Use of the Fluidigm C1 platform for RNA sequencing of single mouse pancreatic islet cells

Significance

Pancreatic islets are complex structures composed of four cell types whose primary function is to maintain glucose homeostasis. Owing to the scarcity and heterogeneity of the islet cell types, little is known about their individual gene expression profiles. Here we used the Fluidigm C1 platform to obtain high-quality gene expression profiles of each islet cell type from mice. We identified cell-type–specific transcription factors and pathways providing previously unrecognized insights into genes characterizing islet cells. Unexpectedly, our data uncover technical limitations with the C1 Fluidigm cell capture process, which should be considered when analyzing single-cell transcriptomics data.

Abstract

This study provides an assessment of the Fluidigm C1 platform for RNA sequencing of single mouse pancreatic islet cells. The system combines microfluidic technology and nanoliter-scale reactions. We sequenced 622 cells, allowing identification of 341 islet cells with high-quality gene expression profiles. The cells clustered into populations of α-cells (5%), β-cells (92%), δ-cells (1%), and pancreatic polypeptide cells (2%). We identified cell-type–specific transcription factors and pathways primarily involved in nutrient sensing and oxidation and cell signaling. Unexpectedly, 281 cells had to be removed from the analysis due to low viability, low sequencing quality, or contamination resulting in the detection of more than one islet hormone. Collectively, we provide a resource for identification of high-quality gene expression datasets to help expand insights into genes and pathways characterizing islet cell types. We reveal limitations in the C1 Fluidigm cell capture process resulting in contaminated cells with altered gene expression patterns. This calls for caution when interpreting single-cell transcriptomics data using the C1 Fluidigm system.

Islets of Langerhans are miniature endocrine organs within the pancreas that are essential for control of blood glucose levels (1). They are composed of four endocrine cell types producing glucagon (α-cells), insulin (β-cells), somatostatin (δ-cells), and pancreatic polypeptide (PP cells). Whole-genome transcriptome analysis has been performed on enriched populations of human and mouse α- and β-cells (2–4). These studies report the ensemble average on the cell populations and do not report variation in expressed genes among cells. The studies also do not allow study of the low abundant δ-cells and PP cells. In addition, data interpretation in these analyses can be affected by the presence of a few contaminating cells. Single-cell RNA sequencing circumvents these problems and has recently been applied to a low number of human pancreatic islet cells (5) as well as to other cell types in complex tissues (6–13). Pancreatic islet cells are suited for single-cell RNA sequencing because they express high levels of a single hormone. This allows for unequivocal identification and unbiased understanding of gene expression in each cell type.

Here we used the C1 Fluidigm system to analyze the transcriptome of dispersed mouse pancreatic islet cells. We also studied how the capture process affected cell quality and contamination. We report identification of all islet cells with high-quality gene expression profiles. Unexpectedly, our data uncover technical limitations with the cell capture, which should be considered when analyzing single-cell transcriptomics data.

Results

Islet Cell Identity.

RNA FISH simultaneously using probes to glucagon (Gcg), insulin (Ins2), somatostatin (Sst), and pancreatic polypeptide (Ppy) showed that 99.2% (n = 15,542) of islet cells used for single-cell RNA sequencing expressed high levels of one hormone (Fig. 1A). The distribution of the cell types is shown in Fig. 1B. The intensity distributions of the fluorescence signal were bell-shaped, suggesting one cell population for each cell type (Fig. 1C). Interestingly, we detected few cells that coexpressed Gcg-Ppy (0.8%; n = 125) (Fig. 1D). Using RNA FISH and immunohistochemistry in pancreas sections from mice we confirmed the existence of rare Gcg⁺-Ppy⁺ cells (SI Appendix, Fig. S1). Consistent with the high sensitivity of RNA FISH (14), we detected low levels (0.02–0.3%) of other endocrine hormones in each single hormone-expressing cell. These data show that the dissociated islet cell preparations used for single-cell RNA sequencing consist nearly exclusively of single hormone-expressing cells.

Fig. 1.

Mouse islet cells rarely express more than one hormone. (A) Representative RNA FISH images of single mouse islet cells expressing glucagon (Gcg), insulin (Ins2), somatostatin (Sst), or pancreatic polypeptide (Ppy). (B) Distribution of islet cells. (C) Intensity distribution histograms of Gcg⁺, Ins2⁺, Sst⁺, or Ppy⁺ cells. (D) Representative RNA FISH images of Gcg⁺-Ppy⁺ cells.

Viability of Captured Cells.

We used two methods to determine viability of the captured cells. The first method is based on LIVE/DEAD staining in the C1 Single-Cell Auto Prep System. We found 77% live (LIVE⁺) cells, 2% dead (DEAD⁺) cells, and 21% cells that stained positive for both (LIVE⁺/DEAD⁺). Viability of the islet cells before capture was 78 ± 16% (n = 9 preparations). The second approach uses unsupervised hierarchical clustering of the top 100 variable genes in the sequenced cells. We used 622 cells from nine preparations for the analysis, after excluding 34 cells where debris or contaminating cells were observed (SI Appendix, Fig. S2). Twelve acinar cells were detected [≥1 reads per kilobase per million (RPKM) for ≥2 of the following genes: Amy2a5, Amy2b, or Pnlip]. This represents an exocrine contamination rate of 1.9% (12/622 cells). Two distinct cell clusters were identified: cells with low (cluster 1) or high viability (cluster 2) (Fig. 2A). Of note, mitochondrial genome-encoded genes are more abundantly expressed in cells in cluster 1. In particular, ATP6, ATP8, COX1, COX2, COX3, CYTB, ND1, Rnr2, and LOC100503946 are highly up-regulated and assigned as the cell viability gene set (Methods). These genes account for >30% of total expression in RPKM. Fig. 2B shows that the median expression of the cell viability gene set is 12-fold higher (P = 5.6e⁻²³) in cluster 1 cells, whereas the expression of all other genes is 285-fold (P = 6.0e⁻²³) reduced. Fig. 2C shows the distribution of the sequenced cells according to their viability score (Methods). Cells with a score >0.3 are likely to be of low quality and were removed from the analysis. We found no pattern of changes in cell quality throughout the C1 Fluidigm circuit (SI Appendix, Fig. S3). In total, the assessments of cell quality resulted in removal of 65 cells (10%; SI Appendix, Fig. S2).

Fig. 2.

Identification of islet cells with low viability. (A) Hierarchical clustering of top 100 variable genes in 622 mouse islet cells. Cluster 1 in the heat map characterizes cells with low viability (high viability score). Cluster 2 shows gene expression of cells mainly with high viability (low viability score). Viability score for each cell is depicted in the top horizontal bar. The gene set defining the viability score is marked by a black vertical bar left to the heat map. (B) Median expression of viability gene set and all other expressed genes in cells in cluster 1 and 2. (C) Distribution of cells according to their viability score. Cells with a viability score >0.3 (indicated by red dotted line) were removed from further analysis.

Characterization of Sequenced Islet Cells.

Each sample was sequenced to an average depth of 1 million read pairs (SI Appendix, Table S1). This sequencing depth was sufficient to detect expressed genes and did not improve when 64 cells were resequenced at 14.8 million read pairs (SI Appendix, Fig. S4A), even when considering low or highly expressed genes (SI Appendix, Fig. S4B). Average single-cell expression agreed well with two matching intact islet samples (r = 0.88 and 0.89) (SI Appendix, Fig. S5). The stochastic nature of single-cell gene expression seems highest for genes with average RPKM values <100 (SI Appendix, Fig. S6). The sequenced cells were also evaluated for technical quality (Methods). Thirty-seven cells failed to meet our criteria and were removed from further analysis.

Islet cell types were defined by their expression of Gcg (α-cell), Ins2 (β-cell), Sst (δ-cell), and Ppy (PP cell). Unexpectedly, of the 520 cells that passed viability and quality control assessments, only 341 cells (66%) expressed one hormone. Among the remaining 179 cells, 10 cells expressed low levels of any hormone (2%), whereas 169 cells (33%) expressed high levels of two or more hormones. These multiple-hormone–expressing cells showed gene profiles reminiscent of fused cells (Fig. 3A). This is supported by principal component analysis because the coexpressing cells typically cluster between their corresponding single-hormone–expressing cells (Fig. 3B). For example, Gcg⁺-Ins2⁺ cells mainly cluster between Gcg⁺ and Ins2⁺ cells (SI Appendix, Fig. S7). This, combined with the RNA FISH data of the input islet cell suspensions (cf. Fig. 1), suggests that nearly all multiple-hormone–expressing cells are artifacts that arise during the cell capture process due to damage or cell–cell fusion. Therefore, the cells that coexpress more than one hormone were excluded from subsequent analysis (SI Appendix, Fig. S2). Fig. 3C shows the distribution of the remaining single-hormone–expressing islet cells. The cells clustered into populations of α-cells (5%), β-cells (92%), δ-cells (1%), and PP cells (2%), matching the distribution in the input islet cell suspensions measured by RNA FISH. Fig. 3C also shows that each cell expresses low levels (0.003–0.27%) of other endocrine hormones. Total number of detected genes varied between 3,900 and 5,300 (SI Appendix, Table S2). These data show that single-cell capture and lower coverage sequencing can be used to profile gene expression of islet cells. The data reveal important technical limitations of the cell capture process resulting in large number of cells with false expression patterns.

Fig. 3.

Gene expression profiles of islet hormone-expressing cells. (A) Top 100 most significant genes in 510 cells with the following hormone-expressing patterns (color coding shown in B): Gcg⁺ (n = 18), Ins2⁺ (n = 313), Sst⁺ (n = 4), Ppy⁺ (n = 6), Gcg⁺-Ins2⁺ (n = 42), Gcg⁺-Ins2⁺-Ppy⁺ (n = 30), Ins2⁺-Sst⁺ (n = 9), Ins2⁺-Sst⁺-Ppy⁺ (n = 32), Ins2⁺-Ppy⁺ (n = 22), Gcg⁺-Ppy⁺ (n = 11), Gcg⁺-Sst⁺-Ppy⁺ (n = 2), Sst⁺-Ppy⁺ (n = 19), and Gcg⁺-Ins2⁺-Sst⁺-Ppy⁺ (n = 2). (B) Principal component analysis of the 510 cells. The first two principal components are depicted and each symbol represents a cell, and cells are color-coded by hormone-expressing pattern. (C) Distribution of Gcg⁺ (α-cells; n = 18), Ins2⁺ (β-cells; n = 313), Sst⁺ (δ-cells; n = 4), and Ppy⁺ (PP cells; n = 6). Each column represents gene expression in one cell. (D) Expression pattern of 721 transcription factors in single mouse islet cells. (E) Average expression of 42 abundant or known islet cell transcription factors in mouse α-cells (n = 18), β-cells (n = 313), δ-cells (n = 4), and PP cells (n = 6).

Transcription Factor Expression.

Previous work suggests that 150–300 transcription factors are expressed in mammalian tissues and constitute 5–8% of all expressed genes (15). Consistent with these data, we detected 372 out of 721 curated transcription factors (7.0–9.5% of expressed genes) with average RPKM ≥1 in at least one cell type (Fig. 3D and Dataset S1). Owing to the low number of identified δ-cells and PP cells and the stochastic nature of gene expression (cf. SI Appendix, Fig. S6), we limited the analysis to 42 abundant or previously reported transcription factors (Fig. 3E). The heat map is sorted by average expression in β-cells. Interestingly, α-cells and PP cells have similar expression patterns for transcription factors. These cells showed enriched expression for Arx, Mafb, Klf4, Atf3, Fosb, and Id3. α-Cells selectively expressed Pou3f4 (Brn4). On the contrary, δ-cells have the most distinct expression pattern. Hhex and Neurog3 are only expressed in this cell type and Pa2g4, Erg1, and Fos have enriched expression. δ-Cells are also characterized by lack of expression of Id3, Hdac2, Sin3a, Hnf1a, and Klf4 (Fig. 3E). β-Cells coexpress Pdx1, Nkx6-1, Nkx2-2, Pax6, and Neurod1, consistent with previous descriptions of genes expressed in mature β-cells (16, 17). Pax4 was not detected and MafA had expression <1 RPKM. These data confirm and expand our understanding of transcription factor expression in islet cells.

Enriched and Abundant α- and β-Cell Genes.

We identified 26 enriched genes in α-cells and 151 genes in β-cells. The average expression is summarized in Datasets S2 and S3. It is important to note that extensive variation in expression was observed for many of the genes (SI Appendix, Figs. S8–S10). Despite the variation in gene expression, we did not observe subpopulations of β-cells (SI Appendix, Fig. S11). The lower number of α-cells precluded meaningful subgroup analysis. Pathway analysis revealed that the genes were enriched in 18 pathways and molecular function gene sets (Fig. 4A). The primary function of the β-cell is to sense glucose and mount an appropriate insulin secretory response. Consistent with this, we found that genes involved in the sensing of glucose, cell signaling, and exocytosis were enriched in β-cells. In α-cells, gene enrichment analysis identified pathways regulating cell proliferation and signaling as well as protein synthesis and modification (Fig. 4A). We also identified abundantly expressed genes (average RPKM >100) in α- and β-cells. The average expression for these genes is summarized in Datasets S4 and S5. Pathway analysis identified the top three functional gene sets as pathways involved in oxidative phosphorylation, mitochondrial dysfunction, and EIF2 signaling (Fig. 4B). Collectively, the pathway gene enrichment analyses confirm function of the identified α- and β-cells.

Fig. 4.

α-Cell and β-cell pathways with enriched or abundant genes. (A) Pathways and functional gene sets with enriched genes in α-cells (n = 18) and β-cells (n = 312). (B) Pathways and functional gene sets with abundant genes in α-cells (n = 18) and β-cells (n = 312).

Discussion

Our data show that the C1 Fluidigm platform can be used for single-cell RNA sequencing, allowing identification of all islet cell types. We also demonstrate that half of the cells were damaged during the capture process, resulting in markedly altered gene expression patterns. Therefore, we have developed a workflow that allows identification of low-quality and contaminated cells. This critical evaluation of each captured and sequenced cell is possible because islet cells express high amounts of one hormone, allowing for unequivocal identification and unbiased understanding of gene expression profiles. The workflow can be adapted to any cell type with a distinct molecular gene signature. This is, however, not always possible, calling for caution when interpreting single-cell transcriptomics data using the C1 Fluidigm system.

RNA FISH analysis revealed that 99.2% of mouse islet cells express high levels of one hormone. Consistent with a previous report (16), we observed few Gcg⁺-Ppy⁺ cells. These double-hormone–positive cells are unlikely to be artifacts arising from the cell isolation procedure because they were also observed in intact islets in pancreas sections using RNA FISH and immunofluorescence staining. It is important to emphasize that islet cells do express very low levels (0.003–0.3%) of other endocrine hormones, consistent with a previous study (18). This could reflect low-level contamination, but if real the functional significance remains to be determined.

Our workflow revealed that 45% of captured cells did not meet our inclusion criteria for final analysis. Because the capture rate was 76% (656 captured cells/864 capture sites), the overall efficiency of the C1 Fluidigm system was 39%. Surprisingly, 27% of sequenced cells (169/622 cells) coexpressed more than one endocrine hormone. These cells are most likely artifacts because the islet cell suspension used for cell capture consisted of 99% single-hormone–expressing cells. The high sensitivity of RNA FISH and the detection of rare Gcg-Ppy double-positive cells make it unlikely that the other double-hormone–positive cells detected by the C1 Fluidigm system are real. The flow or pressure in the microfluidics system of the C1 cell capture circuit might somehow cause transient cell damage or cell–cell fusion. Our gene expression analysis suggests that cell–cell fusion might be frequently occurring in the C1 capture circuit. These liabilities hamper important utilities of the C1 Fluidigm system in islet cell research and possibly in all areas of biology.

Islet cells express between 3,900 and 5,300 genes, yet only 26 genes were enriched in α-cells and 151 genes in β-cells. This implies that a small number of genes control cell identify. This is perhaps not surprising because it has previously been shown that only three transcription factors control endocrine cell fate (19, 20). Our transcription factor analysis revealed few differences in their expression between α- and β-cells, as well as between α-cells and PP cells. δ-Cells had the most distinct transcription factor expression profile. The high degree of similarity between the islet cell types at the mRNA level might be important for cells to undergo, for example, transdifferentiation to meet metabolic demand (21, 22). Despite the great similarity between islet cells, we rarely detected cells that coexpress endocrine hormones. We observed high variability in the expression of many genes, yet we were unable to identify distinct subpopulations of β-cells. This was a surprising finding but might reflect the relatively high degree of stochastic variation in gene expression. In particular, we found that the variation in expression was highest for genes with RPKM <100, which represent the majority of the genes.

In conclusion, we describe the utility of the C1 Fluidigm platform to identify and sequence all mouse islet cell types. The islet cells show overall similar expression profiles, suggesting that few genes are likely to control cell identify and function. We uncovered liabilities in the C1 Fluidigm cell capture process leading to a high number of contaminated cells with markedly altered gene expression profiles. We describe a workflow that allows identification of low-quality and contaminated cells. We suggest adapting this workflow when analyzing single-cell RNA sequencing data for any cell type using the C1 Fluidigm platform.

Methods

Islet Cells.

C57BL/6 mice (males, 3–7 mo of age; Taconic) were housed in a controlled environment (12-h light/dark cycle, 22 ± 1 °C, 60–70% humidity) and fed standard chow for ad libitum consumption (Purina Laboratory Rodent Diet 5001; LabDiet). All animal procedures were conducted in compliance with protocols approved by the Regeneron Pharmaceuticals Institutional Animal Care and Use Committee. Islets were isolated by density gradient separation after perfusing the pancreas with Liberase TL (Roche) through the common bile duct. Following digestion for 13 min at 37 °C, the pancreas solution was washed and filtered through a 400-μm wire mesh strainer and islets were separated by Histopaque gradient centrifugation (Sigma). Islets were cultured in RPMI-1640 medium with 10% (vol/vol) FBS, 10 mM Hepes, 50 μM β-mercaptoethanol, 1 mM sodium-pyruvate, 100 U/mL penicillin, and 100 μg/mL streptomycin at 37 °C with 5% CO₂ in air atmosphere. Following overnight incubation, islets were hand-picked and enzymatically digested at 37 °C for 11 min using TrypLE Express (Life Technologies). Dissociated cells were filtered through a 40-μm cell strainer and suspended in RPMI-1640 medium.

Cell Capture, RNA Isolation, and Library Construction.

Single islet cells in RPMI-1640 medium (300–500 cells/μL) were mixed (3:2 ratio) with C1 Cell Suspension Reagent (Fluidigm) before loading onto a 10- to 17-μm-diameter C1 Integrated Fluidic Circuit (IFC; Fluidigm). LIVE/DEAD staining solution was prepared by adding 2.5 μL ethidium homodimer-1 and 0.625 μL calcein AM (Life Technologies) to 1.25 mL C1 Cell Wash Buffer (Fluidigm) and 20 μL was loaded onto the C1 IFC. Each capture site was carefully examined under a Nikon microscope in bright field, GFP, and Texas Red channels for cell doublets and viability. Cell lysing, reverse transcription, and cDNA amplification were performed on the C1 Single-Cell Auto Prep IFC, as specified by the manufacturer (protocol 100-7168 E1). The SMARTer Ultra Low RNA Kit (Clontech) was used for cDNA synthesis from the single cells. Illumina NGS library was constructed with Nextera XT DNA Sample Prep kit (Illumina), according to the manufacturer’s recommendations (protocol 100-7168 E1). Sequencing was performed on Illumina HiSeq2500 (Illumina) rapid mode by multiplexed single-read run with 50 cycles.

RNA Sequencing Data Analysis.

Raw sequence data (BCL files) were converted to FASTQ format via Illumina Casava 1.8.2. Reads were decoded based on their barcodes. Read quality was evaluated using FastQC (www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were mapped to the reference genome (mouse: GRCm38) using CLC bio Genomics Workbench Version 7.0 (CLC Bio) allowing one mismatch. Reads mapped to the exons of a gene were summed at the gene level. Sequencing statistics including mapped counts, unique exon counts, and intron and intergenic counts were summarized and checked for outliers. Expression correlation between samples was calculated using log2-scale Pearson correlation. Principal component analysis was used to examine outliers and potential batch effect. Depending on variances explained in principal components, we used the first three to five components to select top genes accounting for most of the variances. The top 100 genes were selected based on maximum loading of the first few principal components (8). Cells were removed from further analysis using the following criteria: (i) exocrine pancreas contamination (≥1 RPKM) based on marker genes (Amy2a5, Amy2b, and Pnlip); (ii) permeable or dying cells with >0.3 viability score (discussed below) and/or dead cells based on DEAD/LIVE cell staining; (iii) cells with high expression of nonislet tissue markers; (iv) cells with <100,000 exon counts, low exon-to-mapped ratio (<0.2), or high intergenic-to-mapped ratio (>0.3); and (v) outliers using hierarchical clustering. Mann–Whitney test was used for statistical analysis of data Fig. 2B. Single-cell transcriptome data are deposited in the Gene Expression Omnibus.

Identification of Islet Cells.

Islet cell types were identified using densityMclust (Mclust in R) to estimate bimodal expression distribution of Gcg, Ins2, Sst, and Ppy. To obtain more reliable cell-type identification, cells were excluded if their expression is >2 SDs from the average of the high expression mode.

Cell Viability Score.

We defined the following viability score to measure the quality of cells:

$${\displaystyle \sum }_i{g}_i/{\displaystyle \sum }_j{G}_j,$$

where g_i is one of the abundant genes (ATP6, ATP8, COX1, COX2, COX3, CYTB, ND1, Rnr2, and LOC100503946) and G_j is one of the annotated genes. A viability score >0.3 was used to identify cells with low viability.

Evaluation of Gene Expression Variation.

One-way ANOVA was used to identify gene expression variation in cells with the following hormone expression patterns: Gcg⁺, Ins2⁺, Sst⁺, Ppy⁺, Gcg⁺-Ins2⁺, Gcg⁺-Ins2⁺-Ppy⁺, Ins2⁺-Sst⁺, Ins2⁺-Sst⁺-Ppy⁺, Ins2⁺-Ppy⁺, Gcg⁺-Ppy⁺, Gcg⁺-Sst⁺-Ppy⁺, Sst⁺-Ppy⁺, and Gcg⁺-Ins2⁺-Sst⁺-Ppy⁺. After filtering out genes with average expression ≤5 RPKM in α-cells, β-cells, δ-cells, and PP cells, the top 100 most significant genes were selected based on false discovery rate to illustrate gene signatures.

Cell-Type Enriched and Abundant Genes.

Expressed genes were defined by ≥1 RPKM. Abundant genes were defined by (i) average expression >100 RPKM in the selected cell type and (ii) >50% of cells in the selected cell type with expression >1 RPKM. DESeq2 was used to identity enriched α- and β-cell genes according to (i) expression in the selected cell type is >10-fold compared with the other cell type, (ii) false discovery rate <0.01, (iii) average expression in the selected cell type >10 RPKM, and (iv) >50% of cells in the selected cell type with expression >1 RPKM.

RNA in Situ Hybridization and Immunofluorescence.

Dissociated mouse islet cells were fixed in 10% neutral buffered formalin and centrifuged onto charged slides. Whole mouse pancreata were fixed in 10% neutral buffered formalin. After fixation pancreata were paraffin-embedded and sectioned onto slides. For RNA analysis, pancreas tissue and cells were permeabilized and hybridized with combinations of mRNA probes for mouse Gcg, Ins2, Sst, and Ppy, according to the manufacturer’s instructions (Advanced Cell Diagnostics). A fluorescent kit was used to amplify mRNA signal. For protein analysis, pancreas sections were stained with a combination of an antiglucagon (REGN745, an antiglucagon monoclonal antibody generated in-house), an antiinsulin (Dako A0564), an antisomatostatin (Sigma SAB4502861), or an antipancreatic polypeptide (Sigma SAB2500747) antibody. Fluorescent signal was detected using a microscope slide scanner (Zeiss Axio Scan.Z1). Islet cell types were quantified using the HALO image analysis using the Cytonuclear Fluorescence module (Indica Labs).