylogenies (Weinreb and Klein, 2020). Second, as above, the computational integration of diverse data modalities will be critical to uncover the heritable properties that determine cell fate. Third, new techniques to visualize lineage in concert with state manifolds will be essential to fully interpret ground truth lineage data and reveal the existence of hidden variables that could be leveraged to fully understand the underlying molecular mechanisms that control the specification and maintenance of cell identity.

Trajectory inference

The molecular state of a cell in a tissue or population is rarely static, but rather varies stochastically in response to its environment, and to reflect the stage of biological processes that animate it, such as cell cycle and differentiation. Sampling a population of cells by scRNASeq or other single cell molecular profiling captures this heterogeneity. Trajectory inference (TI) is the computational task of determining the position of single cells on temporally regulated biological processes. TI is powerful in many ways. First, it enables the identification of new transition stages or branch points, as well as stage-specific markers that can be used to prospectively isolate transient populations. Second, TI allows for the identification of clusters of genes correlated in temporal expression, and thereby allows for the determination of the function of unannotated genes via ‘guilt-by-association’. Third, by placing cells in pseudotemporal order, TI allows for the inference of causal regulatory relationships and thus can identify regulators of differentiation and the subsequent cascade of transcriptional events. In practice, TI has most often been used to study differentiation and therefore is akin to lineage tracing in that it explores the lineage relationships. However, trajectories and lineages are fundamentally distinct measurements, with the latter requiring the experimental mapping of cell lineage relationships. Therefore, differences in the methods (inference based on transcriptional similarity vs molecular fingerprinting) and the timescales (hours to days vs days to weeks or longer) have entailed distinct computational approaches. In this section, we give a taste of how TI methods work, and we describe some early, pioneering applications of TI in the context of differentiation and CFE. Then, we touch on more recent advances in this area, discuss existing areas that require improvement, and highlight opportunities for conceptual and methodological advancement that will help TI methods to reach their potential.

Among the first TI methods were Wanderlust and Monocle, which were designed for mass cytometry and scRNAseq data respectively (Bendall et al., 2014; Trapnell et al., 2014). Key aspects of both of these approaches have endured as common features of most subsequent TI methods: i) embed single cell data in a lower dimensional space to provide a more efficient representation of the cells and a basis for a more biologically relevant cell-cell distance metric; ii) create a graph that links cells or group of cells; iii) infer the trajectory based on the topology of the cell-cell graph; iv) place cells on the inferred trajectories, which is also referred to as determining pseudotime. The application of Monocle to differentiating myoblasts revealed the power of the TI approach in several ways. First, it identified transcriptional regulators of differentiation based on enrichment of TF binding sites in genes sharing the same temporal expression pattern. Second, it uncovered alternative, unexpected differentiation trajectories. Finally, it led to the identification and experimental validation of novel regulators of branch point decisions. Another early and pioneering application of TI was to determine the extent to which directed differentiation and direct programming to motor neurons follow the same developmental path (Briggs et al., 2017). Although no formal TI method was used in this study, trajectories and the relative ordering of cell states were inferred from dimensionality-reduced scRNA-seq data. Comparing directed differentiation and direct programming led to the observation that direct programming skipped intermediate stages characterized by the expression of patterning genes, the activation of which was subsequently recovered close to the final motor neuron stage. This study was a potent demonstration of how TI analysis can be used to explore fundamental questions; in this case of convergent development and the concept of cell types as attractors states.

Dozens of TI methods have since been invented that vary primarily in the analysis steps outlined above and in their assumption of the topology of the trajectory of the data (Saelens et al., 2019). For example, diffusion pseudotime uses a nonlinear dimension reduction technique that better reflects the continuous and noisy nature of differentiation than linear methods such as PCA and ICA (Haghverdi et al., 2016). Based on benchmarking analyses, it is clear that there is no single best practice or best method, and therefore we expect that further refinements to this core pipeline will continue. More recently however, very creative approaches that go beyond the core TI pipeline to predict differentiation paths from scRNAseq have emerged. For example, RNA Velocity and subsequent extensions use the ratio of spliced to unspliced mRNA to model transcriptional kinetics and thereby predict the future state of a cell (Bergen et al., 2020; La Manno et al., 2018). One of the benefits of RNA Velocity is that it gives direct and automatic indication of the starting and terminal points of a trajectory. Another new approach is Waddington optimal-transport (WOT), which, unlike most TI methods, is robust to the underlying topology (e.g. linear, branching, cycle) of the developmental trajectory (Schiebinger et al., 2019). WOT employs the mathematical tool of optimal transport to identify cell descendants and ancestors among coupled time points using proliferation- and apoptosis-related gene expression. WOT assumes that cell proliferation is the driving force of developmental processes connecting cell descendants to ancestors along a single trajectory. Another innovation of WOT was the use of ligand-receptor interaction analysis to uncover crosstalk along trajectories and to identify crucial interactions that orchestrate cell fate decisions at branch points during reprogramming.

There are several major open questions and opportunities for advancement in TI. First, because TI methods are so new, they generally do not have strong guidance on how to determine whether optimization is required, and if so, how to perform this optimization. Second, many TI methods rely on user input to identify starting and terminal points, but this is often not known. The integration of other methods such as CytoTRACE or RNA Velocity to determine these starting and end points in a systematic way would be valuable. Third, most methods do not handle disjoint sets of cells well (Cao et al., 2019). Fourth, the assumption that transcriptional resemblance equates to lineage relation might not hold across all developmental contexts as different sources of expression variation could mask the bona fide developmental trajectory. Therefore, approaches to identify such cases should be devised. Fifth, although gold standards to evaluate and compare TI methods have matured, the in vivo data still rely largely on temporal ordering as defined by original publications rather than orthogonal data. Fate-state data is an important dimension to add to growing TI gold standards. Sixth, as cell fate decisions are governed by spatial-temporal signaling, computational methods that infer paracrine signaling during development from in situ sequencing data are needed. Finally, especially in the context of comparing in vitro with in vivo development, it would be valuable to have formal methods to systematically compare trajectories in a quantitative manner.

Inferring and using regulatory networks

Critical to understanding cell identity and its emergence via developmental trajectories is to understand the regulatory networks that drive cellular decision-making. Such networks involve many different biological macromolecules (e.g. genes, proteins, metabolites, and non-coding RNA) that interact in diverse ways to produce the incredible diversity of emergent behavior we see across each of our cell types. For instance, specific configurations of these networks control differentiation (Simões-Costa and Bronner, 2015) or direct responses to external stimuli (Bourret and Stock, 2002). Computational representations of regulatory networks enable the modeling of cell behavior as a system, and the prediction of how cell state changes over time or upon perturbation. In the context of stem cell biology, regulatory networks have been used to identify key mediators of pluripotency, multipotency, self-renewal, cell fate decisions, to name just a few (Chen et al., 2008; Nishiyama et al., 2009; Yachie-Kinoshita et al., 2018; Zhou et al., 2007). Moreover, networks provide a formalism for reasoning about general mechanisms that explain stem cell properties. For example, network concepts have been used to explain cell behavior like cell state stability, cell fate choices and the effect of targeted perturbations (Dunn et al., 2014; Zhou et al., 2011). In this section, we describe the most prevalent experimental and computational methods used to identify and analyze regulatory networks. We briefly mention several seminal applications in stem cell biology, development and CFE. We finish the section by discussing how single cell approaches are revolutionizing this area and the pressing questions that need to be addressed in the future.

There are several classes of networks used to explore and understand cellular decision-making. Arguably, the most frequently used in stem cell biology is the transcriptional network, commonly referred to as a gene regulatory network (GRN) (Erwin and Davidson, 2009; Karlebach and Shamir, 2008). In GRNs, nodes represent genes while edges represent transcriptional regulation and are only present between genes encoding TFs and their predicted target genes, which may include other TFs. Edges can be inferred one TF at a time using Chip-Chip or Chip-Seq (Figure 3A). In fact, a pair of pioneering applications of Chip-Chip identified the regulons of the core pluripotency regulators Oct4, Sox2, and Nanog (Boyer et al., 2005; Loh et al., 2006). In both mouse and human ESCs, these factors form an auto-regulatory loop, a type of network motif, which helps to explain how transient loss of expression of any one factor can be tolerated (Chen et al., 2008). Later work dissecting other targets of these core factors revealed that another network motif, mutual inhibition of targets of these factors, contributes to early lineage commitment (Loh and Lim, 2011; Thomson et al., 2011). A benefit of defining edges with Chip-Seq is that it yields transcription factor binding sites (TFBS), which can subsequently be used to search for TF activity in other experimental settings (e.g. enrichment in gene clusters or in accessible genomic regions identified via DNAse-hypersensitivity or ATACseq). Edges can also be inferred by transcriptional profiling after modulation of TF expression. For example, hundreds of mouse and human PSC lines have been engineered such that induction of a single TF is drug-controllable. These lines have been used to identify the transcriptional consequences of induction of each of hundreds of TFs (Nakatake et al., 2020; Nishiyama et al., 2009). Recently, this concept has been extended to the single cell level with approaches such as Perturb-Seq or Reprogram-Seq, which can in parallel test the effect of many TFs in isolation or combination (Dixit et al., 2016; Duan et al., 2019). Groups have compiled the results of these and other experimental approaches into large databases of predicted regulatory relationships (or edges) (Gheorghe et al., 2019; Szklarczyk et al., 2019), which can then be used to perform network analysis in other contexts or as benchmarks against which to evaluate the performance of GRN inference methods.

Figure 3:

Derivation and application of cellular networks. (A) Experimental strategies to reconstruct transcriptional regulatory networks. Top-left: Chip-Chip and Chip-Seq identify transcription factor binding sites. Top-right: Association in expression patterns between TF and putative target genes across perturbations, or in a time-lagged manner, imply a regulatory relationship. (B) Leveraging networks to improve CFE. GRNs are first constructed from in vivo data. Computational integration of the activity of network components in engineered cells is used to predict regulators to modulate to improve the CFE outcome. (C) Generative networks can be used for either quantitative (using ordinary differentiation equations) or qualitative (using Boolean Networks) dynamic simulations.

Another class of methods to identify GRNs is based on the premise that a variety of perturbations will elicit TF-target associations in expression. By computing the pairwise association in expression between TFs and all genes across a range of perturbations or states, it is possible to detect regulatory relationships (Le Novère, 2015). There are many such network reconstruction methods that differ in the metric of association (e.g. Pearson correlation or Mutual information (Wang and Huang, 2014)), approaches to deal with the preponderance of false positives that come from indirect interactions (Margolin et al., 2006), whether the method predicts direction of regulation (promotes or represses), and whether the method leverages TFBS or epigenomic information, among many other aspects. These methods have been subject to community-driven benchmarking, which has led to the conclusion that when applied to eukaryotic systems any single co-expression based method has a qualitatively low performance, but aggregating results across methods yields better predictions (Marbach et al., 2012). An oft-maligned culprit is the difficulty in distinguishing direct from indirect effects, leading to a high number of false positives. It was initially anticipated that single cell data would improve GRN reconstruction by avoiding Simpson’s Paradox, and thus reducing the false positive rate (Trapnell, 2015). However, initial benchmarking of single cell GRN reconstruction methods suggest that single cell data alone does not improve GRN reconstruction performance (Chen and Mar, 2018; Pratapa et al., 2019). Nonetheless, reconstructing GRNs from single cell data has yielded important insights, including pinpointing the roles of Sox and Hox TFs in the emergence of hematopoietic lineages from mesoderm (Moignard et al., 2015), identifying the targets of pluripotency TFs as mESCs transit from naïve to primed to neurectoderm (Chan et al., 2017; Stumpf et al., 2017), and identifying novel and specific regulators of outflow tract cell differentiation (de Soysa et al., 2019). There are many directions for exploration and improvement of GRN methods from single cell data; one logical next step is to determine how to optimally incorporate pseudo-temporal information both to infer causal relations and to derive dynamic networks (Qiu et al., 2020).

One of the most potent applications of regulatory networks has been in the area of CFE, to predict how to control cell state and fate transitions. The first such algorithms, such as CellNet, Mogrify, and SeeSawPred relied on bulk genomic profiling data and used cell type specific GRNs to identify candidate TFs whose expression could be modulated to engineer desired cell fate transitions (Cahan et al., 2014; Hartmann et al., 2018; Rackham et al., 2016) (Fig 3B). These TFs may not have been identified by a more general differential expression analysis because they often exhibit modest differences in expression. However, GRN-based approaches that account for changes in regulon activity will detect fate influencing TFs and thereby reduce the amount of expensive and time-consuming experimental work required. The advent of single cell approaches has opened the possibility of inventing improved CFE computational methods or refining them to target the ever-growing number of cell types and states. One such scRNAseq-based GRN-based method identified a combination of small molecules that increased reprogramming efficiency by inferring and analyzing the GRN governing the corresponding reprogramming trajectory (Tran et al., 2019). Another application of scRNA-seq and GRNs used concepts from information theory to identify synergistic factors that enabled the conversion of hindbrain neuroepithelial cells into medial floor plate midbrain progenitors (Okawa et al., 2018). These are among the first of what will undoubtedly be many approaches and examples of CFE computational methods that leverage single cell data.

There are many other ways that GRNs can be used to explore and learn about stem cells. An example of a different application of GRNs is the Reasoning Engine for Interaction Networks (RE:IN), which offers an extension of the Boolean Network formalism (Peter and Davidson, 2017) that allows for uncertainty in network topology (Dunn et al., 2014; Yordanov et al., 2016). RE:IN borrows a technique from the field of formal verification (Bartocci and Lió, 2016) to incorporate experimental observations as constraints on the trajectories that a valid network should produce. In this way, it enables the user to identify the set of GRNs consistent with their experimental observations, and subsequently to use these candidate GRNs to explore the dynamic behavior of their system (Fig 3C). Exemplifying this, RE:IN has been used to guide experimental validation of untested, predicted behavior, revealing how the naïve state in mouse is sustained or lost via non-trivial interactions between key pluripotency factors, how a dynamic, evolving network of interacting phosphatases regulates commitment and differentiation in the interfollicular epidermis, and which genetic perturbations accelerate and enhance the efficiency of reprogramming (Dunn et al., 2019; Mishra et al., 2017).

Beyond adapting CFE computational methods for single cell data, there are several crucial, unanswered questions in this area. First, how do we optimally use GRNs to identify the most efficient ways to elicit an intended cell fate conversion? Concepts from network theory and control theory will be helpful to address this question, as will simulation systems that direct in silico screens. It is possible that the optimal use of network analysis will depend on the conversion under study and the nature of the given network itself. Second, how accurate and complete must GRNs be to engineer cell fate with a specified precision? Answering this fundamental question will also require a combination of theory and simulation, and in doing so will allow us to evaluate the practical utility of GRN reconstruction methods. Third, how can dynamic GRNs be leveraged to account for temporal dependence of expression states? In other words, CFE computational methods should not just predict what regulators to modulate, but in what order. Finally, how can GRNs be formally linked to signaling networks? Answering this last question will enable a new class of CFE computational methods that predict sets of small molecules, cytokines and growth factors to enable cell fate conversions.

Synthesis and Outlook

There are many other topics that sit at the intersection of Stem Cell Biology and Computational Biology that we have not discussed due to space constraints. We mention some of them here to indicate the pervasiveness of computation in stem cell biology, and to highlight some current unmet needs. A large, open challenge is to identify both coding and non-coding genetic variants that change the propensity of differentiation into specific lineages, which has implications in understanding both congenital disorders (Zhang and Lupski, 2015) and in vitro differentiation bias (Di Giorgio et al., 2008; Hu et al., 2010; Osafune et al., 2008). Another challenge is developing methods to leverage single cell epigenomics, single cell multi-omics (Macaulay et al., 2017), proteomics (van Hoof et al., 2012; Palii et al., 2019), phosphoproteomics (Kimura et al., 2020) and Hi-C (Di Stefano et al., 2020; Dileep et al., 2019; Kim et al., 2020; Zhang et al., 2020) to gain a more complete understanding of stem cells and their differentiation pathways. We note that some of these technologies have not yet reached single cell resolution. Additionally, another type of biological network beyond GRNs are intercellular signaling networks (Yang et al., 2019b), which combine prior knowledge of ligand-receptor complexes with statistical frameworks to predict tissue-specific cell-cell communication networks that contribute to development and tissue homeostasis (Camp et al., 2017; Efremova et al., 2020; Raredon et al., 2019; Skelly et al., 2018; Vento-Tormo et al., 2018). In the future, predictive cell-cell interaction models will be able to inform key cell-cell interactions responsible for maintaining tissue homeostasis and supporting tissue regeneration. Moreover, the comparison of tissue-specific cell-cell interaction networks with cell-cell interactomes of pathological or injured tissues will identify dysregulated interactions that can guide the development of strategies to restore tissue homeostasis. The integration of these models of tissue-specific cell-cell interactions with imaging-based technologies for spatial transcriptome reconstruction (Halpern et al., 2017; Karaiskos et al., 2017; Rodriques et al., 2019) will enable the characterization of the complete interactome in a spatially resolved manner, and therefore enable the generation of more accurate predictions.

A final area that we did not discuss above is cell-based modeling, which is a technique that uses in silico representations to explore how cells interact and change over time (Ghaffarizadeh et al., 2018; Mirams et al., 2013; Sharpe, 2017). The long history of cell-based modeling of stem cells and development has included the exploration of alternative formulations of regulatory networks that spatially pattern the limb (Uzkudun et al., 2015) and the prediction of embryonic toxicity (Kleinstreuer et al., 2013). We anticipate that more accurate and powerful cell-based modeling systems will be generated by incorporating knowledge gained from single cell sequencing, and that such models will be used to iteratively design multi-cellular behavior and function.

In our exploration of how Computational Biology is being applied to Stem Cell Biology, a set of core values or guiding principles emerged. While this list is not comprehensive, and in some ways it is applicable to Computational Biology more generally, it does include the most discussed and, we feel, most crucial principles that can help to meet the challenges of today and tomorrow:

1. Computational methods should have guidance that clearly describes how to interpret the significance or confidence of the method’s results, and this guidance must go beyond standard software documentation (Lee, 2018). Often, the output of computational methods is an ordered list of testable hypotheses, for example a list of transcription factors regulating a process. Since testing all of the possible hypotheses experimentally is often prohibitive, knowing how to interpret the significance of the method’s results is crucial so that the user can prioritize experimental efforts on the most promising candidates.

2. Computational methods should provide clear guidance on when and how they should be optimized. Most Computational Biology tools have multiple parameters that impact analysis outcomes. While devising the methods, creators typically select default values based on optimization on their test data. A method’s performance - for example sensitivity, accuracy, or execution time - can degrade when a user’s data differs substantially from the test data used to derive the default parameters. Therefore, method creators should describe how to determine whether additional optimization is required.

3. Computational methods should be implemented and used in a way to ensure reproducibility. Data need to be described and stored so that they can be re-analyzed later to facilitate data integration, and to allow for computational inference of phenotype. The widespread recognition of the benefits of FAIR data management (Wilkinson et al., 2016) and the fact that most journals now require deposition of genome scale data in public repositories makes this principle widely adhered to. However, the varying degrees to which journals require code to be freely available has continued to hamper reproducibility of computational analytics (Papin et al., 2020).

4. We value tight, mutually beneficial collaborations between computationalists and experimentalists in which ideas and knowledge flow both ways. Such deep relationships will result in more productive studies because mathematical, statistical and computational considerations are incorporated into experimental design, and because biological input from the experimentalists ensures that the computational models are relevant and incisive (Knapp et al., 2015). These types of deep interactions are also valued because they promote entry to the field by people from nontraditional backgrounds such as physics, economics, or other data sciences, and because they lead to the capture and standardization of metadata specific to stem cell biology. And finally, bringing together distinct sets of expertise can fuel new ideas and break away from entrenched norms.

5. Ideally, the initial publication of a method at the intersection of Computational Biology and Stem Cell Biology will include a prospective, experimental assessment. However, we recognize that this is not always practical. In place of this validation, the publication should include a comparison to already published results or a discussion of how the method could be experimentally assessed. Such a discussion can lead to the development of standardized benchmarks that make the comparison and further improvement of a class of methods fairer and more efficient.

Adhering to these principles will help us to meet our challenges and goals in several ways. First, being tightly integrated with our experimental collaborators will make us more responsive to the wider stem cell field, while at the same time ensuring that our contributions are valued appropriately. Second, by setting a standard of reliable and interpretable methods, we will broaden the use of our tools and establish solid credibility in our field’s work. Third, ensuring that our methods are accessible and modifiable will allow for their efficient improvement and adaptation for newly emerging questions and data-generating technologies. Finally, explicit consideration of prospective validation will help to ensure that methods are honed to address specific hypotheses.