Understanding generation and regeneration of pancreatic β cells from a single-cell perspective

The pancreas is an important internal organ for the digestion of nutrients and maintenance of blood glucose homeostasis, the functions of which are executed by the exocrine and endocrine compartments, respectively. The exocrine compartment consists of acinar cells that secrete various digestive enzymes, and ductal cells that transport these enzymes into the duodenum. The endocrine compartment senses changes in blood glucose and executes regulatory functions via the islets of Langerhans, stereotypically organized structures that are composed of five hormone-secreting cell types: insulin (Ins)-secreting β cells, glucagon (Gcg)-secreting α cells, somatostatin (Sst)-secreting δ cells, pancreatic polypeptide (Ppy)-secreting PP cells, and ghrelin (Ghrl)-secreting ε cells (reviewed by Pan and Wright, 2011; Bastidas-Ponce et al., 2017; Larsen and Grapin-Botton, 2017). It is the dysfunction of insulin-producing β cells that leads to diabetes (DeFronzo et al., 2015; Katsarou et al., 2017).

Affecting approximately 451 million people, diabetes is a worldwide health problem and imposes a heavy burden on patients and societies (Cho et al., 2018). Type 1 diabetes (T1D), which accounts for 5-10% of cases, is caused by insufficient secretion of insulin due to attack of β cells by the immune system. Type 2 diabetes (T2D) is much more common and is caused by insulin resistance of surrounding tissues, such as the liver, muscle and fat. Currently, the major approaches used to alleviate the symptoms of diabetes are the transplantation of cadaveric islets and exogenous insulin injection, but these have the disadvantages of donor shortage and inaccurate glucose control, respectively (Aguayo-Mazzucato and Bonner-Weir, 2018; Tan et al., 2019). Sustainable sources of functional β cells or islets represent a promising method for improved diabetes treatment. The sources for β cells include: endogenous regeneration of β cells; plasticity and trans-differentiation of non-β cells into insulin-secreting cells; and exogenous induction of β cells from human embryonic stem cells (hESCs) (Aguayo-Mazzucato and Bonner-Weir, 2018; Zhou and Melton, 2018). To successfully generate mature β cells, or construct whole islets in vitro, we must understand the pathways and regulatory mechanisms that guide the development and maturation of pancreatic endocrine lineages. To induce endogenous regeneration of β cells through promoting β-cell proliferation, or trans-differentiation of non-β cells into insulin-secreting cells, we must understand the level of cellular heterogeneity in these populations, the mechanisms that regulate cell plasticity and the molecular relationships between endocrine lineages. All these efforts will facilitate development of cell therapies for diabetes.

In the past few decades, research in both animal models and human specimens has furthered our understanding of β-cell development and regeneration to drive exciting advances in diabetes therapy. However, we still do not fully understand the molecular characteristics, cell lineage differentiation pathways and cell fate plasticity of pancreatic endocrine cells. Fortunately, owing to the recent development of single-cell technologies, researchers have been able to re-examine the regulation of development and regeneration of endocrine cells at the single-cell level. Single-cell technology is well suited for the study of complex organs such as the pancreas and facilitates the mapping of the developmental trajectory of each cell lineage, which will further our understanding of the regulation of cell fate determination.

Many single-cell approaches have been developed in recent years (summarized in Table 1). Among these, single-cell RNA sequencing (scRNA-seq) has been commonly used in developmental and regenerative biology. Different scRNA-seq methods (Box 1) have been used to: describe cell subpopulations or cell heterogeneity; identify novel marker genes in cell populations; define the pathways of lineage development and transition; identify regulatory cues for cell fate determination; and trace an organism's cell lineage tree (Kumar et al., 2017; Griffiths et al., 2018; Pijuan-Sala et al., 2018; Potter, 2018; Chan et al., 2019; Chen et al., 2019; Morris, 2019).

In the pancreas field, many studies have used single-cell technology to resolve key issues that have been unresolved by cell population-based studies, and to provide insights into β-cell neogenesis, maturation, regeneration and heterogeneity (summarized in Tables 2-4). This Review summarizes recent pancreatic studies that have increased our understanding of the generation and regeneration of β cells from a single-cell perspective. We largely focus on scRNA-seq studies, but do mention other single-cell approaches where appropriate.

In rodents, the pancreas is derived from two origins (the dorsal and ventral endoderm) that respond to different sets of signals from surrounding tissues to form dorsal and ventral pancreatic buds at embryonic day (E) 9.0 and E9.5, respectively (Pan and Wright, 2011; Bastidas-Ponce et al., 2017) (Fig. 1). These buds are composed of multipotent progenitor (MP) cells expressing Pdx1 and the buds fuse at completion of intestinal rotation at ∼E12.5 (Pan and Wright, 2011; Shih et al., 2013). Subsequently, MP cells differentiate into tip and bipotent trunk cells. Tip cells further specify into acinar cells, whereas trunk cells differentiate into ductal or endocrine cells. Cell fate determination of trunk cells is regulated by Neurog3 (Ngn3), a key transcription factor (TF) for the formation of endocrine progenitors (EPs). It is generally accepted that the endocrine cells originate from Ngn3⁺ EPs and that the specification of each endocrine lineage is temporally distinct (Gradwohl et al., 2000; Gu et al., 2002; Heller et al., 2005; Johansson et al., 2007). Although, as a population, Ngn3⁺ can produce all specialized endocrine cells, individual Ngn3⁺ cells are unipotent precursors to differentiate into one of the endocrine cell types (Desgraz and Herrera, 2009). Pancreas development undergoes two transitionary periods: a first wave transition (from E9.5-E12.5) during which the endocrine cells that are formed are primarily Gcg-expressing cells or multi-hormonal cells (Pan and Wright, 2011; Shih et al., 2013; Bastidas-Ponce et al., 2017), and a second wave transition (from E12.5 to birth), which is the main period of endocrine cell formation. Previous reviews have summarized the developmental process, and related regulatory TFs and signals for each pancreatic lineage (Pan and Wright, 2011; Shih et al., 2013; Bastidas-Ponce et al., 2017; Larsen and Grapin-Botton, 2017).

Genetic studies in rodent models have uncovered the main framework of pancreatic lineage differentiation and have identified many important factors involved in regulating key steps of pancreatic lineage differentiation (Pan and Wright, 2011; Shih et al., 2013; Bastidas-Ponce et al., 2017; Larsen and Grapin-Botton, 2017). However, traditional studies that are based on populations of cells and limited marker genes may mask cell heterogeneity and obscure intermediate progenitor cells that are involved in cell lineage differentiation. These limitations thus affect analyses of cellular differentiation pathways and their underlying regulatory mechanisms. Therefore, decoding developmental process at single-cell resolution can guide us to a more comprehensive understanding of pancreatic organogenesis.

Compared with our understanding of rodent pancreatic development, we know little regarding the development of human fetal pancreas. The human pancreas is also derived from a dorsal and a ventral endoderm domain; the dorsal bud is detected at 26 dpc (day post conception) and the ventral bud at ∼30 dpc. The buds are fused upon gut rotation between 6-8 weeks of gestation (Piper et al., 2004; Jennings et al., 2013). Generally, the lineage hierarchy and molecular features of human pancreas are considered to be very similar to mouse, and the details of human pancreas development have been reviewed elsewhere (Pan and Brissova, 2014; Jennings et al., 2015; Baeyens et al., 2018; Petersen et al., 2018). In brief, PDX1⁺ MP cells segregate into tip and trunk compartments, and the endocrine cells are generated from bipotent trunk cells through a transient NGN3⁺ state. Deficiency of NGN3 leads to the loss or reduction of endocrine cells in humans (Pinney et al., 2011; Rubio-Cabezas et al., 2011; McGrath et al., 2015; Zhu et al., 2016). Importantly, the functions of some key TFs are conserved in regulating cell identity between humans and mice: α-cell formation requires ARX (Itoh et al., 2010), and β cells specifically express NKX6-1 in adult islets (Riedel et al., 2012). However, species differences do exist between human and mouse pancreatic development (Pan and Brissova, 2014; Jennings et al., 2015; Baeyens et al., 2018; Petersen et al., 2018). For example, NKX2-2 is downstream of NGN3 in humans but upstream of Ngn3 in mice (Jennings et al., 2013). Moreover, the order of emergence of the endocrine lineages is different: the insulin-expressing cells appear first in humans, whereas the glucagon-expressing cells appear first in mice (Piper et al., 2004; Jeon et al., 2009). In addition, there has been no clear evidence of first wave transition during human pancreatic development (Villasenor et al., 2008; Jennings et al., 2013).

Notably, researchers should be cautious when applying conclusions drawn from rodent models to human developmental processes, because the differences in cell lineage differentiation pathways and the regulatory mechanisms between the species have not been comprehensively addressed. Therefore, defining a blueprint for the development of the human pancreatic lineage is crucial for our understanding of the developmental process and regulatory mechanisms of the human pancreas. However, because of the scarcity of human tissues, large animal models with similar physiology to humans, such as pigs and non-human primates, can be used as alternative models of human pancreatic research (Zhu et al., 2014; Bakhti et al., 2019; Kim et al., 2019 preprint). Next, we review the single-cell transcriptomic studies of pancreas development in rodent models and humans.

Recent research using well-based scRNA-seq approaches (Box 1), has produced a more comprehensive understanding of pancreatic development, including the differentiation trajectories of lineages and intermediate cell states (Table 2), as well as the key genes of each cell population (Tables 2 and 3).

Although the generation of MP cells occurs at the earliest stage of pancreatogenesis and MP cells are not the direct progenitors of endocrine lineages; the generation of high-quality MP cells is a crucial step for obtaining functional endocrine β cells (Russ et al., 2015). To understand MP cell differentiation in the dorsal and ventral pancreatic regions in mice, Smart-seq2 scRNA-seq has been performed on Pdx1⁺ cells from both regions at E9.5 and E10.5 (Li et al., 2018). These analyses have revealed that Pdx1⁺ cells are heterogeneous, and can be divided into Pdx1^low and Pdx1^high populations in both the dorsal and ventral regions (Fig. 1). In the ventral region, Pdx1^low cells represent intermediate MPs that will become hepatoblasts, extrahepatic bile ducts (EHBDs) and Pdx1^high MP cells. However, in the dorsal region, Pdx1^low cells are the first-wave of endocrine cells and Pdx1^high cells are MP cells that, presumably, directly differentiate from endoderm cells. Therefore, this work has furthered our understanding of the differences between dorsal and ventral pancreatic programs (Rodriguez-Seguel et al., 2013; Li et al., 2018) (Fig. 1).

Subsequently, MP cells initiate stepwise differentiation toward the exocrine and endocrine lineages. Our recent work has comprehensively mapped the multistep developmental trajectory of exocrine and endocrine (α- and β-cell) lineages by Smart-seq2 scRNA-seq from the E9.5-E17.5 dorsal pancreas using various mouse lines (Yu et al., 2019). Along this pancreatic lineage differentiation trajectory, we have identified several intermediate cell states and branch points (nodes) at which progenitor cells undergo fate changes (Fig. 1). From our data, MP cells can be divided into MP-early (E9.5) and MP-late (E10.5) cells according to the developmental time of their appearance. MP-early cells serve as the first node and differentiate into tip cells or the first wave of α cells (α-1st cells) through intermediate states (MP-late and pre-α-1st cells, respectively) (Fig. 1, 1). At the second node, tip cells further specify into trunk and acinar cells (Fig. 1, 2), and trunk cells represent the third node for differentiation into ductal cells and EPs (Fig. 1, 3).

As the endocrine pancreatic lineages, especially α- and β-cell lineages, are responsible for homeostasis of blood glucose and are therefore relevant to translational medicine for therapy, the majority of single-cell studies have focused on endocrinogenesis (Table 2). Several groups have surveyed the process of pancreatic endocrine generation and have revealed that mouse EPs are heterogeneous, with dynamic changes in featured gene expression and a temporal propensity towards differentiation into distinct islet lineages (Byrnes et al., 2018; Scavuzzo et al., 2018; Bastidas-Ponce et al., 2019; van Gurp et al., 2019; Yu et al., 2019). However, the sub-classification of EP cells defined by these different groups are not completely consistent.

Based on our Smart-seq2 scRNA-seq study, EPs were defined as the cells that differentiate from bipotent trunk cells and express a cluster of marker genes including Ngn3 and Neurod1, but do not yet express hormonal genes (Yu et al., 2019). Our study has shown that the rapid fate shift of EPs can be divided into four developmental stages (EP1-EP4) and involves a cascade of gene expression for a series of TFs (Fig. 1). Compared with a fraction of Ngn3 lower-expressing trunk cells, Ngn3 expression increases in EP1, peaks in EP2, decreases in EP3 and returns to background level in EP4, and Neurod1 starts to become expressed in EP2 cells. The EP4 population serves as a branch node to generate α and β cells (Yu et al., 2019) (Fig. 1, 4). Scavuzzo and colleagues have classified Ngn3-expressing cells into four subpopulations at E14.5, namely N14_1-N14_4 (Scavuzzo et al., 2018). However, the N14_2 population expresses markers of mesenchymal cells and deviates from the developmental trajectory formed by other three subpopulations. Considering that droplet-based scRNA-seq inevitably causes doublet contamination, the existence of the N14_2 population needs to be verified. Based on the expression patterns of marker genes, N14_1 roughly corresponds to the Ngn3-expressing trunk cells and the EP1 population; N14_3 is similar to the EP2 population, and N14_4 covers EP3 and a fraction of EP4 cells. In another study, Fev⁺ cells are defined as an intermediate EP population because Fev⁺ cells connect Ngn3-expressing cells and hormone-expressing cells on a pseudotime-ordered pathway. It has been confirmed by lineage tracing that Fev⁺ cells produce the majority of endocrine cells (Byrnes et al., 2018). This finding is consistent with previous studies that have shown that Fev is induced by Ngn3, and is expressed in EP cells and islet cells (Miyatsuka et al., 2009; Ohta et al., 2011). Notably, Fev is highly expressed in EP3 and EP4 cells (Yu et al., 2019), and N14_4 cells (Scavuzzo et al., 2018), and these Fev-expressing EP cells have also been confirmed by other scRNA-seq studies (Bastidas-Ponce et al., 2019; van Gurp et al., 2019). Later, Fev⁺ cells that co-express paternally expressed 10 (Peg10) tend to differentiate into α cells, whereas those that co-express G protein subunit gamma 12 (Gng12) are biased to become β cells (Byrnes et al., 2018) (Fig. 1, 4).

Several studies have provided evidence to support that the development and heterogeneity of EP cells are regulated at the chromatin level (Scavuzzo et al., 2018; Yu et al., 2018; Liu et al., 2019). Our work has uncovered a regulatory role of the histone demethylase Jmjd3 (also known as Kdm6b) during the EP cell transition (Yu et al., 2018). We used an Ngn3-GFP knock-in mouse strain to purify Ngn3-GFP^low (containing Ngn3-expressing trunk cells and EP1 cells) and Ngn3-GFP^high cells (mainly including EP2-EP4 cells). Genetic knockout of Jmjd3 in the pancreas impairs the transition of EP cells from Ngn3-GFP^low to Ngn3-GFP^high states and subsequent islet formation. Using scRNA-seq, we found that Jmjd3 only affects the efficiency of the cell fate transition from Ngn3-GFP^low to Ngn3-GFP^high but does not alter the transcriptomes and developmental pathways of Ngn3^high cells (Yu et al., 2018). Liu and colleagues have proposed that the EPs are transcriptionally uniform but epigenetically distinct (Liu et al., 2019). Using scRNA-seq, the authors have shown that mouse Ngn3⁺ cells heterogeneously express Myt1, which encodes a TF involved in regulating endocrine islet cell differentiation and function (Wang et al., 2007, 2008). Myt1⁺Ngn3⁺ and Myt1⁻Ngn3⁺ cells are biased toward β- and α-cell fates, respectively. Myt1⁺Ngn3⁺ cells highly express DNA methyltransferase 1 (Dnmt1), resulting in a higher level of DNA methylation at the enhancer region of Arx, a key TF for α-cell differentiation. Thus, the expression of Arx is repressed in Myt1⁺Ngn3⁺ cells (Liu et al., 2019). In addition, mouse Ngn3-expressing cells at E14.5 and E16.5 show differences in transcriptome and chromatin accessibility (Scavuzzo et al., 2018). E16.5 cells exhibit lineage bias towards β cells, because motifs of TFs that positively regulate β-cell generation are enriched at the open chromatin regions (Scavuzzo et al., 2018). Supporting this temporal lineage bias model, Sharon and colleagues combined scRNA-seq and 3D microscopic imaging of whole-mount immunostaining to propose a growing peninsula model to describe islet formation in mouse embryonic pancreas (Sharon et al., 2019a). In this model, EP cells are considered to leave the epithelial cord but remain attached and form a growing bud, which depends on the continuous recruitment of the differentiated endocrine cells. Notably, the earlier appearing α cells are pushed outward by the β cells generated later, and eventually the α cells reside in the peripheral region of the mouse islets (Sharon et al., 2019a).

Taken together, these studies have revealed the heterogeneity of EP cells during endocrinogenesis; however, the classifications of EP cells generated by different scRNA-seq platforms are not uniform. Integrated analysis of these datasets might determine the correspondence between these classifications. The heterogeneity of EP cells may reflect their differentiation ability or tendency for a specific endocrine lineage, which might be epigenetically regulated at a distinct developmental time. In addition, combining scRNA-seq analysis with traditional genetic, epigenetic and imaging methods has become a new trend in pancreatic developmental biology, and provides useful tools to address these questions.

Although the developmental trajectories of α- and β-lineage generation have been identified using scRNA-seq (as described above), we know very little about the developmental pathways of ε-, δ- and PP-cell lineages, except that all these endocrine lineages were derived from Ngn3⁺ EPs (Gradwohl et al., 2000; Gu et al., 2002; Heller et al., 2005). Recent scRNA-seq studies have captured a ‘branch’ of ε cells, marked by Ghrl and Irs4, which are generated from Ngn3⁺ EP cells in the mouse embryonic pancreas (Liu et al., 2019; Sharon et al., 2019a). Droplet-based scRNA-seq methods are inefficient in capturing sufficient numbers of δ and PP cells that are required to define developmental pathways on the trajectory map, presumably because of the low abundance of δ and PP cells during the early stage of endocrinogenesis. In addition, owing to high noise and low sensitivity for low-level transcripts, droplet-based scRNA-seq methods may have limited potential in defining cell types with only subtle differences at the transcriptomic level and therefore may have a reduced capacity to define intermediate progenitor cells and precise developmental pathways. Therefore, well-based deep-sequenced scRNA-seq of the enriched ε, δ and PP cells might be a promising approach to uncover the precise pathways of these rare endocrine populations.

scRNA-seq analysis has also uncovered the heterogeneity and developmental dynamics of mesenchymal cells (Byrnes et al., 2018); however, their cell-cell interaction with pancreatic cells requires more extensive investigation.

Because of the scarcity of human fetal samples, research on human fetal pancreatic development has been limited and has led to the overestimation of conservation of lineage hierarchy and transcriptome between mice and humans. Using single-cell qPCR, Ramond and colleagues have profiled the transcription profiles of 91 genes in four populations (population A-D) enriched by cell-surface markers, from 9 weeks of development (WD) in the human fetal pancreas. Through this approach, they revealed the cell composition of each population and branched differentiation pathway of endocrine cells (Ramond et al., 2018). Populations A and B mainly include pancreatic progenitors, population C contains EP cells and population D consists of endocrine cells. Moreover, they identify three lineage trajectories that originate from EP cells: β-track, α and γ-track, and δ-track. Importantly, this work depicted a continuous developmental process from pancreatic progenitors to endocrine cells in 9 WD human pancreas and provides a paradigm to explore the developmental trajectories and regulatory mechanisms of the human fetal pancreatic lineages through a single-cell approach, which allows the effective use of scarce samples. Future efforts should focus on comprehensive and unbiased assessment of human fetal pancreatic development using scRNA-seq in combination with other methods.

Once specified during embryonic development, immature endocrine cells mature into functional endocrine cells (Salinno et al., 2019). Most studies have focused on the maturation of β cells, because of translational medicine potential for the therapy of diabetes. Immature β cells are highly proliferative and respond poorly to glucose stimulation (Blum et al., 2012; Henquin and Nenquin, 2018). Postnatal development is a crucial period for β-cell maturation, during which they acquire physiological functions, such as glucose stimulated insulin secretion (GSIS) in response to extracellular glucose. In both mice and humans, the maturation process of β cells coincides with cell cycle exit (Piccand et al., 2014; Stewart et al., 2015) and is marked by gene expression changes, such as increased expression of urocortin 3 (Ucn3) (Blum et al., 2012) and decreased neuropeptide Y (Npy) (Rodnoi et al., 2017). Moreover, the TFs Mafb and Mafa are expressed in mouse immature and mature β cells, respectively, and this conversion is an important feature of mouse β-cell development (Nishimura et al., 2006). However, unlike in mice, MAFB is constantly expressed in adult human β cells (Dai et al., 2012). During human pancreas development, the ratio of β cells that express glucose-sensing proteins increases, mainly after 15 WD (Richardson et al., 2007). In rodents, the postnatal maturation of β cells is driven by weaning (dietary change from high-fat milk to high-carbohydrate chow) and studies have suggested that miRNAs have a central role in regulating postnatal β-cell maturation (Jacovetti et al., 2015; Stolovich-Rain et al., 2015). Furthermore, signaling pathways such as mammalian target of rapamycin (mTOR) (Ni et al., 2017; Sinagoga et al., 2017), AMP-activated protein kinase (AMPK) (Jaafar et al., 2019) and WNT (Dorrell et al., 2016) have been proposed to regulate β-cell maturation, but need further investigation.

Recent studies at the single-cell level have broadened our understanding of cellular heterogeneity, proliferation and maturation of the β and α lineages (Qiu et al., 2017; Zeng et al., 2017). Single-cell transcriptomic analyses have revealed the genes that are specifically expressed in proliferative versus quiescent cells or immature versus mature cells during endocrine lineage maturation. Excluding cell cycle-related genes, the proliferative β cells exhibit similar transcriptomic profiles to the quiescent β cells, and remain synchronized with the pseudotemporal process of maturation (Qiu et al., 2017). In addition, prenatal and juvenile – but not adult – β cells exhibit heterogeneity in maturation state (Qiu et al., 2017). Moreover, investigation of the maturation pathway and concomitant gene expression dynamics of α cells has shown that β cells and α cells use different regulatory strategies for maturation: α cells tend to downregulate gene expression during maturity, whereas β-cell maturation is accompanied by upregulation of many genes (Qiu et al., 2017). Zeng and colleagues have observed that, during the process of β-cell maturation, the genes related to mitochondrial activity (such as Ndufv1) and amino acid metabolism (such as Slc7a2 and Lamtor5) are downregulated, and the data indicated that Srf is a regulator for maturation-associated genes (Zeng et al., 2017). The gene expression profiles of cells recovered from young and old mice are similar at the transcriptomic level, suggesting that aging is not associated with β-cell dysfunction in mice (Xin et al., 2016b). However, in humans, aging is associated with increased transcriptional noise and cell identity drift (Enge et al., 2017).

In summary, scRNA-seq has been used to systematically investigate the development of pancreatic lineages at various stages in mice, and these findings may provide insights to improve and evaluate the efficiency of hESC induction. Nonetheless, we must be cautious when extrapolating these conclusions to human pancreatic development. Thus, it is important to resolve the developmental processes of human pancreatic lineages, as has been done in mouse.

Stimulating proliferation of β cells is a straightforward approach for the endogenous regeneration of β cells for diabetes therapy. β Cells undergo a burst in proliferation during the early postnatal period, followed by a decrease in proliferation and functional maturation (Finegood et al., 1995; Meier et al., 2008; Gregg et al., 2012). Proliferation of adult β cells is relatively slow (Teta et al., 2005), but it is believed that a burst of proliferation can occur under certain physiological conditions, such as obesity and pregnancy, to increase insulin secretion and meet high metabolic demand (Klöppel et al., 1985; Parsons et al., 1992; Sorenson and Brelje, 1997; Rahier et al., 2008). These findings indicate that at least a portion of β cells possess the flexibility to adapt to the changing microenvironment to maintain homeostasis. Therefore, exploiting the heterogeneity in proliferation and function of β cells may inform the development of therapeutic approaches that specifically and controllably stimulate β-cell proliferation. Research based on single-cell technologies has extended our knowledge on proliferation and heterogeneity of endocrine cells, especially β cells.

β-Cell proliferation primarily occurs during the perinatal period, when β cells are in an immature state. Hence, comparison of immature and mature β cells is expected to provide clues regarding the molecular mechanism of proliferation, including the key regulatory genes and signaling pathways. At the single-cell transcriptomic (Qiu et al., 2017; Zeng et al., 2017) and proteomic levels (Wang et al., 2016a), β-cell proliferation rate declines with age, both in mice or humans. In mice, immature β cells display high expression of the cell cycle inhibitor p57 (Cdkn1c), distinct from the p18 (Cdkn2c) expression observed in mature β cells. This finding suggests distinct regulatory strategies of cell cycle inhibition (Qiu et al., 2017). In addition, amino acid metabolism and mitochondrial reactive oxygen species (ROS) play important roles in promoting postnatal β-cell proliferation in mice (Zeng et al., 2017). The supplementation of amino acids to cultured islets from postnatal day (P)28 mice enhances β-cell proliferation, and inducible overexpression of radical scavenger catalase to clear ROS in the mitochondria of β cells reduces the proliferation rate and the mass of β cells (Zeng et al., 2017). In addition, single-cell mass cytometry analysis, profiling 24 proteins which are associated with β-cell proliferation, function and heterogeneity, has identified three subgroups of human β cells, two of which contain proliferating β cells (Wang et al., 2016a).

β-Cell heterogeneity has been reported by different groups; however, the classification of ‘heterogeneity’ is inconsistent between various studies (reviewed by Nasteska and Hodson, 2018; Mawla and Huising, 2019; Wang and Kaestner, 2019). Antibodies against ST8SIA1 and CD9 can classify human β cells into four antigenic subpopulations with different gene expression profiles and GSIS function (Dorrell et al., 2016). Several scRNA-seq analyses have revealed heterogeneity of human β cells in maturation and in response to metabolic stress (Baron et al., 2016; Muraro et al., 2016; Segerstolpe et al., 2016; Xin et al., 2018). Healthy human β cells exhibit heterogeneity reflecting distinct cell states with variable insulin gene expression and unfolded protein response (UPR) activation. One of the cell states (with low insulin but high UPR-related gene expression) is enriched in proliferative cells, and represents cells that have recovered from stress (Xin et al., 2018). This result is consistent with previous studies that reveal the role of UPR in response to metabolic stress to compensate β-cell failure and trigger proliferation to adapt to cellular environment changes (Rabhi et al., 2014; Sharma et al., 2015). Notably, some studies have not detected heterogeneity in human β cells at the transcriptomic level (Li et al., 2016; Wang et al., 2016b; Xin et al., 2016a; Lawlor et al., 2017), which may be partly because of the sensitivity and throughput of scRNA-seq technology or sample variation between individuals (Table 3). Using single-molecule fluorescence in situ hybridization (smFISH) in intact pancreatic tissues of mouse, Farack and colleagues have identified a small subset of ‘extreme’ β cells that express a higher mRNA level of insulin and other secretory genes. These cells contain higher ribosomal RNA and proinsulin, but lower insulin protein, suggesting they may be basal secretors (Farack et al., 2019). In future investigations, stable and high-throughput approaches and larger sample sizes will be needed to examine β-cell heterogeneity in greater detail. Taken together, these studies show that β cells display heterogeneity in their ability to proliferate and enact functional responses to stress and metabolic demand.

Non-β pancreatic cells, including exocrine and endocrine cells, possess the plasticity to convert into insulin-secreting β-like cells under certain injury or pathological conditions (Xu et al., 2008; Solar et al., 2009; Thorel et al., 2010; Chera et al., 2014). In addition, non-β pancreatic cells can be reprogrammed into β-like cells with the intervention of key cell type-specific TFs (Zhou et al., 2008; Yang et al., 2011; Al-Hasani et al., 2013; Courtney et al., 2013; Li et al., 2014; Miyazaki et al., 2016; Chakravarthy et al., 2017; Druelle et al., 2017; Xiao et al., 2018) or signaling pathways (Klein et al., 2015b; Lemper et al., 2015; Zhang et al., 2016; Ben-Othman et al., 2017; Li et al., 2017) (Fig. 2, Table 4).

Although advancements of in vivo and in vitro cell fate conversion have been well summarized (reviewed by Aguayo-Mazzucato and Bonner-Weir, 2018; Zhou and Melton, 2018), the extent to which the obtained β-like cells resemble primary β cells and how the dynamic conversion process occurs remain unclear. It is unclear whether the endogenous conversion occurs through a process of dedifferentiation to a progenitor state followed by re-differentiation or through a direct switch between lineages. By leveraging single-cell technologies to describe the developmental and transitionary trajectory, researchers can comprehensively reveal the progression of rapid cell fate shifts and the resulting molecular features. Indeed, an increasing number of studies have utilized scRNA-seq to understand the reprogramming of non-β to β cells (Chakravarthy et al., 2017; Li et al., 2017; Furuyama et al., 2019).

Mouse α and δ cells are plastic and are able to convert into β-like cells under extreme β-cell ablation. This plasticity depends on age: δ cells and α cells tend to adopt a β-cell fate and thus contribute to the pool of insulin-producing cells in juvenile and adult mice, respectively (Thorel et al., 2010; Chera et al., 2014). Moreover, in vitro re-aggregated human α or PP cells are able to become glucose-responsive, insulin-producing cells following ectopic expression of PDX1 and MAFA (key TFs for β-cell development and maturation), and this plasticity is maintained in T2D α cells (Furuyama et al., 2019). Using scRNA-seq, the transcriptomic profiles and reprogramming status of converted α cells on the pseudotemporal transitionary trajectory to β cells have been resolved. These converted cells are INS and GCG bi-hormonal and are divided into three stages, in which the late state of converted cells exhibits higher expression of β cell-related genes and lower expression of α cell-related genes (Furuyama et al., 2019).

Chakravarthy and colleagues have delineated the mechanism of α-cell identity maintenance by DNMT1 and ARX, and provided evidence via lineage tracing and scRNA-seq in mice of α-cell conversion to β cells after Arx and Dnmt1 ablation (Chakravarthy et al., 2017). Notably, the converted α cells acquire functional features of native β cells (Chakravarthy et al., 2017). The conversion of α to β-like cells has also been reported to be regulated by anti-malaria drug artemether through activation of gamma-aminobutyric acid (GABA) signaling in zebrafish and rodent models, as well as in human islets (Li et al., 2017). Indeed, scRNA-seq showed downregulation of α cell-specific genes and upregulation of β cell-featured genes in artemether-treated α cells from primary human islets (Li et al., 2017). However, the effect of GABA signaling in stimulating α- to β-cell trans-differentiation is under debate, because although the expression of Arx was indeed decreased after artemether treatment, the conversion from α to β cells was not observed in the Gcg-CreER lineage tracing system or in vitro culture systems of mouse and human intact islets (Ackermann et al., 2018; van der Meulen et al., 2018).

Many studies have described the scRNA-seq profiles of human islets and identified the constituent cell types and their corresponding transcriptional signatures, including the less abundant δ, ε and PP cells (Baron et al., 2016; Li et al., 2016; Muraro et al., 2016; Segerstolpe et al., 2016; Wang et al., 2016b; Xin et al., 2016a; Lawlor et al., 2017; Dominguez Gutierrez et al., 2018). These datasets validated the previously known features of each cell type but have also discovered novel signatures that may play roles in cell identity and function (Box 2).

The heterogeneity and corresponding plasticity of ductal and acinar cells are worthy of further investigation because these cells are considered abundant sources for in vivo cell fate conversion to β cells. Two subpopulations of ductal cells have been identified that display inverse expression patterns of mucin 1 (MUC1, a ductal cell marker), and cystic fibrosis transmembrane conductance regulator (CFTR, a marker for secretory cells; Baron et al., 2016), which is consistent with the finding that CFTR is heterogeneously expressed in ducts (Burghardt et al., 2003). Notably, the MUC1^high- and CFTR^low-expressing population localizes at the ductal terminal, whereas the inverse population is localized at the region connecting acinar and ductal cells. Using the StemID algorithm, Grün and colleagues have described putative subpopulations of ductal cells with distinct potential to differentiate into endocrine and acinar cells. However, the conclusion in this study needs to be carefully verified via other approaches such as lineage tracing (Grün et al., 2016).

Using a well-based Smart-seq2 method, Segerstolpe and colleagues have identified two clusters of acinar cells from humans: cluster 1, with elevated expression of inflammation-related genes, and cluster 2, which highly express secretory, digestive enzyme-encoding genes (Segerstolpe et al., 2016). However, Muraro and colleagues have identified four clusters of human acinar cells using a CEL-Seq2 method (Muraro et al., 2016). One cluster expresses regenerating family member 3 alpha (REG3A), which is involved in proliferation and differentiation of various cell types (and in tumorigenesis of pancreatic cells; Parikh et al., 2012; Xu et al., 2016), but displays lower expression of acinar marker genes, suggesting that this acinar subpopulation is in a less-mature state, with increased proliferative potential. Consistent with this finding, a subpopulation of acinar cells with proliferative activity has been identified in rodents; these cells are capable of long-term self-renewal and display high expression of stathmin 1 (STMN1), a cell division-related gene (Wollny et al., 2016). The distinction between subpopulations of human acinar cells in different studies might be caused by different scRNA-seq platforms, variation of human samples or the limited cell number. It should be noted that this proliferative, progenitor-like acinar population needs detailed investigation because they may contribute to the β-cell population under pathological conditions through trans-differentiation. Taken together, the findings from these studies suggest that understanding the heterogeneity of each pancreatic cell type will support selective and efficient cell fate conversion, in which the subpopulation with greater plasticity is the preferred target for therapy.

The use of scRNA-seq in future studies will aid efforts to re-examine the cell fate conversion process and resolve controversial conclusions from previous studies. scRNA-seq will capture rare converted intermediate cell states, describe their comprehensive dynamic gene expression profiles, enable mapping of the trajectory along cell state transitions, and determine vital regulatory drivers, either in vivo or in vitro. A deeper understanding will enable us to exploit cell heterogeneity and plasticity to optimize strategies for regeneration. In addition, to precisely model the conversion to insulin-secreting β-like cells, we must first understand the molecular features of each cell type in the pancreas and the corresponding dynamic changes that occur under various conditions at the single-cell level. Finally, decoding the transcriptional profiles and heterogeneity of non-β pancreatic cell types will aid establishment of a systematic framework to recapitulate the comprehensive relationship between cell lineages.

Regeneration of β cells from hESCs is a promising and sustainable approach for diabetes treatment and can provide an easily manipulated platform for exploring the mechanism of diabetes pathogenesis and for screening treatable targets. In order for findings to be clinically applicable, this approach should precisely mimic the in vivo process. In past decades, based on knowledge gained from animal models, significant breakthroughs have been made to create an in vitro inductive model that mirrors the in vivo developmental process to induce functional β cells. Although current protocols are able to produce glucose-responsive β-like cells with improved functions in vitro, such as increased GSIS capacity and relief of diabetes after transplantation, these cells are still not equivalent to mature β cells in vivo and require further maturation (D'Amour et al., 2006; Pagliuca et al., 2014; Rezania et al., 2014; Russ et al., 2015; Korytnikov and Nostro, 2016; Petersen et al., 2018; Sneddon et al., 2018; Nair et al., 2019; Sharon et al., 2019b; Velazco-Cruz et al., 2019). A direct comparison between primary and induced β cells is necessary to address the gap between the in vivo process and what we can currently recapitulate in vitro.

Researchers have explored the induced cell-type composition and modeled the dynamic differentiation process at the level of gene expression using single-cell qPCR (Petersen et al., 2017; Ramond et al., 2018) or scRNA-seq (Sharon et al., 2019b; Veres et al., 2019). Researchers have also compared the gene expression and GSIS ability of final induced cells with that of adult human islet cells (Petersen et al., 2017; Veres et al., 2019). Single-cell transcriptional profiling has identified distinct subpopulations and dynamic gene expression changes concomitant with the differentiation process and has shown that the majority of hESC-derived endocrine cells recapitulate the developmental trajectory toward human β cells but do not reach full maturity (Sharon et al., 2019b; Veres et al., 2019).

WNT signaling regulates endocrine cell development (Dessimoz et al., 2005; Pedersen and Heller, 2005; Heiser et al., 2006; Murtaugh, 2008). A recent study shows the inhibitory role of the WNT pathway during endocrine differentiation, verified by deletion of the WNT inhibitor APC in mice as well as by small molecule treatment in vitro (Sharon et al., 2019b). Moreover, in the hESC induction system, Sharon and colleagues have revealed two waves of endocrine differentiation: one tends to produce α cells at stage 4, and the other tends to generate β cells at stage 5 (Sharon et al., 2019b). Veres and colleagues have characterized the composition of cell populations and mapped the lineage trajectory during in vitro β-cell differentiation using high-throughput inDrops scRNA-seq (Veres et al., 2019). The stem-cell-derived islets (SC-islets) mainly contain four cell types: SC-β cells, α-like cells with INS expression, exocrine cells and a population that resembles enterochromaffin cells (SC-EC), which synthesize and secrete serotonin in the gut. The induced β cells express a fraction of maturity-related genes, such as IAPP and SIX2, but not others, such as UCN3, MAFA and SIX3. Pseudotemporal analysis shows that SC-β and SC-EC cells originate from a common NGN3⁺ state. The polyhormonal α-like cells are considered a transient cell state to generate monohormonal α cells. In addition, the induced β cells can be enriched by cell-surface marker CD49a, for re-aggregation to improve functions (Veres et al., 2019).

Taken together, these works provide a blueprint and reference for hESC differentiation at single-cell resolution, which is essential for the improvement of therapeutic strategies. However, comparison of the in vivo and in vitro developmental and maturation processes is limited by the scarcity and availability of primary human tissues at continuous developmental stages.

With the development of high-resolution techniques, we have entered a new era of understanding organogenesis at the single-cell level. In this Review, we have summarized the use of single-cell technologies, primarily scRNA-seq, to inform research in the field of β-cell generation and regeneration (Tables 2-4). These findings provide insights into strategies for β-cell regeneration from hESCs in vitro. scRNA-seq has been widely applied to pancreatic development to comprehensively describe the lineage hierarchy landscape in mice. However, it is necessary to determine whether the inferred developmental trajectory based on pseudotime analysis reflects the real developmental trajectories. These findings, based on single-cell analyses, provide insights into the strategies of β-cell regeneration from hESCs in vitro. In addition, scRNA-seq allows for the identification of pancreatic cell state and heterogeneity in adults, to better understand the proliferative ability and functional potential of certain cell types. Moreover, scRNA-seq studies are expected to extend our understanding of cell plasticity, as well as defining cell fate conversion pathways. However, the different conclusions drawn by different scRNA-seq platforms should be further validated through synergy with other approaches, such as microscopy and CRISPR editing. Overall, studies based on single-cell technologies have reconstituted the framework of pancreatic development and have laid the foundation for future mechanistic studies.

The combined application of new single-cell technologies (Table 1) will provide a clearer understanding of the pancreas from a multi-omics perspective. Although pancreatic cell identities have been defined by scRNA-seq, gene expression does not always indicate cellular function. Hence, protein expression patterns in individual cells can inform cell heterogeneity in function and plasticity. More importantly, during the processes of pancreas development and regeneration, the regulatory mechanisms of multistep cell fate choices and transitions need to be delineated from an epigenetic perspective. Therefore, single-cell proteomic and epigenetic studies will complement the framework of pancreatic lineage differentiation established by scRNA-seq, and inform regenerative medicine.

Decoding tissue architecture by characterizing the spatial transcriptome paired with microscopy and morphology is also important for understanding generation and regeneration of pancreatic endocrine cells. The location of distinct cell types will shed light on the function and phenotype of the cells, as neighboring cells may release differing signals to affect cell behaviors. Currently, several high-throughput spatially resolving approaches have been developed (summarized by Mayr et al., 2019); these approaches can deepen our understanding of stepwise regulation of cell fate determination in pancreas development.

In order to assemble functional islets in vitro, the developmental pathways and regulatory mechanisms of all endocrine lineages must be reconstructed. To characterize the features of rare δ, ε and PP cells, these cells must be enriched using gene reporter mouse strains or fluorescence-labeled antibodies at various time points. Findings from human samples will undoubtedly provide more valuable information for islet production. However, direct observations from human pancreas are limited because of sample scarcity. To maximize use of precious primary human tissues, unbiased and high-throughput methods should be applied to uncover the molecular features and regulatory mechanisms of human endocrine lineages.

Although not the focus of this Review, the implications of single-cell technologies to better understand the pathogenesis of diabetes have been summarized elsewhere (Carrano et al., 2017; Tritschler et al., 2017). However, many of the diabetes-associated genes identified in the scRNA-seq studies have not overlapped (Wang and Kaestner, 2019), which may be because of the pathogenic complexity of diabetes in individuals, the limited number of donor samples or the use of different technical platforms. To best use the datasets generated from single-cell technology, future studies need to rely on the development of more sensitive and higher-throughput experimental methodologies, and novel bioinformatics analytical tools, as well as larger sample sizes.

In conclusion, single-cell technology has extended our understanding of the complex processes of pancreas development and regeneration, and has provided novel insights for drug screening and diabetes treatment. In the near future, with the continuous development of single-cell technologies, we will have a more comprehensive and in-depth understanding of the processes and mechanisms of pancreatic endocrine development and regeneration, thus providing information for the production of human islet tissue in vitro.

We thank the members of the Xu laboratory for their advice and comments. We apologize that we were unable to cite many studies owing to space limitations.