Quantitative single-cell RNA-seq with unique molecular identifiers

Transcription

1 brief communications 2013 Nature America, Inc. All rights reserved. Quantitative single-cell RNA-seq with unique molecular identifiers Saiful Islam 1, Amit Zeisel 1, Simon Joost 2, Gioele La Manno 1, Pawel Zajac 1, Maria Kasper 2, Peter Lönnerberg 1 & Sten Linnarsson 1 Single-cell RNA sequencing (RNA-seq) is a powerful tool to reveal cellular heterogeneity, discover new cell types and characterize tumor microevolution. However, losses in cdna synthesis and bias in cdna amplification lead to severe quantitative errors. We show that molecular labels random sequences that label individual molecules can nearly eliminate amplification noise, and that microfluidic sample preparation and optimized reagents produce a fivefold improvement in mrna capture efficiency. RNA-seq has become the method of choice for transcriptome analysis in tissues 1 3 and in single cells 4 7. The two main challenges in single-cell RNA-seq are the efficiency of cdna synthesis (which sets the limit of detection) and the amplification bias (which reduces quantitative accuracy). Published protocols have been reported to have limits of detection of between five and ten mrna molecules 5 7, corresponding to a capture efficiency of around 10%, and all current methods use amplification, either by PCR or by in vitro transcription. To correct for amplification bias, we 8 and others 9 11 have described how molecules can be directly counted through the use of unique molecular identifiers (UMIs). For single-cell RNA-seq, UMIs have been used as an internal validation control 12 but have not yet been explored as a direct, quantitative measure of gene expression. Molecule counting corrects for PCR-induced artifacts (Supplementary Fig. 1) and provides an absolute scale of measurement with a defined zero level. In contrast, standard RNA-seq uses relative measures such as reads per kilobase per million reads (RPKM), which mask differences in total mrna content. For example, a gene may be upregulated in terms of RPKM and have a decrease in absolute expression level if the total mrna content also changes. Thus an absolute scale of measurement is crucial for interpreting transcriptional dynamics in single cells. We applied molecule counting to mouse embryonic stem (ES) cells and used spike-in controls to monitor technical performance (similar results were obtained in independent experiments on ES cells, and hypothalamic and cortical primary neurons). During reverse transcription, each cdna molecule was tagged with a 5-bp random sequence serving as UMI (Fig. 1a and Supplementary Fig. 2). We counted cdna molecules by enumerating the total number of distinct UMIs aligned to each position (Fig. 1b). Mouse embryonic fibroblasts and damaged ES cells were removed on the basis of criteria established after sequencing (Supplementary Note). UMIs will reflect molecule counts only if the number of distinct labels is substantially larger than the typical number of identical molecules. Approximately mrna molecules are present in a typical single mammalian cell, and up to 10,000 different genes may be expressed. However, many genes are expressed from multiple promoters (Supplementary Fig. 3) or have promoters with diffuse transcription start sites (Supplementary Fig. 4), so that the number of identical mrna molecules is expected to be <100 for most genes. We therefore expected our 5-bp UMI, capable of distinguishing up to 1,024 molecules, to be sufficient. To confirm this, we determined the number of distinct UMIs that we could observe for each unique combination of sample barcode (i.e., cell) and genomic position (Fig. 1c). As expected, the vast majority of cases were represented by only a small number of UMIs (corresponding to a small number of cdna molecules), and we did not find a single case of a fully saturated position. To obtain more accurate estimates of molecule counts at high expression levels, we corrected for the collision probability of UMIs (see Online Methods). To ensure that successfully generated cdna molecules are sequenced, it is crucial to sequence to a sufficient depth after amplification. We typically observed each UMI multiple times (Fig. 1b). Across all genes, the average number of reads per molecule was nine, with a distribution consistent with oversampling of most molecules (Supplementary Fig. 5). To further demonstrate that UMIs labeled individual cdna molecules, we examined genes containing a heterozygous single nucleotide polymorphism (SNP). If UMIs worked as intended, each UMI would be derived from a single molecule and therefore from a single allele. As a consequence, all reads derived from this UMI would carry the same allele. Indeed, the observed allele distribution confirmed that UMIs had correctly labeled single cdna molecules derived from single alleles (Fig. 1d). In total, we found 47 informative SNPs, of which all showed the expected monoallelic pattern across UMIs. To improve cdna synthesis efficiency, we implemented our protocol on a commercially available microfluidic platform (Fluidigm C1 AutoPrep) and carefully optimized the conditions for this device (Online Methods). To directly measure the efficiency of reverse transcription, we introduced a known number 1 Department of Medical Biochemistry and Biophysics, Karolinska Institutet, Stockholm, Sweden. 2 Department of Biosciences and Nutrition, Karolinska Institutet, Stockholm, Sweden. Correspondence should be addressed to S.L. (sten.linnarsson@ki.se). Received 27 September; accepted 25 November; published online 22 DECEMBER 2013; doi: /nmeth.2772 nature methods ADVANCE ONLINE PUBLICATION

2 brief communications 2013 Nature America, Inc. All rights reserved. Figure 1 Molecule counting using UMIs. (a) Overview of tagging single mrna molecules with UMIs. Two cells are shown (top) containing mrnas from different genes represented by distinct colors. UMIs are represented by colored boxes (middle and bottom); untagged mrna molecules (gray, middle) were not reverse transcribed (bottom). (b) UMI alignment and mrna molecule counting on a hypothetical example of reads aligned to Tubb2b. (c) Number of genomic positions that were assigned the given numbers of distinct UMIs. (d) UMIs observed in single ES cells for the Exosc1 gene, which contained a SNP (rs ). Stacked bars indicate the number of reads carrying each possible UMI sequence, colored according to the allele observed in the read. Four distinct UMIs were observed, all monoallelic. Horizontal axis indicates the numbers of each UMI counted. Ref, reference allele. of External RNA Controls Consortium (ERCC) control RNA molecules to each well. Counting the resulting number of detected cdna molecules, we found an efficiency of 48 ± 5% s.d. (Fig. 2 and Supplementary Fig. 6), a fivefold improvement over our previously published protocol 5. We attribute the improvement both to the use of an integrated microfluidic device (reducing losses and minimizing background reactions) and to more optimized reagents, in particular the template-switching oligo design. Interestingly, using a separate set of optimizations, a ERCC control RNA 187ss 187ds d Number of molecules g Number of reads 1, ,000 1, ss ds ,000 Number of molecules ,000 10,000 Number of reads h Cumulative probability enumber of molecules 1 0 b Genes (molecules) 187ss 187ds 1, ss a 187ds Cell 1 Cell 2 Reverse transcription, barcoding and UMI labeling PCR amplification ,000 Number of molecules Number of molecules Sequencing and computation UMI 5 end of transcript Lost mrna Molecules in cell 1 Molecules in cell i Observed molecules fnumber of reads 10 0 c Genes (reads) 187ss 187ds 10,000 1, b c Number of positions d Chr13: Tubb2b 34,130,350 34,130,300 Cell 1 Cell 2 Number of reads UMI 1,000,000 10, Barcode 5 end of transcript Reads A C G T(ref) ,024 Number of UMIs (1) (1,024) the recently published Smart-seq2 method achieved similarly improved capture effiency 13, suggesting that even higher efficiencies could be achieved by combining the protocols. Next, we asked how the use of UMIs affected the overall quantification of gene expression in single cells and controls. For comparison, we analyzed the same 187ss ds ,000 10, Number of reads % Spiked molecules per cell 1 20% UMI data sets both using UMIs (counting molecules) and using reads in the Figure 2 Reproducibility of molecule counting. (a,b) Pairwise correlation coefficients calculated for ERCC spike-in control RNA (a) or endogenous genes (b), using molecule counts (n = 41) and prepared with (187ds) and without (187ss) nine additional cycles of library PCR. (c) Pairwise correlation coefficients as in b but counting reads instead of molecules (n = 41). (d) Scatterplot showing the pairwise comparison of two wells indicated in a. Red squares and blue dots show comparisons within and between libraries, respectively. (e) Scatterplot showing the two cells indicated in b based on molecule counts. (f) Scatterplot as in e but using reads instead of molecules. (g) Scatterplot as in d but using reads instead of molecules. (h) Distribution of molecule counts for a single ERCC spike-in transcript (gray dots) compared with the cumulative density function of the Poisson distribution (red line). (i) mrna capture efficiency shown as observed molecule counts versus number of spiked-in molecules for ERCC control RNA transcripts. The shaded bands indicate efficiencies above (dark gray) and below (light gray) 20%. Each red dot represents the average of a single ERCC RNA across 96 wells. Similar results were obtained in one replicate experiment. ADVANCE ONLINE PUBLICATION nature methods

3 brief communications a Coefficient of variation (%) Sprr2b Dqx1 Fst Lefty1 Endogenous genes ERCC control RNA Lossy Poisson (fit) Poisson (100%) Zfp42 Dppa5a b mrna molecules per cell Zfp42 Lefty1 Hes1 Krt ,000 10,000 Average mrna count (number of molecules) Krt8 mrna molecules per cell 2013 Nature America, Inc. All rights reserved. Figure 3 Transcriptional noise in ES cells. (a) The coefficient of variation as a function of the mean number of mrna molecules detected, for genes expressed in ES cells (n = 41 single ES cells; gray and red circles) and for internal ERCC controls spiked into every well (n = 41 replicates; white circles). The solid black line shows a lossy Poisson curve fit to the ERCC technical controls. The dashed line indicates the pure Poisson sampling noise. Genes with significant (α = 0.05; see Online Methods) excess noise are shown as red dots, and examples are indicated by name. (b) Coexpression of noisy genes. The expression of four genes, across 41 ES cells, is plotted against the expression of Krt8. Two branches were observed, interpreted as epiblast-like (high Krt8, Krt18 and Hes1) and pluripotent-like (high Zfp42 and Lefty1). Similar results were obtained in one replicate experiment. conventional manner. We determined the pairwise correlation coefficients (Fig. 2a c) and visualized typical examples as scatterplots (Fig. 2d,e,g,h). As expected, the quantitative precision was improved, especially at low molecule counts. The technical reproducibility was excellent, as demonstrated by correlation coefficients >0.95 for ERCC spike-in control RNA (Fig. 2a). An important consequence of counting molecules is that the data can be displayed on an absolute and biologically meaningful scale, with a defined zero. The scales for read-count scatterplots (Fig. 2g,h) are arbitrary, as the total number of reads differs between wells (and could be increased arbitrarily through additional sequencing runs). Normalizing to reads per million (RPM) amounts only to scaling by a constant factor, and affects neither correlation coefficients nor scatterplots. (Normalizing to RPKM would distort the results, as we sequenced only the 5 end of each mrna, and thus read number was not proportional to gene length; Supplementary Fig. 7). In contrast, scales for molecule-counting scatterplots (Fig. 2d,e) are absolute and would not change appreciably if the number of reads were increased. We observed a smooth distribution of the number of counted molecules, consistent with accurate counting and approaching the theoretically optimal Poisson distribution (Fig. 2h and Supplementary Fig. 8). Examining noise as a function of mrna abundance, we found that the ERCC spike-in controls closely tracked the expected Poisson distribution (Fig. 3a). However, at low levels of expression, the capture efficiency limited our power to detect noise. This can be seen in the difference between the ERCC spike-in controls and the curve for a fully lossless and perfectly accurate measurement (Poisson). Nevertheless, this result demonstrates that little noise above that caused by inefficient reverse transcription was introduced during sample preparation. Next, we examined endogenous transcriptional noise in ES cells. We would expect noise across these measurements to include the sampling noise introduced by reverse transcription as well as biological noise due to stochastic or bursty transcription and to oscillatory and regulated gene expression. We were surprised to find that most genes were expressed at noise levels approaching the Poisson limit (Fig. 3a). At all levels of expression, the noise measured as coefficient of variation (s.d. divided by the mean) closely tracked the predicted Poisson limit, but with a slight excess above the technical noise (compare ERCC controls with endogenous genes). Although most genes showed low levels of noise, using a conservative threshold (see Online Methods) we found a set of 118 significantly noisy genes in ES cells (P = 0.05; Fig. 3a and Supplementary Table 1). Interestingly, several of these noisy genes have been previously demonstrated to show heterogeneous (Zfp42, also known as Rex) 14 or oscillatory (Hes1) 15 expression in ES cells. Nanog was also heterogeneously expressed (P < 0.05), as expected 16, but was excluded by our stringent noise criterion. Furthermore, pathway analysis identified an over-representation of genes involved in the transforming growth factor β signaling pathway (Id2, Id3, Fst, Lefty1 and Lefty2), which has recently been shown to regulate ES cell heterogeneity and self-renewal 17. These results directly validate our approach and suggest that the noisy genes we discovered were not simply affected by stochastic transcription but reflect more complex biological heterogeneity. We hypothesized that noisy expression partially reflected a kind of resonant state between pluripotency and early differentiation. In agreement with this idea, the genes encoding keratins 8, 18 and 19 (Krt8, Krt18 and Krt19) which are known to be expressed at low levels in mouse ES cells but highly induced upon embryoid body formation and in epiblast stem cells 18 were among the noisy genes and were coexpressed in a small subset of cells (Fig. 3b). Using Krt8 to separate these cellular substates, we found that cells ranged between two extreme states characterized by high expression of Hes1, Krt8 and Krt18 (epiblast-like state) and Zfp42 and Lefty1 (pluripotent-like state), respectively. Thus the observed variations in these genes reflect a stochastic distribution of ES cells in regulatory substates rather than intrinsic fluctuations in transcription. A more detailed analysis of these findings will be published elsewhere. The finding that the majority of genes expressed in ES cells had low levels of noise contrasts with some previous findings of large and widespread intrinsic transcriptional noise 19,20. It is plausible that the discrepancy in findings can be partly explained by methodological differences. It is difficult to estimate the contribution of technical noise in imaging-based methods, and previous authors have generally assumed that all observed noise was of a biological nature methods ADVANCE ONLINE PUBLICATION

4 brief communications 2013 Nature America, Inc. All rights reserved. origin. On the other hand, we lost some mrna molecules during cdna synthesis, which would mask some biological noise, especially at low expression levels. Nevertheless, the notion that gene expression in general is extremely bursty seems to be incompatible with the levels of noise that we observed. In conclusion, we have shown that quantitatively accurate singlecell RNA-seq can uncover oscillatory and heterogeneous gene expression within a single cell type. We anticipate that accurate molecule counting will be an important approach for future singlecell transcriptome analyses. Methods Methods and any associated references are available in the online version of the paper. Accession codes. Gene Expression Omnibus: supplementary data are available under accession code GSE Note: Any Supplementary Information and Source Data files are available in the online version of the paper. Acknowledgments Illumina sequencing experiments were performed by A. Johnsson and A. Juréus (Karolinska Institutet). ES cells were generously provided by the Karolinska Center for Transgene Technologies. Recombinant Tn5 transposase was a generous gift from R. Sandberg (Karolinska Institutet). We are grateful for the advice and support of B. Jones and B. Fowler (Fluidigm). This work was supported by grants from the Swedish Research Council (STARGET) and the European Research Council (261063) to S.L. and from the Swedish Cancer Society, Swedish Research Council and Ragnar Söderberg Foundation to M.K. AUTHOR CONTRIBUTIONS S.L. conceived the study. S.I., A.Z. and P.Z. developed the protocols and performed the RNA-seq experiments. S.J. and M.K. prepared cells. A.Z., G.L.M. and S.I. developed the tagmentation protocol. P.L. wrote the bioinformatics pipeline. P.L., S.L. and A.Z. analyzed data. S.L. and S.I. wrote the manuscript with support from all authors. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Reprints and permissions information is available online at com/reprints/index.html. 1. Mortazavi, A., Williams, B.A., McCue, K., Schaeffer, L. & Wold, B. Nat. Methods 5, (2008). 2. Cloonan, N. et al. Nat. Methods 5, (2008). 3. Wang, Y. et al. Nat. Genet. 40, (2008). 4. Tang, F. et al. Nat. Methods 6, (2009). 5. Islam, S. et al. Genome Res. 21, (2011). 6. Ramsköld, D. et al. Nat. Biotechnol. 30, (2012). 7. Hashimshony, T., Wagner, F., Sher, N. & Yanai, I. Cell Rep. 2, (2012). 8. Kivioja, T. et al. Nat. Methods 9, (2012). 9. Shiroguchi, K., Jia, T.Z., Sims, P.A. & Xie, X.S. Proc. Natl. Acad. Sci. USA 109, (2012). 10. Fu, G.K., Hu, J., Wang, P.H. & Fodor, S.P. Proc. Natl. Acad. Sci. USA 108, (2011). 11. Casbon, J.A., Osborne, R.J., Brenner, S. & Lichtenstein, C.P. Nucleic Acids Res. 39, e81 (2011). 12. Shalek, A.K. et al. Nature 498, (2013). 13. Picelli, S. et al. Nat. Methods 10, (2013). 14. Carter, M.G. et al. Gene Exp. Patterns 8, (2008). 15. Kobayashi, T. et al. Genes Dev. 23, (2009). 16. Chambers, I. et al. Nature 450, (2007). 17. Galvin-Burgess, K.E., Travis, E.D., Pierson, K.E. & Vivian, J.L. Stem Cells 31, (2013). 18. Maurer, J. et al. PLoS ONE 3, e3451 (2008). 19. Raj, A., Peskin, C.S., Tranchina, D., Vargas, D.Y. & Tyagi, S. PLoS Biol. 4, e309 (2006). 20. Dar, R.D. et al. Proc. Natl. Acad. Sci. USA 109, (2012). ADVANCE ONLINE PUBLICATION nature methods

5 2013 Nature America, Inc. All rights reserved. ONLINE METHODS Cell culture. R1 mouse ES cells (not authenticated by short tandem repeat profiling but routinely used for successful generation of chimeras and tested for mycoplasma contamination) were grown in ES1 medium on irradiated mouse embryonic fibroblast feeder cells and harvested at confluency. After trypsinization, the feeder cells were given time to settle to obtain a pure R1 ES cell suspension. To reduce contamination with dead and dying cells, dead cells were subsequently stained with Red Fixable Dead Cell Stain (Life Technologies) and dead and early apoptotic cells were depleted with annexin V conjugated microbeads (Miltenyi Biotec). The cells were resuspended at a concentration of 400 cells per microliter in ES1 medium with 10% DMSO, divided into aliquots and frozen at 80 C. Cell capture and imaging. A 30-µL aliquot of ~12,000 cells was thawed, and 20 µl C1 Suspension Reagent was added (all C1 reagents were from Fluidigm, Inc.). Five microliters of this mix were loaded according to the manufacturer s protocol on a C1 Single-Cell AutoPrep IFC microfluidic chip designed for 10- to 17-µm cells, and the chip was then processed on a Fluidigm C1 instrument using the mrna Seq: Cell Load (1772x/1773x) script. This captured one cell in each of up to 96 capture chambers and took ~30 min. The plate was then transferred to an automated microscope (Nikon TE2000E), and an image was acquired from each site using µmanager ( which took <15 min. Lysis, reverse transcription and PCR. The plate was returned to the lab and 20 µl lysis buffer (0.15% Triton X-100, 1 U/µL TaKaRa RNase inhibitor, 4 µm reverse transcription primer C1-P1-T31, 5% C1 Loading Reagent and 1:50,000 Life Technologies ERCC Spike-In Mix 1), reverse transcription mix (1 SuperScript II First- Strand Buffer supplemented with 3 mm MgCl 2, 1.5 mm dntp, 4 mm DTT, 3.3% C1 Loading Reagent, 1.8 µm template-switching oligo C1-P1-RNA-TSO, 1.5 U/µL TaKaRa RNase inhibitor and 18 U/µL Life Technologies Superscript II reverse transcriptase) and PCR mix (1.1 Clontech Advantage2 PCR buffer, 440 µm dntp, 530 nm PCR primer C1-P1-PCR-2 (Supplementary Table 2), 5% C1 Loading Reagent and 2 Advantage2 Polymerase Mix) were added to the designated wells according to the manufacturer s instructions, but using the indicated mixes in place of the corresponding commercial reagents. The plate was then returned to the Fluidigm C1 and the mrna Seq: RT + Amp (1772x/1773x) script was executed, which took ~8.5 h and included lysis, reverse transcription and 21 cycles of PCR. When the run finished, the amplified cdna was harvested in a total of 13 µl C1 Harvesting Reagent and quantified on an Agilent BioAnalyzer. The typical yield was 1 ng per microliter (Supplementary Fig. 9). Tagmentation and isolation of 5 fragments. Amplified cdna was simultaneously fragmented and barcoded by tagmentation, i.e., using Tn5 DNA transposase to transfer adaptors to the target DNA. Ninety-six different 10 transposome stocks (6.25 µm barcoded adaptor C1-TN5-x, 40% glycerol, 6.25 µm Tn5 transposase, where x denotes a well-specific barcode) were prepared, each with a different barcode sequence (Supplementary Table 2). Six microliters of harvested cdna was mixed with 5 µl tagmentation buffer (50 mm TAPS-NaOH, ph 8.5, 25 mm MgCl 2 and 50% DMF), 11.5 µl nuclease-free water and 2.5 µl 10 transposome stock. The mix was incubated for 5 min at 55 C then cooled on ice. Dynabeads MyOne Streptavidin C1 beads (100 µl) were washed in 2 BWT (10 mm Tris-HCl, ph 7.5, 1 mm EDTA, 2 M NaCl, 0.02% Tween-20) then resuspended in 2 ml 2 BWT. Twenty microliters of beads were added to each well and incubated at room temperature for 5 min. All fractions were pooled, the beads were immobilized and the supernatant removed (thus removing all internal fragments and retaining only the 5 - and 3 -most fragments). The beads were then resuspended in 100 µl TNT (20 mm Tris, ph 7.5, 50 mm NaCl, 0.02% Tween), washed in 100 µl Qiagen Qiaquick PB, then washed twice in 100 µl TNT. The beads were then resuspended in 100 µl restriction mix (1 NEB NEBuffer 4, 0.4 U/µL PvuI-HF enzyme), designed to cleave 3 fragments carrying the PvuI recognition site. The mix was incubated for 1 h at 37 C, then washed three times in TNT. Finally, the single-stranded library was eluted by resuspension of the beads in 100 µl 100 mm NaOH, incubation for 5 min, removal of the beads and addition of 100 µl 100 mm HCl and 50 µl neutralization buffer (200 mm Tris, ph 7.5, 0.05% Tween-20). At this point, the typical yield was 1 5 nm single-stranded library. Optionally (but not recommended), the library was further amplified for nine cycles as previously described 21. This additional amplification was used only for the 187ds sample (Fig. 2). Illumina high-throughput sequencing. The library was quantified by quantitative PCRusing KAPA Library Quant (Kapa Biosystems) and then sequenced on an Illumina HiSeq 2000 instrument using C1-P1-PCR-2 as the read 1 primer, and C1-TN5-U as the index read primer. Reads of 50 bp were generated along with 8-bp index reads corresponding to the cell-specific barcode. Reagent costs were reduced more than fourfold compared to commercial kits (Supplementary Fig. 10). Data analysis. Every read that was considered valid by the Illumina HiSeq control software was processed and filtered as follows: (i) any 3 bases with a quality score of B were removed; (ii) the well-identifying barcode was extracted from the 5 end of the read; (iii) if the read ended in a poly(a) sequence leaving <25 transcript-derived bases, the read was discarded; (iv) the UMI was extracted, and if any of the UMI bases had a Phred score <17, the read was discarded; (v) a maximum of nine template-switching generated Gs were removed from the 5 end of the transcriptderived sequence; (vi) if the remaining sequence consisted of fewer than six non-a bases or a dinucleotide repeat with fewer than six other bases at either end, the read was discarded. After filtering, the reads were aligned to the genome using the Bowtie aligner 22, allowing for up to three mismatches and up to 24 alternative mappings for each read. The genome included an artificial chromosome, containing a concatamer of the ERCC control sequences. Any reads with no alignments were realigned against another artificial chromosome, containing all possible splice junctions arising from the exons defined by the known transcript variants. Reads mapping within these splice junctions were translated back to the corresponding actual genomic positions. The UCSC transcript models 23 were used for the expression level calculation. If a locus had several transcript variants, the exons of these were merged to a combined model that represented doi: /nmeth.2772 nature methods

6 2013 Nature America, Inc. All rights reserved. all expression from the locus. To account for incomplete cap site knowledge, the 5 ends of all models were extended by 100 bases, but not beyond the 3 end of any upstream nearby exon of another gene of the same orientation. The annotation step was performed barcode by barcode. For every unique mapping (genomic position and strand) the number of reads in each UMI was counted. Any multiread that had one or more repeat mappings that was outside exons was assigned randomly as one of these repeats and contributed to the summarized read count of that repeat class. Else, if it had one or more mappings to exons, it was assigned at the exon where it was closest to the transcript model 5 end, even if the sequence was repeat-like. If it had no exon mapping, it was assigned randomly at one of the mappings. After assigning reads, the number of molecules at each mapping position was determined by the number of distinct UMIs observed. To account for UMIs that stem from PCR-induced mutations or sequencing errors, any UMI that had fewer reads than 1/100 of the average of the nonzero UMIs was excluded. The raw UMI count was corrected for the UMI collision probability (important only at high UMI counts) as described 10. The expression level of each transcript model was calculated as the total number of molecules assigned to all its possible mapping positions. MEFs and damaged cells were removed on the basis of quality control criteria (Supplementary Note and Supplementary Figs ). Detecting noisy genes. We searched for noisy genes by selecting genes whose noise (measured as CV) was high compared to the pure Poisson distribution (Fig. 3a, dashed line), i.e., where sgene > 3. 7s Poisson The rationale for this selection was twofold: first, that 4 s.d. was enough to exclude all spike-in controls, thus minimizing the false discovery rate; second, that there remained a residual 30% noise even at high molecule counts (for example, compare endogenous genes with the spike-in molecules at >1,000 molecules per cell in Fig. 3a). We found 167 such noisy genes in total. To validate the findings, we performed the following statistical test. We noted that the technical noise distribution (Fig. 3) closely followed that of a Poisson, but its CV was inflated by a constant factor. This can be modeled as a loss factor f, such that only a fraction f of the molecules are actually observed, which inflates the CV from 1/ l to 1/ f l as f approaches zero. To determine the loss factor, we used maximum likelihood to fit the CV = 1/ f l function to the ERCC spike-in controls, as indicated in Figure 3a, and found f = Note that the loss factor combines the effect of actual losses with all other sources of noise, such as differences in reaction conditions, sequencing depth and pipetting errors. We then analyzed each gene in comparison with the ERCC controls, with the null hypothesis that the observed CV obs of the gene was equal to that of a hypothetical ERCC control with the same mean expression. We obtained an expected distribution of CVs by repeatedly (500 times) calculating the CV of 41 (equal to the number of cells) random samples from the Poisson distribution with mean fλ (where f = 0.20). We then verified by Pearson s chi-square goodness-of-fit test that the sampled CVs were normally distributed (omitting genes where P < 0.05 for this test). Finally, we fit a normal distribution to the sampled CVs and used its cumulative density function to calculate the P value of CV obs along this distribution, applying a Bonferroni correction to control for multiple testing. We considered a gene noisy when its corrected significance was α < On the basis of these criteria, we found a total of 118 noisy genes in ES cells (Supplementary Table 1). We performed gene set enrichment analysis using DAVID Islam, S. et al. Nat. Protoc. 7, (2012). 22. Langmead, B., Trapnell, C., Pop, M. & Salzberg, S.L. Genome Biol. 10, R25 (2009). 23. Meyer, L.R. et al. Nucleic Acids Res. 41, D64 D69 (2013). 24. Huang, D.W. Nat. Protoc. 4, (2009). nature methods doi: /nmeth.2772

7 Correcting PCR-induced artefacts by molecule-counting. The plot shows the fold-change relative to the mean, for all genes. On the vertical axis, the fold-change was calculated using reads, whereas the horizontal axis was based on molecules. The expected result is that all genes should fall along the diagonal. However, a subset of the genes showed anomalous fold changes by reads, seen also as a secondary peak in the histogram (blue, right). These anomalous fold-changes were almost fully corrected by molecule counting (red histogram, top).

8 Overview of the method. Single cells were captured in a microfluidic chamber in the Fluidigm C1 AutoPrep instrument. Cells were washed and lysed, and then cdna synthesis was performed using oligo(dt) primer in the presence of a template-switching oligo carrying a unique molecular identifier (UMI) comprising 5 random nucleotides. The resulting cdna was amplified by PCR for 21 cycles, and then tagmented using Tn% transposase carrying a well-specific barcode. 96 libraries were generated in each run, pooled and sequenced on an Illumina HiSeq P1 and P2 denote the Illumina read 1 and read 2 adapters. Read 1 was 50 bp, including the 5 bp UMI and extending into the 5 end of mrna. A separate index read was used to obtain the well-specific barcode for each fragment.

9 Example of multiple transcription start sites (TSSs). The figure shows reads mapped to the Use1 gene in 48 replicate ES cell samples. Each row shows reads from a single cell. Overlapping reads are shaded from light gray (a single read) to black (nine overlapping reads). Known, annotated transcription start sites are indicated by the asterisks at top. Numbers show genomic positions and the horizontal bars show the exon structure of the gene (first five exons only). For all genes, we took the expression value to be equal to the total number of reads or UMIs across all exons.

10 Quantifying highly expressed genes without saturating the UMI. The figure shows the number of UMIs counted for reads aligned to each position upstream of the transcription start site (TSS) of the HGB1 gene in the MAQC human brain reference RNA sample. The data shown is from a single well with barcode TAGAGA. The gene was expressed at a high level, counting a total of 280 molecules of cdna. However, since the TSS was diffuse, there was not a single position where more than 120 distinct UMIs was observed. The highest number of molecules was observed at the canonical TSS position, but a number of molecules were found up to two bases downstream and five bases upstream. As a consequence, the dynamic range for gene expression can exceed the dynamic range for one individual position (which is limited by the total number of UMIs).

11 Sequencing depth. The plot shows the distribution of the number of reads per molecule ( PCR amplification ratio ) for spikes (blue) and endogenous genes (red). The main peak around 10 reads/molecule indicates that the majority of molecules had been sampled about ten times. The peak near 1 read/molecule indicates that a fraction of molecules had been sampled only once.

12 Efficiency and total mrna counts. a The RNA capture efficiency (defined as the fraction of input RNA molecules actually converted to cdna, amplified, sequenced and identified as unique UMIs) was determined from ERCC Spike-in RNA transcripts, counting only transcripts in the range molecules. The average efficiency was 48%. b The total mrna content of each ES cell was estimated by correcting for the capture efficiency in each well.the average mrna content of these ES cells was 200,739 molecules. c Histogram showing the same data as (b).

13 5-3 distribution of reads on ERCC control RNA. The graph shows the number of reads aligned to each position along the 92 ERCC control transcripts. All transcripts are aligned so that their start sites (TSSs) are at position +1. The total length varies from 250 bp up to 2000 bp. The sharp peak at +1 indicates that template switching happens preferentially at the TSS. Inset shows a magnification of the region -20 to +20 around the TSS, demonstrating the highly precise nature of template switching at the 5 end of transcripts.

14 Comparison of mrna count distributions to the ideal Poisson distribution. Each plot shows the cumulative density function (CDF) for the Poisson distribution overlaid on the observed distribution of mrna counts for individual ERCC Spike-In RNA transcripts. The horizontal scales are mrna molecule counts, and the vertical axes show the cumulative probability.

15 Representative BioAnalyzer traces. Each plot shows the BioAnalyzer (Agilent) electropherogram for the amplified cdna obtaned from single cells. Of these cells, all were subsequently accepted by our quality control criteria, with the exception of C7, which failed the criteria.

16 Cost breakdown. The cost of reagents (blue), microfluidic chip (green) and sequencing (red) was calculated for the protocol in this paper ( STRT/C1 ), the commercial Fluidigm C1 protocol and for our original STRT protocol. The STRT/C1 protocol approximately halves the cost of the commercial protocol, and remaining costs are dominated by sequencing.

17 Cell imaging. Each Fluidigm C1 chip was imaged after cell capture, but before lysis. A grid of thumbnails was generated for each chip to facilitate quality controls, and areas and diameters were automatically calculated. The images were then manually inspected for putatively dead cells (Red Fixable Dead Cell Stain) and to detect cases were the automated image analysis had failed. The image shows cells that passed our quality control criteria (left), and cells that failed those criteria (right).

18 Mapped reads. The top chart shows the number of mapped reads in each individual well. The wells are ordered from A01, B01, C01 etc. (leftmost) to F12 (rightmost). There was substantial variation in total yield of mapped reads, which appeared to fall in two distinct categories of low-yield, and highyield wells. These corresponded to cells of high and low quality as judged by cell images (compare A01 and B01, for example). The bottom chart shows reads mapped to the spike-in controls, which were highly uniform with the exception of two excessively amplified wells, and failed well F12.

19 Cell quality control criteria. The plot shows the total number of genes detected versus the fraction of those genes that are cytoplasmic (defined as all non-mitochondrial, non-ribosomal mrna). We used a threshold of 85% cytoplasmic and 5000 genes to reject dead or damaged cells. Cells that stained positive for Red Fixable Dead Cell Stain are indicated in red.

20 The protocol presented here is based on our previously published STRT protocol, with the following major modifications (Supplementary Fig. 2). We had observed that template switching efficiency was greatly reduced with increasing length of the template switching oligo. (TSO) We therefore removed the well-specific barcode and shortened the TSO to 29 bp (from 36 bp). cdna is not purified prior to PCR, but simply diluted 1:10. To prevent the TSO from acting as a primer during PCR, which would generate artefactual molecules, we changed to an all-rna TSO. In the presence of magnesium, RNA rapidly hydrolyses during the hot-start enzyme activation stage of the PCR. Library preparation was simplified by the use of tagmentation, a one-step procedure, instead of fragmentation, end-repair, A-tailing, and adapter ligation. The increased efficiency of this process also allowed us to eliminate library PCR, thus reducing the total amount of PCR from 33 cycles to 21. To make this possible, we replaced the original TSO sequence with one containing the Illumina read 1 adapter. The read 2 adapter was introduced by tagmentation. The protocol was implemented on the Fluidigm C1 AutoPrep instrument, which includes cell capture, imaging, wash, lysis, reverse transcription and PCR. Tagmentation was performed on a Biomek 2000 liquid handling robot. Although the cost of the Fludigm C1 chips is substantial, this is partially compensated by the very low reagent consumption on this instrument. We estimate the reagent cost at $8/cell for this protocol, compared to $27/cell using commercial reagents and $2/cell using the original STRT protocol (Supplementary Fig. 2). To this should be added approximately $14/cell in sequencing costs. Cells were dissociated, counted, stained with a red dead-cell stain (Red Fixable Dead Cell Stain) and frozen as aliquots. Aliquots were then thawed and inspected under microscope. After loading on the C1, each chip was imaged. Automated image capture was used to obtain images of each cell in brightfield as well as the red channel, resulting in an overview image (Supplementary Fig. 11). We observed a large number of cells with a shrunken appearance, almost devoid of cytoplasm (e.g. A01 and B03). Although these cells were not stained, we suspected that they had suffered rupture during processing after staining and had been partially stripped of their cytoplasm during the microfluidic capture process. However, the distinction from healthy cells was not always straightforward. We also

21 observed cells that, although stained red (e.g. putatively dead) appeared to have normal morphology. For example, cells F05 and B04 We therefore decided not to use the images to reject cells, but rather as validation for our bioinformatic quality controls (below). The 96 tagmented cdna libraries were pooled and directly sequenced on the Illumina HiSeq 2000 without further amplification (samples 187ss and 187ss-2 ). For comparison, we also amplified an aliquot of the pooled library, and sequenced this on a single lane (sample 187ds ). Only reads that passed Illumina s acceptance filters were used for further analysis. In total, 541 million raw reads were generated, of which 310 million reads carried a proper barcode, a 5 bp UMI and at least three Gs next to the template-switching site. Of these, 183 million reads mapped to the genome (197 million to exons). The average number of mapped reads per cell was 1.54 million, and the average number of reads mapped to the spike-in controls was 186,000 (Supplementary Fig. 12). The number of mapped reads varied substantially between cells, again reflecting the fact that some cells were apparently dead. This variation was not seen for spike-in controls, with exception of well F12, which had failed completely. Finally, we found two wells in sample 187ds that were excessively amplified during library PCR (this was the only sample that was amplified after cdna synthesis). These wells were omitted from further analysis. Except where indicated, all analyses were performed on the pooled data from three lanes. As expected from the imaging data, we observed two distinct populations of cells. Since we had suspected that some cells had been ruptured and partially stripped of their cytoplasm, we plotted the total number of detected genes versus the fraction cytoplasmic genes (defined as all non-mitochondrial non-ribosomal RNA). The plot (Supplementary Fig. 13) revealed two clearly distinct populations of cells. Using a threshold of 85% cytoplasmic and a minimum of 5000 detected genes as acceptance criteria, we found 50 acceptable cells. Only these cells were used for analysis in the main text. To validate the acceptance criteria we first checked the concordance in accepted cells between the two sequencing libraries that were independently generated from the same cdna (samples 187ss and 187ds); we found that the same 45 cells were accepted in both libraries, plus five additional cells in sample 187ds. Furthermore, these 50 cells were the same that were accepted in the pooled dataset.

22 Second, we compared images of accepted cells with those of failed cells (Supplementary Fig. 11). We found a clear distinction, where failed cells were smaller (measured area 69±24 µm vs. 110±31 µm), often lacked visible cytoplasm, and appeared less rounded. Third, we examined how putatively dead cells (i.e. those stained by Red Fixable Dead Cell Stain) were partitioned among failed and accepted cells according to our criteria. As expected, there was an enrichment of dead cells among failed cells: of all failed cells, 13% were also stained as dead, whereas out of all accepted cells, only 6% were also stained. The existence of some putatively dead cells with apparently normal mrna content could be explained by a difference in timing between the onset of dead-cell staining and the eventual degradation of cellular mrna. Again, these observations were consistent with the imaging results. Finally, we searched for and found nine suspected embryonic fibroblasts (used as feeders in the culture of the embryonic stem cells). These cells were identified based on their expression of fibroblast markers, and their overall distinct expression profiles, as previously described 1. As a result of these quality controls, we retained 41 high-quality embryonic stem cells for the analysis. In order to confirm that the sequencing depth was sufficient to observe most cdna molecules, we plotted histograms of the number of reads per molecule for all genes, as well as for spike-in controls (Supplementary Fig. 5). The average oversampling was around ten-fold, but there remained a substantial number of molecules represented by only 1-3 reads. We tentatively attribute this to mapping errors caused by sequencing errors in the insert of a read, or chimeras generated during PCR. 1. Islam, S. et al. Genome Res. 21, (2011). 2. Huang da, W., Sherman, B.T. & Lempicki, R.A. Nat. Protoc. 4, (2009).