WorksAgricultural and Biological Sciences
Molecular memory reduces timing noise in stochastic gene expression
Abstract
Fluctuations in gene expression affect the timing precision of downstream events triggered when a gene product reaches a critical threshold. The first-passage time (FPT) to this threshold characterizes the timing of such events. In conventional single-step birth–death models, reaction waiting times are exponentially distributed and hence memoryless. In this setting, FPT noise exhibits a U-shaped dependence on the threshold, attaining its minimum near one-half of the steady-state mean. However, mRNA production and degradation often involve multiple sequential steps, leading to hypoexponential distributions of mRNA production times and mRNA lifetimes. Here, we investigate how molecular memory arising from multistep mRNA production and degradation affects FPT statistics in bursty gene expression. We calculate FPT statistics in the small-noise limit. Both multistep mechanisms reduce FPT noise, but through distinct mechanisms. Multistep burst arrival reduces mRNA copy-number fluctuations, thereby suppressing timing noise. Multistep degradation instead reshapes transient accumulation dynamics, reducing relative timing noise even when steady-state mRNA noise increases. The threshold that minimizes FPT noise remains close to one-half of the steady-state mean, except for multistep burst arrival with constant burst sizes, for which it shifts to lower values. Stochastic simulations support the analytical predictions within their regimes of validity and show that these qualitative conclusions persist beyond the small-noise approximation.
Introduction
First-passage processes provide a unifying framework for stochastic phenomena in which a fluctuating variable first reaches a prescribed boundary. Such problems arise throughout statistical physics, from diffusion-controlled reactions and target search to escape processes and neuronal firing [1, 2]. The same framework is particularly natural for intracellular processes, where a cellular event is often initiated when the abundance of a regulatory molecule crosses a critical threshold [5] [3, 4]. When the regulatory molecule is a gene product, fluctuations in its threshold-crossing time arise from the intrinsic stochastic biochemical reactions underlying gene expression [6]. The noise in the corresponding first-passage time (FPT) therefore quantifies the temporal precision of the downstream event. In simple memoryless birth–death models of gene expression, for which reaction waiting times are exponentially distributed, FPT noise depends nonmonotonically on the normalized threshold and is minimized when the threshold is close to one-half of the steady-state mean [7] [4]. This nonmonotonic behavior has also been observed experimentally in bacteriophage λ lysis timing by altering the threshold through sequence mutations in holin [8].
The study of threshold-crossing dynamics has expanded beyond the minimal birth–death model to encompass more complex biological processes and regulatory mechanisms. FPT approaches have been applied to bacterial cell-division control and to gene-expression models incorporating promoter switching, external disturbances, and cell growth [9, 10, 11, 12]. Experimental studies of bacteriophage λ lysis have further demonstrated how the holin threshold and translational efficiency influence lysis-time statistics [8, 13]. The role of regulatory mechanisms involving feedback loops, post-transcriptional regulation, and molecular sequestration in shaping FPT statistics has also been extensively investigated [14, 15, 16, 17, 18, 19]. Collectively, these studies have advanced our understanding of how stochastic gene-expression dynamics and cellular regulation influence the timing of intracellular events.
Despite these advances, most theoretical treatments of FPT statistics still model molecular production and degradation as single-step, memoryless processes with exponentially distributed waiting times. Experimental observations and mechanistic models, however, support multistep descriptions of both transcription and mRNA degradation [20, 21, 22]. When the intermediate states are coarse-grained, the total waiting time for a sequence of exponential steps is generally hypoexponential, introducing effective molecular memory [23, 24]. At the production level, the intermediate states may correspond to transcription initiation, pre-mRNA processing, and nuclear export; at the degradation level, they may correspond to deadenylation, decapping, and exonucleolytic decay [25, 26, 27]. Queueing-theoretic approaches have related general interarrival-time and lifetime distributions to steady-state copy-number moments [28]. However, how memory in burst arrival and mRNA degradation reshapes FPT statistics remains largely unexplored.
Here, we investigate how two sources of molecular memory—multistep burst arrival and multistep mRNA degradation—reshape FPT statistics in stochastic gene expression. We derive transient moments of the mRNA abundance from the chemical master equation and use them to obtain analytical expressions for the mean and noise of the FPT in the small-noise limit. For multistep burst arrival, the analytical calculation of moment dynamics assumes that the mean interburst interval is short relative to the mRNA lifetime, a regime relevant to many mammalian genes [29, 30, 31, 32]. At a fixed mean burst arrival time and mean mRNA lifetime, both multistep burst arrival and multistep degradation reduce FPT noise, but through distinct mechanisms. Multistep burst arrival regularizes production and lowers copy-number noise, thereby directly suppressing timing noise.
Multistep degradation accelerates the mean mRNA response, increasing its logarithmic sensitivity to time at a fixed threshold and thereby allowing relative timing noise to decrease even when steady-state copy-number noise increases. For constant burst sizes, multistep arrival shifts the optimal normalized threshold below the Markovian value , whereas multistep degradation leaves the optimum close to or above one-half. Stochastic simulations support the analytical predictions within their expected regimes and show that the qualitative trends persist beyond the strict timescale-separation regime. Thus, burst-arrival and degradation memory provide distinct routes for controlling both temporal precision and the optimal triggering threshold.
| Symbol | Description | Symbol | Description |
| mRNA production rate | Burst size | ||
| mRNA degradation rate | # sequential transcriptional steps | ||
| Total mRNA copy number at time | # mRNA degradation steps | ||
| Mean of mRNA at | Mean steady-state of mRNA | ||
| Coefficient of variation of mRNA at time | Steady-state CV of mRNA | ||
| Variance | FF | Fano factor | |
| Threshold number of mRNA | Minimum normalized threshold | ||
| First passage time | Coefficient of variation of FPT |
Table 1. Summary of notation used
First-Passage-Time Framework
We define the first-passage time (FPT) as the minimum time required for the mRNA copy number to reach a critical threshold for the first time. Starting from zero initial conditions, the FPT is
where is the threshold level required to trigger a downstream event. Since gene expression is stochastic, varies between realizations. When copy-number fluctuations are small, the mean FPT, , can be approximated by the time at which the mean mRNA trajectory reaches the threshold,
Timing noise is quantified by the squared coefficient of variation of the FPT, , the ratio between variance and mean squared. Using the small-noise approximation [33] the FPT noise can be written as
is the logarithmic sensitivity of the mean mRNA level with respect to time. We emphasize that the formula given by Eq. (3) is valid in the limit of small fluctuations in copy number. Nevertheless, in the small-noise limit, the FPT noise admits a clear physical interpretation through its decomposition into two contributions: the first is determined by the sensitivity of the mean copy-number dynamics to time, while the second is determined by the relative copy-number fluctuations at the mean FPT. While more general approaches have been developed to approximate FPT distributions directly from master equations or Fokker–Planck equations [34], our analysis focuses on the relationship between FPT noise and mRNA moment dynamics. For the multi-state burst-arrival and multistage-degradation processes considered here, we validate the small-noise approximation directly against exact SSA simulation results.
Multistep Burst Arrival Reduces Timing Noise
We first examine how molecular memory in the burst-arrival process affects mRNA fluctuations and, consequently, first-passage-time statistics. A simple way to introduce such memory is to represent the interburst cycle by sequential transcriptional states, , as shown schematically in Figure 1(a). Here, denotes the transcriptional state, and the completion of the final transition produces a burst of mRNAs, which undergo single-step degradation at rate . For sequential exponential steps, each with the same rate , the interburst time follows an Erlang distribution with mean , and coefficient of variation squared . The states are coarse-grained kinetic states, not necessarily distinct molecular species. Depending on the biological system, they may capture delays associated with promoter remodeling, assembly of the transcriptional machinery, and the transition to productive transcription [22, 24]. The equal-rate irreversible chain is therefore an effective model of the aggregate waiting time and does not imply that the underlying biochemical steps are identical or map one-to-one onto the states.

Figure 1. Multistep burst arrival suppresses both mRNA abundance noise and first-passage-time noise. (a) A schematic of -state transcription process. The nodes , , denote successive transcriptional states, and transitions occur at rate . Completion of the final transition produces an mRNA burst of size and resets the system to . Individual mRNAs undergo single-step degradation at rate . (b,c) Representative stochastic mRNA trajectories and corresponding first-passage-time distributions for (b) and (c) , evaluated at the same event threshold. Increasing reduces mRNA noise and narrows the timing distribution. (d–h) Results for constant burst. (d) Mean mRNA abundance and (e) squared coefficient of variation, , as functions of dimensionless time . (f) Dimensionless mean FPT, , and (g) FPT noise, , as functions of the normalized threshold . (h) Optimal normalized threshold, , where FPT noise is minimum, as a function of for and . (i–m) Corresponding results for zero-inclusive geometrically distributed bursts. (i) Mean mRNA abundance, (j) mRNA noise, (k) dimensionless mean FPT, (l) FPT noise, and (m) optimal normalized threshold. Solid curves show the analytical approximations in Eqs. (3), (4), (5), and (7), and symbols show stochastic simulation results [39]. Parameter values for these plots: , and , and to keep fixed.
To obtain the transient abundance statistics, we derive the moment equations from the joint chemical master equation for the transcriptional state and the mRNA copy number. The detailed calculation is provided in Appendix Sec. A. The resulting mean and noise (coefficient of variation squared) dynamics are given as follows:
In the steady-state limit (), our noise expression becomes
which agrees with the previous queueing-theoretic results for bursty gene expression with Erlang-distributed interburst times [28, 35, 36, 37]. For a fixed mean interburst time, hence for a fixed steady-state value of the mean mRNA number (), the mean dynamics in Eq. (4) do not depend on the number of burst arrival steps (see Figure 1 d, i, respectively, for constant and geometric burst distributions). However, the noise in the mRNA level in Eq. (6) decreases with for any burst distribution. To compare the mean and noise dynamics of mRNA levels with stochastic simulation, we consider constant and a geometric burst size distribution (see Figure 1 d, e, i, and j). For constant burst size, each production event generates a fixed number of mRNAs, , such that . In particular, corresponds to non-bursty production, where each event produces a single mRNA. For geometrically distributed burst sizes, we consider the zero-inclusive geometric distribution, , for , where and .
Using the mean mRNA dynamics in Eq. (4), the mean FPT, the average time to reach a threshold mRNA level , can be obtained from Eq. 2 as
The mean FPT for a given is independent of burst arrival steps , which agrees with exact stochastic simulation as well for both constant and geometric burst distributions (see Figure 1 f, k).
Since the FPT noise in Eq. (3) depends directly on the mRNA noise, the reduction in copy-number fluctuations is expected to improve the precision of threshold-crossing times. This effect is illustrated in Figure 1(b,Figure 1c), where sample trajectories cross the same threshold with substantially less temporal dispersion for than for . The corresponding timing histogram is narrower for larger , indicating reduced FPT variability. The quantitative dependence of FPT noise on the number of burst-arrival steps is shown in Figure 1 for deterministic and geometric bursts, in panels (g) and (l), respectively. For both cases, increasing lowers . However, the dependence of the optimal threshold on is strongly influenced by the burst-size statistics. Figure 1(h) and (Figure 1m) show the optimal threshold , defined as the value of the normalized threshold at which is minimized, as a function of for constant and geometric bursts, respectively. For constant burst, increasing the number of transcription steps shifts below the classical value near 0.5. In contrast, for geometrically distributed bursts, increasing lowers the overall magnitude of the FPT noise while leaving the optimal threshold close to one half of the steady-state mean.
So far, we have considered a fixed mean burst size, . To examine how burst magnitude modifies FPT statistics, we vary both and for constant and geometric burst-size distributions in Appendix Fig. 4. For both distributions, increasing increases FPT noise, , which is also consistent with previous FPT studies [38]. At a given and , geometric bursts produce larger FPT noise than constant burst because of the additional fluctuations in burst size. The optimal threshold also depends on mean burst size, , for both constant and geometric burst distributions. In particular, shifts toward higher values as is increased. Thus, the effects of and compete: increasing the number of burst-arrival steps reduces FPT noise and tends to shift the optimal threshold downward, whereas increasing the burst size enhances FPT noise and shifts the optimum toward higher thresholds. These results extend the fixed- findings and show that the precision of threshold-crossing times and the location of the optimal threshold are jointly controlled by burst-arrival regularity and burst magnitude.
Multistep Degradation Reduces Timing Noise
Having considered molecular memory in burst arrival, we next investigate memory generated by multistep mRNA degradation, as illustrated in Figure 2(a). Standard coarse-grained models represent mRNA degradation as a single first-order reaction, resulting in an exponentially distributed and memoryless transcript lifetime. In cells, however, mRNA degradation can involve several successive biochemical processes, including deadenylation, decapping, and exonucleolytic decay. Such pathways can produce nonexponential, age-dependent lifetime distributions [21]. The resulting molecular memory can substantially modify gene-expression fluctuations, as demonstrated in theoretical and experimental studies of general lifetime distributions and multistage gene-expression models [20, 22, 40].

Figure 2. For bursty production, multistep mRNA degradation increases copy-number noise but reduces FPT noise. (a) Schematic of the multistep-degradation model. Bursts arrive with rate and average size , and each mRNA passes through sequential degradation steps, each with transition rate , before removal. The total mRNA abundance, , is the sum over all degradation states. (b,c) Mean mRNA dynamics and (d,e) corresponding copy-number noise, , for constant and geometric burst size distributions, respectively, with . Solid curves show the exact analytical results in Eqs. (8) and (10), and symbols denote stochastic simulation algorithm (SSA) [39] results for , 5, and 10. (f,h) Mean FPT as a function of the normalized threshold . Solid curves show the threshold-crossing estimates obtained from the exact mean dynamics using Eq. (2), while symbols denote SSA results. (g,i) FPT noise, , obtained from SSA matches well with Eq. (3) under small noise approximation. We keep the mean mRNA lifetime () fixed with by changing . The parameter values are and to keep the mean mRNA lifetime fixed at 10 minutes. The burst-arrival rate is chosen as to keep the steady-state mean fixed.
In our model, transcriptional bursts arrive according to a Poisson process with rate and burst size . Newly produced transcripts enter the first degradation state, , and subsequently pass through sequential states, , before removal. All transitions between degradation states, including removal from the final state, occur with rate . The total mRNA abundance includes transcripts in all degradation states. For equal transition rates, the mRNA lifetime follows an Erlang distribution with mean mRNA lifetime , and coefficient of variation squared . The case recovers the conventional exponential lifetime, whereas produces a nonexponential lifetime with increasing temporal regularity. When all degradation steps have identical rate , the exact mean dynamics of the total mRNA abundance is given as follows:
The first term represents the steady-state mean mRNA abundance, determined by the mean production rate and the average mRNA lifetime. The second term accounts for the transient buildup from the initially empty system, with its time dependence reflecting the sequential degradation steps. As this transient contribution vanishes at long times, the mean mRNA abundance approaches its steady-state value,
The detailed derivation is provided in Appendix Sec. B 1. The mean mRNA dynamics for constant and geometric burst distributions are shown, respectively, in Figure 2(b) and (c) for . For both burst distributions, Eq. (8) agrees well with the stochastic simulations for , , and degradation steps. When the mean mRNA lifetime is held fixed, increasing makes the dynamics faster. Consequently, the mean FPT defined in Eq. (2) reduces with , as shown in Figure 2(f,Figure 2h). The analytical predictions for the mean FPT, obtained using Eq. (2), also agree well with the stochastic simulations for multistep mRNA degradation.
The noise in the mRNA level is quantified by the squared CV, which is the ratio between the variance and the square of the mean. For the multistep-degradation model, we derive the noise dynamics in Appendix Sec B 2, given as follows:
The first term represents the Poisson contribution to mRNA fluctuations, whereas the second captures the additional fluctuations associated with bursty production. The magnitude of the latter depends on both the burst-size statistics and the degradation dynamics, which determine how long transcripts produced within the same burst survive together. Taking the long-time limit (), the steady-state noise becomes
which agrees with the general queueing-theoretic expression [28] when the molecular lifetime distribution is specialized to the Erlang distribution arising from sequential degradation steps. The factor captures the effect of multistep degradation, which increases monotonically with , hence increasing the steady-state mRNA noise. For unit-sized bursts (), this additional contribution vanishes, and the steady-state abundance remains Poissonian, independently of the number of degradation steps [24]. We validate the noise dynamics in mRNA levels for multi-step mRNA degradation with stochastic simulation in Figure 2 (d) and Figure 2 (e), respectively, for constant and geometric burst size distributions. For constant bursts (), with unit burst size, , Eq. (11) shows that the steady-state noise is Poissonian, where , which is independent of . However, for , increases with at fixed mean mRNA lifetime. However, for geometrically distributed bursts, ; consequently, Eq. (11) predicts that the steady-state noise increases monotonically with for any at fixed . For the same mean burst size , geometric bursts produce larger fluctuations than constant ones because of their additional burst-size variability, which is also evident in Figure 2 (d) and Figure 2 (e).
In contrast to the behavior of the steady-state abundance noise, the first-passage-time noise decreases as the number of degradation steps, , increases, as shown by the stochastic simulation results in Figure 2(g) and Figure 2(i), respectively, for constant and geometric burst distributions. This reduction arises through a mechanism distinct from multistep burst arrival: at fixed mean mRNA lifetime, increasing the number of degradation steps accelerates the mean mRNA response and increases the logarithmic sensitivity in Eq. (3), thereby contributing to lower FPT noise. Thus, degradation memory can improve timing precision even when it does not reduce steady-state abundance noise.
The behavior of the optimal threshold further distinguishes multi-step degradation from the multi-step burst arrival process. For multi-step degradation, remains close to or above the classical value of 0.5 over most of the parameter range examined and shifts toward higher values as increases (see Appendix Fig. 5 (g, j)). At the smallest burst sizes and largest values of , the optimal threshold remains close to 0.5. Thus, unlike multistep burst arrival, multistep degradation does not produce a systematic shift of toward lower thresholds. Degradation memory therefore primarily improves temporal precision while preserving a relatively high optimal threshold, whereas burst-arrival memory can modify both FPT precision and the location of the optimal threshold.
Combined Effects of Multistep Transcription and Degradation
We next consider multistep burst arrival and mRNA degradation within a single model Figure 3(a). Burst production requires completion of sequential transcriptional stages, each with rate , whereas each transcript independently traverses degradation stages, each with rate , before removal. Burst sizes are deterministic, , and the mRNA abundance counts transcripts in all degradation stages. To isolate the effects of changing the number of stages, we hold the mean interburst time, , and the mean mRNA lifetime, , fixed.

Figure 3. Combined effects of multistep transcription and mRNA degradation on FPT noise for constant burst. Symbols denote SSA results; curves show analytical predictions or approximations as specified below. (a) Schematic of reaction scheme. Transcription proceeds through sequential stages at rate , producing a deterministic mRNA burst upon completion and resetting the cycle. Each transcript independently traverses degradation stages at rate before removal. (b) Mean mRNA dynamics for , and . Black curves show Eq. (12). The mean dynamics is independent of and approaches the common steady state more rapidly as increases. (c) Transient mRNA noise, , which decreases with increasing . Solid curves show Eq. (13). Inset: steady-state noise in the plane for . Increasing reduces noise for all ranges of . However, increasing raises the steady-state noise for and lowers it for . (d) Mean FPT as a function of the normalized threshold . Solid curves show the threshold-crossing estimates obtained from the exact mean dynamics using Eq. (2). The predicted mean FPT is independent of at fixed , so the corresponding analytical curves overlap. Increasing reduces the mean FPT independent of threshold values. (e) FPT noise, , versus normalized threshold. Curves show the small-noise approximation, Eq. (3), evaluated using the exact transient mRNA mean and noise. FPT noise is nonmonotonic in the threshold, with a minimum whose value and location depend on and . Increasing both and reduces FPT noise. All results use , , and , fixing the mean interburst time at and the mean mRNA lifetime at and mean steady-state mRNA number .
Under these conditions, the exact mean dynamics are
The derivation is given in Appendix C. At fixed mean interburst time, increasing changes the temporal regularity of burst production without changing its mean rate; hence, the mean mRNA trajectory is independent of . Increasing at fixed mean lifetime narrows the lifetime distribution and accelerates the approach to the common steady state. These properties are confirmed by the analytical curves and stochastic simulations in Figure 3(b).
Unlike the mean, the transient noise depends on both and . Writing the mean burst arrival rate , the exact squared coefficient of variation for constant burst can be expressed as
Here, , where is the survival probability of an individual transcript, and is the survival-weighted integral of correlations between distinct burst arrivals. Their explicit expressions and derivation are given in Appendix C. The three terms in the numerator represent the individual-molecule contribution, the contribution from molecules produced in the same burst, and the contribution from correlations between different bursts, respectively. For Poisson burst arrivals (), . Multistep arrival modifies the last term and suppresses the transient fluctuations in the cases shown in Figure 3(c), while leaving the mean trajectory unchanged. The exact transient prediction agrees with the simulations.
The steady-state mean and the large-copy-number approximation to the steady-state noise are
Since increases monotonically with , the effect of multistep degradation depends on the ratio . At fixed , increasing reduces the steady-state noise when , leaves it unchanged within this approximation when , and increases it when . The inset of
Fig. 3(c) illustrates the opposite trends. For , Eq. (15) recovers Eq. (11) for constant burst, .
We then examine the first-passage time to a threshold in total mRNA abundance. The mean-crossing estimate in Eq. (2) depends only on the mean trajectory and is therefore independent of at fixed . Increasing lowers this estimate, consistent with the faster accumulation of mRNA [Figure 3(d)]. The simulations follow the predicted trends, with larger deviations at high thresholds.
The small-noise approximation in Eq. (3) depends on both the transient mRNA noise and the slope of the mean trajectory at the mean-crossing time. Consequently, steady-state abundance noise alone does not determine FPT noise. As shown in Figure 3(e), is nonmonotonic in the threshold, and both its minimum and the threshold at which it occurs depend on and . Increasing and reduces FPT noise. Among the combinations considered, yields the lowest FPT noise. Deviations from the small-noise prediction become more pronounced at high copy-number noise. These results show that temporal precision reflects both production regularity and the transient response shaped by degradation.
Discussion and Conclusions
We have investigated how molecular memory generated by multistep burst arrival and multistep mRNA degradation affects the precision of threshold-crossing times in stochastic gene expression. Previous studies of Markovian models found that FPT noise is minimized at a threshold close to one-half of the steady-state expression level [7, 8] [4]. Our analytical calculations and stochastic simulations show that the optimal threshold is not universal for nonexponential waiting times. Both forms of memory reduce FPT noise, but through distinct mechanisms. At fixed mean burst arrival time, increasing the number of burst-arrival steps regularizes production and suppresses copy-number fluctuations without altering the mean accumulation trajectory. Within the small-noise approximation, this reduction in abundance noise translates into lower FPT noise. Multistep degradation instead alters the transient trajectory and its time sensitivity. When the mean mRNA lifetime is fixed, increasing the number of degradation steps accelerates the mean mRNA response and increases the logarithmic sensitivity, thereby contributing to lower FPT noise. For constant burst, production memory also shifts the optimal normalized threshold below one-half, whereas degradation memory generally leaves the optimum close to or above one-half. Thus, production and degradation memory can reinforce one another in reducing FPT noise while having distinct, and in some regimes opposing, effects on the location of its minimum.
The predicted dependence of timing precision on molecular memory is experimentally testable. Live-cell studies have revealed substantial variation in transcriptional bursting kinetics among mammalian genes, and recent work has emphasized the importance of resolving the relevant timescales underlying transcriptional bursting [29, 32]. For an inducible reporter whose expression activates a tunable downstream threshold, one could simultaneously estimate burst waiting-time distributions, burst-size distributions, and threshold-crossing-time statistics. The present theory predicts that promoters with more regular burst arrival should exhibit reduced FPT noise; moreover, a shift of the optimal threshold below one-half should be most apparent when burst-size fluctuations are small. Phage provides a complementary test bed in which the holin threshold and translational efficiency can be perturbed while lysis times are measured in single cells [8, 13]. Synthetic manipulation of RNA processing or degradation pathways could further test the prediction that degradation memory improves timing precision without reproducing the same threshold shift as production memory.
The analytical treatment relies on a small-noise approximation and on a regime in which burst arrival is fast compared with mRNA degradation. These assumptions allow the FPT noise to be related directly to transient abundance fluctuations and the slope of the mean trajectory. Deviations between the analytical approximation and stochastic simulations are expected when burst sizes become large, when burst-amplitude fluctuations are strong, or when threshold crossing is dominated by discrete overshoot events. Nevertheless, the simulations show that the qualitative conclusions persist beyond the strict approximation regime: multistep arrival regularizes timing, burst-size statistics determine whether the optimal threshold shifts, and multistep degradation provides an additional route to lower timing variability. These passive kinetic-memory effects are complementary to active timing control by feedforward or feedback circuits and to robustness against external disturbances [11, 14, 15, 17, 18]. Extensions incorporating explicit promoter switching, protein-level thresholds, extrinsic noise, cell division and dilution, or correlated production and degradation memory would further clarify how non-Markovian gene-expression dynamics regulate cellular decision times.
In summary, molecular memory is not merely a correction to Markovian models of gene expression; it provides a mechanism for controlling temporal precision. Production memory suppresses fluctuations in stochastic input and can shift the optimal threshold to earlier expression levels when bursts are sufficiently regular, whereas degradation memory improves timing through modified transcript persistence and transient accumulation. These findings establish that the precision and optimal design of intracellular timing are determined not only by mean expression levels, but also by the temporal organization and amplitude statistics of molecular events.
References
Transient Mean and Variance for Multistep Burst Arrival
We derive the transient mean and variance of the mRNA abundance for the multistep burst-arrival model considered in Sec. III. The transcriptional cycle consists of sequential states,
where completion of the final transition produces a burst of mRNAs. Individual mRNAs undergo first-order degradation,
We denote the burst-size distribution by
with moments
The chemical master equation for state is
and, for the remaining transcriptional states (),
where
denote the joint probability that there are mRNAs and that the transcriptional cycle is in state at time . The probability of occupying transcriptional state is

Figure 4. FPT noise with and for constant burst and multi-step transcription. FPT noise with normalized threshold () for different mean burst size, when burst is constant and number of transcription steps, (a) , (b) , and (c) . In the large noise limit, our approximated SNA calculation (solid lines Eq. (3) does not hold with SSA simulation (symbols) [39]. (d) Steady-state Fano factor of mRNA level, (e) minima value of FPT noise, and (f) minima value of normalized threshold in the plane, obtained from SSA simulation [39] over realizations. (g)-(i) Geometric burst distribution: the same analysis is presented in the last row for the geometric burst distribution, where , unlike the constant burst case, where can fall below 0.5. Parameter values: , , , and are tuned to keep and fixed for different and .
and the moments of the total mRNA abundance are
The state probabilities satisfy
Let denote the occupancy of transcriptional state , such that
and we obtain
by assuming burst arrivals are faster than the mRNA lifetime.
Exact mean dynamics of mRNA
Let denote the mean total mRNA abundance. A burst-production event changes by , whereas a degradation event changes by . The exact mean equation is therefore
For an arbitrary initial transcriptional phase and , Eq. (29) gives
Thus, the transient mean generally depends on the initialization of the transcriptional cycle. For the stationary phase initialization in Eq. (28), the mean equation reduces to
Starting from an empty mRNA population, , its exact solution is
where the steady-state mean mRNA number in the limit is
Exact second-moment and variance equations
The exact differential equation for the second moment is
Defining
and using Eq. (29), we obtain
Steady-state covariance and abundance noise
To obtain variance dynamics from Eq. (36) we need to know the mixed moment between the mRNA abundance and the burst-producing transcriptional state. We define
These mixed moments obey the closed linear system
For an initially empty mRNA population,
For the stationary phase distribution, . Setting the time derivatives in Eqs. (38) and (39) to zero gives the steady-state solution
The mixed moment involving the burst-producing state is
Since
the steady-state covariance is
For , this covariance vanishes. For , it is negative because
This negative correlation between the mRNA abundance and the burst-producing state contributes to the suppression of mRNA fluctuations by multistep burst arrival. Using Eq. (44) in the steady-state limit of Eq. (36) gives
Fast-arrival approximation for the transient variance
A closed approximation for the variance dynamics can be obtained when all nonzero relaxation modes of the burst-arrival process are fast compared with mRNA degradation. After the short arrival-process transient, we approximate the combined production and covariance contribution in Eq. (36) by its steady-state value: The approximate transient variance equation is therefore
where
Using the mean dynamics in Eq. (32) and the initial condition , the integrating-factor solution of Eq. (47) is
Finally, dividing Eq. (49) by
gives
Multistep mRNA Degradation
We derive the steady-state mean mRNA abundance for the multistep-degradation model considered in Sec. IV. Bursts arrive according to a Poisson process with rate . Each burst produces a random number of mRNAs, with mean , and all newly produced mRNAs enter the first degradation state, . Each transcript then passes through sequential degradation states,
The total mRNA abundance is defined as
Exact Mean Dynamics
Let us denote the mean mRNA as
Because the reaction network is linear, the equations governing the mean copy numbers close exactly:
We consider the following initial condition
Denoting the Laplace transform of by , Eq. (55) gives
For the subsequent degradation states, Eq. (56) gives the recursion
Inverting Eq. (59), we obtain
Summing Eq. (60) over all degradation states gives the exact mean dynamics of the total mRNA abundance:
Taking the long-time limit yields
The normalized transient mean can therefore be written as
In our simulations, the steady-state mean is held fixed by choosing
Equations (B10) and (B12) are exact and do not require a small-noise or timescale-separation approximation. Moreover, the burst-size distribution affects the mean dynamics only through .
Exact Variance Dynamics
We next derive the exact variance dynamics of the total mRNA abundance. The lifetime of an individual mRNA is Erlang distributed with shape parameter and rate . Its survival probability at age is
Consider a burst that occurred at time before observation. Conditional on the burst size , the number of mRNAs from this burst that survive until the observation time, denoted by , follows a binomial distribution:
Its first two conditional moments are
Averaging over the burst-size distribution gives

Figure 5. FPT noise for multi-step mRNA degradation and bursty single-step production. (a) The schematic of the gene network, where transcription is a single-step process with bursty production of size and frequency , and degradation is an -step process with equal rate . FPT noise with normalized threshold for different mean burst size, when burst is constant and number of degradation steps, (a) , (b) , and (c) . (e) Steady-state Fano factor of mRNA level, (f) minima value of FPT noise, and (g) minima value of normalized threshold in the plane, obtained from SSA simulation [39] over realizations. (e-g) We consider constant burst, i.e. , (h-j) and geometric burst, i.e, . Parameter values: to keep the mRNA lifetime fixed at 10 and to keep mean steady-state mRNA number fixed at 50 for varying and .
Since the bursts arrive according to a Poisson process with rate , the variance of the total mRNA abundance is
Using Eq. (70), this becomes
We define
Substituting Eq. (65) into Eq. (73) gives
Equations (B21) and (B23) provide the exact variance dynamics for an arbitrary burst-size distribution with finite first and second moments.
For a constant burst of size ,
and hence
In particular, for unit bursts, ,
at all times, independently of the number of degradation steps.
Taking the long-time limit of Eq. (74) gives
Using
the steady-state variance becomes
Therefore, the steady-state Fano factor is
For constant burst of size , this reduces to
Finally, the exact transient Fano factor is
These exact abundance moments are used to validate the stochastic simulations.
mRNA Mean and Noise Dynamics with Multistep Burst Arrival and Degradation
We derive the transient and steady-state moments of the total mRNA abundance when both burst arrival and degradation proceed through multiple stages. Throughout, bursts have a fixed positive integer size , so that . Individual transcripts degrade independently, and denotes the total number of transcripts across all degradation stages.
Model and mean dynamics
Each burst is generated upon completion of sequential transcriptional stages, each with rate . The interburst interval therefore has the Erlang density
Thus, the stationary mean burst-arrival rate is . Here, denotes an interburst interval, whereas denotes its mean.
Each transcript passes through degradation stages, each with rate , before removal. Its survival probability at age is
For the transient calculation, we take and sample the initial transcriptional phase uniformly over the stages. The arrival process is therefore stationary from , although the mRNA population is initially empty. In particular, the first waiting time is the stationary residual waiting time, rather than a complete Erlang interburst interval when .
A burst of age contributes surviving transcripts on average. Integrating over burst ages gives mean mRNA number
Taking yields the steady-state mean
Exact transient noise
Correlations between distinct burst arrivals are described by the renewal density
where denotes the -fold convolution of . Conditional on a burst at time zero, is the expected number of bursts in an interval of width at lag . Its Laplace transform is
For a stationary renewal process, the density of ordered pairs of distinct bursts separated by a positive lag is . The excess over the uncorrelated value is therefore .
The number of surviving transcripts from a burst of age is binomial with parameters and . Its second factorial moment is . Separating pairs of transcripts produced in the same burst from those produced in distinct bursts gives
Using , and defining
we obtain the exact transient variance
The corresponding squared coefficient of variation is
The restriction is required because . The term involving accounts for transcripts produced in the same burst, whereas the term involving accounts for correlations between distinct bursts.
For Poisson burst arrivals (), and hence . Equation (C11) then reduces to
Steady-state limit
To evaluate the long-time limit, define the survival overlap
Changing variables in Eq. (92) gives
Thus, the exact steady-state variance is
For Erlang lifetimes,
Dividing Eq. (98) by Eq. (87) yields the exact Fano factor
The steady-state squared coefficient of variation is therefore
Finite-sum representation and limiting cases
The correlation integral in Eq. (100) admits a finite-sum representation. Define
Expanding the survival functions in Eq. (96) gives
with coefficients
Since
the exact Fano factor becomes
For single-step burst arrival (), , and Eq. (106) reduces to
Thus, for , whereas it increases with for . For single-step degradation (), , giving
Large-copy-number approximation
At fixed , , and , the large-copy-number limit corresponds to , because
The approximation below concerns this limit; increasing the burst size alone does not establish . Expanding Eq. (102) about gives
At leading order, only the term in Eq. (106) contributes. Moreover, Eqs. (103) and (99) imply . Substitution therefore yields
or, equivalently,
The remainder in Eq. (111) is stated for fixed , , and ; it is not a uniform error bound when these parameters also increase.
Since , the leading-order change in the Fano factor at fixed and is
provided both parameter sets lie within the approximation regime. Accordingly, the Fano factor decreases with when , is independent of to leading order when , and increases with when . At fixed , the same trends apply to .
For and , the approximation predicts that multistep degradation enhances the noise suppression associated with regular burst arrival. In contrast, for and , the interburst correlation term vanishes, and the contribution from transcripts produced in the same burst causes the Fano factor to increase with , as follows exactly from Eq. (107). These opposite trends reflect the competition between arrival regularity and the shared production times of transcripts within each burst.
References
- [5]L. Ham, M. A. Coomer, K. Öcal, R. Grima, and M. P. H. Stumpf, Nature Communications 15, 5286 (2024).
- [6]J. Paulsson, Physics of Life Reviews 2, 157 (2005).
- [7]A. Dal Co, M. Cosentino Lagomarsino, M. Caselle, and M. Osella, Nucleic Acids Research 45, 1069 (2017).
- [8]S. Kannoly, T. Gao, S. Dey, I.-N. Wang, A. Singh, and J. J. Dennehy, iScience 23, 101186 (2020).
- [9]K. R. Ghusinga, C. A. Vargas-Garcia, and A. Singh, Scientific Reports 6, 30229 (2016).
- [10]K. R. Ghusinga and A. Singh, in 2019 Sixth Indian Control Conference (ICC) (IEEE, 2019) pp. 280–285.
- [11]K. R. Ghusinga and A. Singh, in 2020 59th IEEE Conference on Decision and Control (CDC) (IEEE, 2020) pp. 5867–5872.
- [12]K. Rijal, A. Prasad, A. Singh, and D. Das, Physical Review Letters 128, 048101 (2022).
- [13]S. Kannoly, K. Singh, N. Juman, Z. Marufee Islam, I. Caballero Quiroga, A. Singh, and J. J. Dennehy, bioRxiv , 2025.05.14.654169 (2025), preprint, version 1, posted May 17, 2025.
- [14]S. Dey, S. Kannoly, P. Bokes, J. J. Dennehy, and A. Singh, in 2021 European Control Conference (ECC) (IEEE, 2021) pp. 2127–2132.
- [15]Z. Vahdat, K. R. Ghusinga, and A. Singh, in 2021 29th Mediterranean Conference on Control and Automation (MED) (IEEE, 2021) pp. 450–455.
- [16]K. Biswas and A. Ghosh, Physical Review E 101, 022418 (2020).
- [17]S. Rezaee, C. Nieto, and A. Singh, in 2023 27th International Conference on System Theory, Control and Computing (ICSTCC) (IEEE, 2023) pp. 505–510.
- [18]S. Rezaee, C. Nieto, and A. Singh, in 2024 IEEE 63rd Conference on Decision and Control (CDC) (IEEE, 2024) pp. 7772–7777.
- [19]K. Biswas, S. Dey, and A. Singh, Scientific Reports 14, 27199 (2024).
- [20]J. M. Pedraza and J. Paulsson, Science 319, 339 (2008).
- [21]C. Deneke, R. Lipowsky, and A. Valleriani, PLOS ONE 8, e55442 (2013).
- [22]D. E. Weidemann, J. Holehouse, A. Singh, R. Grima, and S. Hauf, Science Advances 9, eadh5138 (2023).
- [23]J. Zhang and T. Zhou, Proceedings of the National Academy of Sciences 116, 23542 (2019).
- [24]J. Szavits-Nossan and R. Grima, Biophysical Journal 123, 1034 (2024).
- [25]D. L. Bentley, Nature Reviews Genetics 15, 163 (2014).
- [26]N. L. Garneau, J. Wilusz, and C. J. Wilusz, Nature Reviews Molecular Cell Biology 8, 113 (2007).
- [27]M. Ma, J. Szavits-Nossan, A. Singh, and R. Grima, Mathematical Biosciences 373, 109204 (2024).
- [28]T. Jia and R. V. Kulkarni, Physical review letters 106, 058102 (2011).
- [29]D. M. Suter, N. Molina, D. Gaffield, K. Schneider, U. Schibler, and F. Naef, Science 332, 472 (2011).
- [30]K. B. Halpern, S. Tanami, S. Landen, M. Chapal, L. Szlak, A. Hutzler, A. Nizhberg, and S. Itzkovitz, Molecular Cell 58, 147 (2015).
- [31]V. Agarwal and D. R. Kelley, Genome Biology 23, 245 (2022).
- [32]J. V. W. Meeussen and T. L. Lenstra, Trends in Genetics 40, 160 (2024).
- [33]A. Dal Co, M. Cosentino Lagomarsino, M. Caselle, and M. Osella, Nucleic Acids Research 45, 1069 (2017).
- [34]H. Chen, L. Zhou, J. Teng, A. Chen, and T. Zhou, Physical Review E 114, 034122 (2026).
- [35]L. Liu, B. Kashyap, and J. Templeton, Journal of Applied Probability 27, 671 (1990).
- [36]N. Kumar, A. Singh, and R. V. Kulkarni, PLoS computational biology 11, e1004292 (2015).
- [37]Z. Vahdat and A. Singh, in 2024 IEEE 63rd Conference on Decision and Control (CDC) (IEEE, 2024) pp. 7792–7797.
- [38]K. R. Ghusinga, J. J. Dennehy, and A. Singh, Proceedings of the National Academy of Sciences 114, 693 (2017).
- [39]D. T. Gillespie, The Journal of Physical Chemistry 81, 2340 (1977).
- [40]S. J. Park, S. Song, G.-S. Yang, P. M. Kim, S. Yoon, J.-H. Kim, and J. Sung, Nature Communications 9, 297 (2018).