Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

# An Automated Microwell Platform for Large-Scale Single Cell RNA-Seq

## Abstract

Recent developments have enabled rapid, inexpensive RNA sequencing of thousands of individual cells from a single specimen, raising the possibility of unbiased and comprehensive expression profiling from complex tissues. Microwell arrays are a particularly attractive microfluidic platform for single cell analysis due to their scalability, cell capture efficiency and compatibility with imaging. We report an automated microwell array platform for single cell RNA-Seq with significantly improved performance over previous implementations. We demonstrate cell capture efficiencies of >50%, compatibility with commercially available barcoded mRNA capture beads and parallel expression profiling from thousands of individual cells. We evaluate the level of cross-contamination in our platform by both tracking fluorescent cell lysate in sealed microwells and with a human-mouse mixed species RNA-Seq experiment. Finally, we apply our system to comprehensively assess heterogeneity in gene expression of patient-derived glioma neurospheres and uncover subpopulations similar to those observed in human glioma tissue.

## Introduction

Single cell RNA-Seq is a powerful approach to quantifying cellular heterogeneity with both basic and clinical research applications1,2,3,4. As a result, considerable effort has been devoted to increasing the throughput and accuracy of these methods including the introduction of unique molecular identifiers (UMIs)5 and barcoding techniques that facilitate pooled library construction6. Recent advances in single cell RNA-Seq have resulted in dramatically increased scalability with a concomitant reduction in library preparation costs7,8,9,10,11. Microfluidic technology has played a crucial role in the advancement of single cell expression analysis by reducing reagent volumes, allowing high-fidelity single cell isolation and enabling robust and automated workflows for RNA extraction and amplification12,13,14,15. New tools for single cell RNA-Seq exploit highly scalable microfluidic platforms, including aqueous droplets7,8,10 and microwell arrays9,11 and have facilitated miniaturization of split-pool barcoding methods for labeling cDNA libraries from hundreds or thousands of individual cells in parallel. These techniques are leading to new applications of single cell RNA-Seq including large-scale, unbiased analysis of tissues and tumors without the need for cell sorting7.

We recently reported single cell RNA-Seq in a solid-state microwell array platform9. Microwell arrays have several important advantages over droplet-based devices for single cell analysis including low sample and reagent dead volume, short cell loading time and enhanced compatibility with short-term cell culture, cell perturbation assays and optical imaging16,17,18. The last two features are particularly useful in minimizing sample degradation prior to cell lysis and allow the experimenter to examine and tune cell loading, identify multiplets or cell debris and use fluorescence microscopy to determine marker composition and cell viability. In addition, high-efficiency capture of individual cells from a small sample is relatively straightforward with microwells, because cells and beads can be loaded into microwells by repeatedly flowing them over the array until all of them are captured by gravity. While our original system was capable of profiling a few hundred cells per experiment with library preparation costs of $0.10–$0.20 per cell, it suffered from several key drawbacks including low cell and molecular capture efficiency and a lack of automation9. Here, we report significant improvements of microwell-based single cell RNA-Seq in these three areas with no effect on overall cost. In addition, we demonstrate the compatibility of this system with the simple, 3′-end library preparation scheme SCRB-Seq19 and the commercially available barcoded “Drop-Seq” capture beads reported by Macosko et al.7. The level of cross-contamination between wells is critically evaluated by both imaging fluorescently tracked cell lysate in oil sealed microwells and a human-mouse mixed species experiment. To demonstrate the utility of our method, we applied it to patient-derived glioma neurospheres and observed multiple phenotypic subpopulations that resemble features of intratumoral heterogeneity in glioblastoma.

## Results

### An Automated Microwell Platform for Single Cell RNA-Seq

As described above, cell-bead pairing occurs randomly when both entities are loaded by gravity. Once this manual step is complete, cell lysis and reverse transcription occur on a computerized fluidics and temperature control system (Fig. 2D). We use a thermoelectric module for temperature control and an electronic, rotary selector valve to introduce different solutions to the device and reversibly seal and unseal the microwell array9,17. In our original report, we sealed the microwell array after introducing a lysis buffer containing a mild detergent9. We then used freeze-thaw cycles to initiate cell lysis, trap individual cell lysates in the microwells and capture the liberated mRNA on barcoded beads9. This approach is relatively inefficient and requires low temperatures that are incompatible with automation. For efficient cell lysis, a strongly denaturing buffer that can rapidly disrupt cell membranes and deactivate nucleases would be ideal, but rapid sealing of the microwells is essential to minimize material loss and cross-contamination. Our automated system allows multiple fluids to be introduced in rapid succession, enabling the use of efficient lysis buffers without significant material loss prior to sealing. For cell lysis, we introduce a denaturing lysis buffer containing guanadinium isothiocyanate. We then rapidly introduce perfluorinated oil to seal the microwell array22 before cell lysis occurs. Figure 2E shows the lysates of isolated, fluorescently labeled cells in a microwell array following automated cell lysis and sealing. On-chip fluorescence imaging facilitates quality-control of cell viability and lysis and microwell sealing quality while providing a simple means of counting the number of cell-bead pairs and multiplet loading rate in every experiment.

Following cell lysis and mRNA capture, we introduce a detergent-containing buffer to rapidly remove the oil sealant and cell lysates. At this point the barcoded beads with hybridized mRNA are exposed to the microfluidic channel located above the microwells and the automated system introduces all of the reagents required for reverse transcription at the appropriate temperature. Here, we use the SCRB-Seq protocol19 similar to what was reported for Drop-Seq7 and so the reverse transcription reaction also includes a template-switching step to generate full-length cDNA with universal sequence adapters on both the 3′- and 5′-ends. Once the reverse transcription reaction is complete, we disconnect the device, remove the beads from the microwells by gentle sonication, gravity and detergent-containing buffer flow and complete the library construction procedure as described previously7. The empty device is then imaged by microscopy to measure bead extraction efficiency, which typically exceeds 99%. Note that there are still a few steps of the library construction procedure that require human intervention, including two PCR reactions. Further system development is required to fully automate the library construction procedure.

### High-Quality Large-Scale Single Cell RNA-Seq Profiling with an Automated Microwell System

To characterize the performance of our system, we obtained RNA-Seq profiles of ~3,000 individual cells from a mixture of the human glioma cell line U87-MG and the murine fibroblast cell line NIH-3T3 in a single experiment. As shown previously, mixed species analysis is an effective approach to assessing cross-talk and purity, particularly in pooled single cell RNA-Seq experiments7,8. We chose the U87-MG and NIH-3T3 cell lines in order to compare the performance of our system to previous studies. We sequenced U87-MG cells in our initial report of microwell-based single cell RNA-Seq9 and NIH-3T3 cells were sequenced in the original report of Drop-Seq7.

Figure 3A shows a histogram of the fraction of molecules that uniquely aligned to either the human or murine transcriptome but that aligned best to the human transcriptome for each cell. The bimodal distribution indicates that almost all of the molecules detected for roughly half of the cells originate from human mRNA versus murine mRNA for the remaining half. Because the original mixture was comprised of about 50% human and 50% murine cells, this implies that our single cell RNA-Seq profiles are quite pure (median purity of >98.8%). Cell barcodes associated with a significant number of both human and murine transcripts (<90% purity for the species with the most transcripts) likely originate from “multiplets” or instances in which two or more cells of both species were captured in a single microwell (<0.8% of cell barcodes).

An additional indicator of purity and performance is the ability to detect subtle phenotypic subpopulations. For example, expression heterogeneity due to cell cycle asynchrony is a hallmark of single cell RNA-Seq profiles of mitotic cells. Figure 3B,C show heatmaps containing cell cycle state scores for both human U87-MG cells and murine NIH-3T3 cells from this dataset (see Methods). Here, we can clearly distinguish cells in each of five stages of the cell cycle from each other as well as groups of cells transitioning between stages.

Figure 3D–G show the distributions of numbers of molecules and genes detected per cell as well as saturation curves for molecule and gene detection for U87-MG cells. In our original report, we detected an average of <1,000 genes per U87 cell9, but here we detect ~4,800 genes per U87 cell on average. Hence, our automated microwell system has significantly higher molecular capture efficiency than our initially reported system. Similarly, Fig. 3H–K show the same analysis for individual murine NIH-3T3 cells. Our molecular and gene detection efficiencies are similar for the two cell lines. We detect ~25,000 molecules and ~4,600 genes per NIH-3T3 cell on average, similar to what was reported for Drop-Seq for the same cell line7. On average, we obtained ~208,000 raw reads per cell. Importantly, we note that neither our U87-MG nor our NIH-3T3 libraries have been sequenced to saturation. Therefore, this analysis represents an underestimate of our actual molecular and gene detection efficiencies.

We also compared our sensitivity to that of the Fluidigm C1 system for the same cell line using a publically available data set in which individual 3T3 cells were sequenced (Supplementary Fig. S4)7. Because UMIs were not implemented in these experiments, we cannot make a direct comparison of our molecular capture efficiency, but we can compare the number of genes detected per cell. We found that, at full coverage (~1 million uniquely aligned reads per cell), the Fluidigm system detected ~8,800 genes per cell on average. However, when we down-sampled the Fluidigm C1 data to ~42,000 uniquely aligned reads per cell (similar to what we obtained for 3T3 cells in this study), the Fluidigm system detected ~5,300 genes per cell. While this is comparable to the number of genes that we detected in this same cell line, the Fluidigm C1 libraries likely require more reads to reach saturation due to their full gene body coverage than our libraries in which we sequence only the 3′-end. Hence, the Fluidigm C1 library complexity and detection efficiency are most likely considerably higher than those of our platform at saturating coverage.

### Glioma Neurospheres Preserve Key Features of Intratumoral Heterogeneity based on Large-scale Single Cell RNA-Seq

We obtained RNA-Seq profiles of >2,200 individual cells from a patient-derived glioma neurosphere culture in a single experiment. The performance of our automated microwell array platform with these neurospheres is summarized in Supplementary Fig. S3. The mean numbers of molecules (Supplementary Fig. S3A) and genes (Supplementary Fig. S3B) detected per cell are ~14,000 and ~3,300 respectively. On average, we obtained ~303,000 raw sequencing reads per cell for TS543 cells. Saturation analysis of both the numbers of detected molecules (Supplementary Fig. S3C) and genes (Supplementary Fig. S3D) suggests that our current sequencing depth is close to saturation.

Glioma neurospheres represent an important model system for brain tumors because, in many cases, they more effectively preserve the phenotypic and genotypic features of tumors than conventional monolayer cultures23. They have been widely used to study drug response, glioma stem cells and tumor progression as xenograft models23,24,25. However, to our knowledge, glioma neurospheres have not been analyzed comprehensively by single cell RNA-Seq to determine the extent of phenotypic heterogeneity and co-occurrence of cellular subpopulations within a single culture. Expression profiling of surgical specimens from glioma patients by The Cancer Genome Atlas has established classifier gene sets that stratify tumors into distinct subtypes26. Recent studies employing bulk expression analysis of regional heterogeneity27 and single cell RNA-Seq28 have shown that gene signatures corresponding to different patient subtypes co-occur within individual gliomas. We analyzed single cell expression profiles obtained from TS543 cells, a glioma neurosphere line that most closely resembles the Proneural glioma subtype and harbors amplification of PDGFRA, a genetic alteration associated with Proneural gliomas29. We used unsupervised dimensionality reduction and density-based cluster assignment that was uninformed of the identities of the glioma classifier genes (taken from Table S3 of Verhaak et al.26) to show that individual TS543 cells are comprised of at least two clear phenotypic subpopulations (Fig. 4A). For simplicity, we refer to these subpopulations as the red cluster and blue cluster. The median number of molecules detected per cell in the red and blue clusters was 11,382 and 9,771, respectively, suggesting that coverage is not a major driver of the separation between these two subpopulations. As expected, we found that Proneural genes are more commonly expressed in the majority of TS543 cells than genes from either the Classical or Mesenchymal subtypes. However, when we project expression of subtype-specific genes onto our clustering analysis, we find considerable expression heterogeneity among the classifier genes. For example, above-median expression of the Proneural classifier genes (Fig. 4B) is significantly enriched in the blue cluster (p < 10−6, hypergeometric test) whereas above-median expression of both Classical (Fig. 4C) and Mesenchymal (Fig. 4D) genes is significantly enriched in the red cluster (p < 10−6 for both gene sets). This phenomenon is reminiscent of the “hybrid cellular states” observed in by Patel et al. among individual cells in human glioblastoma tissue specimens28. Hence, our results suggest that glioma neurosphere cultures can recapitulate the subtype-specific expression heterogeneity found in human glioma tissue.

## Discussion

In our previously reported system, we achieved library preparation costs of ~$0.10–$0.20/cell9. The data set presented here included >5,000 cells from two experiments and was obtained with library preparation costs of $0.11/cell and sequencing costs of$0.48/cell. Taken together, the improvements described here have resulted in a microfluidic system for single cell RNA-Seq that is compatible with imaging and can detect thousands of genes across thousands of individual cells with a cell capture efficiency >50% and library preparation costs that are almost negligible compared to the cost of sequencing. Due to the enormous barcoding capacity of the Drop-Seq beads7 and the parallel fashion in which cells and beads are loaded into our prefabricated microwells, throughput of our platform, when necessary, can be further scaled up to hundreds of thousands of cells per run simply by increasing the number of microwells in a single lane and the number of lanes on a single device while keeping the time required for cell/bead loading short which is important to minimize sample degradation prior to cell lysis.

Conventional approaches to single cell analysis such as microscopy and flow cytometry are routinely employed to analyze thousands of individual cells from complex tissues. With the development of new microfluidic tools7,8,9,10,11 and an appreciation that important subpopulations can be identified with relatively shallow sequencing coverage14, genome-wide analysis of individual cells is beginning to reach a similar scale. As a result, new applications can be contemplated including comprehensive identification of cell types throughout an organism, simultaneous, unbiased characterization of transformed and stromal cells from solid tumors and detection of rare cellular subpopulations that give rise to drug resistance.

## Methods

### Fabrication of PDMS Microwell Flow Cell Devices for Large-Scale Single Cell RNA-Seq

The devices are fabricated using standard SU-8 soft lithography30. SU-8 wafer molds are designed in Draftsight (http://www.3ds.com/products-services/draftsight-cad-software/). The diameter, height of each well and center-to-center distance between neighbor wells are 50 μm, 58 μm and 75 μm, respectively. The height of the flow cell is 112 μm. Silanized SU-8 silicon wafer molds are obtained from FlowJEM (http://www.flowjem.com/). PDMS (Sylgard 184, Dow Corning) base and curing agent are thoroughly mixed at the ratio of 10:1, degassed under house vacuum in a desiccator (Z354074, Sigma-Aldrich) for 2 hours and poured onto the SU-8 wafer molds in containers made of aluminum foil (01-213-100, Fisher Scientific). The degassed PDMS mixture is then cured in a 90 °C oven (414004-556, VWR) for 2 hours. PDMS slabs are then gently peeled off from the molds. A 1.75 mm OD biopsy punch (15110-15, Ted Pella) is used to create inlet and outlet of flow cells. One PDMS slab with microwells and one PDMS slab with flow cell are treated in a plasma cleaner (PDC-32G, Harrick Plasma) for 30 seconds and then covalently bonded together to form the final microwell flow cell device.

### Computer-Controlled Automation for Microwell-Based Single Cell RNA-Seq

A schematic of the computer-controlled automation system is shown in Fig. 2D. The system consists of both temperature and fluidic control systems. Temperature control of the PDMS device is realized by directly mounting the PDMS device on top of a thermoelectric heater/cooler (CP-031, TE Technology) which is controlled through a bi-polar temperature controller (TC-36-25-RS232, TE Technology). A multi-channel selector valve (MLP777-605, IDEX Health & Science), located at the upstream of the PDMS device, is deployed to control which reservoir is connected to the device. A three-way solenoid valve (EW-01540-11, Cole-Parmer), located at the downstream of the PDMS device, is used as an on/off switch of the flow. Fluid flow is driven by a constant pressure source (3 psi) stabilized by a pressure regulator (AW20-F02, SMC Pneumatics). Because the on/off switch is located at the downstream of the device, the device is under a constant positive pressure during any incubation steps. This feature is crucial for preventing bubble formation in the device, especially at elevated temperatures such as during the reverse transcription step. To minimize dead volume, tubing with small inner diameter (127 μm, 37005T, Fisher Scientific) is used to connect reagent reservoirs and inlet of the device and that the length of the tubing is kept at minimum. This way, we are able to keep the dead volume below 10 μL which is less than the total volume of the device itself (20 to 250 μL depending on the number of wells the device has). The multi-channel selector valve is controlled by a USB digital I/O device (NI USB-6501, National Instruments). The three-way solenoid valve is controlled by the same USB digital I/O device, but through a homemade transistor-switch circuit. A C program is used to control the system.

### Data Processing Procedure for Microwell-Based Single Cell RNA-Seq

Even after the filtering procedures described above, we obtain more cell barcodes than the number of cell-bead pairs loaded in our device. These additional cell barcodes arise from several sources including additional sequencing or synthesizer errors and the beads that are not paired with a cell, which can capture low levels of ambient RNA during the experiment. Nonetheless, as shown in Supplementary Fig. S2, we can readily identify a population of very high coverage barcodes based on the distribution of captured molecules that is consistent with the number of cell-bead pairs imaged in our device. Similar observations have been made in previous studies7,9.

### Cell Cycle Analysis

We adopted the cell cycle analysis method developed by Macosko et al.7. Please refer to the original paper for details. Briefly, the expression level of a set of genes that are known to reflect different phases of cell cycle were used to calculate a phase-specific score for each cell. Each cell is then classified into one of the ten patterns of phase-specific scores (including eight potential patterns along the cell cycle and two patterns for equal scores of all phases (either all active or all inactive)) based on the maximal correlation of the cell’s phase-specific score with these ten patterns. Cells within each class were further ordered based on their relative correlation with the preceding and succeeding patterns. The set of genes used to calculate the phase-specific scores were obtained from the Supplemental Fig. 15 in Whitfield et al. which reflect five phases of cell cycle (G1/S, S, G2, G2/M, M/G1)31. The eight potential patterns along the cell cycle that the cells were classified into are: only G1/S is on, both G1/S and S are on, only S is on, both S and G2 are on, only G2 is on, both G2 and G2/M are on, only G2/M is on, both G2/M and M/G1 are on.

### Clustering Analysis of Single Cell Expression Profiles

We clustered our TS543 single cell expression profiles using a set of highly variable genes identified based on a dispersion analysis of the entire data set. We first normalized the molecular counts for each gene in each cell by the total number of molecules detected in that cell. We considered these normalized molecular counts to be expression levels. Next, we plotted the coefficient of variation vs. mean expression across all genes detected in at least five cells and grouped the genes into 50 evenly-spaced bins based on log-transformed expression levels. We computed a z-score for each bin and took genes with a z-score greater than three to be highly variable given their expression levels as long as they were detected in at least 10% of cells (see Supplementary Table S1 for a complete list). Hence, the variance in these genes is less likely to result from technical noise and more likely to result from real biological variation. We then computed a matrix of Pearson correlation coefficients between the log-transformed expression profiles of each cell using only the highly variable genes. Finally, we used this Pearson correlation matrix as input to the t-stochastic neighborhood embedding (t-SNE) algorithm32 for unsupervised clustering as implemented in the Python scikit-learn package. The results of the t-SNE clustering are displayed in Fig. 4. We assigned cells to discrete clusters by density analysis with the DBSCAN function in scikit-learn using the Euclidean distance metric.

We used the following score, Ssubtype,i, to assess expression of glioma subtype-specific genes in an individual cell i:

where nsubtype,i is the number of subtype-specific genes detected in cell i, Nsubtype is the number of subtype-specific genes detected in the entire dataset and ngenes,i is the number of genes detected in cell i.

### Analysis of Single Cell RNA-Seq Data Generated by the Fluidigm C1 System

As described above, we sequenced NIH-3T3 murine fibroblasts as part of a performance test for our system. This same cell line was sequenced using the Fluidigm C1 system by Macoscko et al.7. We downloaded the raw SRA data for these experiments from GEO accession GSE701151 and converted these data to 192 fastq files, corresponding to 192 single cell profiles using fastq-dump in the SRA Toolkit package. We then aligned each fastq file to a concatenated human-mouse pre-assembled transcriptome using bwa-mem and identified uniquely aligned reads just as described above. Because the Fluidigm C1 data set originated from a mixed species experiment in which human HEK cells were mixed with murine 3T3 cells, we identified cells with >90% of the reads aligned to the murine transcriptome and quantified the number of genes detected per cell at two different read depths (Supplementary Fig. S4).

Accession codes: The RNA-Seq data generated in this study has been deposited in the Gene Expression Omnibus hosted by the National Center for Biotechnology Information under accession GSE85575. http://www.nature.com/srep

How to cite this article: Yuan, J. and Sims, P. A. An Automated Microwell Platform for Large-Scale Single Cell RNA-Seq. Sci. Rep. 6, 33883; doi: 10.1038/srep33883 (2016).

## References

• Eberwine, J., Sul, J. Y., Bartfai, T. & Kim, J. The promise of single-cell sequencing. Nat Methods 11, 25–27 (2014).

• Tang, F., Lao, K. & Surani, M. A. Development and applications of single-cell transcriptome analysis. Nat Methods 8, S6–11 (2011).

• Tang, F. et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nat Methods 6, 377–382 (2009).

• Islam, S. et al. Characterization of the single-cell transcriptional landscape by highly multiplex RNA-seq. Genome Res 21, 1160–1167 (2011).

• Islam, S. et al. Quantitative single-cell RNA-seq with unique molecular identifiers. Nat Methods 11, 163–166, 10.1038/nmeth.2772 (2014).

• Hashimshony, T., Wagner, F., Sher, N. & Yanai, I. CEL-Seq: single-cell RNA-Seq by multiplexed linear amplification. Cell Rep 2, 666–673 (2012).

• Macosko, E. Z. et al. Highly Parallel Genome-wide Expression Profiling of Individual Cells Using Nanoliter Droplets. Cell 161, 1202–1214, 10.1016/j.cell.2015.05.002 (2015).

• Klein, A. M. et al. Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell 161, 1187–1201, 10.1016/j.cell.2015.04.044 (2015).

• Bose, S. et al. Scalable microfluidics for single-cell RNA printing and sequencing. Genome Biol 16, 120, 10.1186/s13059-015-0684-3 (2015).

• Rotem, A. et al. High-Throughput Single-Cell Labeling (Hi-SCL) for RNA-Seq Using Drop-Based Microfluidics. PLoS One 10, e0116328, 10.1371/journal.pone.0116328 (2015).

• Fan, H. C., Fu, G. K. & Fodor, S. P. Expression profiling. Combinatorial labeling of single cells for gene expression cytometry. Science 347, 1258367, 10.1126/science.1258367 (2015).

• Dalerba, P. et al. Single-cell dissection of transcriptional heterogeneity in human colon tumors. Nat Biotechnol 29, 1120–1127 (2011).

• Streets, A. M. et al. Microfluidic single-cell whole-transcriptome sequencing. Proc Natl Acad Sci USA 111, 7048–7053, 10.1073/pnas.1402030111 (2014).

• Pollen, A. A. et al. Low-coverage single-cell mRNA sequencing reveals cellular heterogeneity and activated signaling pathways in developing cerebral cortex. Nat Biotechnol 32, 1053–1058, 10.1038/nbt.2967 (2014).

• Marcus, J. S., Anderson, W. F. & Quake, S. R. Microfluidic single-cell mRNA isolation and analysis. Analytical chemistry 78, 3084–3089 (2006).

• Love, J. C., Ronan, J. L., Grotenbreg, G. M., van der Veen, A. G. & Ploegh, H. L. A microengraving method for rapid selection of single cells producing antigen-specific antibodies. Nature Biotechnology 24, 703–707 (2006).

• Sims, P. A., Greenleaf, W. J., Duan, H. & Xie, X. S. Fluorogenic DNA sequencing in PDMS microreactors. Nat Methods 8, 575–580 (2011).

• Gracz, A. D. et al. A high-throughput platform for stem cell niche co-cultures and downstream gene expression analysis. Nat Cell Biol 17, 340–349, 10.1038/ncb3104 (2015).

• Soumillon, M., Cacchiarelli, D., Semrau, S., Van Oudenaarden, A. & Mikkelsen, T. S. Characterization of directed differentiation by high-throughput single-cell RNA-Seq. bioRxiv (2014).

• Shiroguchi, K., Jia, T. Z., Sims, P. A. & Xie, X. S. Digital RNA sequencing minimizes sequence-dependent bias and amplification noise with optimized single-molecule barcodes. Proc Natl Acad Sci USA 109, 1347–1352 (2012).

• Kivioja, T. et al. Counting absolute numbers of molecules using unique molecular identifiers. Nat Methods 9, 72–74 (2011).

• Zhang, H., Nie, S., Etson, C. M., Wang, R. M. & Walt, D. R. Oil-sealed femtoliter fiber-optic arrays for single molecule analysis. Lab Chip (2012).

• De Witt Hamer, P. C. et al. The genomic profile of human malignant glioma is altered early in primary cell culture and preserved in spheroids. Oncogene 27, 2091–2096, 10.1038/sj.onc.1210850 (2008).

• Laks, D. R. et al. Neurosphere formation is an independent predictor of clinical outcome in malignant glioma. Stem Cells 27, 980–987, 10.1002/stem.15 (2009).

• Niola, F. et al. Id proteins synchronize stemness and anchorage to the niche of neural stem cells. Nat Cell Biol 14, 477–487, 10.1038/ncb2490 (2012).

• Verhaak, R. G. et al. Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR and NF1. Cancer cell 17, 98–110 (2010).

• Sottoriva, A. et al. Intratumor heterogeneity in human glioblastoma reflects cancer evolutionary dynamics. Proc Natl Acad Sci USA 110, 4009–4014, 10.1073/pnas.1219747110 (2013).

• Patel, A. P. et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science 344, 1396–1401, 10.1126/science.1254257 (2014).

• Silber, J. et al. miR-34a repression in proneural malignant gliomas upregulates expression of its target PDGFRA and promotes tumorigenesis. PLoS One 7, e33844, 10.1371/journal.pone.0033844 (2012).

• Xia, Y. N. & Whitesides, G. M. Soft lithography. Annu Rev Mater Sci 28, 153–184, 10.1146/annurev.matsci.28.1.153 (1998).

• Whitfield, M. L. et al. Identification of genes periodically expressed in the human cell cycle and their expression in tumors. Mol Biol Cell 13, 1977–2000, 10.1091/mbc.02-02-0030 (2002).

• Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. Journal of Machine Learning Research 9, 85 (2008).

## Acknowledgements

The authors thank Dr. Sohani Das Sharma for assistance with cell culture and preparation, Erin C. Bush for assistance with library preparation and sequencing and Dr. Harris Wang for the loan of a syringe pump. P.A.S. is supported by K01EB016071 from NIH/NIBIB, R33CA202827 from NIH/NCI and U54CA193313 from NIH/NCI.

## Author information

Authors

### Contributions

J.Y. and P.A.S. conceived and designed the automated microwell array system. J.Y. fabricated the microwell array devices, constructed the automated system, implemented the library construction protocol and generated the single cell RNA-Seq data. J.Y. and P.A.S. analyzed the data and wrote the paper.

## Ethics declarations

### Competing interests

Columbia University has filed a patent application based on this work.

## Rights and permissions

Reprints and Permissions

Yuan, J., Sims, P. An Automated Microwell Platform for Large-Scale Single Cell RNA-Seq. Sci Rep 6, 33883 (2016). https://doi.org/10.1038/srep33883

• Accepted:

• Published:

• DOI: https://doi.org/10.1038/srep33883

• ### Capture and reagent exchange (CARE) wells for cell isolation, labeling, and characterization

• Kevin Loutherback
• Allan B. Dietz

Microfluidics and Nanofluidics (2022)

• ### Deconvolution of cell type-specific drug responses in human tumor tissue with single-cell RNA-seq

• Wenting Zhao
• Athanassios Dovas
• Peter A. Sims

Genome Medicine (2021)

• ### High-throughput and single-cell T cell receptor sequencing technologies

• Joy A. Pai
• Ansuman T. Satpathy

Nature Methods (2021)

• ### Mapping a mammalian adult adrenal gland hierarchy across species by microwell-seq

• Shujing Lai
• Lifeng Ma
• Guoji Guo

Cell Regeneration (2020)

• ### How single-cell immunology is benefiting from microfluidic technologies

• Fabien C. Jammes
• Sebastian J. Maerkl

Microsystems & Nanoengineering (2020)