Abstract
High-throughput single-cell RNA-seq methods assign limited unique molecular identifier (UMI) counts as gene expression values to single cells from shallow sequence reads and detect limited gene counts. We thus developed a high-throughput single-cell RNA-seq method, Quartz-Seq2, to overcome these issues. Our improvements in several of the reaction steps of Quartz-Seq2 allow us to effectively convert initial reads to UMI counts (at a rate of 30%–50%). To demonstrate the power of Quartz-Seq2, we analyzed transcriptomes from a cell population of in vitro embryonic stem cells and an in vivo stromal vascular fraction with a limited number of sequence reads.
Background
Single-cell transcriptome analysis allows us to identify nongenetic cellular heterogeneity, which includes differences in cell type due to differentiation and differences in cell state within a cell population. In previous studies, various methods for single-cell RNA-seq were developed [1-19]. Some of these that generate read coverage across all transcripts have been exploited to detect alternative transcription splicing isoforms [14,17] and others using unique molecular identifier (UMI) have been applied to quantify the number of transcripts expressed in a cell [1-3,6,7,9,11-13,20]. To extract substantial information on a cell population, such as the composition of different cell types or the distribution of cell states, it is necessary to analyze hundreds or thousands of cells. Cell barcoding is a key technology for this, which enables us to deal with samples from numerous cells in a single tube. Cell barcoding technology, which tags nucleotides unique to each cell to target RNA molecules from that cell, is a key technology for increasing the throughput of single-cell RNA-seq [16,18]. Mixing cDNA tagged with cell barcodes before whole-transcript amplification decreases the cost of reaction reagents and the laboriousness of experimental steps. There are two types of cell barcoding technology according to the method of cell sampling used. One method involves single cells being selectively sorted to multi-well plates using flow cytometry, which allows us to remove dead or aggregated cells. Besides, transcriptome data can be linked to cellular information measured in flow cytometry. The other method involves single cells and barcoded beads being captured in water-in-oil droplets using droplet-generation microfluidic devices [6,7]. In this latter method, thousands of cells can probabilistically be captured in half an hour. However, the total number of sequence reads generated by a deep sequencer is still limited. To increase the number of analyzed cells, each cell is assigned a limited number of initial sequence reads. For example, approximately 400 million initial reads were sequenced for 3,000–4,000 cells in several previous studies [6,7]. UMI counts were converted from initial sequence reads for each cell, and the conversion ratio of that was limited to approximately one-tenth [21]. Ideally, greater UMI counts should be generated from limited sequence reads because the increase in UMI count assigned to each cell leads to the detection of low-copy genes and the identification of cell-type-specific genes using statistical tests.
In this study, we developed a novel high-throughput single-cell RNA-seq method, Quartz-Seq2. As Quartz-Seq is a sensitive and reproducible single-cell RNA-seq method, Quartz-Seq2 was developed based on it [15]. Quartz-Seq is based on a poly-A tagging strategy. By the combination of molecular biological improvements including major improvement of poly-A tagging, Quartz-Seq2 resulted in an increase in the effectiveness with which the initial sequence reads were converted to the expression UMI counts (UMI conversion efficiency: 30%–50%). To demonstrate the highly effective use of initial reads in Quartz-Seq2, we analyzed a population of mouse embryonic stem (ES) cells as in vitro cells and the cell population from the stromal vascular fraction (SVF) as in vivo cells.
Results
Outline of Quartz-Seq2 experiment
To increase the UMI conversion efficiency, we improved several steps in the preparation of the single-cell RNA-seq library (Additional file 1: Figure S1), resulting in the development of Quartz-Seq2. Below, we explain the five steps of the Quartz-Seq2 procedure (Figure 1).
The first step is single-cell collection using a cell sorter. We selectively sort living single cells into lysis buffer in a 384-well PCR plate without dead cells. In cell sorting, various types of channel information, such as the intensity of fluorescence, are measured for each cell. This enables us to link the transcriptome to cellular information from a cell sorter for each cell.
The second step is cell barcoding. Each well contains lysis buffer and reverse-transcription (RT) primer, which includes a cell barcode sequence (14- or 15-mer), a UMI sequence (8-mer), and an oligo-dT sequence (24-mer). Using these RT primers, respective RNA from single cells is converted to cDNA with unique cell barcodes. Note that a long RT primer resulted in a severe problem regarding the synthesis of byproducts at the downstream reaction in our system (Additional file 1: Figure S2 and Supplemental Note). Therefore, we use a relatively short RT primer (73- or 74-mer), which allows us to skip the step of removing byproducts using exonuclease I. We design two types of RT primer set (v3.1: 384 barcodes; v3.2: 1536 barcodes). Within each primer set, barcode sequences are designed such that the minimum Sequence–Levenshtein distance between two sequences should be greater than 5, which leads to the correction of two nucleotides of mismatch, insertion and deletion in sequence reads [22]. We also optimize the buffer and temperature in the RT reaction, leading to an improvement of RT efficiency from that of the original Quartz-Seq (Additional file 1: Figure S3 and Supplemental Note).
The third step involves the pooling of cell-barcoded cDNA. By cell barcoding, Quartz-Seq2 can pool cDNA of up to 1,536 individual cells into one mixture. We developed a rapid and high-throughput method for collecting small volumes of cDNA in multiwell plates. This method also achieved higher efficiency of collection than dispensing with pipettes and tips (Additional file 1: Figure S4 and Supplemental Note). As the efficiency of cDNA purification after pooling was 93.77%, we estimated that approximately 80% of cell-barcoded cDNA could be used for subsequent whole-transcript amplification in our system (Additional file 1: Figure S3a and Supplemental Note).
The fourth step is whole-transcript amplification based on an improved poly-A tagging strategy. Poly-A tagging is one of the methods of converting first-strand cDNA to amplifiable cDNA. First-strand cDNA is extended with a poly-A tail by the terminal transferase. Subsequently, second-strand cDNA is synthesized with tagging primers that contain a poly-dT sequence, followed by PCR amplification. Here, we improve the efficiency of poly-A tagging by 3.6-fold. This improvement is a crucial point in the development of Quartz-Seq2. We describe the details of this in the next subsection.
The fifth step is library preparation for deep sequencing. Amplified cDNA is fragmented, ligated with the sequence adapter, and amplified by PCR. In sequencing using Illumina sequencers, a sequence for a cell barcode and a UMI is read in Read1, while a sequence for a region of a transcript (mRNA) is read in Read2.
Quartz-Seq2 consisted of five steps. (1) Each single cell in a droplet was sorted into lysis buffer in each well of a 384-well PCR plate using flow cytometry analysis data. (2) Poly-adenylated RNA in each well was reverse-transcribed into first-strand cDNA with reverse transcription primer, which had a unique cell barcode. We prepared 384 or 1,536 kinds of cell barcode (CB) with a unique sequence based on the Sequence–Levenshtein distance (SeqLv). The edit distance of SeqLv was 5. RT primer also had a UMI sequence for reduction of PCR bias (MB) and a poly-dT sequence for binding to poly-A RNA. (3) Cell barcode-labeled cDNAs from all 384 wells were promptly collected by centrifugation using assembled collectors. (4) Collected first-strand cDNAs were purified and concentrated for subsequent whole-transcript amplification. In the poly-A tailing step, purified cDNA was extended with a poly-A tail by terminal deoxynucleotidyl transferase (TdT). Subsequently, second-strand cDNA was synthesized with a tagging primer, which had a poly-dT sequence. The resulting second-strand cDNA had a PCR primer sequence (M) at both ends of the cDNA. The cDNA was amplifiable at subsequent PCR amplification. (5) For conversion from amplified cDNA to sequence library DNA, we fragmented the amplified cDNA using the ultrasonicator Covaris. Such fragmented cDNA was ligated with a truncated Y-shaped sequence adaptor, which had an Illumina flow-cell binding sequence (P7) and a pool barcode sequence (PB). PB made it possible to mix different sets of cell-barcode-labeled cDNA. Ligated cDNA, which had CB and MB sequences, was enriched by PCR amplification. The resulting sequence library DNA contained P7 and P5 flow-cell binding sequences at respective ends of the DNA. We sequenced the cell-barcode site and the UMI site at Read1, the pool-barcode site at Index1, and the transcript sequence at Read2.
Improvement of poly-A tagging efficiency
We previously reported Quartz-Seq based on the poly-A tagging strategy, which has a significant potential for detecting a large number of genes expressed in a cell [15]. However, the efficiency of poly-A tagging itself for single-cell RNA-seq has not been improved. We hypothesized that the improvement of poly-A tagging would lead to high UMI conversion efficiency. Therefore, we attempted to improve the efficiency of this tagging step. Poly-A tagging is composed of two processes: (1) The first-strand cDNA is modified with a poly-A tail by terminal deoxynucleotidyl transferase. (2) Next, the poly-A-tailed cDNA is annealed with a tagging primer, which has a poly-dT sequence. Then, the second-strand cDNA is extended. The resulted second-strand cDNA is amplifiable cDNA, which has a PCR primer sequence at both ends of it.
It is known that the DNA yield of amplified cDNA generally reflects the quantitative performance of single-cell RNA-seq methods [3,13,14]. Thus, we determined the effects of various buffers for the poly-A tailing step on the amplified cDNA yield. We performed the poly-A tailing reaction with various buffers using purified first-strand cDNA from 1 ng of total RNA (Figure 2a, Additional file 1: Figure S5). Finally, we obtained amplified cDNA. We found that the use of T55 buffer and RH55 buffer in the poly-A tailing reaction efficiently improved the cDNA yield (Figure 2a). The amount of amplified cDNA increased 2.88-fold and 2.55-fold using T55 buffer and RH55 buffer, respectively, compared with the level using Quartz-Seq buffer (Figure 2a). In these buffer conditions, we did not observe any obvious byproducts derived from the RT primer (Figure 2a, Additional file 1: Figures S2 and S5, and Supplemental Note).
a) Improvement of poly-A tagging efficiency. Barplot represents the relative DNA yield in various poly-A tagging conditions using purified first-strand cDNA from 1 ng of total RNA. T55 buffer as the terminal deoxynucleotidyl transferase (TdT) buffer and the temperature condition “Increment” for the poly-A tagging step improved the cDNA yield of whole-transcript amplification. Buffer compositions were indicated in Additional file 4: Table S3. QuartzB represents use of a Quartz-Seq-like buffer as a positive control, in accordance with the approach described in the original Quartz-Seq paper. Finally, we quantified cDNA yield (300–9,000 bp) and byproduct DNA yield (50–300 bp) using Bioanalyzer (Agilent). (b) The improvement of cDNA yield with the combination of T55 buffer and “Increment” (T55+Inc) was confirmed by qPCR assay. We detected eight genes (Eef1b2, Nanog, Pou5f1, Sox2, Utf1, Spp1, Ywhae, Tbp) with amplified cDNA and nonamplified cDNA. We prepared nonamplified cDNA with 200 pg of total RNA and random hexamer. SCC stands for Spearman’s rank correlation coefficient. c) Comparison between Quartz-Seq2 and Quartz-Seq-like conditions regarding sequence performance. We analyzed 384 wells with 10 pg of total RNA and used approximately 0.19 M fastq reads per well on average. We show the UMI count and gene count in barplots. We also show a diagrammatic depiction of the relationship between the cumulative number of detected genes and the gene expression level. d) Gene expression reproducibility between bulk poly-A-RNA-seq (1 µg of total RNA) and Quartz-Seq2 (10 pg of total RNA, averaged over 384 wells). e) The panels show a scatter plot between the mean of gene expression and the variability of gene expression with 10 pg of total RNA in 384 wells. Red lines represent the theoretical variability of gene expression in the form of a Poisson distribution. The left and right panels represent the mean-CV plots before and after UMI filtering. The term “RT100” represents the condition of enzyme concentration in reverse transcription for Quartz-Seq2 (see Material and Methods).
Next, we added an “Increment” temperature condition for the tagging and second-strand synthesis steps (see Material and Methods). In this condition, the reaction temperature of these steps was gradually increased. As a result, the amount of cDNA tended to increase, by approximately 1.2-fold (Figure 2a). Moreover, upon combining T55 buffer and the Increment condition, the amount of cDNA increased approximately 3.6-fold. We also confirmed the reproducibility of this phenomenon of cDNA increment in the additional experiments (Additional file 1: Figure S5). Moreover, we confirmed the amplified cDNA yield of various genes by qPCR analysis as another assay. Specifically, we determined the qPCR scores of eight genes from amplified cDNA and nonamplified cDNA (Figure 2b). Spearman’s rank correlation coefficients (SCCs) between amplification and nonamplification were approximately 0.79 in both conditions (T55+Increment, RH55+Increment). The SCC was approximately 0.66 in Quartz-Seq-like conditions. We also observed clear increments of qPCR scores for almost all genes. These results show that the combination of T55 buffer and this temperature condition improved the efficiency of the poly-A tagging step. We also found that other conditions (NBF40+Increment) improved the cDNA yield. However, under these conditions, byproducts were clearly synthesized (Additional file 1: Figures S4c and S5b). Moreover, the cDNA amount with T55 buffer was slightly greater than the cDNA amount with RH55 (Figure 2 and Additional file 1: Figure S5a). Therefore, we used the combination of T55 buffer and the “Increment” temperature condition for the poly-A tagging strategy.
Evaluation of the quantitative performance of Quartz-Seq2 using single-cell total RNA
Finally, we adopted three molecular-biological improvements [modified RT condition, poly-A tailing buffers (T55), and the “Increment” temperature condition] for Quartz-Seq2 (Additional file 1: Figure S1). To determine the technical variability and specificity of Quartz-Seq2, we performed whole-transcript amplification using 10 pg of diluted mouse total RNA as a single-cell-like averaged sample with the v3.1 384 cell-barcode RT primer in a 384-well plate. In this experiment, we used the “RT100” enzyme concentration for reverse transcription. The effect of enzyme concentration is specifically described in the next subsection of this paper. We analyzed 10 pg of total RNA in all wells at approximately 0.19 M fastq reads on average per well. In the case of Quartz-Seq-like conditions, we detected 19,614±3516 UMI counts and 5726±494 gene counts (n=384 wells). In the case of Quartz-Seq2, we achieved a UMI count of 35,391±8719 and a gene count of 6,621±711 (n=384 wells). We confirmed the reproducibility of the quantitative performance by a further batch experiment with other 384-well plates, which had 10 pg of total RNA in all wells (Additional file 1: Figure S6). These results indicated that the combination of the three molecular-biological improvements for Quartz-Seq2 clearly improved the conversion ratio from target RNA to sequence library DNA. We also confirmed the correlation between single-cell RNA-seq data and bulk RNA-seq data. SCC between Quartz-Seq2 and bulk RNA-seq was 0.85 (Figure 2d). This result indicates that Quartz-Seq2 data are comparable to conventional bulk RNA-seq data. We also determined the variability of expression between technical replicates to investigate the UMI filtering effect. The mean-CV (coefficient of variation) plot of Quartz-Seq2 clearly showed that UMI filtering reduced expression variability derived from PCR bias (Figure 2e). Moreover, technical gene expression variability of Quartz-Seq2 became fairly close to the theoretical variability of a Poisson distribution (Figure 2e).
We also focused on the relationship between sequencing platform and Quartz-Seq2. We did not observe clear differences between the NextSeq500 and HiSeq2500 sequencing platforms for determination of the UMI count and the gene count for Quartz-Seq2 (Additional file 1: Figure S7a). On the other hand, the unique mapping ratio and UMI count clearly depended on the Read2 length for transcript mapping (Additional file 1: Figure S7c, d). In this regard, the number of detected genes slightly depended on Read2 length (Additional file 1: Figure S7d). Long Read2 length led to better UMI counts, but high sequencing cost (Additional file 1: Figure S7b). Therefore, in this study, we mainly used the NextSeq500 sequence platform and a Read2 length of 62 due to the cost performance of this combination. Moreover, we mainly used the adaptor ligation method for sequence library preparation of Quartz-Seq2. We did not observe a clear difference between the two sequence library preparation methods (ligation and Nextera/Tn5 transposase) regarding determination of the UMI count and the gene count. However, we could easily identify specific familiar genes (such as Pou5f1 and Gapdh) using the ligation method (Additional file 1: Figure S7e). Therefore, we mainly used the ligation method for Quartz-Seq2.
Reduction of enzyme concentration in reverse transcription decreased cost and technical noise of Quartz-Seq2
The cost of experimental preparation per cell was approximately ¥2,600 ($23) for our previously reported Quartz-Seq, which does not use cell barcoding (Figure S8c). As mentioned earlier, we used the “RT100” enzyme condition in reverse transcription for Quartz-Seq2. This enzyme concentration in RT reaction is broadly used for various molecular biological application including single-cell RNA-seq methods. By using the cell barcoding strategy, the cost of experimental preparation for Quartz-Seq2 under the “RT100” condition was reduced (¥121 or $1.08 per cell)(Additional file 1 Figure S8c). We found that 65% of the cost in experimental steps is derived from reverse transcription in Quartz-Seq2 under the “RT100” condition (Figure 3a). To further reduce the cost of experimental preparation of sequence libraries on large scale, we investigated the effect of a low enzyme concentration in reverse transcription in Quartz-Seq2. In the assessment assay, we noticed that a low enzyme concentration did not markedly affect the efficiency of reverse transcription and the amount of amplified cDNA yield (Figure S8a right panel and S8b). Therefore, we prepared three technical replicates of a 384-well PCR plate with 10 pg of total RNA with the “RT100” condition or the “RT25” low-enzyme condition. In the “RT25” condition, the amount of cDNA yield showed a tendency for a slight increase of approximately 1.17-fold, which although not being a major improvement did at least not involve a decrease. The cost of experimental preparation per cell became approximately ¥45–¥62 ($0.4–$0.55) in the “RT25” condition.
a) Reverse-transcription step has a highest share in single-cell experimental part of Quartz-Seq2. We present the different proportions of costs associated with each of the experimental steps. The reverse-transcription step accounts for approximately 65% of the total cost in the “RT100” condition. We thus established the “RT25” condition by reducing the enzyme concentration to 25% in reverse transcription (for details, see Material and Methods). b) Barplots present the amount of amplified cDNA from three technical replicates of a 384-well plate with 10 pg of total RNA in all wells for the “RT100” and “RT25” conditions. The y-axis represents the yield of amplified cDNA. The cDNA yields from 10 pg of total RNA were 55.9 ± 5.5 ng for “RT100” and 65.7 ± 2.5 ng for “RT25”. c) We analyzed amplified cDNA from 10 pg of total RNA by Quartz-Seq2 with various initial data sizes (approximately 0.27 M, 0.19 M, 0.171 M, 0.152 M, 0.133 M, 0.114 M, 0.095 M, 0.076 M, 0.057 M, 0.038 M and 0.019 M fastq reads per 10 pg of total RNA on average) from NextSeq500. The Read2 length was 62 nt. In all panels, the x-axis represents the initial data size for one sample. We present the UMI count, gene count, and spike capture ratio in the upper panels. We also show the technical variability of the UMI count or gene count between wells in the lower panels. The y-axis presents the coefficient of variation of the UMI count or the gene count. d) Relationship between initial data size and the number of single cells for sequence analysis in NextSeq500 runs. Typically, one sequence run with NextSeq 500/550 High Output v2 Kit reads out 400–450 M fastq reads. The x-axis represents the input cell number for one sequence run. The y-axis represents the initial data size (fastq reads) per cell on average. If we analyzed 4,000 single cells in one sequence run, approximately 0.1 M fastq reads on average would be assigned to one single cell. Red circle represents the typical range of shallow input read depth for a single cell. e) Capture ratio of ERCC spike mix I in the “RT25” condition. The capture ratio of ERCC spike-in mix I was 16.38±0.66% (n=3) at approximately 0.27 M fastq reads on average per well. The x-axis represents the copy number of input spike RNAs. The y-axis represents the averaged UMI count for ERCC spike RNAs.
We validated the quantitative performance with 10 pg of total RNA in the “RT100” and “RT25” conditions at various input data sizes. Unexpectedly, we found that the “RT25” condition improved the quantitative performance (Figure 3c and Additional file 1: Figure S9). In a typical pattern, approximately 4,000 single cells were analyzed by the NextSeq500 sequencer in the high-throughput single-cell RNA-seq method [6,7,21]. In this case, the input data size for a single cell involves shallow initial reads (0.1 M fastq reads) (Figure 3d). We compared RT25 with RT100 at approximately 0.096 M fastq reads on average per well. In the case of Quartz-Seq2 in the “RT25” condition, we achieved a UMI count of 30,117±789 and a gene count of 6,320±35 (n=3 plates). In the case of Quartz-Seq2 in the “RT100” condition, we achieved a UMI count of 25,671±1020 and a gene count of 5,889±35 (n=3 plates). The capture ratios of ERCC spike were 11.49±0.46% and 10.28±0.31% in the “RT25” and “RT100” conditions, respectively (Figure 3d and e). In addition, the capture ratio of ERCC spike was 16.38±0.66% in the “RT25” condition at approximately 0.27 M fastq reads on average per well (Figure 3c, e). We also investigated the technical variability of UMI count and gene count. The well-to-well technical variability for UMI count and gene count clearly decreased in the “RT25” condition (Figure 3). These results showed that the “RT25” low-enzyme condition clearly reduced the experimental cost and improved the quantitative performance, including that for the UMI count. We thus applied the “RT25” condition in subsequent experiments using real single cells.
Quartz-Seq2 shows high efficiency of UMI conversion
The increase of UMI counts generated from limited initial sequence reads leads to the detection of low-copy genes and the identification of more marker genes using statistical tests. We calculated the UMI conversion efficiency, which indicates how effectively initial reads can be converted to UMI counts (Figure 4a and Additional file 1: Figure S10). The UMI conversion efficiency of Quartz-Seq2 with 10 pg of purified total RNA (from mouse ES cells) ranged from 31% to 39% at approximately 0.096 M initial reads depending on the Read2 length.
a) We defined the UMI counts divided by the raw fastq reads as the UMI conversion efficiency. b) UMI counts, gene counts, and UMI conversion efficiency for Quartz-Seq2 and Drop-seq2 experiments. These values depended on the initial fastq reads. c) Venn diagram of genes that were differentially expressed between the ES cluster and the Dex-treated ES (PrE) cluster, as identified by Quartz-Seq2, Drop-seq, and bulk RNA-seq (left). The number of genes that differed in expression levels between ES and PrE cells by at least twofold was counted (FDR < 0.05). We also present a Venn diagram of the Reactome pathway (right). d) Venn diagram of GO terms for genes with highly variable expression, as identified by Quartz-Seq2 and Drop-seq.
We performed Quart-Seq2 and Drop-seq on a mixture of mouse ES cells and the Dex-treated mouse ES cells (primitive endoderm (PrE) cells), and calculated the UMI conversion efficiency (Figure 4b and Additional file 1: Figure S11a). Drop-seq is one of the high-throughput single-cell RNA-seq methods, which can capture thousands of cells and barcoded beads in half an hour [6]. We performed Drop-seq experiments, which were validated by species-mixing analysis (Figure S11b, c). The effectiveness of Quartz-Seq2 ranged from 25% to 35% depending on the initial read depth (Figure 4), which was higher than that of Drop-seq.
As Figure 4b shows, Quartz-Seq2 detected more genes with high UMI conversion efficiency. To examine the power for identifying marker genes, we performed principal component analysis (PCA) and clustering. The number of genes differentially expressed between two distinct cell types was calculated (ES and PrE cells) (Additional file 1: Figure S11d). The results showed that more genes that were differentially expressed between the two cell types were identified with Quartz-Seq2 (Figure 4b). In addition, more biological pathways particularly associated with the differentially expressed genes were detected in Quartz-Seq2 (Figure 4c). Furthermore, we calculated genes with highly variable expression in Quartz-Seq2 and Drop-seq, which potentially include not only genes that are differentially expressed between cell types but also genes of which the expression changes depending on the cell state in a cell type. The results also showed that more Gene Ontology (GO) terms enriched among the genes with highly variable expression were identified with Quartz-Seq2. Terms on cell cycle state were only enriched in the genes calculated with Quartz-Seq2 (Figure 4d). These results suggest that high UMI conversion efficiency with limited initial reads leads to more biological information being revealed, such as functional terms and biological pathways.
To obtain additional evidence for the superiority of Quartz-Seq2 in terms of the UMI conversion efficiency and gene count, we compared Quartz-Seq2 to other methods. Using mouse ES cells, Ziegenhain et al. systematically compared the quantitative performance of several single-cell RNA sequencing (RNA-seq) methods [CEL-seq2(C1), SCRB-seq, MARS-seq, and Drop-seq] that use the UMI technique [21]. For comparison between our and their data, we reanalyzed our Quartz-Seq2 data from mouse ES cells using their analysis condition. The results indicate that the UMI conversion efficiency of the Quartz-Seq2 method was approximately 34.3 ± 0.7% (n = 4); the UMI conversion efficiency of the other methods ranged from 7 to 23%. The gene count using Quartz-Seq2 was approximately 6,722 ± 57 (n = 4), while the average gene counts for the other methods ranged from 2,700 to 5,200. In particular, the UMI conversion efficiency and gene count of the Quartz-Seq2 method were significantly better than those of the other methods at approximately 0.1 million initial reads (Figure 5a and Additional file 1: Figure S12). We also estimated the UMI and gene counts and the UMI conversion efficiency at various numbers of initial fastq reads (Figure 5b). We found that Quartz-Seq2 is greatly advantageous for detecting the UMI and gene counts from limited initial amounts of data (under 0.25 million fastq reads).
a) We determined the UMI and gene counts from 0.1 million fastq reads with Quartz-Seq2 in the “RT25” condition using a mixture of mouse ES cells and Dex-treated ES cells. In total, we analyzed one set of 384-well plates and three sets of 1,536-well plates using Quartz-Seq2. We also estimated the UMI conversion efficiency of other single-cell RNA-seq methods [CEL-seq2(C1), SCRB-seq, MARS-seq, and Drop-seq] from a previous study that used mouse ES cells [21]. In our comparison, the Read2 length for transcript mapping was 45 nt for all the methods, including Quartz-Seq2. To investigate the throughput capacity for establishing sequence library DNA, we estimated the number of processable single cells per $1,000 spent on each method: Quartz-Seq2 (384 indexes) yielded 1,818 cells, Quartz-Seq2 (1,536 indexes) yielded 2,500 cells, CEL-seq2(C1) yielded 111 cells, SCRB-seq yielded 500 cells, MARS-seq yielded 769 cells, and Drop-seq yielded 10,000 cells. The UMI conversion efficiency was 34.1% for Quartz-Seq2 (384 indexes), 34.4 ± 0.9% (n = 3) for Quartz-Seq2 (1,536 indexes), 22.4% for CEL-seq2(C1), 13.3% for SCRB-seq, 10.6% for MARS-seq, and 7.1% for Drop-seq. The average gene count was 6,673 for Quartz-Seq2 (384 indexes), 6,738 ± 57 (n = 3) for Quartz-Seq2 (1,536 indexes), 5,164 for CEL-seq2(C1), 4,044 for SCRB-seq, 3,252 for MARS-seq, and 2,738 for Drop-seq. b) We also estimated the average UMI and gene counts and the UMI conversion efficiency with various numbers of initial fastq reads for each method. The findings indicate that compared with the other methods, Quartz-Seq2 has a superior ability to detect UMI and gene counts from limited initial amounts of data (under 0.25 million fastq reads).
Quartz-Seq2 achieves high UMI conversion efficiencies at relatively low costs
As mentioned earlier, the experimental cost of Quartz-Seq2 for library preparation ranges from ¥45–62 ($0.4–0.55). In terms of the cost, Quartz-Seq2 is highly competitive with other single-cell RNA-seq methods that use a cell sorter (Additional file 1: Figure S8c). We estimate that the number of processable single cells per 1,000 dollars spent on Quartz-Seq2 ranges from 1,800 to 2,500 (Figure 5a). We also calculated the total cost, including the sequencing and library preparation costs, for each method. The total cost per cell is extremely low (approximately ¥71–131, or $0.63–1.16) for Drop-seq. The total cost per cell is ¥105–165 ($0.94–1.47) for Quartz-Seq2. We note that the total cost of Quartz-Seq2, a cell sorter–based method, approaches that typical of droplet-based methods (Additional file 1: Figure S8d). The Quartz-Seq2 method achieves highest UMI conversion efficiency at limited numbers of initial reads, providing single-cell RNA-seq data in a high-throughput manner.
Quartz-Seq2 analysis on 4,484 mouse embryonic stem cells in maintenance and differentiated conditions
We did not observe any clear differences in UMI counts and gene counts between the v3.1 and v3.2 RT primer sets (Additional file 2: Table S1). To demonstrate the capability of Quartz-Seq2 for quantifying the transcriptome of a large number of cells and identifying rare cell populations, we analyzed approximately 4,484 cells from a mixture of mouse ES cells and differentiation-induced cells. Cells were prestained with Hoechst 33342 and/or Calcein-AM as an indicator of DNA content and culture condition (ES or PrE cells), respectively (Figure 6a). Calcein-AM-positive and -negative cells were sorted to 12 of 384-well plates in a checkered pattern (Figure 6a). In this condition, the cell doublets caused by mis-sorting can be detected at the highest probability. In this analysis, as an average for cells, 0.1 M initial fastq reads were effectively converted to 35,915±1,188 UMI counts in a final gene expression matrix. The UMI conversion efficiency was 35.91±1.18%.
a) We successfully analyzed 97.3% of 4,608 wells. The procedures for cell suspension as used in this assay are shown. Cells cultured under ES-mainstenance and Dex-treatment conditions were separately dissociated into single cells, stained with Hoechst 33342 and/or Calcein-AM, and mixed evenly. Calcein-AM-positive and -negative cells were sorted to 384-well plates in a checkered pattern. White scale bars represent 100 µm. b) Clustering of 4,484 single cells according to the transcriptome. Plotting of cells on t-SNE space with color labeling for each cluster. The percentage indicates the proportion of cells for each cluster relative to all cells analyzed. Numbers in parentheses indicate the numbers of cells making up the cluster. c) Marker genes for each cluster identified by Quart-Seq2. Cluster-specific or cluster-enriched genes were calculated for each cluster, and their expression is displayed as a color in a heatmap. No more than 50 cells are shown for simplicity. d) Reconstructed distribution of Calcein-AM intensity for each cluster. The x-axis represents the intensity of Calcein-AM dye staining. e) Reconstructed distribution of Hoechst 33342 intensity for each cluster. The y-axis represents the density of cells. The x-axis represents the intensity of Hoechst 33342 dye staining.
Dimensionality reduction of the expression UMI matrix resulting from Quartz-Seq2 showed clear separation of six clusters, including the main populations of ES cells and PrE cells and four small populations. We did not observe a clear batch effect between 12 of the 384-well plates (Additional file 1: Figure S13). One cluster contained a relatively high ratio of mitochondrial RNA (Additional file 1: Figure S14), which was judged to reflect low-quality cells [23]. Another cluster showed high values of detected UMI counts and gene counts, and expressed both ES and PrE marker genes, which were judged to represent doublets of an ES cell and a PrE cell. To characterize the identified populations, we determined the genes that were specific or enriched for each cluster using binomial tests (Figure 6c and Additional file 1: Figure S15). Although cells in cluster 3 shared gene expression with the main population of ES cells (cluster 1), some genes including Stmn2 and Rhox5 were additionally expressed, suggesting that cluster 1 and cluster 3 shared the same cell type but had different cell states. Cluster 5 was a small population (0.17%, Figure 6b), but was characterized by many specific marker genes (Figure 6b), including Zscan4c and Zscan4d, which are known to be expressed in a subpopulation of mouse ES cells [24]. However, among our culture conditions, these genes were more expressed in the Dex-treatment condition than in the ES-maintenance condition. Flow cytometry information on cells in cluster 5 also showed no fluorescence of Calcein-AM, indicating that these cells were from the suspension of Dex-treated cells (Figure 6d). These observations were consistent with a previous study demonstrating that Zscan4 was more expressed in a differentiation condition induced by withdrawal of leukemia inhibitory factor [7].
In sampling of cells using a cell sorter, flow cytometry information including the intensity of Hoechst 33342 staining for each cell was collected. Using this information, we reconstructed the distribution of Hoechst 33342 staining intensity for each cluster (Figure 6e). The main population of ES cells (cluster 1) showed an embryonic pattern of distribution of DNA content, which is characterized by a high ratio of cells in G2 and M phases to cells in the G1 phase. On the other hand, the main population of PrE cells (cluster 4) showed a somatic pattern of distribution of DNA content. This was consistent with PrE cells being differentiated from ES cells. The cells in cluster 3, the gene expression of which was similar to that of the ES main population (cluster 1), showed an embryonic pattern rather than a somatic one. As the number of cells in cluster 5 that expressed Zscan4c/d was small, it was difficult to classify the observed pattern as the embryonic or somatic type. These findings indicated the usefulness of single-cell RNA-seq using flow cytometry for the reconstruction of population information after transcriptome-based clustering. For good interpretation of the distribution, a large number of cells for each cluster are required.
Quartz-Seq2 revealed two types of cell state: cell cycle-dependent and -independent heterogeneity
When cells were plotted on t-SNE (t-distributed stochastic neighbor embedding) space using transcriptome analysis and the intensity of Hoechst 33342 staining was colored, the gradient pattern was easily observed for clusters 1, 3, and 4 (Figure 7a), suggesting that Quartz-Seq2 was highly sensitive for detection of the cell cycle state. To examine this, we plotted the gene expression of several cell cycle markers against the intensity of Hoechst 33342 staining, which indicated the strong dependence of such expression on DNA content (Figure 7b). Using the intensity of Hoechst 33342 staining, cell cycle dependence was calculated for all genes detected in Quartz-Seq2 experiments, resulting in the identification of numerous genes for which the expression level changed depending on the DNA content (Figure 7c, d and Additional file 1: Figure S16). We called these genes cell cycle-associated genes. Again, it was confirmed that the ratio of cells in G1 and S phases to those in G2 and M phases was different between ES and PrE cells (Figure 7c). As we showed that the efficiency of UMI conversion was high in the Quartz-Seq2 method, we examined whether 10,000 initial reads are sufficient for cell cycle analysis (Additional file 1: Figure S17). In this analysis, final UMI count was 4,774±62 and the UMI conversion efficiency was 47.74±0.66%. Plotting cells on t-SNE space with coloring the intensity of Hoechst 33342 staining revealed a gradient pattern, which was positively correlated with G2/M marker genes including Ccnb1 and Top2a and was negatively correlated with G1-phase marker genes such as Ccne1 (Additional file 1: Figure S17). These observations demonstrated Quartz-Seq2 with few initial sequence reads could analyze cell cycle due to the high UMI conversion efficiency.
a) Plotting cells on t-SNE space with the colors representing the intensity of Hoechst 33342 staining. b) Dependency of expression of several genes on the intensity of Hoechst 33342 staining [G2/M phase: Ccnb1; G1 phase: Ccne1; housekeeping gene: Gnb2l1; upper panels: ES main population (cluster 1); lower panels: PrE main population (cluster 4)]. c) Identification of cell cycle-associated genes. Average gene expression of cells in a bin is colored in the heatmap. Bins were arranged in order of Hoechst 33342 staining intensity. d) The number of cell cycle-associated genes identified for each z-threshold. e) Several examples for genes whose expression was variable, but not associated with the phase of the cell cycle. f) Single-cell qPCR detection of genes with variable expression.
Within a single cell type, there are two types of cell state. One is the different stages of the cell cycle as mentioned above, and the other refers to cellular heterogeneity that is only remotely related to the cell cycle phases. This means that genes with variable expression in a cell type include cell cycle-associated genes and gene with variable expression that is independent of the cell cycle phases (cell cycle-phase independent variable genes). We determined the number of this latter group of genes by subtracting the cell cycle-associated genes from the genes with variable expression (Figure 7e). For the main population of ES cells (cluster 1), we identified tens of cell cycle-phase independent variable genes. These genes included Spp1, Rhox5, and S100a6, which were previously identified as genes with highly variable expression by scRNA-seq (Figure 7) [7,13,15]. We also identified Stmn2, Dnmt3l, Tmsb4x, and several genes as novel cell cycle phase-independent variable genes. Single-cell RT-qPCR also showed high variability for Stmn2 and Rhox5 compared with that for Nanog (Figure 7f). In previous studies, Sgk1 and Actb were identified as genes with variable expression by scRNA-seq that analyzed cells only in G1 phase or cells without measuring the intensity of Hoechst 33342 [7,15]. In this study, these were classified as cell cycle-associated genes (Additional file 1: Figure S16).
These results show that Quartz-Seq2 globally classified genes with variable expression within a cell type into cell cycle-associated genes and cell cycle phase-independent genes with variable expression by using Hoechst 33342 staining intensity data obtained in flow cytometry.
The global picture of cell-type composition in stromal vascular fraction as revealed by Quartz-Seq2
Next, Quartz-Seq2 was applied to cells from the mouse stromal vascular fraction (SVF), which was taken from adipose tissue (Figure 8a). SVF is thought to be one of the most important sources of mesenchymal stem cells (MSC) due to its potential for cell therapies and the stimulation of endogenous repair [25-28]. However, the global picture of cell-type composition in SVF has not been clarified. To do so, we first observed the distribution of cell size in the SVF population and found it to be broad (5–13 µm, average 6.43 ± 1.35 µm; n=200) (Figure 8a). This was confirmed by flow cytometry data that were obtained using a cell sorter. It is known that the typical amount of total RNA in a mammalian cell is approximately 10 pg [15,17]. Thus, the average amount of total RNA in the SVF population of approximately 1.7 pg is relatively small.
a) Morphology of SVF cells. Adipose tissue from a cell suspension of SVF was prepared. Upper panels present a photograph of adipose tissues and dissociated SVF samples. Yellow scale bar represents 1 cm. White scale bar represents 10 µm. Lower panels represent the distribution of cell size information with different platforms (left: diameter of cell size using photography, right: flow cytometry information using a cell sorter). The diameter of cell size for SVF samples was 6.43 ± 1.35 µm (n=200). b) Clustering of cells included in SVF. The transcriptome of approximately 1,000 cells was quantified by Quartz-Seq2 and clustering on t-SNE space was performed. In accordance with the genes and functional terms enriched in each cluster, the cell type was annotated. The percentage indicates the proportion of cells for each cluster relative to all cells analyzed. Numbers in parentheses indicate the numbers of cells constituting the cluster. c) Marker genes for each cluster were identified by Quart-Seq2. Cluster-specific or cluster-enriched genes were calculated for each cluster, with their expression being displayed as color in a heatmap. No more than 50 cells are shown for simplicity. d) The results of GO-PCA analysis. Functional terms enriched in the genes with high factor loadings of PCA were calculated and the enrichment is displayed as color in the heatmap. No more than 50 cells are shown for simplicity. e) Reactome pathway with genes differentially expressed between cluster 1 and cluster 8.
We collected cells in multiple 384-well PCR plates and analyzed their transcriptome using Quartz-Seq2. In this analysis, 38,450±3513 initial fastq reads were converted to UMI counts in the final digital expression matrix. The UMI conversion efficiency was 26.85±2.70% (n=3). We observed broader distributions of UMI counts and gene counts in SVF than in ES and 10 pg of total RNA. The gene counts correlated well with side scatter (SSC) values, which were associated with cell size (Additional file 1: Figure S19).
Dimensionality reduction and clustering of single-cell transcriptome data showed the clear separation of 11 clusters (Figure 8b). To annotate the cell type on each cluster, we calculated marker genes and functional terms that were either specific or enriched for each cluster (Figure 8c, d and Additional file 1: Figures S18 and S20). As Figure 8c shows, clusters 1, 2, 5, and 8-11 expressed cluster-specific genes, and clusters 3, 4, 6, 7, and 9 shared expressed genes (marker class C) as well as differentially expressed genes of classes D-G. Taking these markers and functional terms as well as previous knowledge together, each cluster was identified as follows: two types of Cd34-positive MSC (cluster 1: 7.8%, cluster 8: 6.4%), two types of Cd4-positive T cells (cluster 3: 45.0%, cluster 4: 1.0%), Cxcr6-positive T cells (cluster 9: 1.71%), B cells (cluster 5: 15.9%), killer T cells (cluster 6: 11.0%), natural killer cells (cluster 7: 3.0%), macrophages (cluster 2: 5.5%), dendritic cells (cluster 10: 0.86%), and plasmacytoid dendritic cells (cluster 11: 0.57%). We noted that the cell sizes of the T cells and B cells from clusters 3, 4 and 5, which were measured using a cell sorter, were smaller than those of MSCs and macrophages from clusters 1, 2, and 8 (Additional file 1: Figure S19). We confirmed that the cell sizes of CD4-positive T cells and CD79-positive B cells were smaller in SVF via immunofluorescence observation (Additional file 1: Figure S19). These results show that Quartz-Seq2 with flow-cytometry data is useful for defining respective cell clusters with cell size information.
Subsequently, we investigated the gene expression of receptors and ligands in stromal vascular fractions. The variety of receptors expressed in each cluster was similar (Additional file 1: Figure S21). However, we found a wide variety of ligands expressed in clusters 1 and 8 (Additional file 1: Figure S21). It is known that MSCs secrete many paracrine mediators, which have a therapeutic effect [29,30]. These results suggest that clusters 1 and 8 are MSCs, which secretes many paracrine mediators.
We focused on two potential MSC clusters (1 and 8) because the heterogeneity of the MSC population has been discussed in many papers [31,32]. By using Quartz-Seq2, clusters 1 and 8 had the Cd31-/Cd34+/Cd45-phenotype (Figure 8, Additional file 1: Figure S22). In a previous study, qPCR-based single-cell transcript analysis of the Cd31-/Cd34+/Cd45-population of SVF was reported. By analyzing 140 genes, the authors identified the Cd55+/Dpp4+/Cd31-/Cd34+/Cd45-cell population and showed that administration of this population was effective for normalizing diabetic wound healing in mice [31]. They identified 13 marker genes for the Cd55+/Dpp4+/Cd31-/Cd34+/Cd45 population in mouse SVF. We compared the expression of those 13 genes between clusters 1 and 8 (Additional file 1: Figure S23). The results showed that almost all the genes were more highly expressed in cluster 1 than in cluster 8, suggesting that the cluster 1 population was similar to the Cd55+/Dpp4+/Cd31-/Cd34+/Cd45-population. To further analyze the relationship between our identified clusters and previously reported heterogeneity of the MSC population, we examined the expression of typical MSC markers (CD90, CD105, PDGFRa, and Sca-1) [31,33] as well as subpopulation specific markers (Pou5f1, Nanog, Sox2, Tnnt2, and Myog) [34,35]. PDGFRa (Pdgfra) and Sca1 were expressed in both clusters 1 and 8, whereas Cd90 was expressed at a relatively higher level in cluster 8 than in cluster 1, and Cd105 was not strongly expressed in either cluster. Cd90 and Cd105 were, however, not specific to these MSC clusters. Neither pluripotent markers (Pou5f1, Nanog, and Sox2) nor skeletal and cardiomyogenic markers (Tnnt2 and Myog) were detected in either cluster (Additional file 1: Figure S22). Collectively, our transcriptome analysis showed that the MSC population is divided into two clusters, suggesting that there is less heterogeneity of MSC in the SVF than expected.
Next, we investigated the difference between the two MSC clusters in SVF. We detected 182 differentially expressed genes between clusters 1 and 8 (Additional file 1: Figure S24, Additional file 5: Table S4). Cluster 8 was characterized by the enriched expression of genes that encode extracellular matrix proteins, including collagens, Hspg2, and Bgn. By contrast, cluster 1 was characterized by the expression of IGFBPs (Igfbp4, Igfbp5, Igfbp6), prostaglandin endoperoxidase synthases (Ptgs2), and secreted factors Fgf2 and Bmp2. Of note, the expression of stem cell characteristic genes Aldh1a3, Fgf2, Bmp2, and Tgfbr2 were identified as important modules in cluster 1, suggesting that cluster 1 had more stem-like phenotypes that secrete medicinal factors.
To leverage the power of the entire transcriptome analysis, we performed an enrichment analysis of biological pathways using the 182 differentially expressed genes between the two clusters (Figure 8e, Additional file 6: Table S5). Cluster 1 was characterized by elastic fiber pathways and IGF transport pathways. Elastic fibers contribute various connective tissues and are related to vascular function [36]. Growth factors including FGF2, BMP2, IGFBPs, and PGE2 regulate growth and differentiation [37]. In cluster 8, a tightly linked network of collagen metabolism pathways was highly enriched. These results suggest that two MSC clusters of SVF have different functions. Cluster 1 could be medicinal cell characterized by the expression of growth and development-related genes. On the other hand, cluster 8 may be characterized by production of extracellular matrix.
Taken together, our transcriptome analysis by Quart-seq2 successfully identified cell populations that consisted of various types of both large and small cells, which were isolated from in vivo tissue, and demonstrated that fresh SVF contains two closely related types of MSCs that have distinct characteristics.
Discussion
The rate of conversion from initial reads to UMI counts was low, especially in high-throughput single-cell RNA-seq. An increase in UMI counts assigned to each cell theoretically leads to the detection of low-copy genes and the identification of cell-type-specific genes using statistical tests. Therefore, in this study, we developed a novel high-throughput single-cell RNA-seq method, Quartz-Seq2. We calculated the UMI conversion efficiency, which indicated how effectively initial reads can be converted to UMI counts. The UMI conversion efficiency of Quartz-Seq2 was higher than those of other single-cell RNA-seq methods (Additional file 1: Figure S12). In other words, Quartz-Seq2 could detect more genes from limited fastq reads and for a lower cost. By cell barcoding, Quartz-Seq2 can pool cDNA of up to 1,536 individual cells into one mixture. Streamlined and simplified experimental processes of Quartz-Seq2 allowed us to handle thousands of single cells within a few days. Quartz-Seq2 also showed high accuracy and high reproducibility, leading to precise measurement of the transcriptome from single cells (Figure 2d and e). We demonstrated Quartz-Seq2 analyses on mouse ES cells as an in vitro sample, and mouse SVF including MSC as an in vivo sample. Moreover, we decreased the cDNA preparation cost per cell of Quartz-Seq2 by 97.6%-98.2% compared with that of Quartz-Seq (Figure S8c). The total cost (cDNA preparation cost and sequence cost per cell) of Quartz-Seq2 as a cell sorter based method approached that of a droplet-based method (Additional file 1: Figure S8d). Furthermore, the use of evaporation-preventing oil might reduce the cDNA preparation cost of Quartz-Seq2 by 75% [2].
We proposed that UMI conversion efficiency could be a useful variable for evaluating performance for the further development of high-throughput single-cell RNA-seq methods with shallow reads. The increase of initial sequence reads by additional sequencing does not cancel out the low efficiency of UMI conversion at equal rate. For example, approximately 24,000 fastq reads were converted to approximately 11,000 UMI counts in Quartz-Seq2, while Drop-seq required approximately 169,000 fastq reads to generate the same UMI counts. At this condition, the difference of UMI conversion efficiency was 2.5-fold but the difference of required fastq reads to generate the same UMI counts was 7-fold. This is because UMI counts do not linearly increase with an increase of initial sequence reads due to UMI filtering. Unfortunately, the rate of increase of sequence throughput was lower than that of the processing ability for single cells. It will thus continue to be important to utilize limited initial fastq reads for high-throughput single-cell RNA-seq methods.
Quartz-Seq2 is based on the poly-A tagging, which is one of the strategies of converting first-strand cDNA to amplified cDNA [19,38]. However, the poly-A tagging efficiency itself has not been improved for single-cell RNA-seq. In this study, three molecular biological improvements on Quartz-Seq2 contributed to the increase of amplified cDNA, which led to high UMI conversion efficiency (Figure 2 and Additional file1: Figure S1). In addition, improvement of poly-A tagging steps was most efficient to increase amplified cDNA for Quartz-Seq2 (up 360%, Additional file 1: Figure S1). Several previous studies improved the efficiency of conversion from an mRNA molecule to amplifiable cDNA, and showed that the increase of amplified cDNA is a good guidepost to improve UMI counts or quantitative performance for single-cell RNA-seq [3,14,15]. These results suggest that improved poly-A tagging was the most important point for high UMI conversion efficiency of Quartz-Seq2.
Several improvements proposed in this study could contribute to the further development of other single-cell RNA-seq methods, as follows. (1) We showed that a decrease of enzyme concentration in RT solution led to decreases in technical error and cDNA preparation cost. (2) We applied the Sequence–Levenshtein distance for the design of cell barcodes containing RT primer sets. The use of primer sets designed in this study allows the user to correct mutations of at most two nucleotides of cell barcode, including substitution, insertion, or deletion; the correction capability of these primer sets is higher than that of barcode sets used in previous studies [1,3]. Such correction of the cell barcode increased the average UMI count by 3%–5%. (3) We hope that the spin-down collection system developed in this study can be applied to other cell sorter based single-cell RNA-seq methods [2,3,11,12]. (4) In Quartz-Seq2, we improved the efficiency of poly-A tagging itself (Figure 2). Several single-cell RNA-seq methods including Quartz-Seq were based on poly-A tagging [4,5,10,15,19,39]. Such improvement may directly contribute to these methods having increased quantitative performance.
We achieved a marked improvement in UMI conversion efficiency from initial reads to UMI counts, allowing us to analyze 3-to 10-fold more single cells in limited sequence experiments. However, it is difficult to sort more than 20,000 single cells using one cell-sorter machine in a day. While there is greater scalability for cell sampling in droplet-based single-cell RNA-seq methods than in cell sorter-based methods, these latter methods can utilize additional information obtained in flow cytometry [40,41], which cannot be determined from transcriptome data alone. These two types of method provide complementary approaches for investigating complex biological phenomena. C1 Single-Cell Auto Prep System (Fluidigm) is another widely used platform for single-cell RNA-seq methods [42]. CEL-seq2 shows high UMI conversion efficiency, so it is a very convenient method for users of C1 platform as a single-cell RNA-seq method based on UMI count [3]. Quartz-Seq2 cannot be performed with C1 platform, but it can assimilate 1,536 cell barcodes in a cell sorter in a high-throughput manner. We think that both methods can be effective as long as the most appropriate one is selected for each situation.
In this study, we showed that Quartz-Seq2 has advantages in gene detection and the identification of biological pathways via high UMI conversion efficiency (Figures 4, 5, and 7). Moreover, we analyzed thousands of single cells from all cell cycle phases within a cell type (Figure 7). These specifications with Quarz-Seq2 allow us to perform the global classification of genes with variable expression in a cell type into those whose expression varies in association with the cell cycle or independently of it.
A previous study utilizing the FUCCI fluorescent reporter system reported that competence to respond to specific differentiation signaling was limited to only an early or a late window of the G1 phase in human ES cells [43]. In the future, combining Quartz-Seq2 and FUCCI/FUCCI2 or other fluorescent reporter systems will lead to an understanding of the global picture of differentiation dynamics on competence, response, transition, and commitment because Quartz-Seq2 can analyze changes of cell state for thousands of cells.
In summary, Quartz-Seq2 can be used to detect continuous cell states because of the large number of cells with which it deals and the high efficiency of use of initial sequence reads. Quartz-Seq2 can facilitate investigation of the cell state within a cell type, such as gradient or stochastic changes of the cell population in organism development and disease progression.
Conclusions
In this study, we developed a high-throughput single-cell RNA-seq method, Quartz-Seq2, which can analyze cells numbering up to 1,536 that are pooled together in a single sample. Quartz-Seq2 allows us to effectively utilize initial sequence reads from a next-generation sequencer. The UMI conversion efficiency in Quartz-Seq2 ranged from 30% to 50%, which is much higher than for other single-cell RNA-seq methods (5%–20%). This was caused by the improvements in several molecular biological steps including poly-A tagging. The technical gene expression variability of Quartz-Seq2 was sufficiently low, and that was close to the theoretical variability of a Poisson distribution. As we showed in the analysis of SVF and the ES/PrE mixture, cell types in the population were identified with marker genes and functional terms that characterized each cell type. We identified two types of Cd34-positive MSC in SVF, namely, those that express numerous transcription factor-coding and secreted protein-coding genes specific to each. Quartz-Seq2 can also be used to provide continuous data on cell states because of the large number of cells with which it deals and the high efficiency in the use of initial sequence reads. Quartz-Seq2 should facilitate investigations of the cell state within a cell type, such as gradient or stochastic changes of cell populations in organism development and disease progression.
Methods
Cell culture
Mouse embryonic stem cells were cultured as described previously [15]. Briefly, 1.0 x 105 cells were seeded on a 60-mm dish coated with gelatin (0.1%). Cells were maintained in GMEM-based medium containing 10% FBS and 1000 units/ml leukemia inhibitory factor (Millipore ESGRO). The cell line used in this study was G6GRGFP, which was established in a previous study [44]. To differentiate ES cells into primitive endoderm-like cells, 1.0 x 105 cells were seeded on a 60-mm dish coated with gelatin and cultured in GMEM-based medium containing 10% FBS, 1000 units/ml LIF, and 100 mM dexamethasone for 72 h.
Cell staining
To identify dead or damaged cells, propidium iodide (PI) was added to the cell suspension (final concentration 1–2 µg/ml). As an indicator of an undifferentiated state in a suspension containing a mixture of cells, Calcein-AM was used. Suspensions for cells cultured under maintenance and differentiation conditions were prepared independently, and cells in the maintenance condition were treated with 1 µg/ml Calcein-AM for 10 min on ice. After washing with PBS, cell suspensions of ES cells and PrE cells at the same concentration were mixed. Hoechst 33342 was used as an indicator of the DNA content in a cell. The procedures were performed as described previously [15].
Single-cell preparation for stromal vascular fraction
These experiments were carried out in accordance with the protocol approved by the National Institute of Advanced Industrial Science and Technology (AIST) Animal Care and Use Committee and the RIKEN Animal Experiment Committee. Subcutaneous fat tissues from three-to four-month-old ICR male mice (n=3 per sample, two biological replicate samples) were minced into small pieces and incubated with 0.4 U/mL collagenase NB4G (Serva) at 37°C for 35 min in a shaking water bath. The digested solution was sequentially filtered through 100-µm and 40-µm cell strainers (Corning), followed by centrifugation at 250 × g for 5 min to remove mature adipocytes. The pellet was treated with erythrocyte lysis buffer (BD Biosciences) and centrifuged at 180 × g for 5 min. The nucleated cells were suspended in HBSS with 0.1% BSA, filtered through a 20-µm cell strainer (pluriSelect), and then kept on ice (Cell solution A). The cell aggregates that did not pass through the 20-µm strainer were further treated with Accutase (Thermo Fisher Scientific) at 37°C for 15 min to dissociate them into single cells, centrifuged at 180 × g for 5 min, and suspended in HBSS with 0.1% BSA (Cell solution B). Cell solutions A and B were mixed and again filtered through the 20-µm cell strainer, followed by centrifugation at 180 × g for 5 min. The pellet was resuspended with HBSS with 0.1% BSA, stained with PI, and used for single-cell analysis.
RNA preparation
Total RNA was purified from cultured cells using Direct-zol RNA MiniPrep kit (Zymo Research) with TRIzol RNA Isolation Reagents (Thermo). We measured the concentration of purified total RNA using a NanoDrop 1000 Spectrophotometer (Thermo). We confirmed that the RNA integrity number of total RNA was over 9.5 using an Agilent RNA 6000 Nano Kit (Agilent). For high-throughput single-cell RNA-seq, we prepared total RNA from a single cell with ERCC spike-in mix I. First, we diluted the ERCC spike mix I by tenfold. We then added 6 µL of 1:10 diluted ERCC spike-in mix I per 10 µg of total RNA. We used diluted 10 pg of total RNA with ERCC spike mix I for the technical validation of Quartz-Seq2.
Bulk RNA-seq methods for populations of cells
We prepared sequence library DNA with 1 µg of total RNA using NEBNext Poly(A) mRNA Magnetic Isolation Module and NEBNext Ultra Directional RNA Library Prep Kit. The total RNA did not contain ERCC spin-in mix I. In addition, we used SuperScript III instead of ProtoScript in the reverse-transcription step and KAPA HiFi DNA polymerase instead of NEBNext High-Fidelity PCR DNA polymerase in the PCR step. The resulting sequence library DNA was analyzed by HiSeq2500.
Design of cell barcodes
The selection of 384 or 1,536 sequences for v3.1 and v3.2 barcode primer sets was performed as described below. First, 1,582 or 4,714 candidate sequences were created using the DNABarcodes package of R Bioconductor for the v3.1 set with 14-mer and the v3.2 set with 15-mer, respectively (version: 1.0.0). To reduce the loss of reads converted into UMI counts, we applied the Sequence–Levenshtein distance as an edit distance in order to maximize the ability to correct errors that occur during the synthesis of oligonucleotides or sequencing [22]. The minimum distance between any two sequences was controlled to 5, leading to the correction of a maximum of two errors of substitution, insertion, or deletion. As the base composition of the created sequences was not uniform, we selected 384 or 1,536 sequences from among 1,582 or 4,714 created sequences so that the variance of base composition decreased. These sequences are listed in Additional file 3: Table S2.
Single-cell collection using flow cytometry
Stained cells were analyzed using flow cytometry SH800 (Sony) or MoFlo Astrios EQ (Beckman Coulter). For SH800, we used 130-µm microfluidic sorting chips. For MoFlo Astrios EQ, we used 100-µm nozzle sizes. Each cell sorter was equipped with a custom-made splash-guard to prevent contamination from unexpected droplets between the target well and neighboring wells. We used two sets of RT primer for Quartz-Seq2. The set of v3.1 RT primers has 384 kinds of unique cell barcodes, with a length of 14 nucleotides (OPC purification, FASMAC). The set of v3.2 RT primers has 1,536 kinds of unique cell barcodes, with a length of 15 nucleotides (OPC purification, Sigma). The set of v3.1 RT primers corresponds to one set of the 384-well PCR plate with lysis buffer, the wells of which have unique barcodes. The set of v3.2 RT primers corresponds to four sets of the 384-well PCR plate with lysis buffer. The RT primer position in the 384-well plate and the sequence were as described in Additional file 3: Table S2. Single cells were isolated in the 384-well PCR plate with 1 µL of lysis buffer (0.1111 µM respective RT primers, 0.12 mM dNTP mix, 0.3% NP-40, 1 unit/µL RNasin plus) containing ERCC spike mix I. During single-cell sorting, the 384-well PCR plate was kept on a 384 aluminum stand at 4ºC. For Moflo Astrios EQ, we used G5498B-060 Insert 384 Eppendorf twin.tec PCR as a 384 aluminum stand (Agilent). For SH800, we used an SH800 384 aluminum stand (Sony). Immediately after the cell collection, the plate was temporarily sealed with LightCycler 480 Sealing Foil (Roche) and the sealed 384-well PCR plate was centrifuged at 10,000 g and 4ºC for 1 min using TOMY MX307, equipped with a Rack-in-Rotor and centrifugation rack PCR96-02. These steps were very important to collect the droplet with a single cell in lysis buffer in an efficient manner. By altering the volume of lysis buffer from 0.4 µL (as described in the original Quartz-Seq paper) to 1 µL (Quartz-Seq2), bubbling of lysis buffer was not required before single-cell sorting and the lysis buffer could easily be handled. We then peeled open the temporary seal and re-sealed it with Agilent PlateLoc Thermal Microplate Sealer (Agilent). We agitated the plate at 2,600 rpm and 4ºC for 1 min using ThermoMixer C (Eppendorf), after which we centrifuged the plate again. The resulting 384-well plate was then immediately cryopreserved at −80ºC and maintained under such conditions until subsequent reverse transcription for cell barcoding. After cryopreservation, we performed subsequent whole-transcript amplification using the cryopreserved 384-well PCR plate within a few months.
Whole-transcript amplification of Quartz-Seq2
Cryopreserved 384-well plates with single-cell lysate were centrifuged at 10,000 g and 4ºC for 1 min. Subsequently, we denatured total RNA in each 384 plate at 70ºC for 90 s and hybridized the RT primer to poly-adenylated RNA at 35ºC for 15 s using the C1000/S1000 thermal cycler. The resulting plates were again centrifuged at 10,000 g and 4ºC for 1 min. Next, the plates were placed on the 384 aluminum plate at 0ºC. We peeled away the seal and added 1 µL of RT premix (2x Thermopol buffer, 5 units/µL SuperScript III, 0.55 units/µL RNasin plus) to 1 µL of lysis buffer for each well using a Mantis microfluidic dispensing system (Formulatrix) or a 384 Transfer Plate system (1859-384S, Watson). The above RT solution was used for the RT25 condition. For the RT100 condition, we used the following RT solution: 2x Thermopol buffer, 20 units/µL SuperScript III, and 2.2 units/µL RNasin plus. We sealed the plates again and agitated them at 2,600 rpm and 4ºC for 1 min. The plates were then centrifuged at 10,000 g and 4ºC for 1 min. We then performed reverse transcription at 35ºC for 5 min and 50ºC for 50 min. The reverse transcription was stopped at 70ºC for 15 min. Then, the plates were placed on a prechilled aluminum block, after which we peeled off their seals. Subsequently, we turned the plates upside down on the assembled collector type A or type B (supplemental figure). We mainly used type A. We centrifuged the plates with an assemble collector at 3,010 g and 4ºC for 3 min with swing-bucket rotors. Subsequently, we collected the cDNA solution into a disposable reservoir. Typically, we obtained 650–700 µL of cDNA solution from one 384-well PCR plate. We purified and concentrated the cDNA solution using the DNA Clean & Concentrator™-5 kit (Zymo Research). We used three purification columns for one 384-well PCR plate in the case of the v3.1 RT primer system (384-cell barcode). Purified cDNA was extracted into 20 µL of nuclease-free water from one column purification and transferred into an eight-linked PCR tube (TaKaRa). The PCR tubes were placed on an aluminum PCR stand at 0ºC. We added 25 µL of TdT solution [1x Thermopol buffer, 2.4 mM dATP, 0.0384 units/µL RNase H (Invitrogen), 26.88 units/µL terminal transferase (Roche)] into 20 µL of extracted cDNA using a pipette at 0ºC. The resulting 45 µL of TdT solution was mixed with a pipette at 0ºC or ThermoMixer at 2,000 g and 0ºC for 1 min. Immediately thereafter, the PCR tubes were centrifuged at 10,000 g and 0ºC for 1 min. We used a C1000/S1000 thermal cycler equipped with the 96-Deep Well Reaction Module for the following steps. The PCR tubes were placed on the block of the thermal cycler, which had been prechilled to 0ºC. We then performed a poly-A tailing reaction at 37ºC for 75 s. The solution was inactivated at 65ºC for 10 min. The PCR tubes were placed on an aluminum PCR stand at 0ºC. We then dispensed approximately 11 µL of solution into four wells from 45 µL of TdT solution. We added 46.16 µL of PCR I premix (1.08492x MightyAmp Buffer version 2, 0.06932 µM Tagging primer, 0.05415 units/µL MightyAmp DNA polymerase) to 11 µL of TdT solution for the respective wells of the PCR tube. We performed gentle inversion mixing on the resulting solution in the PCR tube. The tubes were then centrifuged at 10,000 g and 4ºC for 1 min. Subsequently, the solution was mixed with ThermoMixer at 2,000 rpm and 4ºC for 2 min. Then, we spun down the tube again. Next, we denatured the solution at 98ºC for 130 s and hybridized Tagging primer to poly-A-tailed cDNA at 40ºC for 1 min. After that, we performed the “Increment step” by heating to 68ºC at 0.2ºC every second and performed second-strand synthesis at 68ºC for 5 min. The tubes were placed on an aluminum PCR stand at 0ºC. We added 50.232 µL of PCR II premix (0.99697x MightyAmp Buffer version.2, 1.8952 µM gM primer) to 56.16 µL of PCR I solution. We performed gentle inversion mixing on the resulting solution in the PCR tube. The tubes were then centrifuged at 10,000 g and 4ºC for 1 min. Subsequently, the solution was mixed with ThermoMixer at 2,000 rpm and 4ºC for 2 min, after which we spun down the tube again. We then placed it on the block of the thermal cycler at 68ºC. Subsequently, we amplified the cDNA for 11 cycles under the following conditions: 98ºC for 10 s, 65ºC for 15 s, and 68ºC 5 min. We then incubated the tube at 68ºC for an additional 5 min. Finally, we transferred all of the PCR solution, derived from one 384-well PCR plate, to a 50-mL Polypropylene Centrifuge Tube (Watson). Typically, we obtained approximately 1.2 mL of PCR solution per 384-well PCR plate. We added 32 µL of 3 M sodium acetate (pH 5.2) and 6420 µL PB-Buffer (Qiagen) to the PCR solution. The mixture was then purified using a MinElute Spin Column (Qiagen). Purified cDNA was extracted into 40 µL of nuclease-free water. We additionally purified the cDNA with 32 µL of Ampure XP beads. Finally, we obtained 32 µL of purified cDNA. We checked the length distribution of amplified cDNA with an Agilent High Sensitivity DNA Kit (Agilent). The typical average size of the amplified cDNA in Quartz-Seq2 was approximately 1,400 bp (Additional file 1: Figure S2c). Primer sequences are listed in Additional file 3: Table S2.
In the case of the usage of the v3.2 RT primer in Quartz-Seq2, we modified the above steps as follows. After reverse transcription, we collected cDNA solution into a disposable reservoir from four sets of 384-well plates, which corresponded to 1,536 wells. We purified and concentrated the cDNA solution using eight purification columns for four 384-well PCR plates in the case of the v3.2 RT primer system. In the PCR step, we amplified cDNA for nine cycles with the following conditions: 98ºC for 10 s, 65ºC for 15 s, and 68ºC for 5 min. Finally, we transferred all of the PCR solution derived from four 384-well plates to a 50-mL Polypropylene Centrifuge Tube (Watson). Typically, we obtained approximately 3.5 mL of PCR solution per four 384-well PCR plates. We added 88 µL of 3 M sodium acetate (pH 5.2) and 17.6 mL of PB-Buffer (Qiagen) to the PCR solution. The mixture was purified using the MinElute Spin Column (Qiagen). Subsequently, cDNA was again purified using Ampure XP magnetic beads.
For the Quartz-Seq1-like reaction, we performed the following procedure in accordance with our previous study. We added 1 µL of RT premix (2x PCR buffer, 5 units/µL SuperScript III, 0.55 units/µL RNasin plus) to 1 µL of lysis buffer. We then performed reverse transcription at 35ºC for 5 min and 45ºC for 20 min. This reverse transcription was stopped at 70ºC for 15 min. We added 5 µL of ExoIB solution (1.6x Exonuclease I buffer, 3.2x PCR buffer, 16 mM DTT) and 20 µL of TdT solution [1x PCR buffer, 3 mM dATP, 0.0384 units/µL RNase H (Invitrogen), 33.6 units/µL terminal transferase (Roche)] into 20 µL of extracted cDNA using a pipette at 0ºC. The PCR tubes were placed on the block of the thermal cycler that had been prechilled to 0ºC. We performed a poly-A tailing reaction at 37ºC for 75 s. We then denatured the solution at 98ºC for 130 s and hybridized Tagging primer to poly-A-tailed cDNA at 40ºC for 1 min. After that, we performed second-strand synthesis at 68ºC for 5 min. Subsequently, we amplified the cDNA via a PCR reaction for 12 cycles.
Preparation of truncated sequence adaptor
The truncated sequence adaptor was composed of an rYshapeP5 primer (HPLC-purified) and rYshapeP7LT primers (HPLC-purified), which had a TruSeqLT-compatible pool barcode. We prepared 100 µM respective primers with adaptor buffer (10 mM Tris-HCl pH7.8, 0.1 mM EDTA pH8.0, 50 mM NaCl). We added 5 µL of 100 µM rYshapeP5 primer and rYshapeP7LTxx primer into a single PCR tube. We denatured the solution at 90ºC for 90 s. After that, we achieved annealing by cooling to 10ºC by 0.5ºC every 30 s and then maintaining the sample at 4ºC. We then placed the tube on an aluminum PCR stand at 0ºC. Subsequently, we added adaptor buffer, which was prechilled at 0ºC, to 10 µL of 50 µM truncated adaptor. Finally, we obtained about 50 µL of 10 µM truncated adaptor. We cryopreserved 1 µL of 10 µM truncated adaptor in a PCR tube at −80ºC until usage in the adaptor ligation step. Primer sequences are listed in Additional file 3: Table S2.
Sequence library preparation of Quartz-Seq2
We added 130 µL of nuclease-free water including 5–10 ng of amplified cDNA into a Crimp-Cap microTUBE with AFA Fiber. We then performed cDNA fragmentation using an LE220 Focused-ultrasonicator (Covaris) under the following conditions: duty factor 15%, peak incident power 450 W, cycles per burst 200, and treatment time 80 s. We purified and concentrated the cDNA solution using the DNA Clean & Concentrator™-5 kit. Purified cDNA was extracted into 10 µL of nuclease-free water from one column purification and transferred into an eight-linked PCR tube (TaKaRa).
We then added 2 µL of End-Repair premix [1.4 µL of End repair & A-tailing buffer and 0.6 µL of End repair & A-tailing Enzyme (KAPA Biosystems)] to 10 µL of fragmented cDNA solution. Subsequently, we mixed the solution by pipetting on an aluminum PCR stand at 0ºC. The PCR tubes were placed on the block of the thermal cycler, which had been prechilled to 20ºC. We then incubated the tubes at 37ºC for 60 min and 65ºC for 30 min. After this, we added 2 µL of adaptor buffer (1.5 µM truncated adaptor, 10 mM Tris-HCl pH7.8, 0.1 mM EDTA pH8.0, 50 mM NaCl) and 8 µL of ligation premix [6 µL of ligation buffer, 2 µL of DNA ligase (KAPA Biosystems)] at 4ºC. The 22 µL of ligation solution was well mixed by pipetting at 4ºC. Then, we performed adaptor ligation at 20ºC for 15 min. After ligation, we added 18 µL of Ampure XP beads to 22 µL of adaptor ligation solution and mixed them well. Adaptor-ligated cDNA was extracted into 20 µL of nuclease-free water. We added 32 µL of PCR premix [25 µL of 2xKAPA HiFi ready mix, 1.75 µL of 10 µM TPC2 primer (HPLC-purified), 10 µM P5-gMac hybrid primer (HPLC-purified)] to 18 µL of adaptor-ligated cDNA. We denatured the solution at 98ºC for 45 s. Subsequently, we amplified cDNA for eight cycles under the following conditions: 98ºC for 15 s, 60ºC for 30 s, and 72ºC for 1 min. Finally, we additionally incubated the tube at 72ºC for 5 min.
We added 40 µL of Ampure XP beads to 50 µL of PCR solution. Purified sequence-library DNA was eluted into 20–30 µL of nuclease-free water. We checked the DNA concentration and DNA size of sequence library DNA using the Agilent High Sensitivity DNA Kit (Agilent) and QuantiFluor® dsDNA System (Promega). We also prepared sequence library DNA with Tn5 transposase, in accordance with a modified version of a procedure used in a previous study [6]. Specifically, the modifications were as follows. We used 0.75 ng of amplified cDNA for the Nextera XT library preparation kit. We amplified sequence library DNA with the P5-gMac hybrid primer and the Nextera XT primer, which has the P7 sequence.
Deep sequencing for Quartz-Seq2
We analyzed the sequence library DNA from Quartz-Seq2 using NextSeq500 (Illumina) and HiSeq2500 (Illumina). For the Read1 sequence, we used a custom sequence primer named Read1DropQuartz primer (HPLC-purified). In the case of the v3.1 RT primer, sequence specification was as follows (Read1: 22 cycles, Index1: 6 cycles, Read2: 64–118 cycles). In the case of the v3.2 RT primer, sequence specification was as follows (Read1: 23 cycles, Index1: 6 cycles, Read2: 63 cycles). When Read1 length was within 64 nt we mainly used the NextSeq 500/550 High Output v2 Kit (75 cycles), which can be used for sequencing up to 92 nt. For long sequences (over 64 nt), we analyzed sequence library DNA using the HiSeq Rapid SBS Kit v2. Sequence analysis with HiSeq2500 was supported by the Genomic Network Analysis Support facility (GeNAS) in RIKEN.
Amplification-free single-cell qPCR analysis
Amplification-free single-cell qPCR was performed in accordance with our previous study with the following modifications [15]. We collected a single cell into 1 µL of lysis buffer (2.5 µM random hexamer, 1 mM dNTP mix, 0.3% NP40, 2 units/µL RNasin plus) of a 384-well PCR plate (Eppendorf) using SH800. The resulting 384-well PCR plate was cryopreserved at −80ºC. We denatured total RNA in the 384-well PCR plate at 70ºC for 1.5 min using a C1000/S1000 thermal cycler and hybridized the random primer to RNA at 0ºC for 2 min. We added 1 µL of reverse-transcription solution (2x SSIV buffer, 10 mM DTT, 2 units/µL RNasin plus, 20 units/µL SuperScript IV) to the lysis buffer. Subsequently, we performed reverse transcription at 23ºC for 10 min and 50ºC for 10 min. The reverse transcription was stopped at 80ºC for 10 min. The resulting solution was diluted with qPCR solution (10 mM Tris-HCl pH 8.0, 0.05% Tween-20). The obtained diluted solution was then used for qPCR detection with QuantiTect SYBR Green PCR Master Mix and the LightCycler480 system. For primer sets for each gene, see Additional file 3: Table S2.
Drop-seq experiments and data analysis
Drop-seq was performed as reported previously [6] and in line with an online protocol (http://mccarrolllab.com/dropseq/), but with the following modifications: flow rates for oil and aqueous suspensions were 14,000 µl/h and 4,000 µl/h, respectively. The diameter of droplets was 95–100 µm. Microfluidic devices were fabricated by Fluidware Technologies (Japan). The lot number of barcoded beads was 051415 (ChemGenes). Data analysis for Drop-seq was performed as described online (http://mccarrolllab.com/dropseq/). The versions of the software and databases were as follows: STAR: v2.5.1b; mouse genome: mm10; genome annotation: gencode GRCm38.p4 vM9; and Drop-seq tools: v1.11.
Quartz-Seq2 read alignment and generation of digital expression data
The structure of the sequence library and data processing was designed based on those of Drop-seq ([6] and online as referenced above). BCL files generated by Illumina NextSeq500 were converted to fastq files by bcl2fastq2 (v2.17.1.14) with demultiplexing pool barcodes. The –mask-short-adapter-reads parameter was set to 20. If needed, fastq reads were randomly downsampled by seqtk software (version: sgdp). The last two bases of Read2, the sequence quality of which was relatively low, were removed by FASTX-Toolkit (version: 0.0.14). Fastq files for Read1 and Read2 were converted to a bam file by FastqToSam of Picard tools (version: 1.134). Extracting cell barcodes and UMI (equal to molecular barcode) and filtering out reads with low barcode quality were performed using Drop-seq tools (version: 1.11). The resulting bam files were re-converted to fastq files by SamToFastq of Picard tools, and mapped to the mouse genome (mm10) using STAR (version: 2.5.1b). After sorting the resulting bam files by SortSam of Picard tools, the unaligned bam and aligned bam were merged by MergeBamAlignment of Picard tools. Then, gene names were assigned to each record using TagReadWithGeneExon of Drop-seq tools and gtf file (version: gencode GRCm38.p4 vM9). For the correction of errors of cell barcodes considering the Sequence–Levenshtein distance, a custom-made python program was used, which enabled the correction of up to two errors of substitution, insertion, or deletion. This program used python2 (version: 2.7.11+), PypeR (version: 1.1.2), R (version: 3.2.3), Bioconductor (version: 3.2), and Bioconductor package DNABarcodes (version: 1.0.0). Finally, the UMI for each gene for each cell was counted using DigitalExpression of Drop-seq tools, which generated the digital expression matrix. To generate a non-UMI-filtered matrix, a custom-made python program that counts reads for each gene for each cell was used.
Dimensionality reduction, clustering, and term analysis
Cells with low detected gene counts were removed for further analysis (Quartz-Seq2 on ES/PrE mixture: 4,000 genes; Quartz-Seq2 on SVF: 500 genes). Total counts of each cell were normalized to 10,000 or the mean of total UMI counts for cells. UMI counts had 1 added to them and were then log-transformed (natural logarithm), after which they were scaled to have mean and variance for each gene of 0 and 1, respectively. For PCA, UMI counts for all detected genes (for ES/PrE mixture) or genes with highly variable expression (for SVF) were used. For t-SNE, the top 10 to 40 principal components of matrices produced by PCA were used. For clustering, the DBSCAN algorithm was used. The values of the parameter epsilon were 0.69 for Quartz-Seq2 on the 4,500 mouse ES/PrE mixture, 5 for Quartz-Seq2 on the 384 mouse ES/PrE mixture, 3 for Drop-seq on the 500 mouse ES/PrE mixture, and 2.2 for SVF analysis. Significantly enriched GO terms with the top principal components were calculated using the GO-PCA package [45].
Identification of differentially expressed genes for each cluster
Marker genes for each cluster were identified based on a generalized linear model. The difference in deviance between two models, in which the gene was or was not differentially expressed between two clusters, was calculated for the genes. To filter out noise, pseudocount 1 was added to the averaged gene expression in clusters and only genes with fold change of 2 or more between two clusters were further analyzed. P-values were calculated as 1 − cumulative density function of a chi-squared continuous random variable for difference in deviance. After corrections for multiple testing, genes with FDR of less than 0.05 were identified as differentially expressed genes for the cluster. For comparison between differentially expressed genes identified by scRNA-seq methods and those identified by bulk RNA-seq, this pseudocount was not added.
Identification of cell cycle-associated genes and cell cycle-phase independent variable genes
To identify genes for which the expression changes depending on the phase of the cell cycle, we used the intensity of Hoechst 33342 staining measured using a cell sorter. First, cells were discretized into 40 equal-sized buckets based on the rank of Hoechst 33342 staining intensity. Then, the CV of averaged UMI counts for a gene in each bin was calculated. After z-scaling of this, genes with high z-scores were identified. To identify genes whose expression fluctuates in a manner not dependent on the cell cycle phase, we calculated the CV of UMI counts for a gene in each cell and the CV of averaged UMI counts for a gene in each bin. After z-scaling of such data, the genes of which difference between two scaled CV was large were identified as cell cycle-phase independent variable genes.
Enrichment analysis on pathway and Gene Ontology (GO) terms
Pathways that were particularly enriched for the differentially expressed genes were calculated using the ReactomePA package of R Bioconductor (version: 1.14.4, [46]) and Metascape (http://metascape.org/ [47]). The cut-off parameter for the q value was 0.05. Terms that were enriched for genes with highly variable expression were calculated using DAVID, the Database for Annotation, Visualization and Integrated Discovery (version: 6.8, [48,49]). Ontology used for calculation was GOTERM_BP_FAT, GOTERM_MF_FAT, and GOTERM_CC_FAT. The terms with FDR < 0.05 were identified as enriched terms.
Data resources
Raw and processed data files for Quartz-Seq2 and Drop-seq experiments are available under GEO: GSE99866 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE99866) for single-cell RNA-seq data and DRA002954 for bulk-RNA-seq data.
Declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Availability of data and material
The data have been deposited under GEO accession number GSE99866 and DRA accession number: DRA002954.
Competing interests
The authors declare that they have no competing interests.
Funding
This work was supported by the Projects for Technological Development, Research Center Network for Realization of Regenerative Medicine by Japan, the Japan Agency for Medical Research and Development (AMED), and the Japan Science and Technology Agency (JST). The sequence analyses of bulk RNA-seq were supported by MEXT KAKENHI (No. 221S0002) and JSPS KAKENHI Grant Number JP24651218 and JSPS KAKENHI Grant number JP15K16329. The sequence operations of Quartz-Seq2 using HiSeq2500 were supported by the Platform Project for Supporting Drug Discovery and Life Science Research (Platform for Drug Discovery, Informatics, and Structural Life Science) from AMED.
Author’s contributions
YS, HD and IN designed the study. YS designed and performed experiments of single-cell RNA-seq. HD designed and performed data analysis. HT and AK designed and performed experiments of mouse stromal vascular fraction. ME assisted the experiments. TH assisted the development of spin-down collection system. YS, HD, HT, AK and IN wrote the manuscript. All authors read and approved the final manuscript.
Acknowledgements
We greatly appreciate the Drop-seq team’s useful comments about our technical introduction to Drop-seq. We thank Mr. Akihiro Matsushima and Mr. Manabu Ishii for their assistance with the infrastructure for the data analysis. We appreciate Dr. Christoph Ziegenhain and Dr. Wolfgang Enard for giving us their quantitative data from single-cell RNA-seq methods. We thank Dr. Yutaka Suzuki for his assistance with the bulk RNA-seq. We are grateful to Dr. Mika Yoshimura and Dr. Haruka Ozaki (RIKEN) for their assistance with the bulk RNA-seq analysis. Our thanks also go to Sony’s cell sorter team for specially customizing the 384-well sorting mode of the SH800.
Footnotes
Yohei Sasagawa: youhei.sasagawa{at}riken.jp
Hiroki Danno: hiroki.danno{at}riken.jp, redgrapefruit{at}mac.com
Hitomi Takada: htakada{at}bs.naist.jp
Masashi Ebisawa: masashi.ebisawa{at}riken.jp
Tetsutaro Hayashi: tetsutaro.hayashi{at}riken.jp
Akira Kurisaki: akikuri{at}bs.naist.jp
Itoshi Nikaido: itoshi.nikaido{at}riken.jp
List of abbreviations
- UMI
- unique molecular identifier
- PCR
- polymerase chain reaction
- SCC
- Spearman’s rank correlation coefficient
- CV
- coefficient of variation
- ES cell
- embryonic stem cell
- PrE cell
- primitive endoderm cell
- SVF
- stromal vascular fraction
- RT
- reverse-transcription
- Dex
- dexamethasone
- PCA
- principal component analysis
- MSC
- mesenchymal stem cell
- TdT
- terminal deoxynucleotidyl transferase
- GO
- Gene Ontology
- FDR
- false discovery rate
- t-SNE
- t-distributed stochastic neighbor embedding
- nt
- nucleotide
- bp
- base pair
- SSC
- side scatter
References
