Across all domains of life, cell proliferation requires that the chromosome replication and cell division cycles are coordinated to ensure that every new cell receives one copy of the genetic material. While in eukaryotes this coordination is implemented by a dedicated regulatory system in which genome replication and division occur in well-separated stages, no such system has been found in most bacteria. This suggests that the molecular events that control replication initiation and division might be coordinated more directly in bacteria, through molecular interactions that are yet to be elucidated. The contrast between this efficient coordination and the apparent absence of a dedicated regulatory system is particularly remarkable since most bacteria feature a unique replication origin which imposes that multiple rounds of replication occur concurrently in fast growth conditions. For example, in the specific case of E. coli that we study here, it has long been known that growth rate, cell size, and replication initiation are coordinated such that the average number of replication origins per unit of cellular volume is approximately constant across conditions (Donachie, 1968) or that cellular volume grows approximately exponentially with growth rate (Taheri-Araghi et al., 2017). Although several models have been proposed over the last decades to explain such observations (for a review and a historical perspective see for example Willis and Huang, 2017), so far direct validation of these models has been lacking, due to a large extent to the lack of quantitative measurements of cell cycles parameters in large samples with single-cell resolution.

Thanks to techniques such as microfluidics in combination with time-lapse microscopy, it has recently become possible to perform long-term observation of growth and division in single bacteria. By systematically quantifying how cell cycle variables such as size at birth, size at division, division time, and growth rate vary across single cells, insights can be gained about the mechanism of cell cycle control. Several recent studies have focused on understanding the regulation of cell size, resulting in the discovery that E. coli cells maintain a constant average size by following an adder strategy: instead of attempting to reach a certain size at division (i.e. a sizer mechanism) or to grow for a given time (i.e. a timer mechanism), it was found that cells add a constant length $dL$ to their birth length $L_b$ before dividing (Amir, 2014; Campos et al., 2014; Taheri-Araghi et al., 2017). In particular, while the cell size at division and the division time correlate with other variables such the cell size at birth and growth rate, the added length $dL$ fluctuates independently of birth size and growth rate. A remarkable feature of the adder model is its capacity to efficiently dampen large cell size fluctuations caused by the intrinsically noisy regulation, without the need for any fail-safe mechanism. This efficient strategy has been shown to be shared by various bacterial species as well as by archea (Eun et al., 2018) and even some eukaryotes such as budding yeast (Soifer et al., 2016).

Here, we focus on how the control of replication initiation is coordinated with cell size control in E. coli, a question that has attracted attention for a long time (Helmstetter, 1974; Pierucci, 1978; Koppes and Nanninga, 1980). Several models have been proposed to explain how the adder behavior at the level of cell size might arise from a coordinated control of replication and division. Broadly, most models assume that the accumulation of a molecular trigger, usually assumed to be DnaA, leads to replication initiation, which in turn controls the corresponding future division event (Campos et al., 2014; Ho and Amir, 2015; Wallden et al., 2016). Subtle variations in how the initiation trigger accumulates and how the initiation to division period is set in each model imply distinct molecular mechanisms, and thus fundamentally different cell cycle regulations. Specifically, most models assume that initiation is triggered either when a cell reaches a critical absolute volume (initiation size, see for example Wallden et al., 2016) or alternatively when it has accumulated a critical volume since the last initiation event (see e.g. Ho and Amir, 2015). In order to explain the coordination between cell cycle events, division is often assumed to be set by a timer starting at replication initiation, but recent studies have also proposed that the two cycles might be independently regulated (Micali et al., 2018a; Si et al., 2019). Finally, it is often assumed that the regulation strategy could be different at slow and fast growth where different constraints occur.

We use an integrated microfluidics and time-lapse microscopy approach to quantitatively characterize growth, division, and replication in parallel across many lineages of single E. coli cells, both in slow and fast growth conditions. We show that insights about the underlying control mechanisms can be gained by systematically studying the structure of correlations between these different variables. Our single-cell observations are inconsistent with several previously proposed models including models that assume replication is initiated at a critical absolute cell volume and models that assume division is set by a timer that starts at replication initiation. Instead, the most parsimonious model consistent with our data is a double-adder model in which the cell cycle commences at initiation of replication and both the subsequent division and the next initiation of replication are controlled by the added volume. We show that this model is most consistent with the correlation structure of the fluctuations in the data and, through simulations, we show that this model accurately reproduces several non-trivial observables including the previously observed adder behavior for cell size control, the distribution of cell sizes at birth, and the distribution of the number of origins per cell at birth. Moreover, the same model best describes the data both at slow and fast growth rates. As far as we are aware, no other proposed model at the same level of parsimony can account for the full set of observations we present here.

To test possible models for the coordination of replication and division in E. coli we decided to systematically quantify growth, replication initiation, and division across thousands of single E. coli cell cycles, across multiple generations, and in various growth conditions. To achieve this, cells were grown in a Mother Machine type microfluidic device (Wang et al., 2010) and imaged by time-lapse microscopy. We used M9 minimal media supplemented with glycerol, glucose or glucose and eight amino acids, resulting in doubling times of 89, 53 and 41 min, respectively. The cell growth and division cycles were monitored by measuring single-cell growth curves obtained through segmentation and tracking of cells in phase contrast images using the MoMA software (Kaiser et al., 2018). The replication cycle was monitored by detecting initiation as the duplication of an oriC proximal FROS tagged locus imaged by fluorescence microscopy (Figure 1A). This of course only offers a proxy for initiation timing, as origin splitting can be affected by a cohesion time (Reyes-Lamothe et al., 2008). These measurements allowed us to quantify each single cell cycle by a number of variables such as the growth rate, the sizes at birth, replication initiation, and division, the times between birth and replication initiation and the time between birth and division. As done previously, we assume that cell radius is constant and use cell length as a proxy for cell volume (Adiciptaningrum et al., 2016; Taheri-Araghi et al., 2017). Since we can follow cells over multiple generations, we can also measure quantities that span multiple division cycles such as the total time or total cell growth between consecutive replication initiation events. As we analyze growth conditions spanning cases with both single and overlapping rounds of replication, we defined two alternative ways of defining variables. In particular, while the cell cycle is classically defined from division to division (Figure 1C), we also use an alternative framework proposed recently (Ho and Amir, 2015; Amir, 2017), in which the cell cycle is defined from one replication initiation to the next (Figure 1D). As this framework is centered on origins of replication rather than on cells, we consequently define a new quantity $\mathrm{\Lambda }$, the cell length per origin, which allows tracking of the amount of cell growth per origin of replication. For instance, in a case where a cell is born with an ongoing round of replication which started at time $t$, ${\mathrm{\Lambda }}_i$ for that cell is defined as ${\mathrm{\Lambda }}_i=L_i/4$ where $L_i$ is the length of the mother cell which contains four origins at time $t$ (Figure 1—figure supplement 1). Using these definitions allows us to avoid having to apply artificial cut-offs as for example done in Wallden et al. (2016). In this article, we explore a wide variety of models that take either a division-centric or replication-centric view of the cell cycle. Using specific correlations observed in our data, we show how entire classes of models can be rejected. In addition, we present a general statistical framework for ranking models based on their ability to capture the full correlation structure of the data.

Figure 1 with 1 supplement see all

Download asset Open asset

Figure 1—source data 1

Table with source data for Figure 1B.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig1-data1-v2.csv

Download elife-48063-fig1-data1-v2.csv

Replication initiation mass

A popular idea dating back to the 1960s and still often used today to explain the coupling of division and replication cycles is the initiation mass model. The observations that cell volume grows exponentially with growth rate (Schaechter et al., 1958) and that, across a range of conditions, the time between replication initiation and division is roughly constant (Helmstetter et al., 1968) led Donachie to propose that the volume per origin of replication is held constant (Donachie, 1968). In particular, the model proposes that initiation occurs when a cell reaches a critical volume. A simple prediction of this model is that, for a given cell, the cell length $L_i$ at which initiation occurs should be independent of other cell cycle variables such as the length at birth $L_b$. However, as can be seen in Figure 2A, we observe that the initiation length $L_i$ and birth length $L_b$ are clearly correlated in all conditions, rejecting the initiation mass model. The absence of an initiation mass has already been shown by Adiciptaningrum et al. (2016) and confirmed in a recent comprehensive analysis (Micali et al., 2018a). Note that, by assuming that the concentration of the molecule that controls initiation is the same independent of growth rate, and a critical amount of this molecule has to accumulate per origin to trigger initiation, then this model would still predict that the average volume per origin in bulk is independent of growth rate (Si et al., 2017).

Figure 2

Download asset Open asset

Figure 2—source data 1

Table with source data for Figure 2.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig2-data1-v2.csv

Download elife-48063-fig2-data1-v2.csv

Connecting replication and division cycles

Having validated the multiple origins accumulation model for replication control, we now investigate its relation to the division cycle. A common assumption is that the period $T_{id}$ from initiation to division (classically split into the replication period C and the end of replication to division period D) is constant and independent of growth rate (Cooper and Helmstetter, 1968; Ho and Amir, 2015). Such a constant period might suggest a timer mechanism by which, on average, a fixed time elapses in each cell between initiation and division. However, as visible in Figure 3A, while on average $T_{id}$ is indeed rather constant across growth conditions, within each condition fast growing cells clearly complete this period faster than slow growing cells. One possibility is to assume that the time $T_{id}$ is coupled to growth rate $\lambda$ by an unspecified mechanism in such a way as to recover the empirically observed correlation (Wallden et al., 2016). However, Figure 3B reveals another and arguably simpler solution. We find that $d{\mathrm{\Lambda }}_{id}={\mathrm{\Lambda }}_d-{\mathrm{\Lambda }}_i$, the length per origin added by a cell between initiation and division, has an adder behavior as well: independently of its size at initiation $L_i$, a cell will complete the corresponding division cycle after having accumulated a constant volume per origin $d{\mathrm{\Lambda }}_{id}$.

Figure 3

Download asset Open asset

Figure 3—source data 1

Table with source data for Figure 3.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig3-data1-v2.csv

Download elife-48063-fig3-data1-v2.csv

The double-adder model

These observations motivated us to formulate a model in which the cell cycle does not run from one division to the next, but rather starts at initiation of replication, and in which both the next initiation of replication and the intervening division event, are controlled by two distinct adder mechanisms. In this replication-centric view, the cell cycles are controlled in a given condition by three variables: an average growth rate $\lambda$, an average added length per origin $d{\mathrm{\Lambda }}_{if}$, and an average added length $d{\mathrm{\Lambda }}_{id}$ between replication initiation and division. In particular, we assume that these three variables fluctuate independently around these averages for each individual cell cycle, and that all other parameters such as the sizes at birth, initiation, and the times between birth and division or between initiation and division, are all a function of these three fundamental variables. This double-adder model is sketched in Figure 4 for the case of slow growth conditions: a cell growing at rate $\lambda$ initiates replication at length $L_i$ and thereby starts two adder processes. The cell then divides when reaching a size $L_d=L_i+nd{\mathrm{\Lambda }}_{id}=n({\mathrm{\Lambda }}_i+d{\mathrm{\Lambda }}_{id})$, where $n=2$ is the number of replication origins. Second, the next replication round will be initiated when the total length has increased by $d{\mathrm{\Lambda }}_{if}$ per origin.

Figure 4 with 1 supplement see all

Download asset Open asset

Simulations of the double-adder model

To assess to what extent our double-adder model can recover our quantitative observations, we resorted to numerical simulations. We first obtained from experimental data the empirical distributions of growth rates $\lambda$, the added length per initiation $d{\mathrm{\Lambda }}_{if}$, and the added length between initiation and division $d{\mathrm{\Lambda }}_{id}$. A series of cells are initialized at the initiation of replication, with sizes taken from the experimental distributions. For each cell, a growth rate $\lambda$ is independently drawn from its empirical distribution, and values of $d{\mathrm{\Lambda }}_{id}$ and $d{\mathrm{\Lambda }}_{if}$ are drawn from independent distributions, to set the times of the next division and replication initiation events. This procedure is then iterated indefinitely, that is a new growth rate and values of each adder are independently drawn for each subsequent cycle. As has been observed previously (Campos et al., 2014) the growth rate is correlated ($r\approx 0.3$) between mother and daughter. Accounting for this mother-daughter correlation in growth rate was found not to be critical for capturing features of E. coli cell cycle, but was included in the model to reproduce simulation conditions of previous studies.

As can be seen in Figure 5, the double-adder model accurately reproduces measured distributions and correlations at all growth rates. In particular, the global adder behavior for cell size regulation naturally emerges from it (Figure 5A). Note however that in particular at fast growth (Glucose +AA), the relation between $dL$ and $L_b$ deviates from pure adder behavior, that is $dL$ weakly anti-correlates with $L_b$ (Figure 1B). Interestingly, this deviation is recovered as well in the simulations (Figure 5A), supporting the validity of our model. The specific relation between length at initiation $L_i$ and length at birth $L_b$, which prompted us to reject the initiation mass model, is too reproduced by the model (Figure 5B). Furthermore, the distribution of the number of origins at birth, which reflects the presence of overlapping replication cycles is also reproduced (Figure 5D). Finally, we note that while the average cell length parameters (e.g. for $L_b$ in Figure 5C) are recovered in the simulations, the variances of the simulated distributions are a bit larger in the faster growth conditions, which can be attributed to an overestimation of the variances of the two adder distributions. An exhaustive comparison between experiments and simulations can be found in Figure 5—figure supplement 1 through Figure 5—figure supplement 4.

Figure 5 with 4 supplements see all

Download asset Open asset

Figure 5—source data 1

Table with source data for Figure 5AB.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig5-data1-v2.csv

Download elife-48063-fig5-data1-v2.csv

Figure 5—source data 2

Table with source data for Figure 5C.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig5-data2-v2.csv

Download elife-48063-fig5-data2-v2.csv

Figure 5—source data 3

Table with source data for Figure 5D.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig5-data3-v2.csv

Download elife-48063-fig5-data3-v2.csv

As our model recovers the classic adder from birth to division, that is independence of $L_b$ and $dL$, one could ask whether, instead of assuming division to be controlled by an adder between initiation and birth, one could simply have a classic birth-to-division adder running independently and in parallel with a replication adder. Harris and Theriot (2016) proposed that a homeostatic mechanism maintaining a constant cell surface to volume ratio could produce a birth-to-division adder, but only if division were controlled completely independently of replication. A recent study that emerged during the course of this work (Si et al., 2019) also argued that the division and replication cycles are uncoupled and controlled independently. To test these ideas, we implemented a model in which division and replication are controlled by completely independent adder mechanisms. Although this model recapitulates most experimental observations, it predicts a negative correlation between volume at initiation $L_i$ and the added volume $d{\mathrm{\Lambda }}_{ib}$ between initiation and birth (Appendix 2), which is at odds with the adder behavior we observe in the data (Figure 3). It should also be noted that purely geometric division models such as the Harris and Theriot (2016) model have been recently challenged more broadly by research indicating that surface expansion is coupled to dry-mass increase rather than to volume increase (Oldewurtel et al., 2019).

Recently, Micali et al. (2018b) introduced a wider class of models in which the replication and division cycles are controlled by two concurrently running processes, which are coupled through a check-point that demands a certain minimal amount of time has to elapse between replication initiation and division. This class includes models in which independent replication and division adders are coupled through such a check-point and, if parameters are appropriately tuned, it is possible for such a model to exhibit no correlation between $L_i$ and $d{\mathrm{\Lambda }}_{ib}$, that is reproduce the observed adder behavior for the period between replication initiation and division (Micali and Cosentino Lagomarsino, personal communication). Although such a model is thus potentially consistent with our observations, it requires the parameters to be precisely tuned and, moreover, this tuning has to apply to all growth conditions.

The double-adder model best captures the correlation structure of the data

In this work, we used empirically observed correlations between certain variables, such as those shown in Figure 2 and Figure 3, to motivate our double-adder model, and then used numerical simulations of this model to show that this model also successfully reproduces other features of the experimental data (Figure 5). However, it is conceivable that other models would be equally successful in reproducing the experimental observations. We thus aimed to devise a general approach for systematically comparing the performance of a large class of models.

The basic idea behind this approach is that, if the cell cycle is controlled by independent control processes, then we expect the variables associated with these control processes to fluctuate independently. To illustrate this, let’s consider for the moment the simpler case of the division cycle and cell size control (Figure 6A). If cell size homeostasis were controlled by a ‘sizer’ mechanism, we would expect the size at division $L_d$ to fluctuate independently of size at birth $L_b$ and growth rate $\lambda$. If it were instead controlled by a ‘timer’, then we would expect the time $T_{bd}$ between birth and division to fluctuate independently from size at birth $L_b$ and growth rate $\lambda$. However, the measurements show that the time $T_{bd}$ is clearly negatively correlated with both growth rate and size at birth. Instead, the added volume $dL$ fluctuates almost independently of size at birth $L_b$ and growth rate $\lambda$ (Figure 6A). It is precisely this independence of fluctuations in added volume from fluctuations in the other variables, that constitutes the support of the adder mechanism. Here, we generalize this idea by systematically considering all sets of variables that can be used to describe the cell cycle in single cells, and quantifying, for each set of variables, the extent to which they fluctuate independently.

Figure 6 with 1 supplement see all

Download asset Open asset

Figure 6—source data 1

Table with source data for Figure 6.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig6-data1-v2.csv

Download elife-48063-fig6-data1-v2.csv

A single division cycle needs 3 variables to be fully described and we will refer to such sets of variables as decompositions. For example, the set $(L_b,L_d,\lambda )$ corresponds to a ‘sizer’ decomposition, the set $(L_b,T_{bd},\lambda )$ to a ‘timer’ decomposition, and $(L_b,dL,\lambda )$ to an adder decomposition. For each decomposition, we can now calculate, from the data, the matrix $R$ of observed correlations $R_{ij}$ between each pair of variables $(i,j)$ in the decomposition. For example, Figure 6A shows the correlation matrices $R$ for the adder and timer decompositions. The more independent the variables are in the decomposition, the smaller the correlations will be. As a measure for the overall independence of the variables in a decomposition we use the determinant of the matrix $R$, that is we define $I=\det (R)$. The intuitive meaning of this measure is illustrated in Figure 6B. We can think of each cell cycle as a point in the three-dimensional space spanned by the decomposition, such that all cell cycle observations form a scatter of points in this space. If we normalize each axis by setting its scale such that the variance equals 1 along the axis, then the independence $I=\det (R)$ corresponds exactly to the volume covered by the scatter of points. Note that, if there are no correlations, the matrix $R$ will become the identity matrix, that is $R_{ii}=1$ and $R_{ij}=0$ when $i\ne j$, and the independence will become exactly 1. The larger the correlations between the variables, the smaller the volume $I$ will become, and it will become zero in the limit of perfect correlation between 2 or more variables. Thus, the independence per definition ranges from 0 to 1 (perfect independence).

We can now apply this systematic approach to the complete cell cycle case. In this case, we measured the variables shown in Table 1 (measured variables) either for the replication- or the division-centric view of the cell cycle. As for the division cycle case analyzed in Figure 6, we can use equations relating these variables in order to derive a set of additional variables as shown in Table 1 (derived variables). We need four variables to define both replication and division cycles, and using now all the measured and derived variables, as well as the equations connecting them, we can create all possible decompositions of four variables sufficient to describe the cell cycle, and measure their independence using the decomposition method. Finally, we can rank all decompositions according to their independence to find which ones offer the most accurate description of the data. Such a statistical analysis is only relevant when applied to a large dataset and we therefore focus here on the slow growth condition (M9 glycerol) for which we implemented automatic origin tracking.

The tables in Figure 7A show the five best models ranked by decreasing independence (all decompositions can be found in Figure 7—figure supplement 3). Note that these variable sets include previously proposed sizer and timer models as special cases, for example the inter-initiation model combined with an initiation to division timer (Ho and Amir, 2015) is highlighted in red in Figure 7—figure supplement 3. The most successful decompositions are shown in greater detail as correlation matrices Figure 7B. We find in general that none of the division-centric models accomplishes high independence. For example, as shown in the covariance matrix of Figure 7B right, the best division-centric model is plagued by high correlation between $L_b$ and $d{L}_{bi}$. This strongly suggests that the cell cycle control is better described from a replication-centric point of view. Of all replication-centric models, our double-adder model clearly reaches the highest independence, followed by various derivative models in which one of the adders is replaced by another variable. Notably, the top three decompositions are the same for the real data and for the data from the simulations of the double-adder model, underscoring that these near optimal variants are expected for data produced by a double-adder mechanism. We note that independence of our double-adder model on the real data is a bit lower than on simulated data Figure 7B, that is 0.88 versus 0.98. This residual dependence might either result from correlated errors in the measurements, or it might reflect some small biological dependence not captured by our model. As an additional control, we also applied our decomposition analysis to a dataset from a simulation with the multiple origins accumulation model in which there is an inter-initiation added volume $d{\mathrm{\Lambda }}_{if}$ and a timer from initiation to division $T_{ib}$. As shown in Figure 7—figure supplement 4, the decomposition analysis successfully identifies the model used in the simulation as having the highest Independence $I$. Finally, to verify that the results obtained by analyzing the fluctuations of the division and replication cycles are not primarily driven by differences in experimental details (such as strains or reporters used to measure replication initiation), we also applied our decomposition method to a dataset from a study that appeared in the course of this work (Si et al., 2019). As shown in Figure 7—figure supplement 5, here again the double-adder model is the most successful at explaining the experimental measurements. In summary, this systematic analysis shows that, within a large class of alternative models, the double-adder model best captures the correlation structure of both our data and data recently obtained by another laboratory.

Figure 7 with 5 supplements see all

Download asset Open asset

Figure 7—source data 1

Table with source data for replication-centric decompositions of both experimental and simulation data of Figure 7 and Figure 7—figure supplement 1, Figure 7—figure supplement 2 and Figure 7—figure supplement 3.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig7-data1-v2.csv

Download elife-48063-fig7-data1-v2.csv

Figure 7—source data 2

Table with source data for division-centric decompositions of both experimental and simulation data of Figure 7 and Figure 7—figure supplement 1, Figure 7—figure supplement 2 and Figure 7—figure supplement 3.

https://cdn.elifesciences.org/articles/48063/elife-48063-fig7-data2-v2.csv

Download elife-48063-fig7-data2-v2.csv

Thanks to experimental techniques like the one used here, models of bacterial cell cycle regulation dating back to the 1960s have been recently re-examined in detail in several studies (Campos et al., 2014; Tanouchi et al., 2015; Ho and Amir, 2015; Adiciptaningrum et al., 2016; Wallden et al., 2016; Si et al., 2017; Logsdon et al., 2017; Micali et al., 2018a; Eun et al., 2018; Si et al., 2019). Although these data have shed important new light on the regulation of bacterial physiology, such as identifying the adder behavior in cell growth and division, it has remained somewhat unclear to what extent different models are consistent with the data. In particular, many studies only focus on certain correlations and dependencies between measurable cell cycle quantities, and these can often be explained by multiple models. In this study, we have in a first step empirically built a model which is based on previous ideas and which recapitulates measured cell cycle parameters. This model makes replication the central regulator of the cell cycle, with each initiation round triggering subsequent division and replication events through concurrent adder processes. In a second step, we then designed and applied a statistical method to determine, within a class of models built on biologically relevant cell cycle variables, which set of those variables best explains the correlation structure observed in the measurements. Following this fully independent and more systematic approach, our empirical double-adder model clearly comes out as the most successful. It should be noted that, whereas models based on a timer between initiation and division naturally predict the well-known exponential increase of average cell-size with average growth rate measured in the bulk (Schaechter et al., 1958; Taheri-Araghi et al., 2017), the double-adder does not require such a relationship. However, neither is such a relationship inconsistent with the double-adder model. The difference between a timer and adder model is in the nature of the correlations of different cell cycle quantities across single cells within one condition, and not in the behavior of their averages across conditions. Therefore, it is possible to implement the same dependence of average size on growth rate using different models that have different single-cell correlation structure. In particular, in order for the average bulk cell-size cell size to grow exponentially with the average bulk growth rate, it is enough that the average time between initiation and division is independent of growth rate and this can also be achieved with an adder model. For example, if we imagine that the initiation to division adder is implemented by the accumulation of a key molecule to a critical amount, and we assume that the average rate of accumulation of this molecule is independent of the bulk growth rate, then the average cell-size will grow exponentially with growth rate. At the same time, within such an adder model the single-cell fluctuations in the added volume will be independent of the single-cell fluctuations in growth rate and initiation size, as required by our observations, and in contrast to the predictions of a timer model, which would predict fluctuations in growth rate and added volume to correlate positively.

While the division and replication cycles are seemingly coupled, our analysis demonstrates that two simple adders connected to replication initiation are sufficient to recapitulate both cycles without explicitly enforcing constraints reflecting mechanisms such as over-initiation control by SeqA and nucleoid occlusion which ensures that division only occurs after chromosome replication is completed. The initiation-to-initiation adder mimics SeqA activity by creating a refractory period without initiation, and the initiation-to-division adder ensures that a minimal time is allocated for replication to complete. While the simulations might in rare cases generate unrealistic situations, for example if a large initiation adder is combined with a small division adder leading to premature division, those clashes seem rare enough to not affect the global statistical behavior of the model. Naturally, the model would break down and these controls would need to be explicitly included in the case where cells are subject to stress conditions where these control mechanisms act as fail-safes for example to ensure that division is delayed if DNA repair is needed. In the course of our study, new research proposed that division and replication cycles are only seemingly connected, and used perturbation methods to drive cells to states where the uncoupling is revealed (Si et al., 2019). While such perturbation studies are very informative, more work is needed to understand to what extent hidden compensatory mechanisms might be at play when affecting DnaA or FtsZ expression. Also, that study focuses exclusively on a model which explicitly enforces various correlations between variables unlike our model which naturally produces such relations. Another recent study (Wehrens et al., 2018) using a perturbative approach has shown that cells driven to a filamentous state recover normal sizes through successive single division events seemingly independently from initiation and following an adder behaviour. As division inhibition also hinders replication related processes such as chromosome segregation (Sánchez-Gorostiaga et al., 2016), it is difficult to untangle the different mechanisms that might be at play. Complementing such experiments with a monitoring of replication as done here would surely be a worthwhile future endeavour, a task though which is beyond the purpose of this study, which tries to clarify the normal growth case.

Interestingly, a double-adder mechanism similar to the one that we propose here has been shown to explain cell cycle control in mycobacteria (Logsdon et al., 2017). These mycobacteria have a much more complex cell cycle than E. coli, in particular characterized by a strong asymmetry between daughter cells and a growth rate almost an order of magnitude smaller than that of E. coli. Despite those important differences, it was shown that mycobacterial cell cycles exhibit adder behavior for both division and replication starting at initiation, in a manner highly similar to our observations in E. coli. This suggests that the mechanism connecting replication and division must be quite fundamental and independent of the specifics of available genes and their expression.

Although the single-cell observations provide clear indications of which variables are most likely to be directly involved in the cell cycle control, they of course do not indicate the underlying molecular mechanisms. However, it is not hard to speculate about possible molecular mechanisms that could implement the double-adder behavior. As others have pointed out previously (Ho and Amir, 2015), an adder for the regulation of replication initiation can be easily implemented at the molecular level by having a 'sensor’ protein that builds up at each origin, and that triggers replication initiation whenever a critical mass is reached at a given origin. If this sensor protein is additionally homeostatically controlled such that its production relative to the overall protein production is kept constant, then the average volume per origin will also be kept constant across conditions.

It is more challenging to define a molecular system that can implement the second adder that controls division. The main challenge is that this adder does not run throughout the entire cell cycle, but only between replication initiation and division. It is well known that division is driven by the polymerization of the FtsZ ring, which includes a host of other FtsZ-ring associated proteins, and its progressive constriction. It might seem simplest to assume that the division adder could be implemented directly through FtsZ production, again in the logic of the regulated 'sensor’ mentioned above. However, this would require FtsZ to be produced and accumulating at the division sites only from replication initiation to cell division. Although this is conceivable, that is it is known that FtsZ and other division proteins are heavily regulated at several levels (Dewar and Dorazi, 2000) and that especially in slow growth conditions its concentration varies during the cell cycle (Männik et al., 2018), it is hard to imagine how this model could work under fast growth conditions in which there are overlapping rounds of replication such that FtsZ would be constantly expressed. Moreover recent data (Si et al., 2019) rather suggest that FtsZ concentration is constant during the cell cycle.

Alternatively, rather than FtsZ production, Ftsz polymerization could be regulated. One remarkable observation that is well known within the field (Lau et al., 2004; Nielsen et al., 2006) and that we also observe in our data (see Figure 4—figure supplement 1), is that origins always occupy the position of future division sites (mid-cell, 1/4 and 3/4 positions etc.) when replication is initiated. This observation not only suggests that, at replication initiation, some local molecular event occurs that will eventually trigger division at the same site, but it is also remarkably consistent with the idea of an adder running only between replication initiation and division. One long-standing idea that is consistent with these observations is that some molecular event that occurs during replication initiation triggers the start of FtsZ ring formation, and that the timing from initiation to division is controlled by the polymerization dynamics of the FtsZ ring (Weart and Levin, 2003). At the molecular level, the common triggering of initiation and polymerization might be explained by the accumulation of acidic phosholipids in the cell membrane precisely at future division sites (Renner and Weibel, 2011) where they probably interact with components of the division machinery. At the same time those lipids are known to play a role in promoting replication by rejuvenating the initiator protein DnaA-ADP into DnaA-ATP (Saxena et al., 2013), and might therefore be a 'hub’ coordinating the two cycles. Finally, it remains to be explained how FtsZ polymerization or pole building could result in an adder behavior. For that purpose, future experiments should focus on combining the type of information collected in this study and detailed measures of the dynamics of FtsZ-ring assembly and constriction as done in Coltharp et al. (2016).

Images of growth channels and MoMA segmentations have been deposited on Zenodo.

1. 1. Witz G
   2. van Nimwegen E
   3. Julou T

   (2019) Zenodo

   Analysis of division and replication cycles in E. coli using time-lapse microscopy, microfluidics and the MoMA software.

   https://doi.org/10.5281/zenodo.3149097

1. 1. Adiciptaningrum A
   2. Osella M
   3. Moolman MC
   4. Cosentino Lagomarsino M
   5. Tans SJ

   (2016) Stochasticity and homeostasis in the E. coli replication and division cycle

   Scientific Reports 5:18261.

   https://doi.org/10.1038/srep18261

   • Google Scholar
2. 1. Aguet F
   2. Antonescu CN
   3. Mettlen M
   4. Schmid SL
   5. Danuser G

   (2013) Advances in analysis of low signal-to-noise images link dynamin and AP2 to the functions of an endocytic checkpoint

   Developmental Cell 26:279–291.

   https://doi.org/10.1016/j.devcel.2013.06.019

   • PubMed
   • Google Scholar
3. Software

   1. Allan DB
   2. Caswell T
   3. Keim NC
   4. van der Wel CM

   (2018)

   Trackpy, version v0.4.1

   Zenodo.
4. 1. Amir A

   (2014) Cell size regulation in Bacteria

   Physical Review Letters 112:208102.

   https://doi.org/10.1103/PhysRevLett.112.208102

   • Google Scholar
5. 1. Amir A

   (2017) Is cell size a spandrel?

   eLife 6:e22186.

   https://doi.org/10.7554/eLife.22186

   • PubMed
   • Google Scholar
6. 1. Baba T
   2. Ara T
   3. Hasegawa M
   4. Takai Y
   5. Okumura Y
   6. Baba M
   7. Datsenko KA
   8. Tomita M
   9. Wanner BL
   10. Mori H

   (2006)

   Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the keio collection

   Molecular Systems Biology 2:2:2006.0008.

   • Google Scholar
7. 1. Campos M
   2. Surovtsev IV
   3. Kato S
   4. Paintdakhi A
   5. Beltran B
   6. Ebmeier SE
   7. Jacobs-Wagner C

   (2014) A constant size extension drives bacterial cell size homeostasis

   Cell 159:1433–1446.

   https://doi.org/10.1016/j.cell.2014.11.022

   • PubMed
   • Google Scholar
8. 1. Coltharp C
   2. Buss J
   3. Plumer TM
   4. Xiao J

   (2016) Defining the rate-limiting processes of bacterial cytokinesis

   PNAS 113:E1044–E1053.

   https://doi.org/10.1073/pnas.1514296113

   • PubMed
   • Google Scholar
9. 1. Cooper S
   2. Helmstetter CE

   (1968) Chromosome replication and the division cycle of Escherichia coli B/r

   Journal of Molecular Biology 31:519–540.

   https://doi.org/10.1016/0022-2836(68)90425-7

   • PubMed
   • Google Scholar
10. 1. Datsenko KA
    2. Wanner BL

    (2000) One-step inactivation of chromosomal genes in Escherichia coli K-12 using PCR products

    PNAS 97:6640–6645.

    https://doi.org/10.1073/pnas.120163297

    • PubMed
    • Google Scholar
11. 1. Dewar SJ
    2. Dorazi R

    (2000) Control of division gene expression in Escherichia coli

    FEMS Microbiology Letters 187:1–7.

    https://doi.org/10.1111/j.1574-6968.2000.tb09127.x

    • PubMed
    • Google Scholar
12. 1. Donachie WD

    (1968) Relationship between cell size and time of initiation of DNA replication

    Nature 219:1077–1079.

    https://doi.org/10.1038/2191077a0

    • PubMed
    • Google Scholar
13. 1. Edelstein AD
    2. Tsuchida MA
    3. Amodaj N
    4. Pinkard H
    5. Vale RD
    6. Stuurman N

    (2014) Advanced methods of microscope control using μmanager software

    Journal of Biological Methods 1:10.

    https://doi.org/10.14440/jbm.2014.36

    • Google Scholar
14. 1. Eun YJ
    2. Ho PY
    3. Kim M
    4. LaRussa S
    5. Robert L
    6. Renner LD
    7. Schmid A
    8. Garner E
    9. Amir A

    (2018) Archaeal cells share common size control with Bacteria despite noisier growth and division

    Nature Microbiology 3:148–154.

    https://doi.org/10.1038/s41564-017-0082-6

    • PubMed
    • Google Scholar
15. 1. Harris LK
    2. Theriot JA

    (2016) Relative rates of surface and volume synthesis set bacterial cell size

    Cell 165:1479–1492.

    https://doi.org/10.1016/j.cell.2016.05.045

    • PubMed
    • Google Scholar
16. 1. Helmstetter C
    2. Cooper S
    3. Pierucci O
    4. Revelas E

    (1968) On the bacterial life sequence

    Cold Spring Harbor Symposia on Quantitative Biology 33:809–822.

    https://doi.org/10.1101/SQB.1968.033.01.093

    • Google Scholar
17. 1. Helmstetter CE

    (1974) Initiation of chromosome replication in Escherichia coli

    Journal of Molecular Biology 84:21–36.

    https://doi.org/10.1016/0022-2836(74)90210-1

    • Google Scholar
18. 1. Ho PY
    2. Amir A

    (2015) Simultaneous regulation of cell size and chromosome replication in Bacteria

    Frontiers in Microbiology 6:662.

    https://doi.org/10.3389/fmicb.2015.00662

    • PubMed
    • Google Scholar
19. 1. Hunter JD

    (2007) Matplotlib: a 2D graphics environment

    Computing in Science & Engineering 9:90–95.

    https://doi.org/10.1109/MCSE.2007.55

    • Google Scholar
20. 1. Jones E
    2. Oliphant T
    3. Peterson P

    (2001)

    SciPy: Open source scientific tools for Python

    SciPy: Open source scientific tools for Python.

    • Google Scholar
21. 1. Kaiser M
    2. Jug F
    3. Julou T
    4. Deshpande S
    5. Pfohl T
    6. Silander OK
    7. Myers G
    8. van Nimwegen E

    (2018) Monitoring single-cell gene regulation under dynamically controllable conditions with integrated microfluidics and software

    Nature Communications 9:212.

    https://doi.org/10.1038/s41467-017-02505-0

    • Google Scholar
22. 1. Koppes LJ
    2. Nanninga N

    (1980)

    Positive correlation between size at initiation of chromosome replication in Escherichia coli and size at initiation of cell constriction

    Journal of Bacteriology 143:89–99.

    • PubMed
    • Google Scholar
23. 1. Lau IF
    2. Filipe SR
    3. Søballe B
    4. Økstad O-A
    5. Barre F-X
    6. Sherratt DJ

    (2004) Spatial and temporal organization of replicating Escherichia coli chromosomes

    Molecular Microbiology 49:731–743.

    https://doi.org/10.1046/j.1365-2958.2003.03640.x

    • Google Scholar
24. 1. Logsdon MM
    2. Ho PY
    3. Papavinasasundaram K
    4. Richardson K
    5. Cokol M
    6. Sassetti CM
    7. Amir A
    8. Aldridge BB

    (2017) A parallel adder coordinates mycobacterial Cell-Cycle progression and Cell-Size homeostasis in the context of asymmetric growth and organization

    Current Biology 27:3367–3374.

    https://doi.org/10.1016/j.cub.2017.09.046

    • PubMed
    • Google Scholar
25. 1. Männik J
    2. Walker BE
    3. Männik J

    (2018) Cell cycle-dependent regulation of FtsZ in Escherichia coli in slow growth conditions

    Molecular Microbiology 110:1030–1044.

    https://doi.org/10.1111/mmi.14135

    • PubMed
    • Google Scholar
26. Conference

    1. McKinney W

    (2010) Data structures for statistical computing in Python

    Proceedings of the 9th Python in Science Conference. pp. 51–56.

    https://conference.scipy.org/proceedings/scipy2010/mckinney.html

    • Google Scholar
27. 1. Micali G
    2. Grilli J
    3. Marchi J
    4. Osella M
    5. Cosentino Lagomarsino M

    (2018a) Dissecting the control mechanisms for DNA replication and cell division in E. coli

    Cell Reports 25:761–771.

    https://doi.org/10.1016/j.celrep.2018.09.061

    • PubMed
    • Google Scholar
28. 1. Micali G
    2. Grilli J
    3. Osella M
    4. Cosentino Lagomarsino M

    (2018b) Concurrent processes set E. coli cell division

    Science Advances 4:eaau3324.

    https://doi.org/10.1126/sciadv.aau3324

    • PubMed
    • Google Scholar
29. 1. Morgan-Kiss RM
    2. Wadler C
    3. Cronan JE

    (2002) Long-term and homogeneous regulation of the Escherichia coli araBAD promoter by use of a lactose transporter of relaxed specificity

    PNAS 99:7373–7377.

    https://doi.org/10.1073/pnas.122227599

    • PubMed
    • Google Scholar
30. 1. Nielsen HJ
    2. Li Y
    3. Youngren B
    4. Hansen FG
    5. Austin S

    (2006) Progressive segregation of the Escherichia coli chromosome

    Molecular Microbiology 61:383–393.

    https://doi.org/10.1111/j.1365-2958.2006.05245.x

    • PubMed
    • Google Scholar
31. Preprint

    1. Oldewurtel ER
    2. Kitahara Y
    3. Cordier B
    4. Özbaykal G
    5. van Teeffelen S

    (2019) Bacteria control cell volume by coupling cell-surface expansion to dry-mass growth

    bioRxiv.

    https://doi.org/10.1101/769786

    • Google Scholar
32. 1. Pierucci O

    (1978)

    Dimensions of Escherichia coli at various growth rates: model for envelope growth

    Journal of Bacteriology 135:559–574.

    • PubMed
    • Google Scholar
33. 1. Renner LD
    2. Weibel DB

    (2011) Cardiolipin microdomains localize to negatively curved regions of Escherichia coli membranes

    PNAS 108:6264–6269.

    https://doi.org/10.1073/pnas.1015757108

    • PubMed
    • Google Scholar
34. 1. Reyes-Lamothe R
    2. Possoz C
    3. Danilova O
    4. Sherratt DJ

    (2008) Independent positioning and action of Escherichia coli replisomes in live cells

    Cell 133:90–102.

    https://doi.org/10.1016/j.cell.2008.01.044

    • PubMed
    • Google Scholar
35. 1. Sánchez-Gorostiaga A
    2. Palacios P
    3. Martínez-Arteaga R
    4. Sánchez M
    5. Casanova M
    6. Vicente M

    (2016)

    Life without division: physiology of Escherichia coli FtsZ-Deprived filaments

    mBio 7:5.

    • Google Scholar
36. 1. Saxena R
    2. Fingland N
    3. Patil D
    4. Sharma AK
    5. Crooke E

    (2013) Crosstalk between DnaA protein, the initiator of Escherichia coli chromosomal replication, and acidic phospholipids present in bacterial membranes

    International Journal of Molecular Sciences 14:8517–8537.

    https://doi.org/10.3390/ijms14048517

    • PubMed
    • Google Scholar
37. 1. Schaechter M
    2. MaalOe O
    3. Kjeldgaard NO

    (1958) Dependency on medium and temperature of cell size and chemical composition during balanced grown of Salmonella typhimurium

    Journal of General Microbiology 19:592–606.

    https://doi.org/10.1099/00221287-19-3-592

    • PubMed
    • Google Scholar
38. 1. Si F
    2. Li D
    3. Cox SE
    4. Sauls JT
    5. Azizi O
    6. Sou C
    7. Schwartz AB
    8. Erickstad MJ
    9. Jun Y
    10. Li X
    11. Jun S

    (2017) Invariance of initiation mass and predictability of cell size in Escherichia coli

    Current Biology 27:1278–1287.

    https://doi.org/10.1016/j.cub.2017.03.022

    • PubMed
    • Google Scholar
39. 1. Si F
    2. Le Treut G
    3. Sauls JT
    4. Vadia S
    5. Levin PA
    6. Jun S

    (2019) Mechanistic origin of Cell-Size control and homeostasis in Bacteria

    Current Biology 29:1760–1770.

    https://doi.org/10.1016/j.cub.2019.04.062

    • PubMed
    • Google Scholar
40. 1. Soifer I
    2. Robert L
    3. Amir A

    (2016) Single-Cell analysis of growth in budding yeast and Bacteria reveals a common size regulation strategy

    Current Biology 26:356–361.

    https://doi.org/10.1016/j.cub.2015.11.067

    • PubMed
    • Google Scholar
41. 1. Taheri-Araghi S
    2. Bradde S
    3. Sauls JT
    4. Hill NS
    5. Levin PA
    6. Paulsson J
    7. Vergassola M
    8. Jun S

    (2017) Cell-Size control and homeostasis in Bacteria

    Current Biology 27:1392.

    https://doi.org/10.1016/j.cub.2017.04.028

    • PubMed
    • Google Scholar
42. 1. Tanouchi Y
    2. Pai A
    3. Park H
    4. Huang S
    5. Stamatov R
    6. Buchler NE
    7. You L

    (2015) A noisy linear map underlies oscillations in cell size and gene expression in Bacteria

    Nature 523:357–360.

    https://doi.org/10.1038/nature14562

    • PubMed
    • Google Scholar
43. 1. van der Walt S
    2. Colbert SC
    3. Varoquaux G

    (2011) The NumPy array: a structure for efficient numerical computation

    Computing in Science & Engineering 13:22–30.

    https://doi.org/10.1109/MCSE.2011.37

    • Google Scholar
44. 1. van der Walt S
    2. Schönberger JL
    3. Nunez-Iglesias J
    4. Boulogne F
    5. Warner JD
    6. Yager N
    7. Gouillart E
    8. Yu T

    (2014) scikit-image: image processing in Python

    PeerJ 2:e453.

    https://doi.org/10.7717/peerj.453

    • Google Scholar
45. 1. Wallden M
    2. Fange D
    3. Lundius EG
    4. Baltekin Özden
    5. Elf J

    (2016) The synchronization of replication and division cycles in individual E. coli Cells

    Cell 166:729–739.

    https://doi.org/10.1016/j.cell.2016.06.052

    • Google Scholar
46. 1. Wang P
    2. Robert L
    3. Pelletier J
    4. Dang WL
    5. Taddei F
    6. Wright A
    7. Jun S

    (2010) Robust growth of Escherichia coli

    Current Biology 20:1099–1103.

    https://doi.org/10.1016/j.cub.2010.04.045

    • PubMed
    • Google Scholar
47. 1. Weart RB
    2. Levin PA

    (2003) Growth rate-dependent regulation of medial FtsZ ring formation

    Journal of Bacteriology 185:2826–2834.

    https://doi.org/10.1128/JB.185.9.2826-2834.2003

    • PubMed
    • Google Scholar
48. 1. Wehrens M
    2. Ershov D
    3. Rozendaal R
    4. Walker N
    5. Schultz D
    6. Kishony R
    7. Levin PA
    8. Tans SJ

    (2018) Size laws and division ring dynamics in filamentous Escherichia coli cells

    Current Biology 28:972–979.

    https://doi.org/10.1016/j.cub.2018.02.006

    • PubMed
    • Google Scholar
49. 1. Willis L
    2. Huang KC

    (2017) Sizing up the bacterial cell cycle

    Nature Reviews Microbiology 15:606–620.

    https://doi.org/10.1038/nrmicro.2017.79

    • PubMed
    • Google Scholar
50. Software

    1. Witz G

    (2019) Guiwitz/DoubleAdderArticle

    GitHub.

    https://github.com/guiwitz/DoubleAdderArticle/tree/v1.0

Author details

1. Guillaume Witz

   1. Biozentrum, University of Basel, Basel, Switzerland
   2. Swiss Institute of Bioinformatics, University of Bern, Bern, Switzerland

   Present address

   Mathematical Institute, University of Bern, Bern, Switzerland

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Funding acquisition, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing

   For correspondence

   guillaume.witz@math.unibe.ch

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1562-4265

2. Erik van Nimwegen

   1. Biozentrum, University of Basel, Basel, Switzerland
   2. Swiss Institute of Bioinformatics, University of Bern, Bern, Switzerland

   Contribution

   Formal analysis, Funding acquisition, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

3. Thomas Julou

   1. Biozentrum, University of Basel, Basel, Switzerland
   2. Swiss Institute of Bioinformatics, University of Bern, Bern, Switzerland

   Contribution

   Resources, Formal analysis, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-7123-198X

• 3,860

  views

• 554

  downloads

• 59

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.