Single-cell RNA bursting (genomics)

Genes make RNA in bursts. From RNA counts in thousands of single cells one can infer how often a gene fires and how much it makes each time, provided the statistics are done carefully; time-resolved data then show that some mammalian genes wait between bursts on a regular clock rather than at random. Includes a short film.

Two kinds of time in mammalian promoters (six minutes) — a telling of the second paper on this page. Most classifiable mouse genes switch on at random, but one in seven waits days between bursts on a regular clock whose length tracks the methylation of its promoter; these are the matrix and secreted-protein genes that build the world outside the cell.

Genes fire in bursts

A gene is not transcribed at a steady trickle. Its promoter, the stretch of DNA where transcription starts, switches on, polymerases run off a batch of transcripts, and it switches off again. This transcriptional bursting was followed in real time in single bacteria in 2005, filmed in a slime mold in 2006, and found in mammalian cells the same year by counting RNA molecules cell by cell.

The standard description is the telegraph model of Ko (1991) and Peccoud and Ycart (1995). A promoter flips from OFF to ON at rate $k_{\rm on}$ and back at rate $k_{\rm off}$, makes transcripts at rate $k_{\rm syn}$ while ON, and each transcript decays at a rate that sets the unit of time. Two combinations summarize a gene: the burst frequency $k_{\rm on}$ and the burst size $k_{\rm syn}/k_{\rm off}$, the number of transcripts a typical burst produces. A cell can turn a gene up either way, so whether a given piece of regulatory DNA acts on frequency or on size is a question about mechanism.

Deciding between frequency and size for every gene in the genome means inferring the rates from single-cell RNA sequencing, which counts each gene’s transcripts cell by cell; in hybrid mice from two inbred strains the two copies of every gene are counted separately. Larsson and colleagues did this in 2019 and fitted the telegraph model to every gene with a program called txburst. They reported that burst frequency is encoded mainly in enhancers, regulatory DNA that can lie far from the gene, and burst size in core promoters. The counts are small, zero in most cells, and the whole inference rests on the probability the model assigns to them.

From counts to rates

At steady state the telegraph model predicts a definite distribution for the count $n$ in a cell. The promoter’s recent activity acts as a random dimmer $u$ between zero and one, and given the dimmer setting the count is Poisson:

$$P(n)=\int_0^1 \mathrm{Pois}(n;\,k_{\rm syn}u)\,\mathrm{Beta}(u;\,k_{\rm on},k_{\rm off})\,du .$$

Here $\mathrm{Pois}$ is the Poisson probability of $n$ at mean $k_{\rm syn}u$ and $\mathrm{Beta}$ is the Beta density, shaped by the two switching rates so that a rarely-on promoter puts most weight near $u=0$. The integral has a closed form in Kummer’s confluent hypergeometric function ${}_1F_1$. A gene’s log-likelihoodthe probability that the model, at given rates, assigns to the counts actually observed; the fitted rates are the ones that make it largest is the sum of $\log P(n_c)$ over its cells $c$. The fitted rates maximize it, and the error bars come from how fast it falls away from the maximum. txburst computes the integral with a 50-point quadrature rulea recipe that replaces an integral by a weighted sum of the integrand at fixed points; accurate only while the integrand varies slowly between them and adds $10^{-10}$ to each probability before the logarithm, a routine guard against $\log 0$. Neither txburst nor any other implementation the authors examined reports an error bound on the value.

The likelihood can instead be evaluated exactly. Written as a power series, the closed form alternates in sign and loses its digits to cancellation once $k_{\rm syn}$ is large; Kummer’s transformation turns it into a series with all terms positive, so nothing cancels and the tail past a known index is bounded. Two such sums at the top of the count range then seed a three-term recurrence for every other probability:

$$n(n{+}1)\,P(n{+}1) \;=\; n\,(n+k_{\rm on}+k_{\rm off}+k_{\rm syn}-1)\,P(n)\;-\;k_{\rm syn}\,(n+k_{\rm on}-1)\,P(n{-}1).$$

Willmot and Panjer wrote this recurrence down in the actuarial literature in 1987; as far as the authors could determine, no gene-expression tool had used it. The recurrence has to be run downward; run upward from $P(0)$ it amplifies rounding error until, in one tested regime, no correct digit is left by $n\approx200$. The computation is done in ball arithmeticcomputing with intervals instead of numbers: every quantity is stored as a midpoint plus a radius proven to contain the true value, and every operation updates both; a standard tool of computer-assisted proofs with the Arb library, so each likelihood value comes with a proven error bound. A gene’s probability vector takes about 11.7 milliseconds, five to eight times the quadrature’s cost and cheap enough to use everywhere.

Line chart on logarithmic axes: the maximum relative error of the count probabilities P(n) against the exact reference, from 10^-15 to 10, against the synthesis rate k_syn in units of the mRNA degradation rate, from about 4 to 10,000. Three rising curves. A pink curve labeled BPSC, N = 10 starts near 10^-15 at k_syn of 4, passes 10^-7 near k_syn of 10 and flattens just below 10 beyond k_syn of a few hundred. An orange curve labeled txburst test, N = 40 runs flat near 10^-12 from k_syn of 40 to about 130, then climbs steeply, crossing 10^-6 near 280 and approaching 1 by 10,000. A yellow curve labeled txburst ML, N = 50 runs flat slightly below the orange one until about 200, then climbs the same way, crossing 10^-6 near 400. Dotted vertical lines in the matching colors mark (10/pi)^2 near 10, (40/pi)^2 near 162 and (50/pi)^2 near 253, each at the foot of its curve's climb. Gray dashed horizontal lines mark relative errors of 10^-2 and 10^-6.

Where fixed quadrature gives out. Each curve is the largest relative error in the count probabilities $P(n)$ of an $N$-point quadrature rule, measured against the exact evaluation, as the synthesis rate $k_{\rm syn}$ grows (both axes logarithmic; rates in units of the mRNA decay rate; the promoter shape is the fibroblast median). Pink is the 10-point rule of the program BPSC, orange the 40-point rule of txburst’s test module, yellow the 50-point rule of its fitting routine. The dotted verticals mark $(N/\pi)^2$, about 10, 162 and 253: each rule is accurate well below that scale and loses its digits rapidly above it. About 90% of the published fits have $k_{\rm syn}\le 300$, where the 50-point rule is still good to $3.1\times10^{-9}$, so the damage is confined to the upper tail of synthesis rates. (From Figure 1 of the first paper.)

Rechecking the published fits

When all 32,376 fits of Larsson and colleagues are re-evaluated at their own parameters, about 99% agree with the exact log-likelihood to better than a thousandth of a natural-log unit. Another 205 are off by more than one unit, one by 4,827. Few of these have synthesis rates high enough for the quadrature to fail; in nearly all, instead, some cell carries a count that is almost impossible under the fitted model, and the added $10^{-10}$ caps that cell’s penalty, so the program reports a better fit than the model allows. Refitting on the exact likelihood moves burst frequency or size by more than one error bar (the narrower conditional kind, other rates held fixed) for just 12 of 20,201 quality-filtered genes, so the published point estimates stand.

The published error bars, unlike the point estimates, need recomputing. An interval ends where the log-likelihood has fallen a prescribed amount below its maximum. One line of txburst’s interval routine halves that amount, so nominal 95% intervals are really 83.4% intervals: error bars meant to miss the true value one time in twenty miss it about one time in six. Rescaling is only part of the repair: for 2,066 of the 7,177 genes with reproducible published intervals (28.8%) an endpoint also moves by more than 10% under the exact likelihood, for 309 of them by more than half. The authors did not test whether any biological conclusion of Larsson et al. changes.

Which parameter differs when a gene’s two copies burst differently was tested on deeper data, 682 fibroblasts of the same cross from one animal, sequenced by Johnsson and colleagues in 2022. Of 198 genes tested with exact likelihoods, 77 differ cleanly between the copies, but separate tests of burst frequency and burst size can say which parameter differs for only 19 of them: 17 in frequency, 2 in both, none in size alone. That split says little about how often each mechanism occurs, because at these expression levels a twofold difference in frequency is detected 74% of the time and one in size only 36%. In simulations an even mixture of the two mechanisms gives a split at least this uneven about four times in five. Jiang and colleagues (2017) had concluded from the program SCALE that allelic differences act overwhelmingly on frequency, and Larsson et al. that enhancers set frequency and core promoters size; the calls reported here do not test either claim.

The limits of snapshot data

In live-cell recordings many promoters are refractory: after a burst they pass through more than one slow step before they can fire again, so their silences are more regular than a single random switch allows. The count distribution of such a two-step promoter can also be evaluated exactly.

Detecting a refractory promoter from snapshot counts is limited less by numerical error than by how little information the counts carry. If a gene’s counts come from a two-step promoter, the expected evidence from $n$ cells against the closest telegraph distribution, the average log-likelihood ratio, is $n\,\mathrm{KL}$, where $n$ is the number of cells and $\mathrm{KL}$ is the Kullback–Leibler divergence between the two distributions, a standard measure of how distinguishable they are. Near-certain detection needs $n\,\mathrm{KL}$ to reach about 8. Across 20,216 genes of the 2019 data, each at its own cell count and fitted rates, the largest value is 0.58, for Timp1 with 224 cells. No gene in that data set was measured in enough cells to reveal whether its promoter needs recovery time between bursts. A gene with the median fibroblast parameters would need at least $1.2\times10^{5}$ cells sequenced as these were, and $3.9\times10^{8}$ if the counts were thinned further to a droplet-likedroplet-based sequencing handles far more cells than the plate-based method used here but counts a much smaller share of each cell’s RNA 6% of this experiment’s capture. Even with every transcript counted the same gene would need more than $10^{4}$ cells. These cell numbers hold for a two-step silence only; no bound is given here for silences with more steps.

Adding a clock to the counts

The alternative to more cells is a different kind of data. In metabolic labeling, cells are fed 4-thiouridinea modified uridine that cells build into new RNA in place of ordinary uridine; a chemical treatment before sequencing makes incorporated labels read as C instead of T for a set time, and RNA made in that window carries a mark that sequencing can read, so every molecule is called old or new. The NASC-seq2 data of Ramsköld and colleagues (2024) give old and new counts of every gene in 8,916 primary mouse fibroblasts after a two-hour window, with the new RNA also resolved by gene copy.

Two hours is comparable to a typical fibroblast promoter’s own timescales: a few hours silent and about two hours active at the median. A snapshot constrains the silent period mainly through its mean; a new-RNA count built up from zero over a known interval also depends on its shape, and the competing models differ only in that shape. In the two-state model the promoter leaves its one silent state at rate $k_{\rm on}$; in the multi-step model it passes through $k$ silent states in sequence, leaving each at rate $k\,k_{\rm on}$. As functions of the waiting time $t$, the silent-period densities are

$$f_1(t)=k_{\rm on}\,e^{-k_{\rm on}t},\qquad f_k(t)=\frac{(k\,k_{\rm on})^{k}\,t^{\,k-1}}{(k-1)!}\,e^{-k\,k_{\rm on}t}.$$

Both have mean $1/k_{\rm on}$. The exponential is memoryless: however long a promoter has been silent, its chance of firing in the next minute is unchanged. The second, an Erlang density, is the waiting time for $k$ random steps in a row. Its variance is $k$ times smaller and it has almost no weight near zero, so a promoter that has just fired is unlikely to fire again at once, which is what refractory means. Everything else is shared, including the mean count.

For each gene and trial set of rates, promoter state plus old and new counts form a finite Markov chaina random process whose next move depends only on its present state, here the promoter’s condition plus the running counts of old and new molecules. It starts at stationarity when the label goes on and is propagated across the window by uniformizationa classical way to compute how a continuous-time random process evolves over a fixed time, as a sum of non-negative terms whose omitted remainder is known exactly, with the series truncated once the neglected probability is below $10^{-13}$. Every computed distribution thus carries an explicit error bound and no simulation noise. The two gene copies, a one-in-ten chance that a new molecule reads as old, and each cell’s sequencing depth are folded in exactly. The two models have the same three free rates and neither contains the other, so the $P$ value cannot be read off a textbook chi-square curve; it is computed exactly for each gene under that gene’s fitted two-state model.

Top: a timeline with two square-wave traces labeled allele 1 promoter (on/off) and allele 2 promoter (on/off), a row of RNA molecules drawn as gray dots labeled old and orange dots labeled new (labeled), and an orange shaded band labeled 4sU labeling window (2 h) running from 0 h: label added to 2 h: cells collected. Below, three model diagrams: two-state (telegraph) in blue, circles off and on exchanging at k_on and k_off with new RNA leaving on at k_syn, captioned silent period exponential, mean 1/k_on (memoryless); multi-step (refractory) in red, circles off1, off2 through offk to on, each step at rate k times k_on with return at k_off, captioned silent period Erlang-k, mean 1/k_on (a preferred duration); clustered in green, two parallel off states off_f and off_s with exit rates (2 plus root 2)a and (2 minus root 2)a, each entered at k_off/2, captioned a fast/slow mixture, mean 1/a. At right of each diagram a small density plot of silent period over its mean with a dotted line at 1: blue decaying from its maximum at zero, red rising to a peak below 1 and then decaying, green dropping steeply from a high value at zero.

The labeling design and the three promoter models. Top: during a two-hour 4-thiouridine window each new transcript is marked, so a cell reports old and new molecules of every gene, drawn here for a gene’s two copies. Below: the two-state, multi-step and clustered promoters, identical except in how they return from silence, and at right their silent-period densities at equal mean: memoryless, peaked at a preferred duration, and a fast/slow mixture. Only the shape of the silence separates the models, and a new-RNA count built up over a known window is sensitive to that shape where a snapshot is not. (Model diagrams as in Figure 1A of the second paper.)

The refractory class

The two-state and two-step fits both converged for 16,858 of 23,246 genes, and the exact test rejects the telegraph model for 3,737 of them at a 5% false-discovery ratethe expected share of false positives among the genes called significant, here held to one in twenty. Three further controls, one against cell-to-cell variation in capture and two against the doubled gene copies of cells that have replicated their DNA, leave a refractory class of 186 genes. Another 1,060 clearly favor the two-state model; 329 of those fit better still with a “clustered” silence that mixes a fast and a slow timescale, less regular than memoryless. The refractory class is thus 15% of the 1,246 classifiable genes and 1.1% of those evaluated. The other 15,612 are undetermined at this depth, where a two-step silence shows only for genes with long silences and large bursts. Two steps is a lower bound: for the twelve refractory genes fitted with longer chains the likelihood keeps rising up to eight steps and for most of them is flat beyond that.

Among the refractory class’s 164 genes with functional annotations in the Gene Ontology, cell-cycle genes set aside, 62 encode proteins of the extracellular region, a 4.8-fold enrichment4.8 times as many genes of this kind as a random draw of the same size from all annotated genes would be expected to contain. Extracellular-matrix genes are enriched 7.9-fold; these are the genes for what a fibroblast secretes and builds up around itself. The median silent period is 95 hours against 3.9 for two-state genes, and 171 of the 186 have mean silences longer than a 20-hour cell cycle. With active periods three times longer and bursts twice as large, a refractory gene is transcribed for a few hours roughly every four days.

Scatter plot on logarithmic axes: mean silent period in hours, from below 1 to above 1,000, against mean active period in hours, from 0.01 to 100, one mark per gene. A dense cloud of about a thousand pale blue dots labeled telegraph + clustered (1,060) is centered near an active period of 2 hours and a silent period of 3 to 4 hours, mostly between 1 and 10 hours silent, with a sparse scatter of dots out to very short active periods and a few above 100 hours silent. 186 red diamonds labeled refractory (186) lie higher, between about 15 and 1,300 hours silent and mostly between 1 and 20 hours active, densest near 5 to 10 hours active and 30 to 300 hours silent. A dashed horizontal line labeled 20 h separates nearly all the diamonds, above it, from nearly all the dots, below it.

Days of silence. Every mark is one classified mouse fibroblast gene, placed by the mean length of its active periods (horizontal) and of its silent periods (vertical), both in hours on logarithmic axes. Red diamonds are the 186 refractory genes; blue dots are the 1,060 genes of the two-state class, those that clearly favor the two-state model over the multi-step one, including the 329 with a clustered silence. The dashed line marks a 20-hour fibroblast cell cycle. The refractory genes sit almost entirely above it, 171 of 186, with a median silence of 95 hours against 3.9 for the blue cloud, and their active periods run longer as well. Part of the separation is selection: at this sequencing depth a multi-step silence can be detected only for genes with long silences and large bursts. (Panel C of Figure 2 of the second paper.)

Refractory and two-state genes differ at the promoter too. Most two-state promoters sit in CpG islandsstretches of DNA unusually rich in the two-letter sequence CG and usually left unmethylated; most widely expressed mammalian genes start inside one; refractory promoters are CpG-poor, overlap an island half as often (39% against 81% of expression-matched partners) and carry more DNA methylation, a chemical mark generally associated with silence (mean 43% against 12%). No transcription-factortranscription factors are DNA-binding proteins that turn genes up or down by recognizing short sequence motifs near them motif, of 1,019 tested, is enriched beyond what base composition explains. In public histone-markhistones are the proteins DNA is wound around; chemical tags on them mark a promoter as active or as repressed profiles from embryonic fibroblasts and cell lines, refractory promoters carry the marks of active promoters at reduced levels; only the repressive mark H3K27me3 is more common (53% against 35% of promoters). Within the class the silent period lengthens with promoter methylation (Spearmana correlation coefficient computed on ranks rather than values: +1 when two quantities always rise together, 0 when they are unrelated $\rho=+0.39$) while the active period does not. Pausing of RNA polymerase just past the start site, a known control point of bursting, is an unlikely mechanism for the extra step: measured pauses last seconds to minutes against a median refractory silence of 95 hours, and pausing indicesthe density of polymerase parked just downstream of a gene’s start divided by its density along the gene body, read from published run-on sequencing are lower at refractory promoters.

Regularity and sensing speed

A multi-step silence presumably takes more machinery, which raises the question of what the cell gains from the regularity. Two classical answers, lower protein noise and a higher information rate, come out small at the fitted rates. The third is how quickly a change can be sensed. If a signal raises a gene’s burst frequency, no mechanism in the cell that reads the bursts can detect the change sooner than the optimal sequential detector of statistical theory, an idealized observer of the burst train. A $k$-step promoter’s intervals vary $k$ times less than memoryless ones of equal mean, so each carries $k$ times the information about a rate change. At a fixed false-alarm rate the delay, counted in intervals, then falls roughly as $1/k$.

A worked example is a measured case, the estrogen-responsive human gene TFF1, whose bursts Rodriguez and colleagues (2019) timed at 185 minutes apart at a half-maximal dose and 66 at saturation. If one counts only runs in which the alarm follows the change, a promoter with seven silent steps detects that shift after 114 minutes on average and a memoryless one with the same mean intervals after 444. That is 3.9 times sooner, or 4.7 at a stricter false-alarm setting, with no added expression noise. The gain is per burst cycle, though; in clock time refractory genes are slow, because their fitted burst cycles are an order of magnitude longer than those of two-state genes. The per-cycle gain is offered as a hypothesis about secretory cells: a secreted protein cannot be recalled, so deciding quickly and correctly when to restart transcription should be advantageous.

The papers

References

I. Golding, J. Paulsson, S. M. Zawilski and E. C. Cox, Real-time kinetics of gene activity in individual bacteria, Cell 123 (2005) 1025bursting followed in real time in single bacteria
J. R. Chubb, T. Trcek, S. M. Shenoy and R. H. Singer, Transcriptional pulsing of a developmental gene, Curr. Biol. 16 (2006) 1018bursting filmed in the slime mold Dictyostelium
A. Raj, C. S. Peskin, D. Tranchina, D. Y. Vargas and S. Tyagi, Stochastic mRNA synthesis in mammalian cells, PLoS Biol. 4 (2006) e309bursting in mammalian cells, found by counting RNA molecules cell by cell
M. S. H. Ko, A stochastic model for gene induction, J. Theor. Biol. 153 (1991) 181the telegraph model: a promoter switching on and off at random
J. Peccoud and B. Ycart, Markovian modeling of gene-product synthesis, Theor. Popul. Biol. 48 (1995) 222the telegraph model as a Markov process and its transcript statistics
A. J. M. Larsson, P. Johnsson, M. Hagemann-Jensen, L. Hartmanis, O. R. Faridani, B. Reinius, Å. Segerstolpe, C. M. Rivera, B. Ren and R. Sandberg, Genomic encoding of transcriptional burst kinetics, Nature 565 (2019) 251the genome-wide allele-resolved fits with txburst that the first paper rechecks
T. N. Vu, Q. F. Wills, K. R. Kalari, N. Niu, L. Wang, M. Rantalainen and Y. Pawitan, Beta-Poisson model for single-cell RNA-seq data analyses, Bioinformatics 32 (2016) 2128BPSC, the beta-Poisson package whose 10-point rule appears in the first figure
G. E. Willmot and H. H. Panjer, Difference equation approaches in evaluation of compound distributions, Insur. Math. Econ. 6 (1987) 43the three-term recurrence for the count probabilities, from actuarial mathematics
F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Trans. Comput. 66 (2017) 1281the ball-arithmetic library behind the proven error bounds
P. Johnsson, C. Ziegenhain, L. Hartmanis, G.-J. Hendriks, M. Hagemann-Jensen, B. Reinius and R. Sandberg, Transcriptional kinetics and molecular functions of long noncoding RNAs, Nat. Genet. 54 (2022) 306the 682 deeply sequenced fibroblasts used for the allele-resolved tests
Y. Jiang, N. R. Zhang and M. Li, SCALE: modeling allele-specific gene expression by single-cell RNA sequencing, Genome Biol. 18 (2017) 74SCALE; concluded that allelic differences act mostly on burst frequency
D. M. Suter, N. Molina, D. Gatfield, K. Schneider, U. Schibler and F. Naef, Mammalian genes are transcribed with widely different bursting kinetics, Science 332 (2011) 472live-cell recordings in single fibroblasts showing a refractory period between bursts
D. Ramsköld, G.-J. Hendriks, A. J. M. Larsson, J. V. Mayr, C. Ziegenhain, M. Hagemann-Jensen, L. Hartmanis and R. Sandberg, Single-cell new RNA sequencing reveals principles of transcription at the resolution of individual bursts, Nat. Cell Biol. 26 (2024) 1725NASC-seq2 and the time-resolved counts from 8,916 fibroblasts that the second paper analyzes
J. Rodriguez, G. Ren, C. R. Day, K. Zhao, C. C. Chow and D. R. Larson, Intrinsic dynamics of a human gene reveal the basis of expression heterogeneity, Cell 176 (2019) 213live-cell timing of TFF1 bursts, the worked example of the sensing-speed calculation

← back to the web summaries