The process of single-cell RT-qPCR is summarized in Figure 1. Our approach, which does not involve mRNA purification, requires a lysis solution that is compatible with downstream enzymatic reactions. In addition, it should 1) disrupt the cell membrane; 2) make the mRNA accessible for reverse transcription; and 3) maintain mRNA integrity. Two agents were evaluated for this task: NP-40, a weak, non-chaotropic detergent (also referred to as Igepal CA-630) and guanidine thiocyanate (GuSCN) a strong, chaotropic compound. Five different lysis conditions were tested in terms of their ability to lyse one cell cluster (pancreatic islet of Langerhans, each containing ~1000 cells), assessed by the amount of accessible insulin II (Ins2) transcripts using RT-qPCR (Figure 2A). NP-40 had no effect compared to control (water) when used at concentrations of 0.5% or 4%, indicating that these lysis conditions are too weak to dissociate the islet and lyse the cells. Proteinase K is commonly used in lysis protocols, but had no beneficial effect when added in the presence of 0.5% NP-40. GuSCN based lysis buffer provided efficient lysis of the islet using a concentration of 0.5 M and increased the RNA yield 600-fold compared to control conditions (Figure 2A). Lysis of a compact cluster of cells is clearly more challenging than dissociated cells. We evaluated the latter using another cell type, primary astrocytes, and the RNA yield was here within a 2.5-fold range for both 0.5 M GuSCN and 0.5% NP-40 compared to the water control (Additional file 1, Figure 1). This indicates that any of the tested lysis conditions will suffice for complete cell lysis of dissociated cells, while 0.5 M GuSCN is required to break apart cell clusters. In a sample containing a single lysed cell, the RNA is assumed to be evenly distributed in the solution. We tested this hypothesis by splitting vortexed single-cell lysates into three separate RT-reactions, followed by qPCR using Ins2 primers. Samples lysed in 0.5 M GuSCN showed 80% lower intra-assay variation than cells emptied in 0.5% NP-40, suggesting that GuSCN lyses the cell and efficiently homogenates the mRNA (Additional file 1: Figure 2A).

The presence of potential inhibitors and RNases from the tissue or culture medium may reduce the cDNA yield. We addressed this concern by adding a known amount of purified total islet RNA to tubes containing one of the following: 1) extracellular solution (see the Methods section for details); 2) one cell in 0.5 M GuSCN in water; or 3) ten cells in 0.5 M GuSCN in water (Additional file 1: Table 1). The amount of RNA in the cells was negligible compared to the externally added RNA. The yield was unaffected by the extracellular solution and in fact increased in the presence of cells in GuSCN, a phenomenon investigated further (see below). As an additional control for inhibitors, we added an equal amount (~2000 copies) of in vitro transcribed artificial RNA, based on the cyclophilin E (Ppie) gene, to the lysis solution. Ppie was measured in every single-cell sample and deviation >1% led to exclusion from further analysis (data not shown).

In Figure 2B, an equal amount of purified RNA was reverse transcribed in the presence of 0, 40, 80 and 120 mM GuSCN, followed by qPCR using five different primer pairs. The RT-reaction efficiency was significantly improved (2- to 6-fold) for all tested genes (Ins1, Gcg, Sst, Gapdh and Rps29) using 40 mM GuSCN. By contrast, 80 mM GuSCN had no effect whereas 120 mM GuSCN was strongly inhibitory. We also evaluated the addition of 0.5–1.5% 2-mercapto ethanol, but no effect on the RT yield, either alone or in concert with GuSCN, was observed (data not shown). Formation of correct PCR-products was confirmed by agarose gel electrophoresis. Neither of the other two lysis conditions, NP-40 or proteinase K, had any discernible effect on RT yield (Additional file 1: Figure 2B ). Taken together, we chose to use ~0.5 M GuSCN in the single-cell lysis solution due to superior cell lysis ability and positive effect on the RT reaction at the concentration of 40 mM.

The total technical noise level of RT-qPCR was estimated by diluting purified total RNA from four to six different concentrations (equivalent of ~5-5·10³ copies of chromogranin B (Chgb), ~10-10⁴ copies of ribosomal protein S29 (Rps29), ~50-10³ copies of SRY-box containing gene 2 (Sox2), nestin (Nes) and glial fibrillary acidic protein (Gfap) and ~500-5·10⁵ copies of Ins2 transcripts). For each dilution, 24 identical RT-qPCR reactions were run. The mRNA concentrations were chosen to primarily cover low-medium abundance transcripts, i.e. Chgb, Gfap, Nes, Rps29 and Sox2. Consequently, the concentration of the high abundance Ins2 transcript is 50–100 times higher. The qPCR-specific noise was calculated from 30 qPCR replicates (Rps29) at five different concentrations of mouse cDNA. The measurements were combined to create a model of RT- and qPCR-induced noise, as well as errors introduced by the liquid handling. The technical reproducibility of RT and qPCR is represented here by the noise strength (η² = SD²/mean²), for different initial copy numbers. Assuming that the noise from each reaction is additive, we can write the total noise as η²_Total = η²_Dilution + η²_RT+ η²_PCR and by fitting this model to the data, we can quantify the contribution of each component. η²_Dilution corresponds to Poisson-type noise which arises when a sample with a low number of molecules is imperfectly diluted. Additional noise is introduced by the RT-reaction and the final term corresponds to the noise introduced by the qPCR step. In Additional file 2, we derive functional forms for each component and we show how they can be assessed from replicate measurements.

Figure 3 shows that the total technical noise is very low at initial mRNA copy numbers down to ~100. The noise model allows us to estimate the RT-reaction efficiency (i.e. the fraction between the resulting number of cDNA copies and the initial number of mRNA copies) for each gene by non-linear regression. We find that the RT-reaction efficiency is 99%, 85%, 67%, 28%, 8% and 3% for Gfap, Rps29, ChgB, Nes, Sox2 and Ins2, respectively. The estimate for Ins2 is clearly much lower than what we would expect and we believe that this is due to insufficient data on the qPCR noise for high cDNA copy numbers for the fitting procedure (see Additional file 2). We conclude that for high RT-efficiencies, the noise originating from the RT reaction is comparatively low, 5–20% of total noise, while the qPCR noise dominates (30–90%) at low concentrations. If the RT-efficiency is low (<10%, i.e. Sox2 and Ins2), the qPCR noise is insignificant for all concentrations.

Using a glass capillary mounted on a micromanipulator, we collected pancreatic β-cells incubated in 3, 6, 10 and 20 mM glucose and emptied the intact cell in 1 μl 0.5 M GuSCN. Following reverse transcription, we subjected each single-cell's cDNA to qPCR using primers for Rps29, Chgb and Ins2 (Figure 4). The median expression level across the different glucose concentrations for these three genes were 9900, 230 and 110 mRNA copies per cell respectively. Some cells failed to generate a signal (34, 59 and 2% for Rps29, Chgb and Ins2 respectively). Transcript level heterogeneity is high and the distribution is skewed, which is in agreement with previous studies of eukaryotic cells 51419. One consequence of this heterogeneity is that the top 10% of the cells with the highest Ins2 expression account for 50% of the total Ins2 mRNA. Increasing glucose concentrations altered the median Ins2 expression from 9000 mRNA copies at 3 mM glucose to 8500, 10000, and 13000 mRNA copies for 6, 10 and 20 mM glucose, respectively (a detailed description is found in Additional file 1: Figure 4). The increase was mainly due to an increasing fraction of cells with very high expression levels. Glucose did not have an effect on Rps29 or Chgb.

How much of this variability stems from true, biological variation and how much is attributable to the RT-qPCR? The RT-qPCR noise model applied on the single-cell data suggests that the RT-qPCR noise is ~0.1% of the total variation in the Ins2 expression (Figure 4). For Rps29, which has lower expression, the range is 10–50% and for Chgb 75–200%. Thus, the decreasing accuracy of the RT-qPCR at lower copy numbers heavily influences measurements of low-abundance transcripts such as Chgb. Moreover, the 95% confidence level shows that a single measurement is accurate to within a factor of two when the expression level is >1000 transcripts (Figure 5). We conclude that in the case of abundantly expressed genes, the distribution of the measurement errors is very narrow compared to the biological noise and the fluctuations introduced by the experimental procedure are negligible.

Finally, we investigated PCR inhibition caused by overloading the reaction with cDNA 11222324. We tried to mitigate the PCR inhibition by decreasing the ratio between reverse transcriptase and Taq polymerase, and found that a ratio ≤ 2 (U/U) was required to completely remove inhibitory effects. At low concentrations of RNA, decreasing the RT-enzyme concentration did not have a negative effect on RT-yield (Additional file 1: Table 2).

Pancreatic islets were prepared from healthy female National Maritime Research Institute (NMRI) mice aged 3–4 months (Bomholtgaard) and fed a normal diet ad libitum. The mice were sacrificed by cervical dislocation, and pancreatic islets were isolated by collagenase P digestion (Roche) followed by manual collection of islets 36. To prepare dispersed single cells the collected islets were gently shaken at low extracellular Ca²⁺ concentration to dissolve the structure of the islet 37. Dispersed cells were plated in plastic 35 mm Petri dishes (Nunc) in RPMI 1640 medium (SVA) supplemented with 10% FCS, 100 U ml^-1 penicillin, and 10 μgml^-1 streptomycin (all from Invitrogen) in the presence of 5 mM glucose (Sigma-Aldrich). The cells were maintained in culture for 2–6 hours, allowing them to attach to the surface of the dish

Primary astrocytes were prepared and cultured as described 38. Cells were grown in Dulbecco's modified Eagle's medium (Sigma-Aldrich) containing 10% fetal calf serum, 2 mmol/L L-glutamine, and penicillin-streptomycin (all Invitrogen).

Dispersed attached cells were washed twice with a buffer (referred to as extracellular solution, EC) containing 138 mM NaCl, 5.6 mM KCl, 1.2 mM MgCl₂, 2.6 mM CaCl₂, 5 mM HEPES (pH 7.4 with NaOH) and 3, 6, 10 or 20 mM glucose to remove dead and loose debris prior to cell collection with patch-clamp pipettes. The dish, containing adhered cells and approximately 1 ml EC-solution, was mounted in a standard inverted light-microscope (Zeiss Axiovert 135). Borosilicate glass capillaries (Hilgenberg GmbH) with outer diameter of 1.5–1.6 mm and wall thickness of 0.16 mm were pulled to pipettes using a patch-clamp pipette puller (Heka PIP5). The diameter of the tip was approximately 10 μm on average, substantially wider than standard patch-clamp pipettes and large enough to allow passage of an intact cell. The glass pipette was mounted on a hydraulic micromanipulator (Narishige) on the microscope. By manually controlling the pressure inside the pipette it was possible to collect intact or nearly intact cells with minimum volume of extracellular solution (<<0.1 μl). In total, 158 cells were collected, and 126 (80%) were identified as β-cells based on their expression of insulin while most of the remaining cells were α-cells (16%). Six samples were negative for all measured genes and they were categorized as technical failures (96% success rate).

An artificial RNA control was generated with the T7 RNA Polymerase in vitro transcription system (Takara). A PCR assay for cyclophilin E (Ppie) was used as template for the in vitro transcription. Ppie PCR product was generated using the protocol as for the SYBR Green-based real-time PCR assays (see below), except that all fluorophores were excluded, and purified using QIAquick PCR purification kit (Qiagen). The PCR product was then amplified using an extended forward PCR primer including the promoter sequence for T7 RNA Polymerase. All primer sequences are found in Additional file 1: Table 3. The resulting (extended) PCR product was purified as above and used in the in vitro transcription reaction, according to the manufacturer's instruction. The reaction mixture contained: 50 U T7 RNA polymerase, 40 mM Tris-HCl (pH 8.0), 8 mM MgCl₂, 2 mM spermidine, 5 mM dithiothreitol (all Takara), 2 mM NTP (Invitrogen), 20 U RNaseOut (Invitrogen) and ~40 ng template DNA. The reaction was carried out at 42°C for 1 h.). In samples containing ~2000 copies of the Ppie RNA, the average intra-assay coefficient of variation was 0.3% and the inter-assay variation was 1.2%.

Due to stochastic fluctuations, relative gene expression levels provide little information, leaving absolute quantification using standard curves as the best option to compare transcript levels within and between cells. For all assays, PCR-products were purified (QIAquick PCR purification kit, Qiagen) and quantified spectrophotometrically (NanoDrop ND-1000) using the following molar absorptivity values (in Moles^-1cm^-1): dAMP, 15200; dTMP, 8400; dGMP, 12010; dCMP, 7050. The PCR-products were diluted in series to generate a standard curve ranging 10-10⁶ copies with PCR-efficiencies spanning 77–95% (Additional file 1: Figure 3). qPCR data analysis was performed essentially as described 9. Since PCR-products are double-stranded and cDNA is single-stranded, one cycle was subtracted from single-cell Ct-values for correct absolute quantification.

The reverse transcriptase SuperScript III (Invitrogen) was used throughout the study. 6.5 μl containing total RNA or lysed single cells, 0.5 mM dNTP (Sigma-Aldrich), 2.5 μM oligo(dT) (15 bp, Invitrogen), 2.5 μM random hexamers (Invitrogen) were incubated at 65°C for 5 min. We then added 50 mM Tris-HCl (pH 8.3), 75 mM KCl, 3 mM MgCl₂, 5 mM dithiothreitol, 20 U RNaseOut and 100, 40 or 10 U (lower amount for single-cell-samples) SuperScript III (all Invitrogen) to total volume of 10 μl. Final reaction concentrations are shown. The temperature profiles used were 25°C for 5 min, 37°C for 10 min, 50°C for 80 min. All reactions were terminated at 70°C for 15 min. Throughout this paper we refer to absolute copy numbers estimated by DNA-based standard curves; assuming an RT efficiency of ≤ 100%, the true number of mRNA copies is equal to or higher than our estimate.

ABI PRISM 7900 HT Sequence Detection System (Applied Biosystems) and LightCycler 480 (Roche Diagnostics) equipped with 384-well blocks were used for qPCR measurements. Reactions (10 μl) with SYBR Green as fluorescent reporter contained 10 mM Tris (pH 8.3), 50 mM KCl, 3 mM MgCl₂, 0.3 mM dNTP, 1 U JumpStart Taq polymerase, 1 × Reference Dye (all Sigma-Aldrich), 0.5 × SYBR Green I (Invitrogen) and 400 nM of each primer (TAGC Copenhagen). The Reference Dye was excluded using LightCycler 480. The iQ™ SYBR^® Green Supermix (BioRad) was used for the Nes assay. The primer sequences are available in Additional file 1: Table 3. All qPCR assays were designed using Primer3Plus http://www.bioinformatics.nl/cgi-bin/primer3plus/primer3plus.cgi and spanned at least one intron when possible. They were optimized to not generate any unspecific products, such as primer-dimers, that could interfere with the quantification of the desired target. Formation of expected PCR products was confirmed by agarose gel electrophoresis (2%) for all assays, and melting curve analysis for all samples.

For RT-qPCR noise calculations, 24-plicate RT-qPCR on diluted purified total mouse islet RNA was used and 30-plicate qPCR reactions on total islet cDNA.

The theoretical model is based on an empirical characterization of the noise in the experiments. First, we consider the noise in the qPCR reaction only and determine how it scales as a function of the number of cDNAs based on replicates which have been diluted to different degrees. The qPCR noise is assumed to be the same for all genes. To model the RT-noise a parameter for the RT efficiency is fitted from replicates which have been diluted to different degrees. Using our empirically determined parameters for the model, we can calculate the contributions of the RT and qPCRs and the total measurement noise can be compared to the biological variation in a population of cells. A full description of the model is given in Additional file 2.