Machine learning workflows identify a microRNA signature of insulin transcription in human tissues.

Dicer knockout mouse models demonstrated a key role for microRNAs in pancreatic β-cell function. Studies to identify specific microRNA(s) associated with human (pro-)endocrine gene expression are needed. We profiled microRNAs and key pancreatic genes in 353 human tissue samples. Machine learning workflows identified microRNAs associated with (pro-)insulin transcripts in a discovery set of islets (n = 30) and insulin-negative tissues (n = 62). This microRNA signature was validated in remaining 261 tissues that include nine islet samples from individuals with type 2 diabetes. Top eight microRNAs (miR-183-5p, -375-3p, 216b-5p, 183-3p, -7-5p, -217-5p, -7-2-3p, and -429-3p) were confirmed to be associated with and predictive of (pro-)insulin transcript levels. Use of doxycycline-inducible microRNA-overexpressing human pancreatic duct cell lines confirmed the regulatory roles of these microRNAs in (pro-)endocrine gene expression. Knockdown of these microRNAs in human islet cells reduced (pro-)insulin transcript abundance. Our data provide specific microRNAs to further study microRNA-mRNA interactions in regulating insulin transcription.

Introduction

Insulin production in pancreatic β-cells is an essential process required for the maintenance of normal glucose metabolism. Any decline in β-cell function is a common denominator to type 1 diabetes (T1D), as well as type 2 diabetes (T2D). Transcription of specialized genes, including (pro-)insulin, is dependent on several factors including chromatin organization, transcription factor assembly, as well as the expression of regulatory microRNAs (miRNAs/miRs). MicroRNAs are short, non-coding RNA molecules known to fine-tune gene expression via targeting multiple messenger RNAs (mRNAs). Biologically active (mature) microRNAs are generated through processing by dicer, an enzyme critical for the biogenesis of both canonical and non-canonical microRNAs (Wong et al., 2018). In mice, pancreatic Dicer1 deletion demonstrated the essential role of microRNAs in the generation of pancreatic β-cells (Lynn et al., 2007) during embryonic development. Islet β-cell-specific (RIP-Cre) deletion of Dicer1 in mice did not affect islet cell development but significantly reduced insulin transcription/production (Melkman-Zehavi et al., 2011) and promoted the development of diabetes (Kalis et al., 2011) in adult mice. Although these loss-of-function studies demonstrate, in general, the essential role of microRNAs in mouse islet biology, it is important to identify microRNAs associated with human (pro-)endocrine gene expression and more specifically those associated with (pro-) insulin gene transcription.

One of the very first studies in microRNA islet biology was the demonstration of a key role for miR-375 in mouse insulin secretion (Poy et al., 2004), followed by validation of its role in zebrafish (Kloosterman et al., 2007) and mouse (Poy et al., 2009) pancreatic islet development. These studies laid the foundation for identification of miR-375 as a microRNA associated with and necessary for islet beta-cell function in zebrafish and mice. In human studies, miR-375 expression was shown in the developing and adult pancreatic islet cells (Klein et al., 2013; Joglekar et al., 2009a, 2009c; Correa-Medina et al., 2009) but also in islet non-β cells (Joglekar et al., 2009a) as well as in multiple non-pancreatic human tissues (Latreille et al., 2015). These studies confirmed that miR-375, although enriched in human islets, is not an “islet-specific” microRNA. It became apparent that a set or combination of microRNAs, rather than a single microRNA, supports/promotes endocrine pancreatic cell function. The scarcity of human islets for research and the availability of elegant molecular tools in non-human model systems have resulted in a large body of islet microRNA research in zebrafish and mice. Studies comparing microRNA profiles of cells/tissues that naturally transcribe the (pro-)insulin gene (e.g. thymus, gallbladder, brain (Choi et al., 2019; Devaskar et al., 1994; Heller et al., 2010; Mehran et al., 2012; Sahu et al., 2009; Dutton et al., 2007; Joglekar et al., 2021) are also lacking. Here, we sought to identify microRNAs that fine-tune the expression of the human (pro-)insulin gene using an unbiased, machine learning discovery approach, followed by wet-lab validation. Our data present a signature of microRNAs that is not only associated with (pro-)insulin transcript abundance and predictive of (pro-)insulin gene expression but also regulates the expression of pancreatic transcription factors and islet (pro-)hormones.

Results

Study samples

To generate a signature of microRNAs associated with (pro-)insulin gene transcription, we isolated RNA from a total of 353 human samples across nine tissue types (Figure 1A). We then assessed (pro-)insulin as well as the housekeeping 18S rRNA gene transcripts in these samples and classified these 353 tissue samples as insulin-positive or insulin-negative tissues (Table S1). A subset (discovery set; n = 92 of 353) of these samples were then subjected to microRNA discovery analyses using the OpenArray high-throughput TaqMan-based real-time quantitative PCR (qPCR) platform (Farr et al., 2015, Wong et al., 2015). The discovery set consisted of two groups of tissues that are at the ends of the insulin transcript expression spectrum: one with very high insulin expression (pure islets) and another with undetectable insulin transcripts. The selection of samples in each group was random. Along with the 754 different microRNAs, we also assayed candidate mRNAs for islet hormones (INS, GCG, SST), transcription factors (PDX1, MAFA, NEUROG3; “referred herein as” NGN3, HES1), and the housekeeping gene (18S rRNA; Figure S1A) in this discovery set of 92 samples that include 30 human islets and 62 insulin-negative samples (Table S1). The remaining samples were then used in validation (n = 50) and prediction (n = 202) sets of human tissues (Table S1). A separate set of nine human organ donor islets from individuals with T2D were also used to assess microRNA expression.

Figure 1

Identification of (pro-)insulin transcript-associated microRNAs in the discovery sample set

(A) Schematic presentation of insulin-positive (green outline) and insulin-negative (red outline) human cell/tissues used. Human cell/tissues (such as the gallbladder and brain, yellow outline) which have low to undetectable insulin expression were also used in this study.

(B) An unsupervised Euclidean (average) bidirectional hierarchical plot on the 754 discovery microRNAs in human islets (n = 30; all insulin positive) and insulin-negative tissues (n = 62). Tissues included in the discovery set are color coded, and the key to color and their corresponding tissue is provided in the bottom right corner of this figure. Heatmap represents the normalized qPCR Ct values (color bar) with low Ct values/high miRNA levels (dark red) and higher Ct values/low to no expression of microRNA (yellow to white).

(C) Volcano plot for the 754 discovery microRNAs presented as Ct value differences (on X axis) and statistical significance (−log10 p value; two-tailed Welch's t-test) on Y axis for insulin-negative vs islet (insulin positive) samples. The dashed horizontal line represents p value = 0.05, whereas dotted vertical lines represent a 2-fold difference (1 Ct value). Significantly altered microRNAs are in red.

(D) A Circos plot for mRNA and microRNA expression in human islets is divided into six sections (a1, a2, b, c, d, and e). a1: larger part of the outermost circle representing miRNA expression; a2: colored rectangles representing mRNA transcripts (as labeled); b: scatterplots illustrating the Ct value (ranging 6.5 to 39; greater values are closer to the center) for the respective microRNA or mRNA. The glyphs are color coded with greater Ct values (low expression) in red, while smaller Ct values (high expression) are in green (median of the data = 22.5; thin white line). c: line plots presenting the z-scores of the transcript expression data (ranging from −3.4 to 3.5; smaller values are closer to the center of circle, the calculation of the z-scores is as described in the methods section), and d: histogram plots showing the log10 transformed relative abundance (to insulin-negative tissues, as described in the methods section in detail), ranging from −7.3 to 10.7. Lesser values are closer to the center of circle. The thin gray lines are placed at every 5% of the data. Section “e” is the innermost plot which displays correlation between miRNAs and the seven mRNAs (Pearson's r > 0.6). The correlations are plotted in the form of colored link (as per the labeled colors of the mRNA).

(E) Bootstrapping involves random deletion (red dots) of multiple samples and replacing them with duplicates from the remainder set (green dots). Each iteration provides a new set, thus eliminating sampling bias.

(F) Frequency/importance (%) of microRNAs (X axis) detected following penalized logistic regression and 1000 machine learning/resampling validation (bootstraps). Highlighted top eight microRNAs represent those common to logistic and linear regression analyses

(G) A t-SNE plot of discovery set using the top eight microRNAs (H) or random eight miRNAs with similar distribution. Each point represents a sample from the discovery set, while the color of the point indicates tissue type.

See also Figure S1 and Tables S1, S2, and S3.

Identifying microRNAs associated with human pancreatic islet (pro-)insulin expression

We performed unsupervised bidirectional hierarchical clustering on global normalized microRNA as well as candidate mRNA expression data from our discovery sample set (n = 92). All human islet samples (insulin-positive tissues) cluster together indicating similarity in their microRNA (Figure 1B) as well as candidate mRNA expression profiles (Figure S1A), relative to those from insulin-negative tissues. Interestingly, a total of 241 different microRNAs were found at significantly higher levels (≥2-fold, p < 0.05) in human islets, and 91 other microRNAs were significantly higher in insulin-negative tissues (≥2-fold, p < 0.05; Figure 1C and Table S2). All of these 332 differentially expressed microRNAs were expressed between 2- to over a hundred thousand-fold higher abundance (Cycle threshold [Ct] value difference of 1–17; Figure 1C and Table S2) and at significantly low p value (p < 0.05 to p < 3.4 × 10−49), relative to the insulin-negative tissues and remain significant after adjustment for multiple testing. Several of these microRNAs show a high correlation with the candidate mRNAs assessed, as presented by connecting lines in the center of the Circos plot (correlation coefficient (r) ≥0.6; Figure 1D).

Since conventional comparisons of microRNA expression data sets (Figures 1B–1D) identified a large number of significantly differentially expressed microRNAs, we decided to employ machine learning workflows to decipher microRNAs that are highly associated with (pro-)insulin (mRNA) levels. Penalization algorithms with bootstrap machine learning workflow (Figure 1E) were used to perform microRNA selection. Good quality human islets (insulin Ct value ≤16.8; n = 30) were marked as the insulin-positive (1) group and all insulin-negative tissues (insulin Ct value = 39; n = 62) were categorized under the insulin-negative (0) group. Validation of the model was carried out using bootstrapping to confirm the signature of microRNAs that were highly associated with (pro-)insulin gene expression. Bootstrap validation workflow of penalized logistic regression included resampling data for 1000 times (Figure 1E). During these analyses, ∼37% of the samples were randomly eliminated at each iteration, while samples from the remaining set were duplicated to maintain the same number of samples at each step (Efron and Tibshirani, 1997; Chernick and Labudde, 2011). After 1000 iterations of resampling, a frequency plot representing the number of times a specific microRNA was found to be highly associated with (pro-)insulin expression was derived (Figure 1F). Such a resampling validation (bootstrapping) procedure eliminates sampling bias. Penalized logistic regression models offer variable (microRNA) selection based on categorical assessment (presence or absence) of (pro-)insulin transcript, while penalized linear regression models facilitate the prediction of (pro-)insulin transcript levels (Goeman, 2010). These analyses (Figure S1B) identified the top eight microRNAs from logistic regression analyses to be common with linear regression outputs (Figure S1C). Although all of the microRNAs derived through our analysis based on the machine learning workflows (Figures 1F and S1B) are important for human (pro-)insulin transcription, we focused our analysis on the most important top eight microRNAs (Figure 1F).

Dimensionality reduction analyses (t-distributed stochastic neighbor embedding; Figure 1G) using the top eight microRNAs from the logistic regression analyses could segregate the insulin-positive islet samples from the insulin-negative tissues. However, three other random sets of eight different microRNAs could not efficiently segregate islet samples from other insulin-negative tissues (Figure 1H and Table S3). In summary, these studies identified a set of microRNAs that are associated with (pro-)insulin gene expression in human islets.

Levels of insulin-associated microRNAs correlate with (pro-)insulin in human tissues that naturally transcribe the insulin gene

Apart from pancreatic islets and thymus, the gallbladder (Choi et al., 2019; Sahu et al., 2009; Dutton et al., 2007) and brain (Devaskar et al., 1994; Mehran et al., 2012; Heller et al., 2010) are known to (naturally) transcribe the insulin gene. We observed insulin, glucagon/somatostatin immunopositivity in majority of the human gallbladder and brain samples that we have assessed (Figure 2A). These samples contained detectable levels of (pro-)hormone gene transcripts (INS, GCG, and SST; Figure 2B). We also measured the abundance of pancreatic transcription factors (MAFA, NGN3, HES1, and PDX1) in these islet (n = 53), gallbladder (n = 75), brain (n = 14), and insulin-negative tissue samples (n = 55; endothelial cells, muscle, liver, skin, and spleen; Figure 2B). All data are presented as normalized Ct value measured by TaqMan-based real-time qPCR. As expected, the level of (pro-)insulin gene transcripts in the pancreatic islets (mean Ct value ± SD; 15.74 ± 0.74) was significantly higher than that in the gallbladder (mean Ct value ± SD; 36.72 ± 2.24, p < 0.0001 vs islet) or in the brain (mean Ct value ± SD; 38.50 ± 0.96, p < 0.0001). We measured the absolute copy number of (pro-)insulin gene transcripts using digital droplet PCR (Maynard et al., 2019) (Figure 2C), which confirmed our qPCR findings (Figure 2B) that the number of INS mRNA copies was around a hundred thousand-fold higher in human islets than in human gallbladders and up to a million-fold higher than that in the brain. As anticipated, (pro-)insulin gene transcripts were not detectable in insulin-negative tissues (Figures 2B and 2C).

Figure 2

Validation and prediction of insulin transcript-associated microRNAs

(A) Immunostaining of insulin (green), glucagon (red), and/or somatostatin (pink) in freshly isolated/cultured human islet, gallbladder epithelium, and brain neurospheres. Nuclei (DNA) are shown in blue. Scale bar is 20 μm.

(B) Real-time TaqMan qPCR expression profile for pancreatic islet (pro-) hormones and transcription factors in human islets (n = 53), gallbladders (n = 75), brains (n = 14), and the insulin-negative tissues (endothelial cells, muscle, liver, skin, and spleen; n = 55). Results are presented as cycle threshold (Ct) values normalized to 18S rRNA. As a lower Ct value represents the higher abundance of gene transcripts, the Y axis is reversed. Each dot within the polygons represents a different sample. The horizontal solid red line represents the median, the horizontal black dotted line within each polygon represents quartiles, and the polygons represent the density of individual data points and extend to min/max values.

(C) Absolute copies for (pro-)insulin gene transcripts in human islet (n = 11), gallbladder (n = 63), brain (n = 14), endothelial (n = 2), muscle (n = 4) and spleen (n = 8) samples, as measured by digital droplet (dd)PCR. Each dot in the floating bars represents a different sample. Horizontal solid lines represent the mean and boundaries extend to min/max values.

(D) MicroRNA abundance of the top eight microRNAs (from the logistic regression analysis, Figure 1F) in human islets (I; n = 30), gallbladders (G; n = 50); brains (B; n = 4), and spleens (S, n = 34) is presented in an aligned dot plot with mean and SEM. Significant differences were computed using one-way analysis of variance (ANOVA) with Dunn's multiple comparisons test. ∗: p < 0.05, ∗∗: p < 0.01, ∗∗∗: p < 0.001, ∗∗∗∗: p < 0.0001; compared to islets.

(E) Receiver operator characteristic (ROC) curve analysis for individual and combined top eight microRNAs obtained through the penalized logistic regression and bootstrap analysis. A separate validation set (n = 50) of human islets (n = 23) and insulin-negative samples (n = 27; consisting of human endothelial cells n = 14, gallbladder n = 11, and muscle n = 2) was used. Area under the curve (AUC), sensitivity, and specificity are shown.

(F) Correlation plot between the Ct value for laboratory-measured (pro-)insulin mRNA (X axis) and the computationally predicted (pro-)insulin Ct value (Y axis) from the islets within the validation set (n = 23), impure islet fraction (n = 85), and pancreas (n = 62) from the prediction set (Spearman r = 0.6, p < 0.0001). Each dot represents a different sample, and the color of the point indicates the tissue type in the correlation plot.

See also Figure S1 and Tables S4 and S5.

We then quantified the abundance of the top eight microRNAs identified earlier (green-shaded area in Figure 1F) across human islets, gallbladders, brains, and spleen samples. A significantly lower level of expression was seen for all of the top eight microRNAs (p < 0.001; Figure 2D) in gallbladders, brains, and spleens (compared to islets). The levels of all eight microRNAs positively correlated with (pro-)insulin transcript (Table S4) across the three tissue types that naturally produce/transcribe insulin and the insulin-negative (spleen) tissues. These data indicate a potential interdependence of the most important microRNAs and (pro-) insulin gene expression.

Insulin-associated microRNAs are predictive of insulin gene expression

Although insulin-associated microRNAs were able to identify insulin-producing cells in the discovery sample set (Figure 1G), we interrogated their predictive potential in different sets of samples. We generated receiver operating characteristic curves for 50 samples (from the validation set; n = 23 islets vs n = 27 insulin-negative tissues; Table S1) to test the specificity and sensitivity of these microRNAs in classifying the presence or absence of (pro-) insulin expression. Interestingly, each of these microRNAs demonstrated high specificity (85–100%) and sensitivity (87–100%) in discriminating insulin-positive vs insulin-negative tissues, with an area under the curve of more than 92% (Figure 2E). A combined model, testing all of the top eight miRNAs, returned the highest predictive power (Figure 2E). Our results validate the role of these microRNAs in accurately stratifying insulin-producing tissues in an independent validation set of primary human tissue samples. We then used the prediction set (n = 202) of human tissues that naturally express (pro-)insulin gene transcripts at varying levels to further elucidate the predictive power of these microRNAs. We derived the coefficients from the penalized linear regression analysis (Figure S1B) to obtain the odds ratios that constituted a mathematical formula to predict the level (qPCR Ct value) of (pro-)insulin gene expression. Two wet-lab biologists independently carried out the measurements of microRNAs and (pro-) insulin mRNA on the prediction set of human tissue samples consisting of the pancreas, impure islet fractions, gallbladder, and brain (n = 202; Table S1). A de-identified list containing the microRNA qPCR Ct values was provided for the computational prediction of (pro-)insulin Ct values. We observed that the levels of these microRNAs could reliably predict the real-time qPCR (laboratory-measured) (pro-)insulin Ct value (Spearman r = 0.8, p < 0.0001, n = 202) in the prediction set of samples (Figure S1D). When the pancreatic tissue was independently analyzed, a significant correlation (Spearman r = 0.6, p < 0.0001, n = 170) between the computationally predicted and laboratory-measured (pro-)insulin Ct values was observed (Figure 2F). In summary, the levels of bootstrapped microRNAs (including top eight microRNAs; Figure S1C) separated categories (Figure 2E) as well as computationally predicted (pro-)insulin Ct values (Figures 2F and S1D). Together, our data demonstrated that the top eight (pro-)insulin-associated microRNAs identified using machine learning workflows were not only associated with (pro-)insulin expression but also predictive of (pro-)insulin transcript abundance in separate sets of primary human tissues.

Forced expression of insulin-associated microRNAs alters pancreatic gene expression

Since associations do not prove causality, we investigated if regulated expression of insulin-associated microRNAs in human pancreatic progenitor cells can promote (pro-) endocrine gene expression. In order to address this question, we forced the expression of individual microRNAs in human pancreatic progenitor cells (Figure S2A) using a puromycin-selective doxycycline-inducible lentiviral vector system (Figure 3A). Experiments with individual microRNA vectors were carried out as attempts to co-transduce cells with all of the top eight microRNAs resulted in high mortality. Since we were interested in assessing the potential of these microRNAs in differentiation of progenitor or precursor cells (rather than enhancing insulin expression in β-cell lines such as EndoC-βH1), we selected human islet-derived progenitor cells (hIPCs) and commercially available human pancreatic duct (PANC1) cells (Figure S2A) for these studies. A total of 16 different microRNA overexpression and one non-targeting/scramble control vectors were used to generate microRNA-overexpressing lines (Figure S2A). These comprised 14 different microRNAs from our penalized logistic regression bootstrap analysis, including the top eight (Figure 1F). Two other microRNAs (miR-30e-5p and -30d-5p) were also selected for these overexpression studies as they have been previously reported by us (Joglekar et al., 2009c) and others (Tang et al., 2009) in modulating (pro-)insulin gene expression. We first confirmed that each of the 16 PANC1 lines generated for this study demonstrated overexpression of the relevant microRNAs at levels closer to those detected in human islets (Figure 3B). Then, using TaqMan real-time qPCR we assessed the ability of each of these 16 microRNAs to instill a measurable and/or significant change in the expression levels of a select set of 51 human pancreatic mRNA/gene transcripts (Table S6) known to be important in pancreas development/function (Jennings et al., 2015; Chakrabarti and Mirmira, 2003). Only 23 of these 51 mRNAs (shown in Figure 3C) were significantly altered following overexpression of individual microRNAs. Surprisingly, two microRNA-overexpressing lines (miR-217-5p and miR-429-3p) did not significantly change the expression of any of these 51 mRNAs assessed. On the other hand, three or more microRNA-overexpressing lines induced a significant increase in IGF2, HES3, NKX2.2, and HNF1β levels compared to non-targeting/scramble control (Figure 3C). At the same time, the expression of PDX1, CDX2, TLR3, and HB9 was significantly reduced in several other microRNA-overexpressing lines (Figure 3C). Interestingly, certain microRNAs including miR-7-2-3p and miR-7-5p were responsible for significant reduction of pancreatic genes, whereas overexpression of miR-375-3p, miR-183-5p, miR-183-3p, miR-184-3p, miR-141-3p, and miR-200c-3p only led to a significant increase in pancreatic gene expression. None of the microRNA-overexpressing lines induced any significant change in (pro-)insulin and SST gene expression. We also mixed PANC1 lines overexpressing each of the top eight (pro-)insulin-associated microRNAs in equal proportion to understand if this would lead to (pro-)insulin expression. A modest (2.2-fold) but insignificant (p = 0.4) increase in (pro-)insulin gene transcripts was observed along with significant and more robust increase in the levels of other (pro-)hormone and transcription factors including GCG, NGN3, MAFA, PAX2, PAX6, NKX2.2, and UCN3 (Figure S2B).

Figure 3

Pancreatic gene expression analysis in microRNA-overexpressing PANC1 lines

(A) A schematic presentation for generating microRNA overexpressing cell lines using lentiviral-mediated, doxycycline-inducible overexpression system.

(B) MicroRNA abundance in human islets (green, n = 4) and microRNA-overexpressing PANC1 cell lines (orange, n = 6 separate experiments) was calculated relative to the respective microRNA levels in non-targeting/scramble control (NTC) PANC1 cell line. A total of 16 microRNAs (presented on X axis) were individually overexpressed in PANC1 cells using doxycycline-inducible lentiviral vectors. The microRNA data are presented as scatter dot plots with mean ± SEM. Each green dot in the plot represents a different islet sample, and each orange dot presents a different preparation/experiment.

(C) Categorical bubble plot displaying effect of overexpression of microRNA (listed on Y axis) as either significant increase (red) or significant decrease (blue) in the expression of a specific transcription factor or (pro-)hormone (presented on X axis) in PANC1 cells. The absence of a circle across any of the pancreatic gene (e.g. INS and SST) indicates a non-significant effect of corresponding microRNA overexpression. The size of the circles represents the p value (p < 0.05, n = 6 separate experiments) based on a two-tailed Welch's t-test. Values are presented in relative abundance compared to NTC.

(D) UMAP: Uniform Manifold Approximation and Projection plots generated using Seurat (version 3.2.2) from pancreatic single-cell RNA sequencing (Panc8) data set (SeuratData version 0.2.1). Single-cell data set from four of the five technologies (see methods for details) in Panc8 is presented here. The single cells were clustered into different pancreatic cell types, and expression levels for HNF1β, PDX1, HB9, NKX2.2, and IGF2 across these 13 clusters are highlighted in individual UMAP plots. Only nine of the most relevant clusters are labeled. Data are presented in log-transformed normalized values and are overlaid in shades of blue. Dark blue color shows higher expression in the specific cell clusters, while gray represents absence/low expression of the demonstrated pancreatic gene in these cells.

See also Figure S2 and Table S6.

To identify the pancreatic cell type that each of these miRNAs promoted, we analyzed the publicly available pancreatic single-cell sequencing (scSeq) data sets (Butler et al., 2018; Stuart et al., 2019). As expected, these human pancreatic scSeq analyses indicated that NKX2.2 was predominantly localized to all endocrine pancreatic (α-, β-, δ-, γ-, ε-) cell lineages (Figure 3D). On the other hand, IGF2, one of the most common and significantly upregulated genes (Figure 3C), is localized to human pancreatic β cells (Figure 3D). These data suggested that overexpression of microRNAs such as miR-433-3p significantly increased abundance of HNF1β and NKX2.2, which are localized to endocrine pancreatic cells (Figures 3C and 3D). We then created a microRNA regulatory map for different cell types observed during human pancreas development (Figure 4A). We identified a set of genes that have been previously reported in pancreas development (reviewed in (Jennings et al., 2015; Chakrabarti and Mirmira, 2003)) to define endocrine pancreatic sub-types. We then applied our observations of single microRNA overexpression studies in regulating these pancreatic genes (Figure 3C) to construct an mRNA-microRNA interaction map (Figure 4A). Here, microRNAs are ranked based on the level of significant change (p value) that each of the microRNAs induced to either increase (shown in green) or decrease (shown in red) target mRNA gene expression. As an alternate validation, we assessed the expression levels of these microRNAs and islet mRNAs in our bulk RNA-seq data sets (GSE134068; n = 18 (Wong et al., 2019)) and correlated selected microRNAs (miR-375, miR-141, miR-30d, and miR-216b) with key pancreatic gene transcripts. As shown in (Figure 4B), several pancreatic genes positively correlated with miR-375, while negatively correlating with the levels of miR-216b (or miR-30d) in human islets. These analyses provided additional evidence in the selection of microRNA candidates for future overexpression and knockdown studies. In summary, forced expression of specific microRNAs induced pancreatic genes (Figures 3C, 4A, and 4B) that map to sub-populations of human pancreatic hormone-producing cells (confirmed through scRNA sequencing; Figure 3D) and microRNA-mRNA expression in human islets provided supportive evidence to identify candidates for future studies.

Figure 4

A microRNA regulatory map of human pancreatic cell fates

(A) A map of microRNA regulation was generated based on our data (see Figure 3) of microRNA overexpression in human pancreatic duct cells and the regulation of a select set of lineage-enriched (α-, β-, or δ-) genes known to be important in pancreas development. Key pancreatic genes and transcription factors for a particular developmental stage are shown. The microRNA labels are colored in green or red based on our data confirming a statistically significant (p ≤ 0.05) change in gene expression (shown by arrows that each microRNA points to). MicroRNAs that significantly increase gene transcript levels are in green color and those that significantly decrease gene expression are in red. All microRNAs that significantly regulate the expression of a specific gene are listed based on the p value (calculated using a two-tailed Welch's t-test), with the miRNAs at the end of the set being closest to the statistical cutoff value of p = 0.05. Since no single microRNA was seen to significantly regulate INS or SST gene transcript levels, these genes are not listed in the β- and δ-cell lineage.

(B) Correlation between representative microRNAs (miR-141, miR-30d, miR-375, and miR-216b) and pancreatic endocrine transcripts. Correlation analysis was performed using the human islet bulk RNA-seq data set (GSE134068; n = 18). The Pearson correlation r coefficients are presented in black text in each square. Only significant correlations (Pearson p ≤ 0.05) are presented with a colored fill (depicted in the scale bar) displaying their correlation coefficient.

A combination of microRNAs directs differentiation to endocrine pancreatic lineage

Since islets are heterogeneous cell clusters, we hypothesized that protocols involving co-culture, aggregation (and differentiation) of multiple microRNA-overexpressing lines would help in assessing the role of select sets of microRNAs. We previously reported the potential of human pancreatic progenitor cells (PANC1 [Hardikar et al., 2003] and hIPCs [Gershengorn et al., 2004]) to form cell aggregates that contain several transcripts of mature islet-like cells. Our microRNA overexpression data (Figure 3C) identified that certain microRNAs (miR-141-3p, miR-183-3p, miR-433-3p, and miR-183-5p) promote the expression of IGF2 and NKX2.2, which are characteristic of human pancreatic endocrine cells (Figure 3D). Since miR-375-3p has been well known to be an important islet-enriched microRNA and has been observed to promote early pancreatic genes (HNF1β and HNF3β), we selected it along with the above four microRNA over-expressing lines (miR-141-3p, miR-183-3p, miR-433-3p, and miR-183-5p) in further studies. Other microRNAs (miR-30d-5p, miR-30e-5p, miR-200a-3p, and miR-124-3p) suppressed the expression of the pancreatic master regulator transcription factor PDX1. The microRNA miR-7-2-3p was seen to reduce several (pro-)endocrine gene transcripts such as NKX2.2, NKX6.1, as well as the (pro-)hormone GCG. The five microRNAs (miR-30d-5p, miR-30e-5p, miR-200a-3p, miR-124-3p, and miR-7-2-3p) were therefore inhibited in further studies. Since miR-216b-5p was observed to suppress NKX6.1 that is known to control a gene regulatory network for maintaining human islet β-cell identity (Schaffer et al., 2013), we inhibited miR-216b-5p at later stages of our differentiation protocol (Figure 5A). We used an equal proportions of five different microRNA-overexpressing lines (miR-183-5p, miR-183-3p, miR-375-3p, miR-433-3p, and miR-141-3p), in doxycycline-containing media for 3 days before induction of differentiation in serum-free defined media (Figure 5A). We then inhibited the expression of other microRNAs (miR-7-2-3p, miR-124-3p, miR-200a-3p, miR-30e-5p, miR-30d-5p, miR-216b-5p) using Locked Nucleic Acid (LNA) inhibitors in a stage-specific manner (Figure 5A). We assessed changes in expression of pancreatic genes in non-targeting control/NTC line (referred to as “N”; Figure 5B), a mix of the five selected miRNA-overexpressing cell lines (referred to as “M”; Figure 5B) or the mix of these five miRNA-overexpressing lines with LNA inhibitors (referred to as “ML”; Figure 5B). MiRNA, mRNA, and insulin release/content assessments were carried out for all the three experimental conditions (N, M, and ML).

Figure 5

A microRNA overexpression and inhibition model of human pancreatic duct cell differentiation

(A) An outline of the microRNA overexpression and inhibition study to assess the potential combination of microRNAs promoting differentiation to endocrine pancreatic lineage. A non-targeting control microRNA overexpressing line (N) or five selected individual microRNA-overexpressing lines (shown in green) combined in equal proportions in the absence (M) or presence (ML) of LNA power inhibitors targeting microRNAs (shown in red) were differentiated (see methods). All cells were maintained in serum-containing (growth promoting) media for three days before inducing differentiation in a serum-free medium (SFM), with doxycycline-regulated microRNA expression at all times. Differentiation promoted aggregation and formation of islet-like cell aggregates in serum-free medium. LNA power inhibitors were added again at day seven of the differentiation process in the ML group. Islet-like cell aggregates were harvested for downstream mRNA and microRNA analysis, as well as glucose-stimulated insulin secretion (GSIS) assays at the end of 14 days.

(B) Violin plots displaying transcript abundance of select pancreatic transcription factors, genes, and (pro-) hormones in differentiated human pancreatic duct cells. Transcript abundance was calculated using the fold-over detectable method (Hardikar et al., 2014). On the x axis are data for the non-target/scramble control (N) (n = 5), the combination of five selected individual microRNA-overexpressing cell lines without LNA inhibitors (M) (n = 5), or with LNA inhibitors (ML) (n = 3). Each data point represents a different experiment. The horizontal red line marks the median and dotted black lines within each polygon represent quartiles. All polygons represent the density of individual data points and extend to min/max values. Three different primer probes used to measure the (pro-)insulin gene are shown in three different colors with the TaqMan microRNA assay ID listed in Table S5 and S6. Statistical significance was calculated compared to N using Dunn's multiple comparisons test. Ordinary one-way ANOVA was used for the trend analysis between N, M, and ML. Exact p values for comparison and trend are reported for significantly different comparisons and trend (p ≤ 0.05).

See also Table S6.

Combined microRNA overexpression alone (M) or with LNAs (ML) induced a significant increase in GCG, GLUT1, HNF4α, and IGF2 transcripts (Figure 5B), as compared to the non-targeting control (N). A significant trend was observed for PDX1, HNF4α, IGF2, and GLUT1, suggesting the added roles of microRNAs (M) and with LNAs (ML) in regulating their expression. Interestingly, HES1 and P48 were reduced in ML, suggesting progression to endocrine (rather than exocrine) lineage. (Pro-)insulin expression was comparable in all the three conditions but was consistently higher when five different microRNA-overexpressing lines and six LNA inhibitors were combined (ML) during differentiation (Figure 5B). Although differentiated control (N) cells showed comparable C-peptide concentrations (20.0 ± 0.7 pmol/L/μg of DNA; mean ± SEM, n = 3) to that in the microRNA overexpressing lines (15.5 + 5.1 pmol/L/μg of DNA; mean ± SEM, n = 3), the latter demonstrated over 4-fold efficient glucose-stimulated insulin secretion (0.8 ± 0.6% of C-peptide content in N vs 4.2 + 1.1% of C-peptide content in M; mean ± SEM, n = 3-4/group). No significant difference in C-peptide content was observed between M and ML groups (data not shown).

We further tested if the effects of these (pro-)insulin transcript-associated microRNAs in human pancreatic duct cells could be replicated in primary human islet cells. We used primary hIPCs that we described earlier (Gershengorn et al., 2004; Joglekar and Hardikar, 2012;Joglekar et al., 2009b, 2016) to create lentiviral-mediated, doxycycline-inducible, microRNA-overexpressing hIPC lines (Figure S2A). Limited availability of freshly isolated human islets, as well as the reduced proliferative capacity after lentiviral transduction, were some major constraints in selecting only three candidate microRNAs as against all of the top eight. Compared to the non-targeting control (NTC), miR-375-3p overexpression led to an increase, although insignificant, in INS, SST, and PDX1 transcripts (Figure 6A), while GCG transcript levels were reduced in all the three microRNA-overexpressing hIPC lines. Our studies in differentiating human pancreatic (duct and islet derived) microRNA-overexpressing lines demonstrate the potential of these microRNAs in promoting the expression of genes that are known to be restricted to specific endocrine pancreatic sub-sets.

Figure 6

Interdependence of insulin-associated microRNAs and (pro-)insulin gene expression in primary human islet-derived cells

(A) Real-time TaqMan qPCR-based expression analysis of the islet (pro-)hormones and pancreatic transcription factors in microRNA-overexpressing human islet-derived progenitor cells (hIPCs) following in vitro differentiation (see methods). Results are normalized to 18S rRNA and presented as transcript abundance, calculated using fold above detectable method (Hardikar et al., 2014). Selected microRNAs (and non-target control/NTC) used in generating overexpressing hIPCs lines are listed on X axis. Each dot in the floating bar plot represents data from 2–3 different islet donors and two separate experiments. The horizontal black line represents the mean and bars extend to min/max values.

(B) Gene expression analysis in hIPCs transduced with the Furin-cleavable (pro-)insulin vector (Ins-Fur; n = 5) or an empty vector (n = 5). Box plots show mean values (indicated by “+”), line at the median and whiskers extending to the minimum and maximum values in the set. Statistical significance based on a two-way ANOVA. ∗∗:p < 0.01 and five biological (islet donor) preparations of hIPCs.

(C) (Pro-)insulin mRNA expression in human islet cells (from different donors, n = 3–5) following selected power LNA inhibitor-mediated knockdown of microRNAs. The levels of normalized (pro-)insulin gene transcript at day 3 and day 6 are presented as percent of scramble control. The horizontal red line represents the median, and the horizontal dotted black lines within each polygon represent quartiles. The polygons represent the density of distribution of the data and extend to min/max values. Statistical significance was calculated using a two-way ANOVA, and exact p values (compared to control) are presented.

(D and E) A summary of our findings understanding the interdependence of regulatory microRNAs (small colored wheels) and (pro-)insulin (large green wheel) in human islet/pancreatic progenitor cells (see text for details).

See also Figures S3–S5.

Interdependence of (pro-)insulin expression and microRNAs

To understand the regulatory hierarchy and the interdependence between (pro-)insulin transcripts and insulin-associated microRNAs, we first forced the expression of a furin-cleavable insulin gene (Ren et al., 2007) in five biological replicates (five different pancreas donors) of hIPCs. Furin-cleavable insulin was used to enable the processing of (pro-)insulin in the absence of desired levels of (pro-)hormone convertase in undifferentiated hIPCs. Although overexpression of the furin-cleavable insulin led to a 100- to ∼50,000-fold increase in its mRNA as compared to the empty vector transduced cells (Figure 6B), no significant changes were observed in the level of expression of any of the bootstrap signature of microRNAs (including the top eight microRNAs; Figure S3A). These data confirmed that although overexpression of insulin-associated microRNAs (Figures 3, 4, and 5) drives (pro-)endocrine gene transcripts, the forced expression of insulin gene does not change the abundance of microRNAs associated with and predictive of (pro-)insulin gene transcription. We then assessed if the loss of these microRNAs had any impact on human (pro-)insulin gene transcription. We obtained freshly isolated human islets and maintained these in a defined Connaught Medical Research Laboratories (CMRL) medium with or without the specific microRNA LNA power inhibitors for the next 3–6 days. As compared to the control, there was 63% (range: 26–98%) reduction in microRNA expression in just three days of incubation, which further decreased to 72% (range: 29–98%) over the next three days (Figure S3B). This decline in human islet microRNAs significantly reduced the expression of (pro-)insulin gene in miR-375-3p, miR-7-5p, miR-183-3p, miR-183-5p, miR-217-5p, or miR-7-2-3p knockdown islet cells as compared to control at day 6. These data support the necessity of these microRNAs in human (pro-)insulin transcription (Figure 6C).

Since selective depletion of microRNAs reduced (pro-)insulin gene transcription under these in vitro conditions, we investigated if lower expression of our bootstrap microRNAs were associated with reduced (pro-)insulin transcription in islets from human subjects. As shown earlier (Figure 2D), low levels of the top eight microRNAs were associated with low levels of (pro-)insulin in human gallbladders and brains. It is well recognized that the progressive loss of pancreatic islet β-cell function is a major cause leading to T2D (Weyer et al., 1999; Kahn, 2003). We therefore screened human islets that were isolated from individuals with T2D. The levels of (pro-)insulin transcripts were significantly lower in all (n = 9) T2D islets (Figure S4A) as compared to islet samples (n = 53) from donors without diabetes. Analysis of all 754 discovery microRNAs in these T2D islets confirmed that several microRNAs were significantly differentially expressed (Figure S4B). Intriguingly, the levels of multiple insulin-associated microRNAs that we identified in our bootstrap analysis were significantly reduced in T2D islets (Figure S4C). Thus, a pathological reduction in islet (pro-)insulin gene transcript was associated with the depletion of insulin-associated microRNAs, suggesting their potential role in functional β-cell loss in diabetes.

Indeed, these top eight microRNAs regulate several human pancreatic genes (Figure S5A) and target multiple pathways (Figure S5B) related to pancreatic islet development/differentiation (WNT, MAPK, VEGF, TGFβ signaling), insulin production/release (calcium signaling, gap junctions, pancreatic secretion), and islet dysfunction (T1D, T2D, MODY, RNA degradation pathways). Overall, our study results using unbiased machine learning analyses identified a set of microRNAs that are associated with, predictive of, and necessary for (pro-)insulin gene transcription in human islets.

Discussion

Subtle changes in microRNA expression and/or processing can have important biological consequences on (pro-)insulin gene transcription. Conventional genetic manipulation strategies in non-human systems have used knockouts, reporters, and transgenic overexpression models. Such approaches (Lynn et al., 2007; Latreille et al., 2015), although useful to identify and understand the functional role(s) of individual microRNAs, do not fully recapitulate physiologically relevant changes in human pancreatic gene expression networks. The goal of our study was to identify microRNAs associated with human (pro-)insulin gene transcription via profiling human cells and tissues using objective analytical workflows. Machine learning approaches that take out the sampling bias through simulation of a thousand sets of insulin-positive and insulin-negative human tissues were used to derive a signature of insulin-associated microRNAs. We then correlated the abundance of these microRNAs in multiple human tissues representing physiological and pathological depletion of (pro-)insulin, followed by an assessment of their regulatory potential in human pancreatic progenitors and islet cells.

We observed that the overexpression of any single microRNA from our analysis was unable to restore (pro-)insulin expression in pancreatic duct progenitor cells. However, depletion of a single microRNA (such as miR-183-5p, -375-3p, -183-3p, -7-5p, -7-2-3p, or -217-5p) was responsible for a significant reduction but not total absence in (pro-)insulin transcript (Figure 6C). These data suggest that the insulin-associated microRNAs that we identified are not direct regulators of (pro-)insulin gene transcription but possibly act through the coordination of regulatory elements and transcription factors (Figure 6D). Indeed, microRNA target prediction across seven different platforms using miRSystem (Lu et al., 2012) confirmed the absence of any targeting microRNAs for human (pro-)insulin gene transcripts. Our data from human pancreatic progenitor and islet cells demonstrate that overexpression of microRNAs associated with (pro-)insulin abundance (Figure S1C) regulates several pancreatic transcription factors and a combination of these factors fine-tune (pro-)insulin expression, while losses of these miRNAs impact significantly on (pro-)insulin mRNA levels (Figure 6E).

Several microRNAs (including miR-375, miR-7, miR-200, miR-184, miR-124, miR-429) from our bootstrap analysis were previously reported to have specific roles in human β-cell maintenance and function such as (pro-)insulin transcription, glucose uptake, calcium signaling, insulin exocytosis, and β-cell apoptosis (Esguerra et al., 2018; Bolmeson et al., 2011; Wang et al., 2013; Martinez-Sanchez et al., 2018; Filios et al., 2014; Belgardt et al., 2015; Baroukh et al., 2007). In addition to this, we identified several other microRNAs that were not previously associated with human (pro-)insulin transcript levels (e.g. miR-217-5p, miR-183-3p, miR-7-2-3p, miR-183-5p). These microRNAs are, however, reported in other context; miR-217 was noted in developing zebrafish intestine, liver, and pancreas (Stuckenholz et al., 2009) and was shown to inhibit proliferation and invasion of pancreatic adenocarcinoma cells (Chen et al., 2017). MiR-7-2-3p was described in human blood cells following saturated fatty acid feeding (Lopez et al., 2018), while miR-183-5p was reported through in silico studies as a T2D gene set targeting microRNA (Baran-Gale et al., 2013).

We also recognize that few previously reported microRNAs were not selected through our analyses. MiR-204, a β-cell-enriched microRNA, was shown to target MAFA and block insulin production (Xu et al., 2013, 2016). However, another study demonstrated that although miR-204 is highly abundant in human islets, it did not significantly change (pro-)insulin gene transcript levels in human islets as well as in EndoC-betaH1 (Marzinotto et al., 2017). Another such candidate microRNA is miR-30d, a member of the miR-30 family that we reported to be islet enriched in human fetal islets (Joglekar et al., 2009c). MiR-30d was also reported to regulate (pro-)insulin gene transcription in mouse cells (Tang et al., 2009). In our analyses, both miR-204 and miR-30d were identified as highly significant (p < 2.349 × 10−19) microRNAs and expressed at higher levels in adult human islets (8.4-cycles or 28.4 = 338-fold higher for miR-204 and 2.5-cycles or 22.5 = 5.7-fold higher for miR-30d; Table S2). These microRNAs, although islet enriched, were not selected through the unbiased machine learning analytical workflow that we implemented in the identification of microRNAs associated with presence/abundance of (pro-)insulin transcript.

Forced microRNA overexpression was previously reported to enhance differentiation of human insulin-producing cells, most commonly using miR-375 (Nathan et al., 2015; Lahmy et al., 2014, 2016; Piran et al., 2017; Shaer et al., 2014; Williams et al., 2020). Similarly, miR-7 overexpression increased human embryonic stem (ES) cell differentiation (Lopez-Beas et al., 2018), whereas a cluster of microRNAs (miR-375, let-7a, let-7g, and miR-200a) was overexpressed using a polycistronic construct (Jin et al., 2019) during the differentiation of human ES cells into insulin-producing cells. The microRNAs that were previously reported to induce/improve differentiation to insulin-producing cells (miR-375, miR-200a, miR-7, and miR-9) are observed in our bootstrap analysis (Figure 1F). Interestingly, while our data confirm that miR-375 is significantly abundant in human islets (p = 1.7 × 10−21), miR-9 was identified to be significantly abundant (p = 0.02) in human insulin-negative tissues (Table S2). This is in line with the approach of miR-375 overexpression along with miR-9 inhibition, which was reported in the differentiation of human bone-marrow-derived stromal cells to insulin-producing lineage (Jafarian et al., 2015). MicroRNA-7 is reported to target Pax6 in mouse islet cells and reduce (pro-)insulin expression (Kredo-Russo et al., 2012), corroborating with our observations that miR-7 overexpression (in PANC1 or hIPCs) did not induce (pro-)insulin expression. The modest increase in (pro-)insulin transcripts or C-peptide production observed in our study may be due to the choice of microRNA overexpressing line(s), as well as the timing and dosage of microRNA overexpression or inhibition achieved in the study.

Targeted transient depletion of candidate microRNAs (Figure 6C) in human islets reduced (pro-)insulin expression, validating the previously reported role of microRNAs (Melkman-Zehavi et al., 2011) and miRNA processing (Kalis et al., 2011; Lynn et al., 2007) in reducing (pro-)insulin transcription in rodent models. Indeed, several of these microRNAs (Figure S1C) were significantly reduced in islets from organ donors with T2D, in agreement with data previously published using a smaller set of human islet samples and other (small RNA sequencing) technologies (Kameswaran et al., 2014). Even though knockdown of miR-216b-5p and miR-429-3p did not result in significant reduction of (pro-)insulin transcript, this could be the result of inefficient knockdown or higher interdependence/correlation between some families of microRNAs affecting function. Future studies with multiple replicates of human islet cells to confirm our observations are merited.

Our study strengths reside in the generation of this resource of microRNA expression profiles in a large number and mixture of human tissues using a previously verified sensitive and reliable OpenArray microRNA profiling platform. Compared to other studies on profiling microRNAs in human pancreas/islets (Correa-Medina et al., 2009; Joglekar et al., 2009a; Bolmeson et al., 2011; van de Bunt et al., 2013; Sebastiani et al., 2015; Klein et al., 2013; Grieco et al., 2017; Fred et al., 2010; Kameswaran et al., 2014), our microRNA data present one of the largest human tissue set (n = 353) of insulin-positive and insulin-negative tissues. Our design incorporates three separate sets of samples for use in discovery, validation, and prediction studies, as well as machine learning resampling approaches to take out sampling bias. Our study revealed correlations between this specific set of microRNAs and human (pro-)insulin transcripts in other tissues that naturally transcribe the (pro-)insulin gene. To our knowledge, this is the first report demonstrating an association of microRNA levels to (pro-)insulin transcript across a range of insulin-positive and insulin-negative primary human tissues. Our study also demonstrates the potential regulatory role of these microRNAs in human diabetes. Since functional loss of β cells is common to T1D (Atkinson et al., 2014; Kim et al., 2016) as well as T2D, our observations open future avenues for assessing the roles of specific microRNAs in diabetes progression. Another strength is the generation of this resource of microRNA overexpressing human pancreatic progenitor lines that allows the study of microRNA-target interactions. The use of analytical and validation tools led to the identification of microRNAs including miR-217-5p, miR-7-2-3p, miR-183-3p, and miR-183-5p that were not previously associated with human (pro-)insulin gene transcription. Without such an approach, the relevance and importance of these microRNAs in islet cells would not have been deciphered.

Limitations of the study

A limitation is the under-representation of some tissue types within our diverse sample set (Table S1), which was difficult to control due to its human origin. We had to work with available tissue sets while trying to include a diverse array of insulin-positive and insulin-negative tissues available through our collaborating investigators. The use of machine learning algorithms with resampling validation (Efron and Tibshirani, 1997, Chernick and Labudde, 2011) was therefore useful to eliminate any sampling bias. Since 10,000 resampling iterations delivered the same signature as with 1000 iterations, we used the latter to save on computing time. However, we acknowledge that the tissue specificity/enrichment of these microRNAs needs to be further examined using more numbers and different types (e.g. thymus) of insulin-producing tissues. The other study limitation is the focus on a selected set and combination of microRNAs presented in this study. We selected the top eight microRNAs from our logistic regression bootstrap analyses as they were also common to the linear regression workflow (Figure S1C). Since the top eight microRNAs (n = 8) could generate up to 255 different combinations of microRNA overexpressing cell lines, we were unable to plan for all of these combinations in the current study. Therefore, a single combination of most relevant microRNA lines was planned and analyzed. Our study presents an example (from 255 combinations) and offers this resource of human pancreatic microRNA-overexpressing lines and data sets for future studies. The association of microRNA levels to (pro-)insulin transcripts in human gallbladder and brain is intriguing. Although lower microRNA levels in these tissues do not establish causality, we also observe a significant reduction in several of our bootstrap microRNAs in islets from human T2D donors (with reduced (pro-)insulin levels). Mixing top eight microRNA-overexpressing lines in equal proportion did not significantly increase (pro-)insulin abundance, which suggests that although the machine learning approach identified the important microRNAs, the understanding of the timing and dosage of their expression/repression and the proportion of their individual contribution in regulating insulin transcription is necessary.

Methods

All methods can be found in the accompanying transparent methods supplemental file.

Resource availability

Lead contact

Anandwardhan A. Hardikar; email: A.Hardikar@westernsydney.edu.au.

School of Medicine, Western Sydney University, 30.2.27, Narellan Road & Gilchrist Drive.

Campbelltown, NSW 2560, Australia.

Materials availability

MicroRNA overexpressing pancreatic duct cell lines can be obtained with reasonable request to the lead contact.

Data and code availability

RNA-seq data used and analyzed in this study were deposited on the Gene Expression Omnibus (GEO) database and are available through the study accession number GSE134068. Our script used for penalized regression and bootstrap analyses carried herein are available through https://github.com/Isletbiology/Penalized-regression.