Central dogma at the single-molecule level in living cells

Abstract

Gene expression originates from individual DNA molecules within living cells. Like many single-molecule processes, gene expression and regulation are stochastic, that is, sporadic in time. This leads to heterogeneity in the messenger RNA and protein copy numbers in a population of cells with identical genomes. With advanced single-cell fluorescence microscopy, it is now possible to quantify transcriptomes and proteomes with single-molecule sensitivity. Dynamic processes such as transcription factor binding, transcription and translation can be monitored in real time, providing quantitative descriptions of gene expression and regulation, and the demonstration that a single-molecule event can determine the phenotype of a cell.

This year marks the thirty-fifth anniversary of single-molecule optical detection and imaging. In 1976, Thomas Hirschfeld successfully detected single molecules at room temperature using an optical microscope to reduce probe volume and hence the background signal¹. Figure 1a shows his one-dimensional (1D) fluorescence image of individual immobilized protein molecules, each labelled with tens of fluorophores. The use of tightly focused laser beams eventually allowed single-fluorophore detection in solution phase at room temperature in 1990². Imaging of single fluorophores in ambient environments was first reported with a scanning probe method in 1993³, followed by much easier and improved methods^4–8 akin to Hirschfeld’s that remain the methods of choice for imaging single molecules until the present. In the past decade, improvements in photodetectors and optical components have enabled extensive single-molecule fluorescence studies on a variety of biological problems first in vitro and more recently in living cells.

Figure 1. Stochastic nature of single-molecule processes.

a, Optical imaging of single protein molecules at room temperature. In his 1976 work, Hirschfeld demonstrated the detection of single protein molecules using a fluorescence microscope. A line scan of eight protein molecules was recorded. Adapted from ref. 1. b, Stochastic turnovers of a single enzyme molecule. The fluorescence signal of a cholesterol oxidase molecule exhibits stochastic switching between a fluorescent (reduced flavin) and nonfluorescent (oxidized flavin) state as enzymatic turnovers take place. The waiting time before each switching event is highly variable owing to a single rate-limiting step. Adapted from ref. 9. c, Single-molecule DNA sequencing. A single DNA polymerase is used to sequence DNA by incorporating fluorescently labelled nucleotides of four different colors. Although each incorporation happens stochastically with variable waiting times, the overall time for DNA replication, which is a sum of many sequential steps, is narrowly distributed. a.u., arbitrary units; adapted from ref. 11.

In a single-molecule experiment, one often observes stochastic behaviour, which would be otherwise obscured in an ensemble measurement. Figure 1b shows an early real-time observation of enzymatic turnovers of a single enzyme molecule, cholesterol oxidase⁹. The enzyme contains a flavin moiety that is naturally fluorescent in its oxidized form, but not fluorescent in its reduced form. Each on/off cycle of fluorescence emission corresponds to an enzymatic turnover. This time trace resembles the electrical signal of a single ion channel recorded using a patch clamp — the first single-molecule technique in biology¹⁰. However, in this case, stochastic chemical reaction events of a single enzyme molecule are seen. Here, stochastic means that each fluorescence on/off time is probabilistic. Unlike the deterministic chemical kinetics of ensembles, each time trace is different, although their statistical properties are reproducible. On a single-molecule basis, when a chemical reaction occurs, the formation of a chemical bond completes in less than 1 ps, the event of which cannot be resolved in a single-molecule experiment. However, the waiting time for the event is much longer and probabilistic. When the kinetic scheme of a reaction includes a rate-limiting step, the distribution of the waiting times follows a single exponential, and the number of events in a fixed time interval follows a Poisson distribution.

By contrast, if the overall reaction does not have one rate-limiting step but instead consists of identical sequential steps, the total waiting time is less stochastic. An example of this is DNA replication by a single DNA polymerase, which is the basis of single-molecule sequencing¹¹ and a key application of single-molecule enzymology in biotechnology. A stochastic time trace of individual nucleotides incorporated into a single-stranded DNA template by a single DNA polymerase molecule is shown in Fig. 1c. Although the waiting time for each base incorporation step is exponentially distributed, the total waiting time for replicating the long DNA is narrowly distributed¹², a consequence of the central limit theorem. Bacterial cell-cycle time, when limited by chromosome replication, is not stochastic for this reason¹³. The experiments in Fig. 1b and c were conducted under non-equilibrium steady-state conditions, in which the substrate concentration (thermodynamic driving force) does not change while substrate molecules are continuously converted to product molecules. This is similar to many non-equilibrium processes in a live cell, such as gene expression.

The central dogma of molecular biology states that genetic information encoded in DNA is transcribed to mRNA by RNA polymerase, and mRNA is translated to protein by ribosomes. In a living cell, DNA exists as individual molecules from which gene expression regulation originates. But our knowledge of gene expression has come mainly from genetic and biochemical studies conducted with large populations of cells and purified biomolecules, which often obscures the single-molecule nature of gene expression. In recent years, single-molecule experiments in vitro have provided mechanistic insight into the functions of macromolecules involved in gene expression, including transcriptional and translational machineries^14–18. Compelling areas of further investigation involve the observation and quantitative description of gene expression and regulation in a living cell.

Not only is there only one copy (or a few copies) of a particular gene, but the copy number of a particular mRNA is also small owing to short intracellular mRNA lifetimes¹⁹, at least in a bacterial cell. By contrast, the copy number of particular proteins range from zero to 10,000 (refs 20, 21); many important proteins, such as transcription factors that regulate gene expression, have small copy numbers²². Consequently, single-molecule sensitivity of mRNA and protein is needed to quantify gene expression in individual cells.

Because of the stochasticity associated with the single or low copy-number macromolecules, gene expression of individual cells cannot be synchronized. It is therefore necessary to make real-time observations of gene expression and regulation in single living cells. In particular, stochastic binding or unbinding of transcription factors on a particular gene, when rate limiting, must result in stochastic mRNA production, just like the single enzyme traces in Fig. 1b, c. Stochastic degradation of individual mRNA molecules further contributes to fluctuations in protein production. These temporal fluctuations of the mRNA and protein numbers (see sketches in Fig. 2) result in cell-to-cell variation of the copy numbers, or gene expression ‘noise’. Under the steady-state condition, the connection between temporal fluctuations and variation within the population is similar to ergodicity in statistical physics — the time average of a system equals the ensemble average of identical systems.

Figure 2. Central dogma at the single-molecule level.

In a living bacterial cell, there is usually one copy of a particular gene, which is regulated by transcription factors (TFs), and transcribed into mRNA and translated into protein. A rate-limiting event, such as TF binding and unbinding on DNA, in this single molecule process results in stochasticity. The expression levels of mRNA (middle panel) and protein (bottom panel) show temporal fluctuations in a single cell lineage. This gives rise to variations of mRNA and protein copy numbers among a population of cells at a particular time (right panels).

Here we review recent single-molecule experiments that provide quantitative descriptions of the central dogma in living bacterial cells, although the strategies and technical advances highlighted are applicable to future studies in eukaryotic cells. We show that single-molecule stochastic events have important biological consequences, such as determining the phenotype of a cell.

Imaging single molecules in living cells

To image a particular biomolecule in a live cell with fluorescence microscopy, specific labelling is required. The advent of genetically encodable fluorescent proteins has provided the highest specificity so far, with minimal perturbation for live-cell imaging²³, allowing real-time observations of fusion proteins of interest. Although the weak signal of a single-fluorescent-protein molecule is detectable in vitro using a fluorescent microscope together with a combination of laser excitation and modern charge-coupled-device detectors (Fig. 3a), the use of single fluorescent-protein reporters in live cells is challenging due to strong cellular autofluorescence. This obstacle can be partly overcome by selecting fluorescent proteins that are spectrally separate from the autofluorescence, which is generally blue-green²⁴. Yellow- or red-emitting fluorescent proteins are therefore favourable for live-cell single-molecule imaging. Furthermore, in the same spirit as Hirschfeld’s experiment, signal can be improved by reducing the detection volume to minimize autofluorescence background,. For example, total internal reflection fluorescence microscopy (TIRFM) can limit the axial depth by illuminating with an evanescent wave that penetrates only a few hundred nanometres into a sample (Fig. 3b). TIRFM is therefore ideal for studying membrane protein dynamics^25,26, but it does not allow imaging of the whole cell body. However, single-fluorescent-protein imaging using wide-field illumination is possible in bacterial cells, and their compact cell bodies make them ideal for single-molecule studies in vivo (Fig. 3a).

Figure 3. Methods for imaging single molecules in live cells.

Single-molecule fluorescence can be imaged using multiple laser illumination geometries that reduce the probe volume. a, In wide-field illumination, the entire cell is subject to laser exposure. For bacterial cells that have small volume, no further probe volume reduction is necessary. b, With total internal reflection, only the region within a few hundred nanometres from the coverslip is illuminated. This method is often used to image single membrane proteins, but cannot detect molecules deep in the cells. c, Two-photon excitation suppresses out-of-focus background, but suffers from slower time resolution owing to the need of point scanning d, Sheet illumination has reduced background, as well as increased time resolution because it does not require point scanning.

In living eukaryotic cells, imaging a single fluorescent protein is more difficult. A typical mammalian nucleus is 5–10 µm in diameter, compared with 1 µm for a bacterial cell. In a wide-field microscope, such a large cell volume gives rise to a strong out-of-focus autofluorescence signal, which overwhelms the signal of a single fluorescent protein. Probing DNA–protein interactions therefore requires 3D sectioning. Although confocal fluorescence microscopy with one-photon excitation could be used, it also causes photobleaching outside the focal plane²⁷. One solution to this problem is to use two-photon fluorescence microscopy^28,29 (Fig. 3c), which allows localized excitation only at the laser focus, considerably reducing out-of-focus photobleaching while providing 3D sectioning in living eukaryotic cells. But, like confocal microscopy, it requires point scanning, thus limiting its time resolution. Alternatively, sheet illumination^30,31, in which a thin light sheet illuminates only the image plane (Fig. 3d), provides low fluorescence background and high sensitivity, as well as high temporal resolution, because it does not require point scanning. These techniques are being adapted for single-fluorescent-protein imaging in living eukaryotic cells.

In a bacterial cell, a freely diffusing protein is difficult to image because its fast diffusion spreads the signal throughout the whole cell^32,33. However, if a single fluorescent protein is localized, it can be imaged above the cellular autofluorescence background³⁴. This method, termed detection by localization (Fig. 4a), works as long as there is only one immobilized molecule in a diffraction limited volume (less than 10 molecules within a bacterial cell). Detection by localization can be done by tethering on membrane ³⁴, specific or even transient non-specific binding on chromosome³⁵.

Figure 4. Real-time measurements of gene expression with single-molecule sensitivity.

a, Detection by localization. The cellular autofluorescence makes it difficult to detect a freely diffusing fluorescent protein. However, a localized single molecule can be imaged above the autofluorescence background. b, Detection of single transcription factors in live cells. A lac repressor (LacI) labelled with YFP can be imaged when bound to its operator site on DNA. The localized fluorescence disappears after dissociation caused by the inducer IPTG. DIC, differential interference contrast microscopy; adapted from ref. 35. c, Target search by a transcription factor. When the inducer IPTG is removed (dilution), the lac repressor begins to search for its operator site. After rebinding to the target, localized fluorescence appears, allowing measurement of the search time. Adapted from ref. 35. d, Real time observation of protein synthesis at low expression levels. Individual YFP fusion protein molecules are visualized after being immobilized to the cell membrane, are synthesized in bursts after intrinsic noise. Adapted from ref. 34. e, Copy numbers of mRNA and protein of the same gene, measured in the same cell, show little correlation, which is mostly due to the differences in the lifetimes of mRNA and protein. The protein copy number distribution follows a gamma distribution. Adapted from ref.21.

In cases where the frame rate of the camera is insufficient to detect transient localization (<10 ms), a shorter pulse of laser excitation can be provided with each imaging frame^33,35, an idea borrowed from strobe photography. Detection by localization therefore allows single-molecule observations with millisecond time resolution.

The width of a single molecule image is about half of the optical wavelength due to the diffraction limit. However the accuracy of determining the center position of a single isolated fluorescent protein can be as high as a few nanometer ^34,36. To image more concentrated samples, higher spatial precision can also be achieved by selectively observing only one molecule at a time using photo-activatable fluorescent proteins. This is the idea behind recent developments in single-molecule-based super-resolution imaging, such as stochastic optical reconstruction microscopy³⁷ and photoactivated localization microscopy^38,39, in which high-resolution images are reconstructed from many single-molecule images. Future applications of super-resolution techniques will probably change the way we view intracellular processes⁴⁰ such as gene expression. Single-fluorophore detection as discussed above remains a prerequisite for super-resolution imaging.

Transcription factor dynamics

As the first step of gene expression, transcription factors must bind to or unbind from DNA in response to environmental signals. Because transcription factors interact with DNA at one location, gene expression is stochastic when the binding and unbinding of a transcription factor becomes rate-limiting (Fig. 2). In the classic example of the lac operon, the transcription factor known as the lac repressor (LacI), which is expressed at less than five copies per cell³⁵, binds to or unbinds from operator sites to control transcription. With detection by localization, a single lac repressor fused to yellow fluorescent protein (YFP) can be visualized when bound to its operator in the lac operon³⁵. When the inducer isopropyl-β-d-thiogalactoside (IPTG) is added to the cell, localized fluorescent spots disappear as a result of LacI dissociation (Fig. 4b). This live cell assay allows single-molecule measurements of transcription-factor dissociation kinetics.

In addition, the binding kinetics can be measured. When IPTG is removed from the medium, the localized signal reappears, indicating the rebinding of LacI (Fig. 4c). This experiment allowed the first measurement of the time required for a LacI molecule to find a vacant operator site on DNA. It takes less than 360 s for one repressor to search for one specific binding site³⁵.

The search time of 360 s is a result of complex molecular processes. The protein-DNA search problem was extensively studied in the 1970s and 1980s^41,42. It was observed that the DNA-binding rate constant of transcription factors significantly exceeds that expected from the 3D diffusion limit for bimolecular binding^42,43. This observation led to the prevailing model of facilitated diffusion. For a transcription factor or any DNA-binding protein to find a target sequence on DNA, it first binds somewhere along the DNA non-specifically and undergoes 1D diffusion in search of the target. If the target is not found, the transcription factor dissociates from the DNA to avoid long search time imposed by 1D diffusion. The 3D diffusion through cytoplasm is much faster, allowing the transcription factor to reach other far away segments of DNA quickly. This combined 1D and 3D search is repeated until the transcription factor finds the DNA segment containing the target sequence. With single-molecule experiments, one can probe these phenomena in real time and quantify the process.

In a series of single-molecule studies in vitro, 1D diffusion has been directly observed for fluorescently labelled transcription factors and other DNA-binding proteins along non-specific DNA under a microscope^44–49. The observed 1D diffusion rate (on the order of 0.05 µm² s⁻¹) is much slower than the 3D diffusion in a live cell (~3 µm² s⁻¹) because the 1D diffusion of the transcription factor is coupled to simultaneous rotation around the DNA, such that the transcription factor tracks the pitch of the DNA double helix^50,51. In the in vitro experiments, low salt concentrations were used to assure long non-specific residence time in order to record long trajectories of 1D diffusion. In a living cell, high salt concentration shortens the residence time, but the diffusion constant often remains the same⁴⁶. Consequently the number of bases inspected in each 1D segment is reduced.

A key question is whether such facilitated diffusion occurs in live cells. Recent single-molecule experiments suggest that it does. We found that, during the search process, a transcription factor spends 90% of its time on non-specific DNA, and the residence time of non-specific binding is less than 5 ms (ref. 35). Given the 1D diffusion constant in vitro, the protein inspects ~100 base pairs (bp), which implies a 100-fold acceleration of target search compared to the case with no 1D diffusion⁵². This observation is consistent with mounting evidence^53–57 that the length of DNA segment a transcription factor inspects is shorter than 1,000 bp, the value estimated from early in vitro experiments⁴³. This 100-bp range indicates that for a 5 × 10⁶ bp genome, a transcription factor must inspect 5 × 10⁶/100 = 5 × 10⁴ segments before reaching the target site. Therefore, the total search time for one transcription factor in a cell is ~5 × 10⁴ × 5 ms = 250 s, in close agreement with the measured search time³⁵.

The combination of different single-molecule approaches have resolved the search problem and led to a quantitative understanding of the facilitated diffusion of transcription factors in bacteria. Similar single-molecule experiments should be able to address the same search problem in mammalian cells, which are complicated by nucleosomes⁵².

Translation in real time

Transcription-factor binding or unbinding leads to transcription and translation. Although the central dogma has been well established, real-time observation and quantitative description of transcription and translation in a single cell, at the single molecule level as motivated above, have only become possible in recent years. These studies have yielded unexpected observations of these fundamental processes in live cells^34,58–62.

We first discuss protein translation, as it can be relatively easily studied under repressed (uninduced) conditions, on a single-molecule basis. Under these conditions, single-molecule experiments have shown that proteins are synthesized in bursts³⁴, and that the characteristics of the bursts can be understood quantitatively at the molecular level. The production of individual molecules of a YFP-fused membrane protein was monitored in real time in E coli.³⁴ (Fig. 4d). Newly synthesized YFPs were visualized one by one as diffraction-limited spots through detection by localization, and they were purposely photobleached after being detected. A fast-maturing YFP, Venus⁶³, was used to achieve seven-minute time resolution in the observation of translation. Using this approach, translational bursting from the lac operon under repressed conditions was observed³⁴. Each burst creates four proteins on average, at a frequency of about one burst per generation time (although not synchronized to the cell cycle). The number of bursts per cell cycle follows the Poisson distribution -- a consequence of the exponentially distributed waiting times between each burst.

Because it was shown that each burst results from transcription of a single mRNA (generated due to the occasional dissociation of the LacI repressor), the observed translational burst must therefore be due to several rounds of ribosomal initiation on the same transcript. This transcript is degraded by nucleases with a stochastic cellular lifetime that is exponentially distributed with a time constant of 1.5 min. The longer an mRNA lives, the more proteins it produces. Consequently, as theoretically predicted in the 1970s^64–66, the burst size is exponentially distributed. This observation of exponentially distributed protein copy numbers per burst was independently confirmed by another single-molecule assay using β-galactosidase activity as a reporter⁶². As we discuss later, such stochastic expression due to transcription factor unbinding can be important in determining how a cell is induced in the presence of external stimuli⁶⁷.

Transcription in real time

As discussed above, under repressed conditions, proteins are generated in bursts. At high expression levels, fluctuations of protein production in single cells have been extensively studied^68,69. The origin of these fluctuations has been partly attributed to stochastic transcription and degradation of mRNA from the gene locus, but has not been directly visualized at the molecular level. Whereas the high copy number of proteins prevents the resolution of single copies, recent single-molecule studies of mRNA have shown that transcription is also burst-like^58–61. But, in contrast to translational bursts, the origin of transcriptional bursts is largely unknown.

One widely used method to detect single mRNA molecules in living cells uses the bacteriophage coat proteins (MS2) that stably bind to specific RNA sequences⁷⁰. To visualize single copies of mRNA, cell lines are engineered to express both MS2–GFP and mRNA containing several MS2-binding sites. First developed by the Singer group, this method allows real-time observation of transcript production, and is ideal for probing transcriptional dynamics in living cells by tracking and counting single mRNA–MS2–GFP complexes^71–72. A caveat is that the secondary structure associated with the binding sites and MS2 binding often interfere with the native mRNA degradation pathways⁷³, preventing the profiling of endogenous mRNA expression levels.

When MS2-containing mRNA is expressed under fully-induced conditions, the production of transcripts is found to be intermittent⁵⁸. If transcript production were to have a single rate-limiting step, such as RNA polymerase binding or initiation, the waiting time between the birth of each mRNA would be exponentially distributed, and the copy-number distribution would be Poissonian (with a variance equal to the mean). Surprisingly, short bursts (average 6 min) of mRNA synthesis followed by long periods (average 37 min) of inactivity have been observed⁵⁸. The burst-like transcription is similar to that shown in Fig. 2, even though there is no known transcription factor binding or unbinding in this case. This burst-like transcription was also observed using fluorescence correlation spectroscopy on MS2-bound mRNA in Escherichia coli⁵⁹, as well as in eukaryotic cells^60,61. Although the overall waiting time between each mRNA synthesis event is not exponentially distributed⁵⁸, the waiting times for transition between the active and the inactive states are. Accordingly, the copy-number distribution is super-Poissonian, meaning that the variance of the distribution is greater than the mean. In other words, the cell-to-cell variation is significantly greater than would be expected from a single, rate-limiting process.

This important finding pointed out that transcription from a supposedly constitutive promoter is not as simple as RNA polymerases transcribing with a constant flux. Rather, it is a much noisier process, and the origin of this noise is unknown. Possible candidates include nucleoid-structuring proteins that are analogous to eukaryotic chromatin, global fluctuations of chromosome supercoiling states and RNA polymerase availability. In vivo single-molecule approaches are poised to further reveal the workings of these fundamental processes.

Single-cell trajectories related to cell-to-cell variation

Under steady-state conditions, temporal fluctuations of gene expression in each cell lineage, as discussed in the previous sections and Fig. 2, lead to variation in copy number in an isogenic population of cells. A typical copy number distribution, which is often asymmetrical, is shown in Fig. 4e. A rigorous mathematical relationship between fluctuations in expression and the distribution of protein copy number in a population of cells was needed. A lognormal function has often been used as a convenient phenomenological fit, but it offers no physical insight.

For each gene, the dynamics of the central dogma can be described by two parameters — the burst frequency, a, which is the number of bursts per cell cycle, and the burst size, b, which is the average number of molecules produced per burst. Experimentally, a and b can be determined by single-cell trajectories, such as in Fig. 2. Alternatively, the fact that temporal fluctuations in a cell lineage are related to cell-to-cell variation of copy numbers implies that a and b can also be inferred from for a population of isogenic cells at a particular moment as observed with a microscope or flow cytometer.

To find the relationship, one seeks a governing equation for gene-expression dynamics — the chemical master equation, which was first used by Delbruck⁷⁴ in 1940. In the late 1970s, the chemical master equation was applied to obtain protein copy-number distributions resulting from stochastic gene expression^64,65. It was not until a decade ago that this approach regained attention ^66,75,76, prompted by new experimental capabilities. Given the chemical kinetics scheme and rate constants connecting all the macromolecules involved in the central dogma, one can, in principle, solve the chemical master equation, which naturally yields time-dependent fluctuations. In practice, this can be simulated numerically using the Gillespie algorithm⁷⁷. Under certain conditions, analytical results can be obtained. For example, under steady-state conditions with uncorrelated and exponentially distributed bursts, the chemical master equation can be solved⁷⁸, and the copy-number distribution, p(n), can be approximated as gamma distribution when the copy number is approximated as a continuous variable⁷⁹:

• p(n)=n^a−1e^−n/b/b^aГ(a)

The gamma distribution has two kinetic parameters — a and b, as defined earlier — providing a clear physical interpretation of the copy-number distribution.

This mathematical relationship allows extraction of intrinsic kinetics parameters (a and b) from fitting a gamma function to the measured copy-number distribution. At low expression levels, the values for a and b determined this way are consistent with those derived from the single-cell trajectories^34,62. As shown in the next section, the cell-to-cell variation at high expression levels is more complicated but remains well described by a gamma distribution.

Global profiling of expression variation

The ability to image single molecules in bacteria has offered an opportunity to profile protein expression globally, at any abundance levels. Pioneering work using a yeast green fluorescent protein (GFP) fusion library⁸⁰ surveyed the cell-to-cell variation of more than 2,500 genes under various growth conditions, yielding several important observations^81,82. First, the noise, or the variance divided by the mean squared, scales inversely with abundance. Second, the deviation of noise in a particular gene away from the global trend reflects its protein’s function and perhaps the underlying regulation. However, because single-molecule sensitivity in yeast cells had not been achieved at the time, 30% of the genes that were weakly expressed in the GFP library were left undetectable.

To profile global variation at all expression levels, an E. coli YFP fusion library was constructed, which included more than 1,000 genes with expression levels ranging from 0.1 to 10⁴ per cell²¹. Of all the tagged proteins, approximately 99% of the copy-number distributions are well fit with the gamma distribution. 50% of the proteins are expressed at an average level of less than ten molecules per cell, which argues for the necessity of single molecule sensitivity in single cell analyses. Protein-expression noise has two distinct scaling properties relative to the mean. Below ten molecules per cell, the noise is inversely proportional to protein abundance. This scaling is the same as what was observed in yeast, indicating that the noise from random birth and death of molecules, also known as intrinsic noise^83,84, dominates the expression variation for low-abundance proteins. By contrast, at abundances above ten molecules per cell, the noise reaches a plateau of 30% and does not decrease any further. This noise plateau is common, or ‘extrinsic’ to most high-abundance proteins, as the expression levels of different proteins have a large covariance from cell-to-cell. Notably, time-lapse movies have shown that the extrinsic noise fluctuates at a timescale much longer than the cell cycle, suggesting that a slow global-regulation process is at work⁶⁸.

At the transcriptional level, the same YFP library was used to simultaneously survey mRNA and protein variation for 137 genes²¹. Instead of labelling with MS2, which requires further cloning steps, mRNA was visualized using single-molecule fluorescence in situ hybridization (FISH)⁸⁵ in fixed cells. Unlike conventional approaches that use several hybridization probes against the mRNA, the YFP mRNA was targeted using a universal singly labelled FISH probe optimized for both hybridization efficiency towards its targets and specificity against off-targets. It was found that, even for highly expressed genes, the average mRNA copy number is less than five per cell. Among a population of genetically identical cells, every mRNA species has a super-Poissonian distribution, which is related to the transcriptional bursts observed in the real-time experiments and suggests that this is a general phenomenon for most genes.

The simultaneous profiling of mRNA and protein²¹ also showed that the mRNA and protein copy numbers of a single cell for any given gene are uncorrelated; that is, a cell that has more mRNA molecules than average does not necessarily have more proteins (Fig. 4e). This perhaps counterintuitive result can be explained by the fact that mRNA has a much shorter lifetime than protein in bacteria¹⁹. This finding argues for the necessity of single-cell proteomics analyses, and offers a warning for interpretations of single-cell transcriptome analyses, at least for bacteria. A mammalian cell, on the other hand, has comparable mRNA and protein lifetimes, and hence is expected to have more correlated mRNA and protein levels than in a bacterial cell.

Single-molecule event triggers phenotypic switching

How cells with identical genomes have different phenotypes is an interesting question. Phenotypes are the physical, chemical and physiological states of the cell as related to function, determined by both the genome and environment. Given the ubiquitous and substantial noise described earlier, it is evident that the phenotype of a cell cannot be solely defined by its transcriptome and proteome. Cells can tolerate rather large noise of protein and mRNA abundance while tightly maintaining their phenotypes. A compelling question is what molecular actions dictate the transition between phenotypes.

In some cases, the cell phenotype can be clearly defined when there are bimodal or multimodal distributions of proteins, in contrast to the unimodal copy-number distribution that are most often observed²¹. As shown in Fig. 5b, a population of isogenic E. coli cells, in which lac permease is labelled with YFP, shows bistability. The lac operon in E. coli, consisting of lacZ, lacY and lacA genes, is normally repressed by the lac transcription-factor repressor (LacI) in the absence of inducer (Fig. 5a). When the inducer is present, it inactivates LacI and triggers expression of the lac operon. The synthesis of the permease increases the inducer influx that inactivates more LacI repressor, creating a positive feedback on permease expression⁸⁶. Without inducer, no cells are induced, whereas with high inducer concentrations, all cells are induced. At moderate inducer concentrations, only a fraction of the cells is induced. This bistability is controlled by the positive feedback of the lac operon.

Figure 5. Phenotype switching due to a single-molecule event.

a, Bistability of the lac operon. The positive feedback by the normally repressed lac permease (LacY) results in bimodal distribution at intermediate inducer concentration with two distinct phenotypes, fluorescent or not. b, Threshold for positive feedback. The copy-number distribution of LacY in uninduced cells is the same as that without inducer, suggesting that the typical leaky expression of LacY is not sufficient to trigger the positive feedback. c, A large burst of expression originates from the stochastic event of the complete dissociation of a single transcription factor from DNA. A large expression burst of LacY (~300 molecules) is necessary to trigger the positive feedback. The switching of the phenotype is attributed to the complete dissociation of a single transcription factor, LacI. This experiment shows that a low probability single-molecule event can determine cell fate. Adapted from ref. 67.

Bistability is commonly exploited by bacteria to generate alternative phenotypes⁸⁷, such as persistence against antibiotics⁸⁸, lysis or lysogeny after phage infection⁸⁹ and induction of the lac operon in E. coli⁸⁶. Although much is known about the genetic switches, what drives the transition between two phenotypic states is unclear in many cases. How does a single cell make a decision on which phenotype to choose? With single-molecule imaging, uninduced E. coli cells have been shown to contain 0–10 copies of the permease enzyme, which is below the threshold for positive feedback (more than 300 molecules per cell)⁶⁷. Transition to the fully induced state therefore requires a large burst of protein synthesis (Fig. 5c).

The transcription factor controlling permease synthesis, the LacI repressor, is a tetramer that binds to two DNA-binding sites, creating a DNA loop. Partial dissociation of LacI and rapid rebinding to DNA result in a single copy of mRNA and a small burst of permease, as was observed in the aforementioned real-time studies of the repressed lac promoters. When the repressor completely dissociates from both operators on DNA, a large burst of permease arises, because it takes a few minutes for the repressor to rebind³⁵. Indeed, bistability was eliminated in strains without DNA looping⁶⁷. It is the stochastic single-molecule event of complete repressor dissociation from DNA that triggers the cell’s phenotypic switching.

Looking forward

We have shown that in the case of lac operon, the workings of the genetic switch can be quantitatively understood at the molecular level. This is an example that low probability, stochastic events of a single molecule have important biological consequences. Another trivial example is point mutations in the course of evolution.

It is well recognized that such stochastic events are connected to cell-fate determination of other systems⁹⁰ For example, there is considerable evidence that bacterial persistence against antibiotics is a stochastic process involving gene expression⁸⁸. Persisters are not drug resistant but are drug tolerant. Drug resistance is related to a changing genome, whereas persisters have identical genomes, but different phenotypes. The phenomenon exists for many bacterial species and antibiotics. The molecular mechanism behind persistence is largely unknown, partly because the tools are not available. Understanding the molecular mechanism of persistence may be crucial to drug development, especially for diseases such as tuberculosis, caused by the bacterium Mycobacterium tuberculosis, which kills almost two million people every year worldwide. Single-cell gene-expression profiling may shed light on the mechanism of persistence.

Similarly, the reprogramming of somatic cells into induced pluripotent stem cells in the presence of certain transcription factors is also stochastic⁹¹. There are no elite cells and every cell has a certain probability of being reprogrammed,in the presence of some transcription factors, which is analogous to stochastic switching in E. coli lac operon at low inducer concentrations. Yet, unlike the lac operon, the molecular mechanism is unknown. Extension of single-molecule approaches into mammalian cells and stem cells will allow real-time monitoring over long periods so that low probability events with considerable biological consequences can be observed directly. We anticipate that the single-molecule approaches summarized in this Review will lead to more biological discoveries for many years to come.