Cycling alpha cells in regenerative drug-treated human pancreatic islets may serve as key beta cell progenitors

Summary

Diabetes results from an inadequate number of insulin-producing human beta cells. There is currently no clinically available effective means to restore beta cell mass in millions of people with diabetes. Although the DYRK1A inhibitors, either alone or in combination with GLP-1 receptor agonists (GLP-1) or transforming growth factor β (TGF-β) superfamily inhibitors (LY), induce beta cell replication and increase beta cell mass, the precise mechanisms of action remain elusive. Here we perform single-cell RNA sequencing on human pancreatic islets treated with a DYRK1A inhibitor, either alone or with GLP-1 or LY. We identify cycling alpha cells as the most responsive cells to DYRK1A inhibition. Lineage trajectory analyses suggest that cycling alpha cells may serve as precursor cells that transdifferentiate into beta cells. Collectively, in addition to enhancing expression of beta cell phenotypic genes in beta cells, our findings suggest that regenerative drugs may be targeting cycling alpha cells in human islets.

Graphical abstract

Highlights

• •All previously annotated cell types are present in regenerative drug-treated human islets
• •Cycling alpha cells uniquely fulfill characteristics for regenerative drug-responsive cells
• •Regenerative drugs increase beta cell phenotypic markers in cycling alpha cells
• •Conversion of cycling alpha cells to beta cells contributes to beta cell mass expansion

Diabetes results from an absolute or functional reduction in human pancreatic beta cells. Small-molecule DYRK1A inhibitors exemplified by harmine induce human beta cell proliferation and enhance human beta cell differentiation and function. Karakose et al. suggest an additional beneficial effect of harmine: alpha cell to beta cell transdifferentiation.

Introduction

Diabetes affects 537 million people globally.¹ While type 1 diabetes (T1D) and type 2 diabetes (T2D) differ in their etiology, they share an important common feature, a marked reduction in the number of functional, insulin-producing pancreatic beta cells, suggesting that beta cell replenishment would be of therapeutic value in both diseases. Indeed, strategies to expand the beta cell pool in T1D and T2D are currently in use or in development, including whole pancreas transplantation, pancreatic islet transplantation, and transplant of human stem cell-derived beta cells.² Although great progress in these approaches has been made, none is scalable to the millions of people with diabetes, due to high cost, scarcity of donor organs, and their intrinsic interventional nature. These considerations have prompted efforts to develop drugs that may be capable of increasing endogenous beta cell mass. To this end, progress has been made with the discovery of the small-molecule drugs that inhibit the enzyme, dual tyrosine-regulated kinase 1A (DYRK1A), when used alone,^3,4,5,6,7 or in combination with synergistic partners such as GLP-1 receptor agonists (GLP-1RAs)^8,9 and transforming growth factor β (TGF-β) superfamily inhibitors.¹⁰

Small-molecule inhibitors of DYRK1A, such as harmine, INDY, Leucettine, 5-Iodotubericidin (5-IT), GNF4877, and 2-2c, are able to induce human beta cell proliferation in vitro, at a rate (or labeling index) of ∼2% as assessed by Ki67, bromodeoxyuridine (BrdU), or 5-Ethynyl-2′-deoxyuridine (EdU).^{3,4,5,6,7,11,12,13} The addition of a GLP-1RA such as exendin-4, semaglutide, or others, or a TGF-β inhibitor such as LY364947 or GW788388, synergistically increases proliferation to 5%–8%.^8,9,10,11 Importantly, multiple studies have shown that harmine and other DYRK1A inhibitors also induce human alpha cell proliferation in vitro.^3,7,10,14 Recently, we have shown that human beta cell regeneration can also be induced in vivo. Specifically, continuous infusion of harmine and exendin-4 for three months into immunodeficient mice bearing small human islet grafts in the renal capsule resulted in a 300% (harmine alone) or 700% (harmine + exendin-4) increase in human beta cell mass.¹⁵ These remarkable increases in beta cell mass are accompanied by reversal of diabetes, as evident by a return to euglycemia and a marked increase in human insulin secretion. In addition, harmine alone or in combination with a GLPRA agonist reproducibly increases expression of key beta cell transcription factors (PDX1, NKX6.1, MAFA, MAFB, NeuroD1, SIX2, SIX3, etc) and phenotypic markers (GLUT2, PCSK1, GLP-1R, SLC30A8/ZnT8, ENTPD3, etc).^7,8 Curiously, in contrast to the in vitro setting in which the harmine-exendin-4 combination causes an increase in human beta and alpha cell proliferation, proliferation in engrafted human beta cells and alpha cells in vivo is modest (∼0.2%–0.5%) and cannot explain the 300%–700% increase in beta cell mass. Moreover, despite the dramatic increase in human beta cell mass, there is no accompanying increase in alpha cell mass.¹⁵ Collectively, these observations suggest that harmine alone or in combination with exendin-4 may lead to the generation of new beta cells through a mechanism unrelated to replication, such as transdifferentiation from non-beta cell types. Since alpha cells, along with beta cells, have been repeatedly shown to be activated by harmine and other DYRK1A inhibitors,^{3,7,8,9,10,14} since they are the most abundant human islet cell type, and since rodent and human alpha cells are capable of transdifferentiating into beta cells under certain circumstances, we suggest that alpha-to-beta cell transdifferentiation might explain the striking in vivo increase in beta cell mass and function occuring in response to harmine.¹⁵

Along these lines, others have previously proposed the existence of endogenous beta cell precursors, progenitors, or stem cells. For example, in the alpha-to-beta transdifferentiation paradigm, Thorel et al. showed that a near-complete loss of beta cells in the adult mice leads to conversion of alpha cells to beta cells.^16,17 However, Nir et al., reported that enhanced proliferation of surviving beta cells in mice after the induction of a milder, diphtheria toxin-induced beta cell death was the key mechanism for beta cell regeneration that was observed after the withdrawal of diphtheria toxin expression.¹⁸ Importantly, Furuyama et al. have suggested that a general plasticity exists among human islet endocrine cells, including alpha cells, which allows them to transdifferentiate from one endocrine cell type to another.¹⁹ Huising et al. also have identified a subset of urocortin-negative, Ins-high, Neurog3-high self-replenishing “virgin” beta cells that appear to be capable of conversion to mature beta cells.^20,21 Kushner et al. have suggested the existence of SOX9⁺/ARX⁺/GCG^+/− cells in the human pancreas that might serve as a renewable source of beta cells in people with T1D.²² Finally, it is also possible that beta cells can arise from other less well-characterized cell types of the pancreas, including murine acinar cells,^23,24 and, although debated in the field,²⁵ PDX1⁺/ALK3⁺/CAII⁻ ductal cells^26,27 in pancreas samples from humans with T1D and T2D.²⁸ Collectively, these studies suggest that a variety of normal beta cell precursors may exist in the human pancreas.

Unfortunately, to date, there are no small-molecule drugs that induce the generation of any human beta cell precursor in numbers that are clinically relevant. In this study, we suggest that a cluster of “cycling alpha cells,” identified through single-cell human islets transcriptomics, may serve as a beta cell regenerative reservoir and may be readily transdifferentiated into functional human beta cells upon treatment with beta cell regenerative drugs of the DYRK1A inhibitor class.

Results

Single-cell transcriptomic analysis of human islets

Human islets from four adult islet donors were treated with 10 μM harmine alone, 10 μM harmine together with 5 nM of the GLP-1RA GLP-1(7-36), or 10 μM harmine together with 3 μM of the TGF-β inhibitor, LY364947 (LY), for 96 h and subjected to single-cell transcriptomics (Figures 1A–1C and S1A). The four human islet preparations and four different drug treatments yielded 16 sets of single-cell RNA sequencing (scRNA-seq) data and comprised 109,881 well-curated single cells. We identified 21 unique cell type clusters and confirmed the presence of all previously identified cell types in human islets^29,30 (Figures 1A–1C and S1B). All annotated cell types were present in all four islet donors, minimizing residual batch effects driven by individual donors as data were integrated and batch-corrected using the “harmony” tool (Figure S1C). We also defined the cell-cycle stage for each cell type (Figure S1D). Finally, we also demonstrated the presence of hormone-producing cells by generating feature plots showing the expression level of islet cell type hormones, insulin (INS), glucagon (GCG), somatostatin (SST), and pancreatic polypeptide (PPY) (Figure S1E). We also provided a high-level overview or “road map” for our overall approach to clustering (Figure S2). Collectively, these findings indicate that the islet cells and subtypes identified here correspond closely to those identified in prior human islet scRNA-seq and single-nucleus RNA sequencing (snRNA-seq) datasets.^31,32

Figure 1.

Regenerative drug treatment on human islets increases the abundance of cycling alpha cell population

(A) Uniform Manifold Approximation and Projection (UMAP) plot showing clustering of single-cell RNA-seq profiles from 109,881 cells from the integrated dataset that includes islet preps from four human donors. Clusters are labeled based on cell type-specific marker gene expression.

(B) Dot plot showing average expression level and percent expressing cells of the top two upregulated genes in each cluster compared to all other clusters.

(C) Split UMAP plots showing single-cell RNA-seq profiles from each treatment group. See also Figure S1.

Dataset annotations are congruent with the HPAP dataset and the deep learning model, UniCell

To further benchmark our annotations with previously published datasets, we first projected our integrated dataset onto the Human Pancreas Analysis Program (HPAP) reference dataset from the HPAP PANC-DB database website using the publicly available Azimuth tool.^30,31,33 This provided high-confidence alignment of multiple cell populations (Figures 2A, 2B, and 2E) with high mapping and prediction scores (Figures 2C and 2D). Furthermore, we also predicted cell types in our integrated dataset using the unsupervised deep learning method called UniCell.³⁴ UniCell analysis confirmed our initial cell type annotations for endocrine, beta, alpha, ductal, acinar, and connective tissue cells (Figures S3A–S3F). Intriguingly, a small fraction of the cycling cells were annotated as connective tissue cells (Figure S3F).

Figure 2.

Concurrence of the dataset annotations with the conventional Azimuth database annotations

(A) UMAP plot showing predicted cell annotation from HPAP human pancreas reference dataset on the integrated dataset that includes islet preps from four human donors.

(B) Split UMAP plots depicting predicted cell annotations from HPAP human pancreas reference dataset on individual treatment groups.

(C) UMAP plot showing mapping score on label transfer.

(D) UMAP plot depicting predicted cell type score on label transfer.

(E) Dot plot showing normalized average expression level and percent expressing cells for selected marker genes in each cluster from (A). See also Figure S3.

Cycling alpha cells uniquely fulfill characteristics for the regenerative drug-responsive target cells

Among endocrine cell clusters, two beta (labeled as “Beta” and “Ins High Beta”), two alpha-beta, and two alpha cell clusters were identified (Figure 1A). Between the two alpha cell clusters, Alpha1 appears to represent more mature or differentiated alpha cells, showing higher expression of GCG, TM4SF4, CRYBA2, and CHGB genes (Figure S1B). Among the two alpha-beta clusters, Alpha-Beta1 selectively expressed alpha cell identity genes as well as INS and PCSK2. In contrast, the Alpha-Beta2 cluster expressed a broader group of alpha and beta cell identity genes (Figure S1B). With respect to cell-cycle activity, two distinct cell types showed high levels of proliferating cell markers such as MKI67, TOP2A, CENPF, and NUSAP1 expression (Figure S1B). Of these two cell types, “cycling alpha” cells expressed mainly alpha cell identity genes (GCG, CHGB, and CRYBA2) as well as the insulin granule-associated gene, SCG2 (Figure S1B), suggesting the presence of a proliferating endocrine cell cluster in all four donors. Importantly, “cycling alpha cells” are also present in the HPAP dataset³⁵ (Figures 2A and 2B).

Regenerative drug treatment leads to expansion of cycling alpha cells

We next compared the proportions of each cell type across samples (Figure 3A). To correct for the cell number variability in different donors, we assessed the proportion of each given cell type by normalizing it to the total cell number. As previously reported,^35,36 alpha cells represent the largest proportion of islet cells, followed by beta cells. Cycling alpha cells were the eighth largest population. Notably, treatment with human regenerative drugs consistently increased the percentage of cycling alpha cells as compared to control treatment (Figure 3A).

Figure 3.

Cycling alpha cell cluster is the only cluster that becomes more abundant with regenerative drug treatment

Cell-cycle regression places cycling alpha cells in the endocrine subset.

(A) Proportion of cells from each cell type, grouped by treatment. Integrated dataset that includes islet preps from four human donors was used.

(B) Differential abundance of islet cell types in different contrast groups, assessed by Milo statistical framework.

(C) UMAP plot depicting unsupervised clustering of single-cell RNA-seq profiles after cell-cycle regression. Cells are colored by initial annotation.

We also noted a possible increase in the percentage of other cell types, exemplified by acinar cells, in Figure 3A. To statistically quantify the changes in cell abundances, we used MiloR³⁷ to perform differential abundance testing by assigning cells to partially overlapping neighborhoods (Figure 3B). Among all cell types, the cycling alpha cluster demonstrated the strongest increase in abundance after drug treatment, underlining the impact of regenerative drug treatment on cycling alpha cells (Figure 3B). Taken together, these results indicate that a four-day treatment of human islets with beta cell regenerative drugs substantially increased the abundance of this previously described but understudied cycling alpha cell cluster.

As seen in Figures 1A and 1C, although cycling alpha cells express high levels of alpha cell markers, they cluster more closely to mesenchymal cells than to endocrine cells. This could potentially be explained by the effect of the cell-cycle phase rather than the cell type. To explore this possibility, we repeated the identical unsupervised clustering, but we first regressed the cell-cycle genes to eliminate their effect on clustering. This approach clearly showed that after cell-cycle gene regression, cycling alpha cells now clustered among the other endocrine cell populations and connected to alpha, beta, and alpha-beta cell clusters (Figure 3C).

Cycling alpha cells expand and increase beta cell identity marker expression following regenerative drug treatment

To better delineate the sub-populations and associated molecular signatures in the cycling alpha cluster, we next removed all cell types except those annotated as endocrine cells from the cell-cycle regressed integrated dataset and performed dimensional reduction and unsupervised clustering again to focus our analysis on the endocrine compartment (Figures S2, S4A, and S4B). This yielded twelve separate cell clusters at resolution 0.4, ten of which perfectly matched our previous cell identity annotations. The two new clusters reflected the division of Alpha1 and Alpha-Beta1 into three clusters (clusters 0, 2, and 10) and the sub-division of the cycling alpha cluster into two subclusters (clusters 6 and 11) with distinct molecular signatures. Remarkably, cluster 11 was separated from all endocrine clusters despite having been clustered together with other cycling alpha cells before cell-cycle regression and showing high expression of proliferation markers. Instead, cluster 11 was enriched for high levels of mesenchymal marker genes, such as COL1A1 and IGFBP5 among others (Figures S4C and S4D). Moreover, regenerative drug treatment led to attenuation or disappearance of this cluster in three of four islet donors when compared to DMSO treatment (Figure S5). Accordingly, due to its clear mesenchymal character, we excluded cluster 11 from the endocrine cell populations and repeated previous unsupervised clustering and analyses.

This new analysis (Figure S2) yielded eleven separate clusters at resolution 0.4, all of which, as expected, were endocrine cells (Figures 4A and 4B). Projection of cell type annotations on these endocrine cells indicated a clear overlap between our previous annotations and the new clustering (Figure 4B) with the exception of, as expected, the absence of a mesenchymal cell cluster in this dataset (Figures 4C, S6A, and S6B). We also verified that the abundance of cycling alpha cells was significantly higher with regenerative drug treatment vs. control treatment (Figure 4D). We next investigated differentially expressed cluster-specific genes and found that 878 genes were upregulated in the cycling alpha cluster (cluster 6) vs. the rest of the clusters (Table S1). Hallmark gene set enrichment analysis on these genes demonstrated that cell cycle-related pathways, exemplified by “E2F targets” and “G2M checkpoint,” were highly upregulated in cycling alpha cells (Figure 4E). In addition, we found that the dimerization partner, RB-like, E2F and multivulval class B (DREAM) complex transcription factors are highly activated in cycling alpha cells, based on the gene regulatory network analysis performed by single-cell regulatory network inference and clustering (pySCENIC) and on the enrichment of DREAM complex members gene expression in cycling alpha cells (Figure S7). Overall, in addition to the increase in cell abundance (Figure 4F), the most important observation was that while cells present in the cycling alpha cluster under basal conditions had a predominant alpha cell phenotype, cycling alpha cells appeared to acquire a beta-like phenotype in response to regenerative drug treatment (Figures 4F and 4G). Interestingly, cells in the beta cell cluster also upregulated beta cell marker genes with regenerative drug treatment (Figure S7C). To determine the degree of beta cell identity increase with drug treatment, we calculated the enrichment of beta cell signature genes (INS, IAPP, PDX1, MAFA, NKX6-1, PCSK1, RBP4) in each cell using the AddModuleScore function from the Seurat package. Our results indicate that the acquisition of beta cell identity is visually and quantitatively comparable in beta cells and cycling alpha cells treated with harmine. However, the scores for H + GLP-1 and H + LY conditions are higher in beta cells than in cycling alpha cells (Figure S7C). Importantly, the degree of beta cell identity acquisition does not reach significance in cycling alpha cells but does in beta cells, perhaps reflecting the far smaller number of cycling alpha cells vs. beta cells.

Figure 4.

Apparent conversion of cycling alpha cells into beta cells

(A) UMAP plot showing unsupervised clustering of single-cell RNA-seq profiles of only endocrine cells after the removal of subcluster 11.

(B) Same UMAP plot as in (A), depicting cell annotation projections.

(C) Dot plot showing average expression level and percent expressing cells of the top two upregulated genes in each cluster.

(D) Proportion of cells from each cell type, grouped by treatment. Note that cycling alpha cells are the only cell type with a significant increase after regenerative drug treatment. Significance was calculated by Dunnett’s test. ∗p < 0.005, ∗∗ p < 0.0001.

(E) Hallmark gene set enrichment analysis on the upregulated genes in cycling alpha Cells (cluster 6) vs. rest of the endocrine cell types.

(F) Split UMAP plots showing single-cell RNA-seq profiles from each treatment group. Arrows depict the cycling alpha cluster.

(G) A heatmap showing the expression level of selected alpha and beta cell marker genes under drug treatment in cycling alpha cells. See also Figures S4–S7.

Intrigued by this observation that cycling alpha cells acquire a beta-like phenotype in response to regenerative drug treatment (Figures 4F and 4G), we further sub-clustered the cycling alpha cells (cluster 6) from the integrated dataset shown in Figures 4A and S2 and demonstrated that it yielded a total of three subclusters at resolution 0.2 (Figure 5A). Exploring the cell-cycle phase of the cells in the cycling alpha cluster, we found that cells pertaining to all three cell-cycle phases were present (Figure 5B). Moreover, we assessed the cell identity annotations by assigning alpha cell, beta cell, and cycling module scores (Figure 5C). This demonstrated that while the alpha cell module score was high in all of these cells, cells with a higher alpha cell score were found in cluster 0, and cells with a high beta cell module score were restricted to cluster 1. We further confirmed cell identities by assessing differentially regulated genes in each cluster. This differential gene expression analysis confirmed that beta cells marker genes such as INS, IAPP, and RBP4 were upregulated in cluster 1 (Figure 5D and Table S2). Hallmark gene set enrichment analysis on the differentially upregulated cluster genes also revealed that cluster 2 had high enrichment for cell cycle-related terms while beta cell-related terms were highly enriched in cluster 1 (Figure 5E). In addition, the cycling module score was high in cluster 2 and partly in cluster 0, but not in cluster 1 (Figure 5C). We found that cluster 2, as well as part of cluster 0 with a high cycling module score, contained the fewest cells in the DMSO condition, and these two clusters expanded with each drug treatment, indicating that regenerative drug treatment induces an important expansion among cycling alpha cells (Figure 5F). Since cluster 1 had the highest beta cell module score and did not display proliferative capacity based on the differentially expressed genes as well as the cycling module score, we hypothesize that the cells present in other clusters may be capable of transdifferentiating into beta cells. Collectively, this analysis suggests that cycling alpha cells can be encouraged to acquire a beta cell phenotype with regenerative drug treatment.

Figure 5.

Cycling alpha cell cluster in the integrated dataset is composed of alpha, beta, and cycling cell subclusters

(A) UMAP plot showing unsupervised clustering of single-cell RNA-seq profiles of only cluster 6 (cycling cells and endocrine cells clustered closely together with cycling cells) present in the endocrine cell population.

(B) UMAP plot showing same cells as in (A), labeled based on cell-cycle phase.

(C) UMAP plots showing cells colored based on alpha, beta, and cycling module scores.

(D) Heatmap depicting top ten upregulated genes in each of the subclusters in the cycling alpha cells. These subclusters are color-coded and labeled on the x axis on top of the heatmap, and the gene names are listed on the y axis. Last three genes (GCG, TM4SF4, and CRYBA2) are manually included in the heatmap as alpha cell markers.

(E) Hallmark gene set enrichment analysis on subcluster 1 and subcluster 2 in the cycling endocrine cells.

(F) Split UMAP plots showing scRNA-seq profiles from each treatment group.

Trajectory inference and immunocytochemistry experiments suggest that cycling alpha cells have the potential to transdifferentiate into beta cells in response to regenerative drug treatment

To test the hypothesis that a subpopulation of cycling alpha cells may transdifferentiate into beta cells or other endocrine cells, we performed RNA velocity analysis to quantify cellular transitions using the integrated endocrine cell dataset from all four donors. To provide an interpretable graph-like map of the data, we used partition-based graph abstraction (PAGA). The results suggest that cycling alpha cells are indeed capable of differentiating into Alpha-Beta2-type cells, which in turn can give rise to beta cells (Figure 6A). Importantly, through coarse-grains cell-cell transition matrix onto the macro-state level, CellRank identified the cycling alpha cluster as the initial state (Figure 6B). In line with the velocity results, diffusion pseudotime analysis also predicted independently that cycling alpha cells can transit into alpha-beta cells and beta cells when cycling alpha cell were selected as the initial cluster (Figure 6B). These findings collectively suggest that a cycling Alpha-to-Alpha-Beta2-to beta cell transdifferentiation axis is present in human islets subjected to regenerative drug therapy. To gain a deeper understanding of the molecular pathways involved in these presumed transdifferentiation events, we also identified the genes that are expressed along the pseudotime axis (Figure 6C; Table S3). This supported the notion that cycling alpha cells (represented in brown) have the potential to serve as a reservoir to give rise to other endocrine cell types.

Figure 6.

Velocity and pseudotime analyses reveal cell fate determination after regenerative drug treatment on human islets

(A) Velocity analysis and partition-based graph abstraction (PAGA) on endocrine cells.

(B) Pseudotime trajectory analysis on endocrine cells. Note that cycling alpha cells are picked as root cells by the pyGPCAA algorithm.

(C) Heatmap depicting genes expressed along the pseudotime axis.

(D) Representative immunocytochemistry images on DMSO-, harmine-, or harmine + GLP-1-treated human islets. Seven independent islet donors were used for these experiments. Note that in harmine- and harmine + GLP-1-treated islets, C-peptide (green) and glucagon (red) are found in the same cell.

(E) Quantification of C-peptide+/glucagon+ cells (significance was calculated by one-way ANOVA). See also Figure S8. Scale bar: 5 μm.

Seeking experimental evidence that the postulated alpha-to-beta cell transdifferentiation can take place in the presence of regenerative drugs, we performed immunocytochemistry to identify and quantify double-positive C-peptide+/glucagon+ alpha-beta cells. These experiments indicated that occasional C-peptide+/glucagon+ alpha-beta cells are present under basal conditions. More importantly, these double-positive cells significantly increased in number following exposure to regenerative drugs (Figures 6D, 6E, and S8, Videos S1 and S2). Taken together, trajectory inference by RNA velocity, pseudo-temporal ordering, and immunocytochemistry studies strongly suggest that cycling alpha cells have the potential to transdifferentiate into beta cells in response to regenerative drug treatment. Further, the presence of alpha-beta cells under basal conditions and the significant increase in their abundance with regenerative drug treatment support the notion that the cycling Alpha-to-Alpha-Beta2-to-beta cell transdifferentiation axis in Figure 6A may underlie the harmine (+/−GLP-1)-induced beta cell mass increase.

Video S1. Video of the same z stacks shown in Figure S8, related to Figure 6D

Depiction of clear colocalization of C-peptide and glucagon within the same cell in human islets treated with harmine.

Video S2. Video of the same z stacks shown in Figure S8, related to Figure 6D

Depiction of clear colocalization of C-peptide and glucagon within the same cell in human islets treated with harmine + GLP-1.

Discussion

DYRK1A inhibitors are able to drive human beta cell proliferation in vitro and in vivo, an effect that can be further augmented by combined treatment with a number of GLP-1RAs, or TGF-β superfamily antagonists.^{3,4,5,6,7,8,9,10,38,39} For example, the DYRK1A inhibitor, harmine, increases beta cell mass in human islets transplanted into immunodeficient mice by 300% over three months, and the beta cell mass is further increased to 700% by the addition of the GLP-1RA, exendin-4 or exenatide.¹⁵ Harmine treatment also leads to enhanced beta cell function in vitro and in vivo, evidenced by enhanced glucose-stimulated insulin secretion, enhanced expression of key beta cell transcription factors and functional markers (e.g., PDX1, MAFA, NKX6.1, SLC2A2, and INS itself), and a rapid (days) return to euglycemia in diabetes models.^{3,7,8,9,10,15} Yet, one conundrum remains unexplained: in contrast to the remarkable 3%–8% increases in beta cell proliferation observed in cultured beta cells in vitro, as assessed by Ki67 or BrdU labeling, beta cell proliferation assessed using Ki67 labeling in transplanted human islets in vivo is modest (0.3%–0.6%) and cannot explain the 300%–700% increases in human beta cell mass that occur over three months in vivo.¹⁵ Here, we report that beta cell regenerative drug treatment of human islets leads to a substantial increase in the abundance of the islet cell subtype called “cycling alpha cells,” suggesting that transdifferentiation of these cycling alpha cells into beta cells is an important contributor to the dramatic increase in new beta cell mass in human islets in response to treatment with DYRK1A inhibitors.

Here, we report single-cell transcriptomics data for regenerative drug-treated human islets. From a data quality and analysis standpoint, the large datasets in this report, the islet cell subtypes, clusters, and annotations correlate closely with datasets from multiple prior studies,^{29,30,32,40,41,42,43,44,45,46} suggesting that cell-type identification is accurate and analyses are technically in line with prior studies. In particular, the “cycling alpha cell” cluster has been reported in multiple prior studies.^{30,36,40,47,48,49,50,51,52,53} However, the abundance of this cluster among other endocrine cells is small in the absence of regenerative drug treatment. Critically, deeper analysis of this cluster in human islets treated with vehicle, harmine alone, or GLP-1 or the TGF-β inhibitor LY364947 (Figures 5A–5D) supports the concept that subsets of these cycling alpha cells acquire beta cell characteristics. In addition, RNA velocity and pseudotime analysis of the cycling alpha cells points to a transdifferentiation event, converting cycling Alpha to alpha-beta cells, which then convert into beta cells (Figure 6A). Alpha-to-beta transdifferentiation is further corroborated by the significant changes observed in the transcriptomic profile of cycling alpha cells upon regenerative drug treatment (Figure 4G; Table S1). Furthermore, the most significantly upregulated genes in the cycling alpha cells were related to “E2F targets,” “Myc targets,” “G2M checkpoint,” and “mitotic spindle” pathways, all of which are consistent with cell-cycle initiation and harmine-mediated inhibition of the DREAM complex⁵⁴ (Figures 4E and S7). Finally, we demonstrate at the immunocytochemical level that C-peptide+/glucagon+ alpha-beta cells are rare but detectable in control DMSO-treated human islets and that their number increases with regenerative drug treatment (Figures 6D, 6E, and S8, Videos S1 and S2). Taken together, in addition to other mechanisms of beta cell expansion with regenerative drug treatment, such as beta cell proliferation and beta cell mass expansion, we propose here that another mechanism for the large and remarkable increase in human beta cell mass observed in response to these drugs in vivo is likely the conversion or transdifferentiation of cycling alpha cell precursors into beta cells.

Spontaneous alpha-to-beta cell transdifferentiation has not previously been reported to occur in response to any type of beta cell regenerative drug therapy. However, previous studies have reported alpha-to-beta cell transdifferentiation, as described in the introduction. As additional examples, Collombat et al. showed that overexpressing Pax4 in mouse alpha cells leads to beta cell conversion and reverses hyperglycemia in juvenile streptozotocin-treated mice.⁵⁵ In another study, Zhang et al. demonstrated that adenoviral delivery of Pax4 in an alpha cell line suppressed glucagon and induced insulin synthesis.⁵⁶ In a separate study, inactivation of Aristaless related homeobox (Arx) and Dnmt1 in adult mouse pancreatic alpha cells led to conversion of alpha cells to beta cells with the ability to secrete insulin upon glucose stimulation, although the insulin secretory capacity of these cells was less robust than that of mature beta cells.⁵⁷ Pharmacological attempts at alpha-to-beta cell conversion include artemisinins that stimulate γ-aminobutyric acid (GABA) signaling, translocating ARX to the cytoplasm, and promoting its degradation,^58,59 although these findings are controversial.^60,61 Xiao et al. demonstrated that adeno-associated virus delivery of Pdx1 and Mafa into the murine pancreatic duct or to human pseudoislets leads to the conversion of alpha to beta cells.⁶² Importantly, Thorel et al. showed that loss of beta cells leads to alpha-to-beta cell conversion in diabetic mice, confirmed using lineage tracing.^16,19 It has also been suggested to occur in fluorescence-activated cell sorting (FACS)-sorted human alpha cells transplanted into mouse models.¹⁹

While the possibility of human alpha-to-beta cell transdifferentiation has appeal from the current report as well as from prior mouse studies, its occurrence is impossible to unequivocally document at present because appropriate lineage tracing technologies do not yet exist for human alpha cells. More specifically, although in theory alpha cell lineage tracing tools used in mice could be transferred to humans, there is currently no reliable human alpha cell-specific promoter, an essential tool for such studies. Indeed, multiple human alpha cell promoters (e.g., glucagon, ARX, TM4SF4, GC) that are effective in mice fail to mark human alpha cells with efficiency or specificity. Similarly, FACS or magnetic bead sorting of human alpha cells can be performed, but yields of alpha cells are poor and they are impure, contaminated by other cell types. Thus, unequivocal demonstration of drug-induced human alpha-to-beta cell transdifferentiation remains to be proven and awaits rigorous human alpha cell tracing using lineage tracing methodologies.

In addition to alpha-to-beta conversion, additional cell type conversion scenarios appear to occur with regenerative drug treatment of islets. For example, conversion of the Alpha2 cluster to other alpha and beta cell types, or that of insulin-high beta cells to beta cells, or that of delta cells to beta cells all appear to occur, but as the cell abundances of these clusters did not show clear difference between drug treatment and DMSO, we focused on cycling alpha cells (Figure 6A). Thus, future studies will undoubtedly shed light on the lineage dynamics of the endocrine cell types in the presence of regenerative drugs. Yet, among these, cycling alpha cells were predicted to be the “root” cells from the transition model constructed by Generalized Perron cluster cluster analysis (GPCAA) using RNA expression and velocity data, and a cycling Alpha-to-Alpha-Beta2-to beta cell transdifferentiation axis is readily detectable (Figure 6B). Along the pseudotime trajectory, delta, gamma, and insulin-high beta cells appear to be in the very end, which may suggest that cycling alpha cells also have the potential to convert into these cells. Further studies are needed to confirm these findings in human islets.

Several prior authors, including ourselves, have suggested that DYRK1A inhibitors induce human alpha cells to proliferate in human islets in vitro.^3,4,5,6,7,39 In retrospect, taken together with our current findings, we suggest that DYRK1A inhibitors induce proliferation in the alpha cell precursors of eventual beta cells. This scenario would explain Ki67 or EdU labeling of alpha and beta cell populations that we and others have reported,^{7,8,9,15,52,54} as well as their conversion to beta cells with an ultimate expansion of beta cell, but not alpha cell numbers.¹⁵ Importantly, if this “alpha cell progenitor scenario” is correct, attempts to develop diabetes regenerative drug-targeting approaches directing drugs exclusively to mature beta cells may be misdirected.

Others have previously suggested the existence of additional putative human beta cell precursor cell types, as described in the introduction.^{20,21,22,23,24,26,27,28} How cycling alpha cells might compare in retrospect to virgin beta cells, SMAD7⁺, BMP7⁺, SOX9⁺, NGN3⁺, MAFA⁺, PDX1⁺, CK19⁺, or other putative beta cell progenitors described by others remains uncertain. None of the aforementioned markers is differentially expressed in the cycling alpha cell populations upon regenerative drug treatment (data not shown). Interestingly, we have also observed a small mesenchymal cell cluster (cluster 11 in Figures S4A and S5) that disappeared with regenerative drug treatment in three out of four islet donors. This may suggest that some mesenchymal cells may be able to change their phenotype into other cell types.

We were surprised not to observe a population of proliferating beta cells following DYRK1A inhibitor treatment, since we have observed this in almost all of our prior studies.^{7,8,9,10,11,15} It is possible that Ki67⁺ beta cells observed in islets by ourselves and others were bihormonal alpha-beta cells. It is also possible that the beta cells in the four human islet preparations used were less responsive than the average human islet donors, although the donor islets appeared to be of high quality. Although we consider this an unlikely scenario, in future studies, it will be important to assess Ki67 immunolabeling and assessment for bihormonal alpha-beta cells prior to single-cell experiments.

Limitations of the study

This study has important limitations, most of which reflect technical limitations in human islet research. First, we included only four human islet donors in this study, yet human islet preparations are highly variable in composition and quality.^63,64 On the other hand, given four different treatment conditions on four separate human islet donors, the overall dataset comprises 16 individual scRNA-seq experiments and yielded 109,881 high-quality single cells. In addition, we paid great attention to the selection of high-quality islets and included islets from donors similar in age, gender, and BMI (Figure S1A; Table S4). It is important to note that two of the four islet donors included in this study had higher levels of H1Abc: 5.7 and 5.9. Reassuringly, our cell type composition closely matched much larger donor sets from human islet scRNA-seq and snRNA-seq.^30,40 More importantly, the results were highly reproducible, with the key findings being observed in islets from each of the four individual donors (Figures S1C and S5).

A second limitation is that these studies were performed on human islets in vitro. Alpha and beta cell proliferation rates are substantially higher (∼2%–8% Ki67 or BrdU labeling) in vitro than they are in vivo (0.2%–0.6% Ki67 labeling), yet in vivo is where a 300%–700% increase in beta cell mass is observed.¹⁵ Thus, we believe that alpha cell transdifferentiation is even more important in regenerative drug-treated human islet grafts in vivo. Accordingly, performance of parallel studies in human islet grafts in vivo will be critical in the future.

A third limitation is that the study necessarily lacks human alpha cell lineage tracing confirmation, as noted earlier. This must be performed, if and when such technologies become available. Indeed, our current efforts focus on generating the tools required to perform alpha cell lineage tracing experiments. Along these same lines, in our experimental design we selected a single time point (96 h) for regenerative drug treatment. This makes it hard to predict the precise lineage trajectory of proliferating cycling alpha cells. Clearly, drug treatment studies at multiple time points (e.g., 3, 6, 9, and 12 days) may be of interest, but human islets survive and retain phenotype poorly over more than a few days. In this regard, Title et al. reported recently that reaggregated pseudoislets can be cultured and subjected to DYRK1A inhibitor treatment for up to 15 days.⁶⁵ These reaggregated pseudoislet models hold promise for longer-term human islet studies. Nonetheless, with the current absence of reliable alpha cell lineage tracing tools, such longer-term studies still cannot provide unequivocal proof of lineage of origin.

A fourth limitation is that we failed to identify differentially expressed genes in cycling alpha Cells that allow them to be distinguished from “ordinary” alpha cells. This absence of a “signature” gene or protein hinders the ability to identify or fractionate a pure population of cycling alpha cells for deeper study.

A fifth limitation is that while scRNA-seq studies evidently provide a clear insight into the transcriptomic profiles, a better understanding of transcriptional regulation and developmental lineage likely can be achieved through the study of the chromatin dynamics of the same cell types and subtypes identified in a given dataset. Although our results suggest that transcriptional regulators are important for the proliferative effects of DYRK1A inhibitors (Figure S7), these findings are computational and require experimental validation. A more complete and comprehensive understanding of the mechanisms that control transcriptional regulation will require multiome studies that combine scRNA-seq and scATAC-seq methodologies.

A sixth limitation or unexplained observation exists: in the apparent absence of GLP-1 receptors in alpha cells, how might cycling alpha cells respond to the harmine-GLP-1 combination? Does this small alpha cell subset begin to express GLP-1 receptors in response to harmine? Is the synergy mediated by a second islet cell type that expresses GLP-1 receptors? Again, future studies are required here.

Finally, how fast are these “cycling alpha cells” cycling? Or are they cycling at all? They are annotated as “cycling” because they express certain cell-cycle genes and are annotated as being in G1, S, or G2M cell-cycle phase: it is telling that they exist in presumably quiescent, vehicle-treated islets as well (Figure 4F). If they are really cycling, how “fast” they might be cycling, and how that might compare to prior reports using in vitro and in vivo Ki67 or BrdU or EdU measurements, remains unknown.

Taken together, these studies suggest a mechanism of action through which DYRK1A inhibitors may be able to expand human beta cell numbers. We suggest that they act via a combination of alpha cell precursor replication followed by alpha-to-beta cell transdifferentiation. If and when critical human alpha cell lineage tracing tools become available, future studies will be required to confirm or exclude DYRK1A inhibitor-induced alpha-to-beta transdifferentiation. Very importantly, this alpha-to-beta mechanism eliminates, in theory, the need for residual beta cells in order for the beta cell regenerative drugs to reverse diabetes. Thus, beta cell regenerative DYRK1A inhibitor drugs may be effective even in people with T1D and T2D with few or no residual beta cells. Conversely, if alpha cells are the principal target for “beta cell regenerative drug therapy,” it raises the question as to whether beta cell drug-targeting strategies are necessary for beta cell regeneration or instead are misdirected. These issues and limitations should be the focus of future studies.

Resource availability

Lead contact

The lead contact for further information and request of resources and reagents should be directed to the lead contact Esra Karakose (esra.karakose@mssm.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All transcriptomic data are available in GEO under the accession number GEO: GSE266427.
• •This paper does not report original code. All analysis is performed by publicly available packages, and no custom method is implemented. All the packages used have been cited throughout the manuscript.
• •Any additional information related to this paper is available from the lead contact upon reasonable request.

Acknowledgments

The authors wish to thank Bonnie and Joel Bergstein, Lonnie and Thomas Schwartz, and Martha and Fred Farkouh families for their constant support of this research. We also thank the NIDDK-supported Human Islet and Adenovirus Core (HIAC) of the Einstein-Mount Sinai Diabetes Research Center (ES-DRC), the NIDDK Integrated Islet Distribution Program (IIDP), and Prodo Laboratories for supplying human organ donor islets. We also thank the Genomics Core at the Icahn School of Medicine. This work was supported by NIH grants K-01 DK128378, P-30 DK 020541, R-01 DK130300, R-01 DK126450, R-01 DK125285, R01 DK105015, and R-01 DK129196; by the Bioinformatics for Next Generation Sequencing (BiNGS) shared resource facility within the Tisch Cancer Institute at the Icahn School of Medicine at Mount Sinai, which is partially supported by the NIH grant P30CA196521; by the computational resources and staff expertise provided by Scientific Computing at the Icahn School of Medicine at Mount Sinai; by the Clinical and Translational Science Awards (CTSA) grant UL1TR004419 from the National Center for Advancing Translational Sciences; and by the NIH Office of Research Infrastructure grant S10 OD 026880. Confocal microscopy was performed in the Microscopy and Advanced Bioimaging CoRE at the Icahn School of Medicine at Mount Sinai, supported with funding from NIH Shared Instrumentation Grant (FAIN: S10OD021838).

Author contributions

E.K., P.W., and O.W. performed experiments. E.K., X.W., S.C., D.H., D.D., L.L., R.K., G.L., and C.A. analyzed data. E.K., C.A., A.G.-O., D.K.S., and A.F.S. conceived of the studies. E.K. and A.F.S. wrote the manuscript with comments from X.W., S.C., and D.H.

Declaration of interests

P.W. and A.F.S. are inventors on patents filed by the Icahn School of Medicine at Mount Sinai. A.G.-O. consults for Sun Pharmaceutical Industries. R.P.S. is a co-founder of Panacent Bio, Inc.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rat monoclonal anti-C-peptide	DSHB	GN-ID4, RRID: AB_2255626
Mouse monoclonal anti-Glucagon	Abcam	ab10988; RRID:AB_297642
Rabbit monoclonal anti-Glucagon	Abcam	ab108426; RRID:AB_10887227
Goat anti-rat Alexa Fluor 633	Life Technologies	A21094, RRID: AB_2535749
Goat anti-rat Alexa Fluor 488	Life Technologies	A11006, RRID: AB_2534074
Goat anti-rabbit Alexa Fluor 594	Life Technologies	A11037, RRID: AB_2534095
Goat anti-rabbit Alexa Fluor 488	Life Technologies	A11008, RRID: AB_143165
Biological samples
All information on cadaveric islet donors can be found in Table S1.	This paper.	N/A
Chemicals, peptides, and recombinant proteins
Harmine	Sigma	286044 CAS 442-51-3
GLP-1 (Glucagon-Like Peptide 1 Fragment 7–37 human)	Sigma	G9416
Critical commercial assays
Chromium 3′ Gene Expression V3 Kit	10X Genomics	PN-1000268
Novaseq 6000 S1 Reagent Kit v1.5	10X Genomics	20028317
Deposited data
Cycling alpha cells in Regenerative Drug-Treated Human Pancreatic Islets May Serve As Key Beta Cell Progenitors	NCBI Gene Expression Omnibus	GEO: GSE266427
Software and algorithms
Imaris	Oxford Instruments	Version 10.2.0
Leica X	Leica Microsystems	Version 3.7.4
GraphPad Prism	GraphPad	Version 10.2.2
R	The R Software Foundation	Version 4.2.2
Python	Python Software Foundation	Version 0.2.5

Experimental model and study participant details

Human pancreatic islet studies

HIPAA-compliant, de-identified adult human pancreatic islets from four donors (three male and one female) were obtained from Prodo Laboratories (Aliso Viejo, CA), Alberta Diabetes Institute Islet Core (Edmonton, Canada), and the Integrated Islet Distribution Program (IIDP). In all cases, informed consent was provided at the institutions where the organs were harvested. The mean age of the donors was 52.75 years (range 47–58 years), and the mean BMI was 30.1 kg/m² (range 27.4–32.3 kg/m²). Purity ranged from 85 to 90%. Additional details are provided in Table S4.

Method details

Dispersion of human islet cells into single cell suspension and compound treatments

Islets were first disassociated with Accutase (Fisher Scientific).^7,8,9,10 Briefly, islets were centrifuged at 200g to obtain pellets of whole islets. Pellets were then washed twice with phosphate-buffer saline (PBS), incubated in Accutase for 15–17 min at 37^oC, and dispersed into single cell solution by pipette trituration. Islet single cell solution was then pelleted by centrifugation at 700g and resuspended in RPMI islet culture medium containing 5.5mM glucose. Subsequently, islet single cell solution was separated and subjected to different compound treatments for 96Hr; DMSO (0.1%), Harmine (10μM), Harmine + GLP-1RA (GLP), and Harmine + LY364947 (LY). Roughly 1000 islet equivalents (IEQs) were present in each of these treatment conditions.

Immunocytochemistry, antisera and microscopy imaging

Immunocytochemistry was performed on seven human islet donors that were accutase dispersed and 4% paraformaldehyde-fixed (15 min) on coverslips.⁷ The following antisera were used: C-peptide (DSHB GN-ID4); Glucagon (Abcam, ab10988 and ab108426). Confocal microscopy was performed in the Microscopy and Advanced Imaging CoRE at the Icahn School of Medicine at Mount Sinai. Leica SP8 STED 3X microscopy was utilized for imaging. Imaris (v10.2.0) software was used to perform 3D rendering and to generate videos presented in Figure S8.

Processing of human islet cells for scRNA-seq

After the completion of 96-h compound treatment, islets were washed once with PBS and then dispersed with Accutase for 10 min at 37^oC. Dispersed islets were then collected in 5.5mM glucose containing islet culture medium and centrifuged at 700g to obtain cell pellets. These pellets were then resuspended in 0.5% bovine serum albumin (BSA) in PBS at 1 x 10⁶ cells/ml. Cells were processed according to Chromium 3′ Gene Expression V3 Kit (10X Genomics) using the manufacturer’s guidelines followed by sequencing on an S1 Novaseq chip (Illumina Inc.) at the Genomics Core of Icahn School of Medicine at Mount Sinai.

Single cell RNA-seq analysis, quality control and preprocessing

FASTQ files were aligned to the GRCh38 human genome reference, filtered, barcoded and UMI counted using Cell Ranger version 7.0.0 (10X Genomics). Empty droplets and substrate ambient RNA remove from count matrices with remove-background function from cellbender (version 0.2.2).⁶⁶ Each dataset was then filtered to retain cells with ≥1000 UMIs, ≥400 expressed genes, and <20% of reads aligned to the mitochondrial genome. UMI counts were then normalized so that each cell had a total of 10,000 UMIs across all genes and log-transformed with a pseudocount of 1 using the “LogNormalize” function in the Seurat package (version 4.0.3, RRID:SCR_016341).³³ The top 2000 most highly variable genes were identified using the “vst” selection method of “FindVariableFeatures” function and counts were scaled using the “ScaleData” function. Principal component analysis was performed using the top 2000 highly variable features (“RunPCA” function) and the top 30 principal components were used in downstream analysis. Datasets for each donor from DMSO, Harmine, Harmine + GLP and Harmine + LY treatments were integrated by using the “RunHarmony” function in the harmony package (version 0.1.0)⁶⁷ where sample name was used as the group for batch correction. K-Nearest Neighbor graphs were obtained by using the “FindNeighbors” function and the UMAPs were obtained by the “RunUMAP” function. Louvain algorithm was used to cluster cells based on expression similarity. Cell cycle regression of the integrated data is performed by first calculating the cell cycle phases using the CellCycleScoring function of Seurat and then doing the regression by using the G2/M and S phase scores as variables to regress for scaling data prior to harmony integration. Endocrine cells were selected from cell-cycle regressed integrated dataset, and clustering analysis was performed again to generate endocrine-only UMAP. Similarly, cluster 6 (Cycling Alpha concentrated cluster) was isolated from endocrine only integrated data and then clustering analysis was performed again to generate Cycling Alpha UMAP. The resolution was set at 0.4 for the integrated dataset containing all cells and endocrine cells, and at 0.2 for integrated dataset containing only the Cycling Alpha cells for optimal clustering.

Cell type annotation

Differential markers for each cluster were identified using the Wilcox test (“FindAllMarkers” function) with adjusted p-value <0.01 and absolute log2 fold change >0.25, and minimum 10% of cells expressing the gene in both comparison groups using 1000 random cells to represent each cluster. The top up-regulated genes and curated genes from the literature were used to assign cell types to clusters in the all cells integrated dataset, and the expression of marker genes are visualized using DotPlot function. Gene expression signature scores were calculated using AddModuleScore function from Seurat package. To further confirm our cell annotations, first, we performed reference-based mapping using Azimuth R packages (version 0.4.6)³³ and employed previously annotated human islet cell dataset as ref.³⁰ Predicted cell annotation was projected on our integrated data UMAP with all cells, and the expression level of annotation markers used by Elgamal et al. was visualized by DotPlot.³⁰ We further performed cell type fractions prediction using Unicell deconvolve (version 0.0.1).³⁴ Cell type probability of selected cell types were visualized on UMAP. The initial cell type annotation in the initial integrated dataset was applied to all subsequent datasets and UMAPs.

To examine the difference of beta-cell gene signature expression in drug-treated group and vehicle control, we fitted the beta-cell module score into a linear mixed-effect model with a random-intercept across donors. The marginal effect was calculated at group level using the model, and the marginal effect in each treated group was contrasted with the marginal effect in vehicle control.

Cell abundance analysis

Cell type abundance was visualized using barplot which shows the percentage of each cell type within each treatment condition. To explore changes in abundance of the different cell types between treatment and DMSO samples, MiloR (version 1.2.0)³⁷ R package was used to perform differential abundance test in integrated dataset with both endocrine and exocrine cells. Milo graph and neighborhoods were generated using k = 20, d = 30 and harmony embeddings as dimensionality reduction in integrated dataset of all cells. Cell count differences across samples were normalized, and comparison was carried out using testNhoods function. Neighborhood with cell type fraction <0.7 is considered as ‘mixed’ neighborhood. Results of cell abundance test was visualized using plotDAbeeswarm function in MiloR. After selecting only endocrine cells, percentage of cell type in each treatment was calculated again as number of cells of each cell type in each treatment divided by total number of cells in the treatment. Unstacked barplots were used to visualize the percentage of cell type across treatments, and Dunnett’s test was applied to assess the significance of change in abundance between H, H + GLP and H + LY against DMSO.

Pseudotime, velocity and PAGA analysis

RNA velocity analysis was performed using the scvelo Python package (version 0.2.5) on integrated endocrine cells data.⁶⁸ The unspliced and spliced count matrices were generated from Cell Ranger output using the run10x function from velocyto package (version 0.17.17).⁶⁹ For all-sample-integrated endocrine cell dataset, Seurat object was converted into anndata (version 0.8.0)⁷⁰ and merged with the unspliced and spliced count matrices using scanpy (version 1.9.3).⁷¹ Moments for velocity estimation were calculated using 30 principal components and 30 neighbors. RNA velocity was estimated using stochastic model with default parameters using scvelo. Cell type labels from the Harmony integrated endocrine cell dataset were projected to the all-sample-integrated dataset, and velocity streams were visualized. To explore the connectivity between cell clusters, partition-based graph abstraction (PAGA) was calculated based on velocity results using the PAGA wrapped in scvelo.⁷² pyGPCAA wrapped in CellRank was utilized to find initial and final state clusters in integrated all cell dataset, using the abovementioned velocity information (version 1.5.1).⁷³ To infer progression of cells through geodesic distance along the graph, diffusion pseudo-time was calculated using the dpt tool from scanpy, with a cell from the initial cluster selected by CellRank as the root cell.

Gene regulatory network analysis

To infer gene regulatory networks in each cluster, pySCENIC (v0.12.1, RRID: SCR_017247) was applied to the integrated endocrine cells dataset.⁷⁴ The list of transcription factors for hg38 human genome was downloaded from https://resources.aertslab.org/cistarget/tf_lists/and was used for querying transcription factors of the pySCENIC analysis. Motif annotation and the cisTarget datasets were obtained from https://resources.aertslab.org/cistarget/motif2tf/and https://resources.aertslab.org/cistarget/databases/homo_sapiens/hg38/refseq_r80/mc_v10_clust/gene_based/respectively. Briefly, co-expression of genes and transcription factors was first inferred from gene expression matrix using grn function and the co-expressed genes and transcription factors together define gene regulons. The regulons were further refined by finding only cis-regulatory target of each transcription factor using ctx function for motif discovery. Finally, the cellular enrichment of each refined regulon was calculated using aucell function from pySCENIC. In each annotated endocrine cell type cluster, regulon specificity score based on the Jensen-Shannon divergence was calculated to evaluate the activity of each regulon. Regulon specificity score was visualized using clustermap function from seaborn (v 0.12.1, RRID: SCR_018132).⁷⁵

Quantification and statistical analyses details

For single cell RNA-seq experiments n = 4 independent biological replicates were used. For immunocytochemistry experiments n = 7 independent biological replicates were used p-values were calculated by one way ANOVA with a common control unless otherwise indicated. n.s. indicates a non-significant difference. ∗p < 0.005, ∗∗p < 0.0001.

Published: December 2, 2024