Matters arising: Insufficient evidence that pancreatic β cells are derived from adult ductal Neurog3-expressing progenitors

SUMMARY

Understanding the origin of pancreatic β cells has profound implications for regenerative therapies in diabetes. For over a century, it was widely held that adult pancreatic duct cells act as endocrine progenitors, but lineage-tracing experiments challenged this dogma. Gribben et al. recently used two existing lineagetracing models and single-cell RNA sequencing to conclude that adult pancreatic ducts contain endocrine progenitors that differentiate to insulin-expressing β cells at a physiologically important rate. We now offer an alternative interpretation of these experiments. Our data indicate that the two Cre lines that were used directly label adult islet somatostatin-producing ∂ cells, which precludes their use to assess whether β cells originate from duct cells. Furthermore, many labeled ∂ cells, which have an elongated neuron-like shape, were likely misclassified as β cells because insulin-somatostatin coimmunolocalizations were not used. We conclude that most evidence so far indicates that endocrine and exocrine lineage borders are rarely crossed in the adult pancreas.

In brief

Magenheim et al. challenge a recent study claiming that pancreatic beta cells are generated at a physiologically important rate from progenitor cells within ducts. They show that the lineagetracing systems used in that study do in fact pulse label islet cells, precluding conclusions about duct cells giving rise to islet cells.

INTRODUCTION

The debate on the cellular origin of pancreatic islet β cells is almost as old as knowledge of the existence of pancreatic islets. The importance of the problem is obvious, as β cell deficiency is central to type 1 and type 2 diabetes, and there are hopes for a cure based on regenerative therapies targeting cells that have an innate ability to replenish β cell mass.

At the turn of the 20^th century, Laguesse and Bensley proposed that new islet endocrine cells arise during adult life from pancreatic ducts.^1,2 This notion subsequently led to the proposal that stem cells that reside in ducts, or else differentiated duct epithelial cells, have the capacity to acquire a β cell fate and migrate to form new islets or to join existing islets.^3–7 The hypothesis that duct cells give rise to β cells in the adult pancreas is grounded on several observations. First, there is unequivocal evidence that pancreatic duct cells and endocrine cells have a common embryonic origin. Both lineages arise from embryonic bipotent progenitors that form tubular duct-like structures and express many characteristic ductal markers, such as HNF1B, SOX9, HES1, and MUC1. During embryogenesis, a subset of these bipotent progenitor cells transiently activate the transcription factor NEUROGENIN3 (NEUROG3, also named NGN3), and subsequently form endocrine cells, whereas other bipotent progenitors give rise to duct cells.^8,9 Second, hormone-producing cells are often seen lining the adult ductal epithelium, and some such cells co-express duct and endocrine markers.^2,6,10–13 Third, pancreatic islets visibly aggregate in the vicinity of ducts within the three-dimensional structure of the mouse and human pancreas. Finally, experiments using a Neurog3 reporter revealed co-expression of duct cytokeratins and NEUROG3 after pancreatic injury and further demonstrated that cells that were purified with a Neurog3 reporter could form β cells in vitro and vivo.¹⁴

All of these observations provided a strong rationale to support a pancreatic duct origin of adult pancreatic β cells, although they did not offer direct proof for this lineage relationship. For nearly two decades, several groups have tested this hypothesis using genetic lineage-tracing experiments and largely failed to reveal a contribution of duct cells to endocrine cells across a broad range of conditions. The first studies used an insulin promoter inducible Cre transgenic reporter system to pulse-label β cells and showed that once β cells are formed, the β cell pool does not show a progressive reduction of labeling during postnatal life or after partial β cell ablation, indicating that it is not replenished by other cell types.^15,16 Similar conclusions were reached from chemical labeling studies¹⁷ or with independent genetic-tracing tools.^18–22 Overall, β cell labeling experiments argued against the existence of β cell neogenesis during physiological and various regenerative settings.

The conclusions from β cell lineage-tracing experiments were largely confirmed by lineage tracing of adult pancreatic duct cells. One study used a Carbonic Anhydrase II proximal promoter transgene to drive Cre in duct cells and observed profuse labeling of β cells and acinar cells in unchallenged mice or after pancreatic duct ligation, suggesting that duct cells continuously replenish β cells throughout postnatal life.²³ This report, however, was followed by genetic lineage-tracing studies from various laboratories using inducible duct Cre lines driven by Hnf1b, Muc1, Hes1, or Sox9 gene regulatory regions, all of which failed to show postnatal duct-to-b cell or acinar transitions in a broad range of physiological and injury settings, including prolonged pulse-chase experiments, alloxan-induced β cell ablation, or pancreatic duct ligation.^9,24–27 Arguing against the possibility that these Cre lines might have labeled most duct cells but selectively spared a progenitor subpopulation, tamoxifen treatment of the same Cre lines during embryonic development has readily marked embryonic duct cells that serve as bipotent progenitors with endocrine potential.^9,24–26 Thus, most duct lineage-tracing experiments consistently showed that adult duct cells do not make a significant contribution to β cells nor to other endocrine cells or to acinar cells. Collectively, lineage-tracing studies have supported the view that postnatal β cells arise from pre-existing β cells, as comprehensively reviewed.²⁸ A corollary is that hormone-producing cells that line ducts are most likely stable duct residents rather than neogenic endocrine cells derived dynamically from duct epithelium.

Against this backdrop, Gribben and colleagues have recently revisited the topic of β cell neogenesis with mouse lineage-tracing experiments, advanced microscopy, and single-cell transcriptomics.²⁹ These studies were performed with genetic lineagetracing tools used in previous studies yet arrived at different conclusions. Gribben et al. concluded that under homeostatic conditions there is a physiologically important population of NEUROG3-expressing cells within adult ducts that give rise to β cells and then migrate to islets. Nonetheless, in the response to this Matters Arising, the authors now acknowledge that the physiological relevance, if any, is unclear. Moreover, they reported that Akita mice—a strain with a point mutation in insulin causing β cell endoplasmic reticulum (ER) stress, apoptosis, and diabetes—exhibit enhanced β cell neogenesis from NEUROG3+ duct cells. An accompanying commentary concluded that the debate regarding the existence of duct stem cells and neogenesis should be over now.³⁰

The conclusions from Gribben et al. have the potential to transform the field of regenerative medicine in diabetes mellitus and thus merit close attention. We re-examined key data from this study, along with prior evidence, and obtained new experimental results. Using the same Cre lines as Gribben et al. combined with different reporters, we demonstrate that reagents used to label and chase duct cells and NEUROG3+ cells in fact directly label endocrine cells within adult islets. We argue that the direct pulse labeling of islet cells in these systems precludes the inference of duct-to-islet transition claimed by Gribben and colleagues.

RESULTS

Hnf1b-CreERT2 labels duct cells but also directly labels islet cells

A central experiment reported by Gribben et al. is the tamoxifen-inducible, duct-specific pulse-labeling of cells, followed by a chase period to demonstrate endocrine differentiation of labeled cells and their migration to islets. This was achieved using an existing Hnf1b-CreERT2 BAC (bacterial artificial chromosome) transgenic mouse model⁹ crossed to a Cre-excisable tdTomato reporter (Rosa-CAGLSL-tdTomato-WPRE).

The original study that used the same Hnf1b-CreERT2 line marked less than 50% of duct cells.⁹ Although duct cells express HNF1B uniformly, it remained possible that the transgene systematically failed to express Cre in a subset of duct cells that have a capacity to act as progenitors. This could hypothetically explain the discordance between the original Solar et al. study and the new findings from Gribben et al. A more recent report by Rovira et al., however, also used the same Hnf1b-CreERT2 transgenic with a more easily excisable Rosa26-TdRFP (tandem-dimer red fluorescent protein) reporter line and pulse-labeled up to 90% of duct cells after 24 h.³¹ This experiment with a more efficient reporter made a new observation: ~50% of islet-resident ∂ cells—a minor population of islet cells that produces somatostatin—were also directly labeled 24 h after tamoxifen treatment. Importantly, immunolocalization analysis of cell-type specific expression of CreER expression in the same Hnf1b-CreERT2 transgenic used by Gribben et al. shows that this is due to widespread expression of Cre in ∂ cells rather than to rapid duct-to-∂ cell transitions³¹(Figure 1A).

Figure 1. Lineage tracing with Hnf1b-CreERT2 mice.

(A) δ cells from Hnf1b-CreERT2 mice 24 h after tamoxifen treatment show nuclear Cre expression in somatostatin-positive cells (arrowheads). Dolichos Biflorus Agglutinin (DBA) is used to label duct cells, which also shows profuse Cre expression.

(B) Low β cell labeling by Hnf1b-CreER 6 weeks after tamoxifen in mice that underwent partial pancreatectomy or sham surgery. Top left shows schematic of the experimental protocol. PPX, partial pancreatecomy. Top right shows quantification of RFP+ β cells (sham n = 4, 2 males and 2 females; PPX n = 5, 2 males and 3 females; Student’s t test p > 0.4). Each dot represents quantification of RFP in 500–2,000 insulin+ cells from >40 islets per mouse. RFP labeling efficiency in CK19 + duct cells was 72%–89%. Bottom shows representative coimmunolocalization images for insulin, RFP, and DAPI. High magnification insets are used to highlight examples of insulin-negative RFP+ cells (arrowheads) with a characteristic rim surrounding the nucleus that lacks insulin immunoreactivity. Note an example of an islet from a sham mouse with an unusually large number of RFP+ cells, all of which are insulin-negative and located in the islet periphery, consistent with the location of somatostatin cells.

(C) Example of lineage-traced CK19 + RFP+ duct cells and an adjacent insulin+ β cell that is not RFP labeled and thus apparently not duct derived. Hnf1b-CreER mice were labeled as described in (B).

(D) Hnf1b-CreERT2 pulse labels a high fraction of somatostatin-expressing cells and a low fraction of β cells. Hnf1b-CreERT2/Rosa26-TdRFP mice were treated with Tamoxifen and analyzed after 72 h by co-immunofluorescence for RFP, insulin, somatostatin, and DAPI to unequivocally resolve the identity of RFP+ insulin or somatostatin-expressing cells. Separation of either insulin-RFP-DAPI or somatostatin-RFP-DAPI channels shows that many somatostatin+ cells, which typically have a smaller cytoplasm than insulin+ cells, are easily misclassified as insulin+ cells in insulin-RFP-DAPI stains if they are surrounded by insulin+ cells. White arrowheads illustrate somatostatin+ RFP+ cells and hollow arrowheads show other examples of somatostatin+ cells that appear to be insulin+ in the insulin-RFP- DAPI channel. These images also confirm that true RFP+ insulin+ cells are rare. RFP, insulin, somatostatin, and DAPI are represented in different colors in each sub-panel to facilitate visualization.

(E) Rare insulin+ tdTomato+ cells from Gribben et al. Shown is a representative example from 5 weeks post-tamoxifen, which was a representative image for average ~10% co-labeling of insulin+ cells in Gribben et al., Figure 1H. To re-quantify, we used the ImageJ Channels tool to remove the tdTomato channel, counted all DAPI-stained nuclei that are directly surrounded by insulin—therefore omitting those with an insulin-negative cytoplasmic rim—and marked them to define the number of insulin+ cells as a denominator. To create the numerator, we added the tdTomato channel and counted those with tdTomato fluorescence covering the nucleus and cytoplasm of insulin+ cells. This yielded 3 tdTomato+ insulin+ cells out of 112 insulin+ cells, which corresponds to a 2.7% co-labeling rate. This is an upper bound since insulin positive cells without DAPI nuclear staining or equivocal perinuclear insulin staining were not included in the denominator. The hollow arrowheads show tdTomato+ cells that have a visible insulin-negative cytoplasm, including one magnified example that is surrounded by insulin+ cells and is easily misclassified as an insulin+ cell. The white arrowheads mark three insulin+ tdTomato+ cells.

Error bars show standard error of the mean. Scale bars as specified in each panel.

In keeping with the expression of CreER in ∂ cells, Gribben et al. also showed that Hnf1b-CreERT2 induces excision and activation of the tdTomato reporter in ∂ cells. In their first analysis timepoint, one week after tamoxifen injection, they found labeling of ~10%–30% of ∂ cells, which increased to ~50%–70% at 5 weeks and then decreased to ~40% at 12 weeks (Figure 1I in Gribben et al.). The fact that a considerable proportion of islet endocrine cells was labeled at the pulse (i.e., in the first analysis that was performed shortly after tamoxifen treatment) makes it difficult to assign duct origins to labeled islet cells after a chase. The cells that were found labeled after the chase could have simply come from pulse-labeled islet cells. Gribben et al. stated that because of islet ∂ cell labeling, they cannot rule out a contribution of islet ∂ cells to the traced β cell population over time. Our interpretation of the data is that pulse labeling of islet cells precludes inferences about islet cell neogenesis from ducts.

Insufficient evidence that pulse labeling ducts leads to time-dependent labeling of β cells

Both Gribben et al. and previous studies used the same Hnf1b- CreERT2 tracer along with several Cre-activated reporters and observed that a proportion of β cells were pulse labeled after tamoxifen treatment.^9,31 Gribben et al. reported that 5%–15% of islet β cells were labeled 1 week after the pulse and then 10%–20% of β cells after a 12-week chase (Figure 1J in Gribben et al.). Rovira et al. reported a much lower rate, as only 2.34 ± 0.24% of β cells were pulse labeled 24 h after tamoxifen injection. Although the timing of analysis differed, Figure 1B now shows that differences in β cell labeling rates in both studies are unlikely due to differences in time elapsed since tamoxifen treatment because 6 weeks after tamoxifen treatment of the same Hnf1b-CreERT2 driver line, the percentage of labeled β cells remained low in mice that underwent partial pancreatectomy or sham intervention (2.91 ± 0.78%, n = 5, and 3.06 ± 0.77%, respectively; Student’s t test, p > 0.4).

Interestingly, we observed similar labeling rates (3.09 ± 0.65%, n = 4) in 281 insulin-positive β cells that were in direct contact with cytokeratin-stained duct cells (Figure 1C). These findings indicate that such cells are not neogenic β cells derived dynamically from duct epithelium but are instead endocrine cells that reside in this well-characterized location for endocrine cells.^10,11

In conclusion, our own findings as well as those of Gribben et al. show that Hnf1b-CreERT2 mice labels β cells, albeit with substantially different labeling rates.

The reasons for the discrepancy in β cell labeling rates are unclear given that all of these studies achieved very high labeling efficiencies in duct cells (Figure 1B).^29,31 We observed, however, that while the numbers differ, the representative images of β cell labeling provided by the two studies do not differ substantially. This suggested that the two studies may have counted β cells differently. One potential confounding factor is that somatostatin-producing ∂ cells have a distinct neuronlike elongated morphology.^32,33 Due to their thin cytoplasmic rim, ∂ cells surrounded by β cells can be misclassified as β cells when performing insulin immunofluorescence studies without simultaneous co-labeling of somatostatin, which is how β cells were quantified in the original report and in the response to this Matters Arising. We illustrate this by simultaneous co-labeling of insulin, somatostatin, RFP, and DAPI-stained nuclei from Hnf1b-CreERT2; Rosa26-TdRFP mice, which show how many somatostatin+ cells that are surrounded by insulin+ cells, are easily misclassified as insulin+ cells when omitting the somatostatin immunofluorescence channel (Figure 1D). This is particularly likely in a setting where most cells are insulin+, whereas most RFP+ islet cells express somatostatin (Figure 1D). In support of this possibility, we examined representative images from Gribben et al. that contain more than 100 insulin+ cells. We used the ImageJ channel function to segregate tdTomato, insulin, and DNA fluorescence channels and observed that <3% of insulin+ cells express tdTomato, whereas many other tdTomato+ cells that are surrounded by insulin+ cells exhibit a thin perinuclear cytoplasm that does not exhibit insulin immunoreactivity (Figure 1E). These findings indicate that in a setting of high ∂ cell labeling by Hnf1b-CreERT2, accurate quantitation of labeled β cells requires simultaneous coimmunolocalization of ∂ and β cells. We propose that the lack of such coimmunolocalization studies has most likely led to inaccurate estimates of β cell labeling rates by Gribben et al.

Regardless of the discordance in β cell labeling rates, ifa large number of β cells were truly labeled at the pulse as proposed by Gribben et al., this should hamper the ability to infer duct-to-β differentiation, which is the central conclusion of that study. The authors focused on the extent to which the fraction of labeled β cells changed with time in their experimental system. They found ~10% labeling at 1 week, a similar percentage at 5 weeks, and ~17% at 12 weeks, using 4 mice in each group and largely overlapping values across all three groups (Figure 1J in Gribben et al.). They compared different timepoints with Student’s t test, which yielded p = 0.034 when comparing 1 vs. 12 weeks. Given that three groups were analyzed, this required correcting for multiple comparisons, which was not performed. Doing so would have logically increased the borderline p value. Thus, this statistical analysis cannot be used to support a progressive increase in the fraction of labeled cells. Moreover, the high labeling rates in δ and β cells reported by Gribben et al. mean that even if the fraction of labeled β cells had truly increased after a chase, such cells could derive from pulse-labeled ducts, from pulse-labeled δ cells, or even from late Cre-based excision recombination in β cells caused by long-acting tamoxifen.³⁴ This data, therefore, cannot be used to conclude that duct cells or δ cells transition to β cells.

In summary, Hnf1b-CreERT2 pulse labels a large fraction of δ cells, which precludes the inference of ductal origins. Furthermore, this system pulse labels a low proportion of β cells, but this labeling rate does not increase throughout a chase period in our studies or in a statistically significant manner in the studies reported by Gribben et al. Thus, our own experiments as well as the reanalysis of data from Gribben et al. do not support the emergence of newly formed β cells from ducts after pulse labeling cells expressing the Hnf1b-CreERT2 transgene.

Neurog3-CreER labels differentiated islet cells

Equally crucial to the interpretation of the Gribben et al. study is the pulse-chase experiment using Neurog3-CreER transgenic mice. These experiments were used to infer the presence of NEUROG3+ cells within ducts, which after a chase period become β cells. Here again, we interpret the specificity of the cells labeled during the pulse as inconsistent with the conclusions. The study does not report the identity of labeled cells immediately after pulsing with tamoxifen, so it is unclear if they are endocrine cells or not. However, all images shown by Gribben et al. reveal profuse labeling of islet cells one month after tamoxifen injection, as well as scarce somatostatin or C-peptide+ cells in ducts (Figures 2B–2E in Gribben et al.). This raises the possibility that the Neurog3-CreER transgenic pulse-labeled differentiated endocrine cells reside in both ducts and islets, as predicted by previous observations of low NEUROG3 expression in adult pancreatic endocrine cells.^20,35 If this were correct, it would not be possible to conclude that any labeled endocrine cell was pulse labeled as a hormonenegative NEUROG3+ progenitor that resided in a duct.

Figure 2. CreER knocked into the Neurog3 locus pulse labels somatostatin+ islet cells but not ducts.

Three month-old Neurog3-CreER+/−;Rosa-LSL-YFP+/+ mice were injected subcutaneously with tamoxifen (8 mg in corn oil, 3 times over one week) and sacrificed 2 weeks after the last injection.

(A) Formalin-fixed paraffin-embedded sections were stained for GFP using a goat anti-GFP antibody (Abcam Ab6673, 1:500) and a secondary antibody conjugated to HRP (Horseradish peroxidase). Extensive staining is seen within peripheral islet cells, while ducts and acinar cells are not labeled. Scale bar, 50 micrometer.

(B) Co-immunostaining reveals that pulse-labeled islet cells are almost entirely somatostatin+ δ cells. Scale bar, 50 micrometer.

To test this hypothesis, we examined a knockin Neurog3-CreER mouse strain. The Neurog3-CreER mouse strain used by Gribben and colleagues is based on a randomly integrated transgene driven by a fragment of the Neurog3 gene promoter.³⁶ This transgenic Cre line sometimes elicits ectopic recombination in adult mice, depending on genetic background (Guoqiang Gu, unpublished observations), which prompted the generation of a new strain whereby CreER was knocked into the genomic Neurog3 locus (Neurog3-Cre^KI).³⁵ We used adult Neurog3-Cre^KI mice to pulse label cells expressing Neurog3 and observed that a considerable proportion of somatostatin+ δ cells within islets was labeled at 14 days (Figure 2), consistent with the observations of Gribben and colleagues. However, no recombination was seen in ducts (Figure 2). This argues strongly against a scenario of duct-resident NEUROG3+ progenitor cells that differentiate to islet cells in adult mice. Instead, the findings support the possibility that the Neurog3 gene is active in pancreatic islet δ cells, and thus Neurog3-CreER directly labels endocrine cells.

Gribben et al. show images with examples of duct-resident C-peptide+ cells that carry a TdTomato label after pulse labeling with Neurog3-CreER transgenics (Figures 2B and 2C). However, many Cre transgenic lines that are not intended to label β cells can label β cells at low frequency, and many studies have previously noted the existence of duct-resident endocrine cells.^10,11 The analysis that is provided does not rule out that those were rare duct-resident endocrine cells that were directly pulse labeled by the Neurog3-CreER transgenics.

Labeling of NEUROG3+ cells in Akita mice is tamoxifenindependent

Gribben et al. combined the Neurog3-CreER transgene with the Akita mutation, which causes severe diabetes and β cell death at an early age. The authors reported a significant increase in somatostatin+ cells in ducts, although this was based on small absolute numbers; an average number of ~3 cells were quantified in each wild-type mouse vs. ~8 cells per Akita mouse (Figure 3C in Gribben et al.—this has been corrected now to be the average number of cells per section). No significant changes were observed in insulin+ duct cells between both genotypes (Figure 3C in Gribben et al.). Likewise, pulse-chase experiments showed an increased number of labeled insulin+ or somatostatin+ cells in Akita mice without separating both endocrine cell types. If Neurog3-CreER directly pulse labels somatostatin+ cells, this means that the authors have pooled pulse labeled cells with putative chased cells in their analysis, which precludes drawing meaningful conclusions on lineage relationships. In the original publication, this increase was tamoxifen independent (endocrine cells from Neurog3-CreER mice were similarly labeled even without injection of tamoxifen). This has also been corrected in the response to this Matters Arising, although the pooled analysis of insulin+ or somatostatin+ cells remains insufficient to suggest that insulin+ cells are derived from duct progenitors.

Figure 3. scRNA-seq of mouse islets reveals Neurog3 and Hnf1b are not exclusive ductal markers.

(A and B) Combined scRNA-seq of islets extracted from pancreata of 8 individual non-treated mice at 3.5, 7, and 10–11 weeks of age (our primary data). (A) t-distributed stochastic neighbor embedding (t-SNE) representation of 14,502 high-quality cells obtained using 10x platform reveals the major endocrine cell types of the pancreas and several other islet-resident cell types. We present gene-expression profiles of select genes. (B) Dot representation of the expression level (represented by color) and proportion of expressing cells (represented by dot size) shows that Neurog3 is highly expressed in δ cells (detected in 43% of cells), whereas Hnf1b is detected in all endocrine cell types, albeit at lower levels and proportions compared with ductal cells (where it is detected in 54% of cells). (C–F) Re-analysis of the scRNA-seq data presented by Gribben et al. shows that in islets of Neurog3-CreERT2 traced mice, tdTomato is most abundant in δ cells. (C) Violin plots represent tdTomato transcript levels in endocrine cells. (D) Bar plots show the relative proportion of different cell types among the top 365 tdTomato-expressing cells (top) and among all other, tdTomato-negative cells (bottom). (E) Numerical representation of the abundance of the top 365 tdTomato-expressing cells and the top 227 Ngn3-expressing cells out of all cells, according to the different cell types identified. (F) Pseudotime analysis of “cherry-picked” top tdTomato+ cells (350 cells total) and top Neurog3+ cells (371 cells total), based on the data from Gribben et al. See also Figure S1.

Gribben et al. also examined Akita mice with germ-line heterozygous Neurog3 null mutations and found that they were more severely hyperglycemic than Akita mice proficient for Neurog3 (Figure 3I in Gribben et al.). This finding can be interpreted in several ways. For example, it is plausible that germ-line Neurog3 haploinsufficiency causes subtle impairment of β cell mass or function,³⁷ both of which could have a small influence on the severity of hyperglycemia in Akita mice. This, however, does not argue for the presence of neogenesis from adult ducts.

Neurog3 is expressed in δ cells

While definitive demonstration of neogenesis must rely on lineage tracing, the endogenous expression pattern of key markers is important to interpret findings. Gribben et al. addressed this by single-cell RNA sequencing (scRNA-seq) of purified mouse islet cells. We re-analyzed their data and also generated independent mouse islet scRNA-seq datasets.

We isolated islets from three different mouse strains and obtained 14,502 cells for analysis (see STAR Methods). Our data showed that Neurog3 mRNA is overwhelmingly expressed in δ cells within pancreatic islet cells in a much higher proportion compared to alpha and β cells and only in scattered cells that cluster with contaminating duct cells (Figures 3A and 3B). These results align with the Neurog3-CreER studies reported in Figure 2, which showed direct labeling of δ cells, but no labeling of duct cells or other major endocrine cell types. Thus, scRNA-seq analysis shows that in the adult pancreas the Neurog3 gene is active in endocrine cells, most notably in δ cells. The use of isolated islets, with only few residual ductal cells, evades the question of whether hormone-producing Neurog3+ cells originate from duct cells. However, this analysis provides further demonstration that in the mouse pancreas, Neurog3 is largely expressed in endocrine cells.

Fate tracking based on scRNA-seq

We next assessed whether scRNA analysis from tamoxifen-treated Neurog3-CreERT mice supported the existence of transitions from islet Neurog3+ cells into β cells. To this end, we re-analyzed scRNA-seq data from islet cells reported by Gribben et al., which were isolated from Neurog3-CreER R26-CAG-tdTomato mice, 10 days after tamoxifen administration. This design lacks data from a time immediately after the pulse and therefore rests on the untested assumption that the Neurog3-CreERT tracer transgene exclusively marks Neurog3+ cells, and tdTomato+ cells are offspring of these cells. Gribben et al. analyzed this data and concluded that Neurog3+ cells and tdTomato+ cells were distinct cell types, while tdTomato+ cells were primarily found among β cells. This led them to propose that Neurog3+ cells give rise to δ cells, which in turn give rise to β cells. Contrary to Gribben’s claim, our analysis of the data reported by Gribben et al. shows that tdTomato reporter expression predominates in δ cells (Figures 3C–3E). Thus, tdTomato was detected in 28% of δ cells vs. 0.7% of β cells. We propose that this discrepancy results from misclassification of β cells in Gribben’s report. Figure 4D and Figure 4G in Gribben et al. show that tdTomato+ putative β cells have low insulin mRNA expression and often express other hormones including somatostatin, potentially indicative of doublets.

Likewise, Neurog3 was detected in 21.62% of δ cells vs. 0.2% of β cells (Figure 3E). These findings, therefore, demonstrate that the Neurog3-CreER transgenic line used by Gribben et al. predominantly labels δ cells, consistent with our own Neurog3-CreER labeling studies and with the scRNA analysis of endogenous Neurog3 mRNA. The results also show that Neurog3+ and tdTomato+ cells were not distinct subpopulations. The data, therefore, failed to show that Neurog3+ cells—which are predominantly δ cells—give rise to β cells.

Next, we explored whether pseudotime analysis of the Neurog3-CreER R26-CAG-tdTomato islet scRNA-seq data from Gribben et al. could shed additional light on islet cell-type transitions. It should be noted that a pseudotime trajectory can be imposed on any group of cells, whether or not they share a temporal relationship, and that the starting point of the tree is arbitrary.^38,39 In the pseudotime analysis performed by Gribben et al., they assumed a developmental relationship between tdTomato+ and Neurog3+ cells, whereas all other cell types were excluded from the analysis, and Neurog3+ cells were arbitrarily chosen as the starting point. We used data from Gribben et al. to perform pseudotime analysis using the same cell types (tdTomato+, Neurog3+ cells) (Figure 3F). This analysis assigned predicted trajectories, as expected for cellular populations with similar transcriptomes, yet these could be equivocally interpreted as leading from β cells to δ cells, or conversely in the opposite direction, and lacked transition states to support the existence of bona-fide conversion processes (Figure 3F and S1J). This analysis ruled out a trajectory from Neurog3+ to δ cells, as proposed by Gribben et al., because Neurog3+ cells and δ cells were overwhelmingly the same cells. Thus, pseudotime analysis of scRNA data from isolated islets does not provide proof for the existence of a line of descent between Neurog3+ cells and β cells in adult mice. Instead, the analysis of islet single-cell transcriptomes suggests that in adult mouse pancreatic islets, Neurog3 is predominantly expressed in δ cells.

The progressive dilution of directly labeled β cells is not supported by a substantial number of analogous lineage-tracing studies

Gribben et al. propose that every week 0.68% of β cells are derived from ducts, and support their model with data suggesting a reduction in labeled β cells over time, after using Insulin-CreER to pulse label differentiated β cells. The authors provide graphs showing ~72%–86% labeling at 1 month and ~59%–76% at 6 months, which was nominally significant (p = 0.046, n = 4 mice per group, Student’s t test, Figure 1C in Gribben et al.) Such findings could theoretically indicate a contribution to β cells from a non-β cell source, or alternatively reflect experimental variation in tamoxifen-induced labeling. An extensive set of experiments from multiple labs have already used orthogonal insulin gene-driven Cre lines or chemical labeling strategies to test the same hypothesis and failed to detect a dilution of the fraction of labeled β cells.^15,17–22 For example, two very recent studies labeled nearly 100% of β cells at embryonic, postnatal, or adult stages and found no dilution of β cell labeling after 6 to 10 month periods.^21,22 One of these studies used dual fate-mapping of β cells and non-β cells (including >97% of duct cells) and failed to see any traces of non-β cell marks in β cells after 6 months.²¹ Therefore, we conclude that although the β cell labeling data provided by Gribben et al. could suggest the existence of a non-β cell source of β cells if it is interpreted in isolation, the evidence is insufficient to question a large body of experiments that used more robust models.

DISCUSSION

Our interpretation of the experiments reported by Gribben et al. does not support the conclusions that ductal Ngn3-expressing progenitors contribute to adult β cell neogenesis in the pancreas and that neogenesis is a physiologically important mechanism that is enhanced in diabetes. Instead, we submit that a large of body of evidence demonstrates that undifferentiated endocrine progenitor cells, if they do exist, do not make a quantifiable, physiologically relevant contribution to the adult β cell pool. Further, recent fate-mapping experiments that achieved nearcomplete labeling of pancreatic cell lineages, including a study that selectively labeled islet δ cells (as well as γ, α, and β cells),²² did not detect interconversions between endocrine cell types during normal postnatal life. On the other hand, various stressors and injuries loosen endocrine cell identity and may infrequently lead to partial or complete reprogramming to another type of islet cell under specific circumstances, although never to an exocrine cell.^21,40,41 Existing evidence does indicate that acinar cells may interconvert to duct cells (acinar-to-ductal metaplasia), although exocrine acinar and duct cells do not easily acquire endocrine identity unless forced by ectopic expression of master genes. Collectively, the weight of current evidence supports that homeostatic maintenance of pancreatic endocrine cells, and most likely other pancreatic epithelial lineages, is primarily dependent on proliferation of differentiated cells.

Limitations of the study

Although the weight of evidence indicates that neogenesis from ducts does not occur in the adult pancreas, there could be scenarios, not tested so far, that do recapitulate embryonic development and trigger NEUROG3 expression in ducts followed by endocrine differentiation. In addition, lineage tracing in humans has not yet been reported, making it theoretically possible that neogenesis does persist to some extent in the adult human pancreas.

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Yuval Dor (yuvald@ekmd.huji.ac.il).

Materials availability

This study did not generate new unique reagents.

METHOD DETAILS

Mice

Mouse strains used in this study were: Ngn3-CreER,³⁵ ROSA-LSL-YFP,⁴² Hnf1b-CreERT2⁹ and Rosa26-LSL-RFP.⁴³ All experiments were approved by the joint ethics committee (IACUC) of the Hebrew University and Hadassah Medical Center, or the Institutional Animal Care Committee of the University of Barcelona. Genotyping of Hnf1b-CreERT2 and Rosa26-LSL-RFP mice was performed as described.^31,44

For lineage tracing experiments using the Hnf1b-CreERT2 model tamoxifen was given by gavage to adult mice in three doses (20 mg, 20 mg, and 10 mg) every other day over 1 week, to 8– to 12-week-old mice, and 70% partial pancreatectomy was performed 3 weeks after the last tamoxifen dose. Lineage tracing analysis was performed 3 weeks after surgery. For pulse tamoxifen treatment analysis, one dose of 20 mg tamoxifen was given by gavage to adult mice and analysis was perform 24–72 h later. For lineage tracing of Ngn3-CreER mice, tamoxifen (Sigma-Aldrich) was dissolved in corn oil to 20 mg/mL and injected subcutaneously over 10 days in 5 doses of 8 mg each. In all experiments we used both males and females, and did not observe differences between the sexes.

Immunostaining

Immunostaining of Ngn3-CreER mice was performed on 5 micron thick paraffin sections of isolated pancreata. Tissue was fixed in 4% buffered formalin for 2 h at room temperature, dehydrated in an ethanol series, cleared in histoclear and embedded in Paraplast (Kendall). When required, antigen retrieval was performed using a pressure cooker. Primary antibodies included: mouse anti-somatostatin (1:400; from Beta Cell Biology Consortium, #1985), goat anti-GFP antibody (1:500; Abcam, Ab6673). For DNA counterstain we used DAPI (Roche). Fluorescent secondary antibodies were from Jackson immunoresearch, used at 1:200. For double staining we used only affinity-purified secondary antibodies, suitable for multiple labeling immunofluorescence. For immunohistochemistry we used a secondary antibody conjugated to HRP (ExtrAvidin–Peroxidase, E2886, Sigma). Images were captured on a Nikon C1 confocal microscope or Olympus FV10-ASW.

Hnf1b-CreERT2 studies were performed as described.⁴⁵ Stainings were performed on 4-micron thick paraffin sections of isolated adult mouse pancreata. Tissue was fixed in 4% paraformaldehyde overnight at 4C, dehydrated in an ethanol series, cleared in xylol, and embedded in paraffin. With the exception of Cre antibody antigen retrieval was performed using pH6 citrate buffer. Primary antibodies included: Cre recombinase (Sigma-Aldrich 69,050, 1:1000), RFP (Rockland #600-401-379S, 1:500), Insulin (Linco #4011-01F, 1:1000) and Somatostatin (Santa Cruz sc-7819, 1:50). DAPI (1:1000, Vector) was used for DNA staining. Labeling efficiency was determined by dual immunofluorescence for RFP and CK19 (Hybridoma Bank #Troma-III, 1:50). Fluorescent secondary antibodies were from Jackson Immunoresearch, used at 1:200–300. Duct were also labeled using Fluorescei-labeled Dolichos Biflorus Agglutinin (DBA) (Vector FL-1031, dilution 1:200). Images were captured on a Leica SP5 confocal microscope. Co-localization analysis and quantification was performed with ImageJ.

To quantify RFP+ cells in each pancreatic lineage, we first we examined insulin-DAPI or somatostatin-DAPI channels to quantify all DAPI-stained nuclei that are directly surrounded by either insulin or somatostatin — thereby omitting those with an insulin-negative or somatostatin-negative cytoplasmic rim, respectively. These were marked to define the number of insulin+ or somatostatin+ cells as a denominator. To create the numerator we added the tdTomato or TdRFP channel and counted those with tdTomato/RFP fluorescence covering the nucleus and cytoplasm of hormone+ cells. We analyzed sections separated by five 4-micron sections, and quantified RFP in 500–2000 insulin+ cells per mouse. This was done manually with ImageJ, by a scientist who was blinded to the identity of sample. Student’s T-test was performed for comparisons between groups of four and five mice.

Reanalysis of images from Gribben et al.

To reanalyze and quantify co-labeling percentage we selected images that contain >100 insulin+ cells, and used the ImageJ channel tool to segregate channels. In short, the red channel (tdTomato) was first removed to count all DAPI+ nuclei that were completely surrounded by insulin staining, therefore omitting those with an insulin-negative cytoplasmic rim and mark them to identify and count insulin positive cells. Then we included the red channel again and counted tdTomato fluorescence covering the nucleus and cytoplasm of those insulin+ cells.

Single cell RNA sequencing analysis

Analysis of new single cell RNA sequencing dataset

Sample preparation.

Islets were obtained from five 10–11 weeks old NOD-SCID (non-obese diabetic, severe combined immunodeficiency) mice, and from four mice carrying either a Cre recombinase without an excisable allele (three 3.5–7 week old Tg(Pdx1-Cre)6Tuv mice), or a LoxP allele without a Cre recombinase (one 7 week old Haster loxP/loxP mouse).⁴⁶ This choice ensured that results were not specific to any strain or experimental conditions. Islets were extracted through collagenase perfusion of the pancreas, and dissociated into single cells using TrypLE either immediately after a short (less than 3 h, for NOD-SCID mice) or long (24–48 h, for the other specimens) culture, and dissociated with Accutase as described.⁴⁶ In samples where dying cells were removed, they were first marked with propidium iodide staining, and the live cells were FACS-purified. Single cell sequencing libraries were then prepared using 10X Chromium kits (all samples were processed using 3′ V3 or V3.1, except for two of the NOD-SCID samples which were processed using the 5′ library and gel kit).

Sequencing and analysis.

10x-generated libraries from NOD-SCID mice were sequenced using Illumina NovaSeq6000 system. The sequenced reads were aligned and processed using 10X’s dedicated Cell Ranger software, including removal of low-quality cells. Further analysis was performed using BioTuring’s BBrowser 3.0, after combination with the previously published data from the Cre-bearing mice.^46,47 In this framework, cells with mitochondrial gene content higher than 15%, as well as cells with fewer than 1,500 features, were removed. Altogether 14,502 cells passed our rigorous QC (quality control) parameters. To ensure that cells that share features from different lineages are not lost, no attempt was made to actively remove doublets from the analysis. We did not apply additional filters and labeled potential doublets as such, rather than remove them. Canonical correlation analysis (CCA) was used to remove batch effects between the samples, the top 2,000 variable genes were used for dimensionality reduction using principle component analysis (PCA), and t-distributed stochastic neighbor embedding (t-SNE) used for visualization.

Reanalysis of Gribben et al. single cell RNA sequencing data

Generation of single cell expression matrix.

The 10X Genomics Single Cell software Suite, Cell Ranger 6.1.2 was used to process the scRNaseq data. Cell Ranger Count aligned reads to the mm10 reference genome via STAR.⁴⁸ To enable mapping of tdTomato-derived transcripts, the sequence for tdTomato was added to the reference genome and GTF file. Aligned sequence reads with valid cell barcodes mapping to exons (Gencode GTF release M23) were used to generate the gene expression matrix. scRNaseq Quality Control and Cell Filtering: Using the Bioconductor package Seurat 4.0.2⁴⁹ low quality cells were removed. We discarded cells with high mitochondrial read proportion (<10%), low cell features (<200) and high expressed gene transcripts (≥ 99.5%). Genes present in less than 3 cells were also removed. Finally, log 10 transformed number of cell features were plotted against log 10 transformed number of total reads and cells that substantially diverged from the regression line showing less features than expected were removed.

Data processing for t-SNE representation.

Following Cell Ranger’s initial alignment and QC benchmarks, data was analyzed using BioTuring’s BBrowser 3.0. In this framework, we added further QC parameters and removed cells with mitochondrial gene content higher than 15%, as well as cells with fewer than 200 genes. Canonical correlation analysis (CCA) was used to remove batch effects between the samples, the top 2,000 variable genes were used for dimensionality reduction using PCA, and t-distributed stochastic neighbor embedding (t-SNE) used for visualization. Following initial examination of the data, we noticed that poor-quality cells with low counts were still present in the dataset, and so cells with fewer than 3,000 counts were removed, and the remaining 14,457 cells re-clustered and analyzed.

Data processing for pseudotime trajectory analysis.

1) Dimensionality Reduction, Uniform mani-fold Approximation and Projection (UMAP) Visualization and Clustering Analysis:

The Bioconductor package Seurat 4.0.2⁴⁹ was used for downstream analysis. Normalization was performed using ‘regularized negative binomial regression’ and the data scaled based upon the top 5000 variable genes. Dimensionality reduction was performed via principal-component analysis (PCA) to enable cell clustering. The top 17 principal components were selected and passed to Uniform Mani-fold Approximation and Projection (UMAP) for clustering visualization.⁵⁰ The clustering analysis was performed with a resolution value of 0.4 and cluster number validated by scree plot.

2) Differential Expression Analysis and Cluster Identification:

Differentially expressed genes between clusters were identified by comparing average intra-cluster expression values against expression in cells of all other clusters using the Seurat DeSeq2 workflow⁵¹ with standard parameters used. The top 10 differentially expressed genes per cluster were used for identification and compared to known marker genes⁵² for validation.

3) Pseudotime Trajectory Analysis:

Pseudotime trajectory analysis of Ngn3+ and tdTomato+ cells was performed using the Bioconductor package Monocle3 0.2.3.³⁸ In brief, raw counts for Ngn3+ and tdTomato+ cells were subsetted and pre-processed prior to dimensionality reduction via PCA. Cells were clustered to form a minimal spanning tree enabling cell ordering along their predicted trajectory as previously described.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.stem.2023.03.003.

Highlights.

• Hnf1b-CreERT2and Neurog3-CreER tracers pulse label adult somatostatin δ cells

• Pulse labeling of islet cells precludes inference of ductal origins in a chase

• δ cells are prone to misclassification as b cells in immunofluorescence analysis

• Weight of evidence indicates that islet cells are generated by simple duplication

Data and code availability

• scRNA-seq data have been deposited (Database: GEO, GSE225739) and are publicly available as of the date of publication, or were obtained from EMBL-EBI: E-MTAB-11477. Accession numbers are listed in the key resources table.

• This paper does not report original code.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Somatostatin	Beta Cell Biology Consortium	Cat# AB1985, RRID:AB_10014609
GFP	Abcam	ab6673, RRID:AB_305643
Alexa Fluor 488 Donkey Anti-Goat	Jackson ImmunoResearch	Cat# 705-545-147, RRID:AB_2336933
Cy3 Donkey Anti-Mouse	Jackson ImmunoResearch	Cat# 715-165-140, RRID:AB_2340812
Rabbit anti-RFP	Rockland	Cat# 600-401-379S, RRID:AB_11182807
Insulin	Dako (Agilent)	Cat# A0564, RRID:AB_10013624
Insulin	Linco (Millipore)	Cat# 4011-01F, RRID:AB_433703
Somatostatin	Santa Cruz Biotechnology	Cat# sc-7819, RRID:AB_2302603
Cytokeratin (CK)19	Hybridoma BankVector (DSHB)	Cat# TROMA-III, RRID:AB_2133570
Cre recombinase	Novagen	Cat# 69050-3, RRID:AB_2314229
Chemicals, peptides, and recombinant proteins
DAPI	Roche	Cat# 10236276001; CAS: 28718-90-3
ExtrAvidin–Peroxidase	Sigma	Sigma-Aldrich Cat# e2886, RRID:AB_2620165
DBA	Vector Laboratories	Cat# FL-1031, RRID:AB_2336394
Deposited data
scRNA seq data	This manuscript	Database: GEO (GSE225739) and database: EMBL-EBI (E-MTAB-11477)
Experimental models: Organisms/strains
Ngn3-CreER	Gu lab	Wang et al.35
ROSA-LSL-YFP	Jax	Srinivas et al.42
Hnf1b-CreERT2	Ferrer lab	Solar et al.9
Rosa26-LSL-RFP	Jax	Luche et al.43