Main

Many biological systems show remarkably precise and robust dynamics. For example, the circadian clock in cyanobacteria uses a combination of transcriptional and post-translational control mechanisms2,3 to keep phase for weeks without entrainment, while displaying robustness to changes in temperature and growth rate3,4,5. Synthetic circuits built from well-characterized parts can also exhibit a wide range of dynamical features—including arithmetic computations6,7, oscillations1,8,9,10,11,12,13, logic gates14 and edge detection15—but often with lower accuracy. For example, the repressilator1, a now iconic device that helped to jump-start the field of synthetic biology 15 years ago, showed clear signs of oscillations using a simple design in which three genes inhibit each other’s production in a single loop (A B C A). However, only about 40% of cells were found to support oscillations, and those oscillations were irregular. Subsequent synthetic oscillators evaluated different control topologies or repression mechanisms8,9,10,11,12,13, but most were again irregular in both phase and amplitude despite being mathematically designed to display sustained oscillations in a wide variety of parameters.

The challenge when designing synthetic circuits to operate reliably in single cells is that biochemical noise can do more than just create different rate constants in different cells. Simple intrinsic noise can, in principle, enhance control16 and even create high-quality oscillations in systems that could not display limit cycles for any rate constants in the absence of noise17,18. However, any component present in low numbers can also randomize behaviour of the whole system, and a single stochastic signalling step can introduce fundamental constraints19 that cannot be overcome by any control system. This suggests that simplicity could even help to achieve accurate oscillations as long as stochastic effects are accounted for in the design, and that minimal control topologies may not only be elegant and interesting but also very effective. We therefore revisited the original repressilator to reduce error propagation from the reporter system, from core cellular processes, and from within the circuit itself.

The repressilator consists of three genes—tetR from the Tn10 transposon, cI from bacteriophage λ and lacI from the lactose operon—and each repressor has a C-terminal ssrA tag20 that targets it for degradation (Fig. 1a). The whole circuit was encoded on a low-copy pSC101 plasmid in an Escherichia coli strain lacking lacI, and a second high-copy ColE1 reporter plasmid that encoded green fluorescent protein (GFP) under the control of TetR with a modified degradation tag1,21. We first re-evaluated this circuit using a microfluidic device in which cells are trapped in short channels and newborn cells are washed away by fresh medium22,23 (Fig. 1a). Tracking reporter levels under the microscope for hundreds of consecutive generations across hundreds of single-cell traces (Methods) revealed clear oscillatory dynamics in all cells (Fig. 1b, Extended Data Fig. 1a), particularly in the rate of production (Extended Data Fig. 1a). This shows that the simple design was sound and that some of the erratic behaviour originally reported was due to the limited imaging platforms available at the time.

Figure 1: Reducing reporter interference.
figure 1

a, Schematics of the original repressilator plasmids and microfluidic device in which E. coli cells are diffusively fed in growth channels and daughter cells are eventually washed away. b, Typical time trace of a single cell for original repressilator (NDL332). The GFP concentration (green trace) oscillates noisily, whereas expression of the red fluorescent protein (RFP; red trace) remains constant. Both traces were normalized to their means. c, Autocorrelation functions (ACF) and power spectral densities (PSD) were calculated over the whole population (2,706 generations) and demonstrate oscillations with a mean period of 2.4 average division time. d, Top, oscillations are more regular when the reporter is expressed on the repressilator plasmid rather than on a separate high-copy plasmid (Extended Data Fig. 2). Some cells irreversibly shift period from approximately 2.5 to 5.5 generations. Bottom, the period change was invariably connected to a loss of the separate mCherry-ASV-expressing reporter plasmid. Analysis of, for example, empty plasmid vectors, various reporter proteins and reporter degradation tags, and circuits with and without repressor degradation (Supplementary Information 3.1 and 3.3) shows that the interference was caused by the reporter ssrA degradation tag in which the last three amino acids were substituted to ASV. e, ACF and PSD for the YFP-expressing repressilator without separate reporter plasmid (LPT25), calculated over all 8,694 total cell divisions observed. Average period was 5.6 generations. Reporter protein close to fluorescence detection limit at troughs, and the actively degraded repressors should be much lower yet. The PSD was normalized by peak frequency, with width of the window function indicated by a red line. f, Histograms of interpeak distances for one, two and three periods in blue, red and black, respectively. Orange or grey lines were obtained by summing two or three samples, respectively, from the blue distribution. Consecutive periods are thus independent. Right panel shows that the variance in period grows linearly with the number of periods elapsed (LPT25).

PowerPoint slide

Full size image

We next evaluated how much of the noise reflected error propagation from the reporter system. Mathematical predictions have suggested that high-copy ColE1 cloning vectors fluctuate substantially and slowly, owing to poorly controlled self-replication, and therefore effectively transmit fluctuations to encoded proteins24. Moving the fluorescent protein reporter onto the low-copy repressilator plasmid indeed reduced the relative standard deviation in amplitude greatly, from 78% to 36% (Fig. 1d, Extended Data Fig. 2a, b). Degradation-tagged reporter proteins have also been predicted to potentially have ‘retroactivity’ effects on oscillations25 owing to competition for shared proteases. Protease competition has in fact been cleverly exploited for improved control in synthetic circuits13, but stochastic theory24,26 suggests that saturated degradation enzymes can also create effects related to the dynamic instability of microtubules, with large random fluctuations in single cells. Comparison of several constructs showed that the synthetic degradation tag caused fluctuations to propagate from the reporter proteins to the repressors via the proteolysis system (Supplementary Information 3.1.1). Notably, however, the ‘competing’ reporter proteins accelerated rather than decelerated the degradation of ssrA-tagged substrates (Extended Data Fig. 3). Removing this interference created very regular oscillations, with periods increasing from around 2.4 to 5.7 generations (Fig. 1d, e, Extended Data Figs 1b and 2c–e). Characterizing the phase drift (the statistical tendency of oscillations in individual cells to go out of phase with each other) showed that on average this circuit oscillates for 5.5 periods before accumulating half a period of drift (Methods).

In cell-free extracts, that is, without the low-copy noise of single cells, the repressilator has been shown to display the sinusoidal curves expected for harmonic oscillators27. However, analysing the highly asymmetric shape of the time traces in single cells shows that it effectively operates as a relaxation oscillator28, with a characteristic build-up phase sharply followed by an almost pure dilution and degradation phase until concentrations reach very low levels (Fig. 1d, Extended Data Fig. 4 and Box 1). The mathematical conditions for sustained harmonic oscillations—cooperative repression and similar mRNA and protein half-lives1—are then less relevant, but it becomes crucial to reduce the heterogeneity in the build-up and dilution phases (Box 1, Supplementary Information 4.2) because stochastic effects otherwise fundamentally compromise the ability of the system to keep track of time. Specifically, if the production phase for each of the repressors involves a low number of stochastic production events, statistical variation in that number will cause heterogeneity in peak amplitude, which then to some extent creates heterogeneity in the subsequent dilution and decay period. If peak protein abundances are low, random degradation events or uneven partitioning of molecules at cell division will in turn cause heterogeneity in the decay and dilution process. However, increasing peak abundances should only help marginally unless the repression thresholds are also increased appropriately (Box 1), since the last few steps contribute disproportionally to the variance (Box 1, Supplementary Information 4.2.2).

Motivated by these results, we eliminated repressor degradation by removing the ssrA degradation tags from the repressors, by using a ΔclpXP strain, or both (Supplementary Information 3.3). These circuits oscillated in all cells, with a period of approximately 10 generations. However, as predicted, the noise in the period was only slightly reduced (Fig. 2a, Extended Data Figs 5c and 6). To pinpoint the reason we built a circuit with compatible fluorescent reporter proteins for each repressor. Analysing the variance in the three interpeak distances showed that the noisiest phase was when TetR levels were low (Fig. 2b). We then estimated the protein abundances from the partitioning errors at cell division (Supplementary Information 3.5), and found that the derepression of the TetR-controlled promoter occurs at an extremely low threshold.

Figure 2: Identifying and eliminating inherent sources of error.
figure 2

a, Typical time trace in ΔclpXP cells (LPT61) in which repressors are not degraded. ACF and PSD calculated over 5,356 cell divisions. Average period was 10 generations, and correlation coefficient was 0.1. Dashed vertical lines are separated by an average period to illustrate periodicity in a–c. b, Left, time trace of multireporter repressilator (ΔclpXP, LPT113). TetR represses YFP production (yellow trace), LacI inhibits CFP production (blue trace) and cI represses RFP production (red trace). Peak indicated by asterisk not shown owing to its high amplitude of 11.5 units. Right, interpeak distances evaluated for YFP to CFP (YtoC, red), CFP to RFP (CtoR, yellow) and RFP to YFP (RtoY, blue), without (LPT113, n = 163, 150 and 173) and with (LPT117 and LPT127, combined, n = 109, 86 and 116) the titration sponge (plasmid with PLtet-binding sites). Respective contributions to the average and variance shown by bar plot. The RtoY part of the oscillation (induction of YFP, low TetR levels) represents 27% of the period, but contributes 44% of the variance. Addition of the PLtet titration sponge reduces the variance almost fourfold. c, Example time trace of single-reporter repressilator with PLtet-mCherry-ASV (ΔclpXP, LPT64), along with ACF and PSD calculated over 3,695 generations. Oscillations have an average period of 14 generations and a correlation coefficient of 0.5 after one period. Inset shows a time trace from the triple reporter repressilator without degradation and with titration sponge (LPT127, colour scheme as in b).

PowerPoint slide

Full size image

The theory suggests that the regularity could be greatly improved if this threshold was raised, for example, using a ‘sponge’ of repressor-binding sites that soaks up small numbers of TetR molecules. The high-copy reporter plasmid included in the original repressilator design in fact already carried binding sites for TetR, and simply reintroducing it greatly reduced the noise in all steps (Fig. 2b), whereas similar sponges for the cI and LacI repressor proteins had minor effects (Extended Data Fig. 7d) as expected. Titration may in fact be a particularly useful way to increase the thresholds because it can also help to create sharp switches29 (Supplementary Information 4.3), which may or may not be necessary for oscillations in single cells but should generally increase accuracy.

These changes created a streamlined repressilator with highly regular oscillations that peak around every 14 generations (Fig. 2c). Each repressor spends several generations at virtually undetectable concentrations (Supplementary Information 3.5) followed by several generations at concentrations that completely saturate repression. The amplitude still displays some variation (Extended Data Fig. 8d), but because the time it takes to dilute levels from a peak amplitude of N to a threshold of S depends logarithmically on N/S, little variation in amplitude is transmitted to the timing (Box 1, Extended Data Fig. 8e). The phase drift was only about 14% per period (Extended Data Fig. 5d) and on average the circuit should oscillate for approximately 18 periods before accumulating half a period of drift (Methods). The theory shows that similar accuracy should be possible in systems in which dilution in growing cells is replaced by first-order degradation, and that it is not the slowness itself that creates accuracy, but the absolute number of proteins at peaks and troughs.

The periods of circadian clocks, as measured in hours, are often robust to changes in growth conditions. However, other intracellular oscillators may need periods that instead are robust relative to internal physiological time scales such as the generation times. Synthetic circuits generally do not display either type of quantitative robustness because periods depend on so many different parameters that change with conditions. That is in principle also true for the circuits above: as conditions change, the plasmid copy numbers, RNA degradation, gene expression, and cell volume change in non-trivial ways. However, the logarithmic dampening that makes individual periods insensitive to fluctuations in peak amplitudes (Box 1) is also predicted to make the number of generations per period insensitive to growth conditions (Box 1). We found that the circuit retained the 14-generation period under all conditions tested (Fig. 3a), including growth at 25–37 °C and in conditioned medium from early stationary phase cultures in which the cells become much smaller and almost spherical. The combination of robustness to conditions and great inherent precision suggest that cells could display macroscopic, population-scale oscillations without cell–cell communication. We therefore synchronized (Extended Data Fig. 9) a liquid culture and maintained it in early exponential phase (Methods). We found that whole flasks oscillated autonomously, with a period of the expected 14 generations (Fig. 3b, Supplementary Video 1). We also imaged the growth of large colonies originating from single cells containing the triple reporter repressilator. Because only the cells at the edges of the colonies grow considerably cells in the interior were arrested in different repressilator phases, creating ring-like expression profiles much like the seasonal growth rings seen in tree stumps (Fig. 3c, Extended Data Fig. 7a). The regularity originated in the autonomous behaviour of single cells—no connections were introduced and cells kept their own phase when merging into areas where the neighbouring cells had a different phase (Extended Data Fig. 7c).

Figure 3: The modified repressilator shows great robustness to growth conditions.
figure 3

a, The repressilator without degradation and with titration sponge (LPT64) has a period of 14 generations at different temperatures (blue bars, division time of 27, 40 and 59 min for 37 °C, 30 °C and 25 °C, respectively) and in conditioned media (OD600 nm value of 2, doubling time of 44 min). Repressilator with repressor degradation (LPT25) shows a varying period (yellow bars, doubling time of 26, 34 and 52 min for 37 °C, 30 °C and 25 °C, respectively). Error bars indicate s.d. on the first maximum of the ACF obtained by bootstrapping. b, Cells containing multireporter repressilator without repressor degradation and with PLtet-peptide-ASV plasmid (ΔclpXP, LPT117) were grown in liquid culture in 25-ml flasks. After the culture was initially synchronized with isopropyl-β-d-thiogalactoside (IPTG), it was kept in exponential phase via dilution. Average YFP intensity shown for coloured square area, with unsychronized culture for comparison. c, A 5-mm diameter colony of cells with the triple reporter repressilator (LPT117) reveals tree-like ring patterns in fluorescent protein levels. The average YFP intensity is reported for the slice in the white rectangle. The decrease in RFP levels towards the edge of the colony is probably due to different response to stationary phase of its promoter.

PowerPoint slide

Full size image

These results (summarized in Extended Data Fig. 10) illustrate the importance of understanding genetic networks at the level of stochastic chemistry, particularly for synthetic circuits in which the noise has not been shaped by natural selection and where the heterologous components and reporters used may interfere with cells. We propose that if statistical properties are systematically measured and the mechanisms are iteratively redesigned based on general stochastic principles, the next generation of synthetic circuits could rival or even surpass the precision of natural systems.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment.

Data and materials availability

Segmented and assembled single-cell traces files are accessible online. The masks files used for the microfluidic device master fabrication are available on request. The plasmids have been deposited to the Addgene plasmid repository, and the plasmids maps are available online.

Imaging protocol: chip preparation

Dimethyl siloxane monomer (Sylgard 184) was mixed in a 10:1 ratio with curing agent, defoamed, poured onto the silicon wafer, degassed for 1 h and cured at 65 °C for 1 h. Individual chips were then cut and the inlets and outlets were punched with a biopsy puncher. Bonding to KOH-cleaned coverslips was ensured using oxygen plasma treatment (30 s at 50 W and O2 pressure at 170 mTorr) on the day the experiments start. The chips were then incubated at 95 °C for at least 30 min to reinforce the bonding.

Cell preparation

Escherichia coli strains were grown overnight in LB with appropriate antibiotics and diluted 1:100 approximately 2–3 h before the beginning of the experiments in imaging media, consisting of M9 salts, 10% (v/v) LB, 0.2% (w/v) glucose, 2 mM MgSO4, 0.1 mM CaCl2, 1.5 μM thiamine hydrochloride and 0.85 g l−1 pluronic F-108 (Sigma Aldrich, included as a passivating agent). The cells were loaded into the device at OD600 values 0.2–0.4, and centrifuged on a custom-machined holder that could fit into a standard table-top centrifuge at 5,000g for 10 min to insert them into the side-channels. The feeding channels were connected to syringes filled with imaging media using Tygon tubing (VWR), and media was pumped using syringe pumps (New Era Pump System) initially at a high rate of 100 μl min−1 for 1 h, to clear the inlets and outlets. The media was then pumped at 5–10 μl min−1 for the duration of the experiment and cells were allowed to adapt to the device for several hours before imaging was started.

Microscopy and image acquisition

Images were acquired using a Nikon Ti inverted microscope equipped with a temperature-controlled incubator, an Orca R2 CCD camera (Hamamatsu), a 60× Plan Apo oil objective (numerical aperture (NA) 1.4, Nikon), an automated xy-stage (Ludl) and light engine LED excitation source (Lumencor). All experiments were performed at 37 °C. Microscope control was done with MATLAB (Mathworks) scripts interfacing with μManager31. Typical exposure was low (50–100 ms) to reduce photobleaching, and the reporter channels were acquired using 2 × 2 binning (CCD chip dimension of 1,344 × 1,024 pixels, effective pixel size of 129 × 129 nm). Then 16-bit TIFF images were taken every 5–8 min, and focal drift was controlled via the Nikon PerfectFocus system, as well as a custom routine based on z-stack images of a sacrificial position (a position that was not used for further analysis). The following filter sets were used for acquisition: GFP (Semrock GFP-3035B), RFP (Semrock mCherry-A), YFP (Semrock YFP-2427A) and CFP (Semrock CFP-2432A).

Conditioned medium

The conditioned medium was obtained by growing the strain used in the experiment until OD600 nm = 2.0, and then the culture was rapidly sterilized with a 0.2-μm filter and kept at 4 °C until the experiment. When indicated, the imaging medium was supplemented with IPTG for the duration indicated in the figure.

Data processing: segmentation

Image segmentation and single-cell trace assembly were performed similarly to a previously described procedure23. In brief, the segmentation was done using images from a bright, constitutively expressed (PRNA1 promoter on the chromosome or on the plasmid) CFP or RFP. The rough channel boundaries were estimated, to reject out-of-channel cells, with a simple threshold followed by erosion, opening and dilation of the mask. The contrast of the fluorescent image was enhanced using a ‘unsharp mask’. Then, the edges of the cells were detected using the Laplacian of Gaussian method. Cells joined by their poles (as indicated by objects with definite constrictions) were separated, and spurious non-cell objects were rejected using their size, orientation and shape. Finally, the boundaries were refined using opening, thickening and active contours. The parameters used for these functions were optimized specifically for the combination of the strain, growth conditions and microscope setup. We will share the code used on request, but the specific parameters will need to be re-optimized depending on the exact setup.

We chose to follow only the cells at the top (closed end) of the channel (the ‘mother’ cell), as it made compiling the single-cell traces much easier as these cells stay in place for the duration of the experiment. Owing to physical limitations of the setup, the segmentation mask was slightly mis-registered with respect to the ‘data channel’ (that is, GFP or YFP). Each object was then registered to the proper channel before the data were extracted.

For the triple reporter strains (that did not contain a specific segmentation fluorophore), we combined the three reporters channels according to their signal-to-noise ratio to obtain an effective segmentation channel. In some cases, we used an alternative procedure in which the whole channels were segmented, and the top 50 pixels were used as a ‘cell’ that represented one to two cells. Because the oscillations for these analyses were very slow and the cells in the channels are very close in phase (for example, see Extended Data Fig. 5g), the two methods gave very similar results to the normal procedure, but in general the ‘channel’ analysis worked more reliably. This analysis was only used in Fig. 2b.

The background fluorescence was corrected by subtracting the median value of the fluorescent images (the cells represent a very small fraction of the image). We then estimated the concentration of fluorophore using the average of the background-substracted intensities inside the segmentation mask. Nearly identical results were obtained using the ‘peak’ intensity (median of the top 10% of the pixel in the segmentation mask), but for simplicity we only report the results obtained with the mean.

Single-cell trace construction

The temporal information of the cell data (such as intensities and area) was then compiled into single-cell traces by matching the centroid of the cells from frame n to frame n + 1. As there was little drift, this procedure was very reliable, and we prevented spurious matches by setting an upper limit on the centroid distance. We identified cell divisions by sudden decreases in cell area; if the cell area dropped to less than 60% of its previous value, a division was called.

Production rate estimation

To estimate the production rate of fluorophore, we used the derivative of the concentration, as it was more robust to errors in segmentation. Let T be the total intensity, A the area of the cell and C the concentration. Since T(t) = A(t)C(t),

in which is the normalized production rate and g(t) is the growth rate (for example, g = ln(2)/τdiv for exponential growth, , τdiv being the doubling time). Equation (1) was used for estimating the production rate in the paper, with g(t) estimated for each cell cycle with the initial and final area and dC/dt using Tikhonov regularization to enforce smoothness32. The normalization factor was kept as small as possible, but similar results were obtained for factors one order of magnitude smaller or larger. In practice, photobleaching or degradation of the fluorescent protein can affect the estimation of the production rate. One can account for these effects by using an effective half-life instead of τdiv (for example, , in which τphoto and τdeg are the photobleaching and degradation half-lives, respectively). These effects were negligible for GFP (even with the asv degradation tags), so we chose to report the production rate without correction.

Autocorrelation function and power spectrum estimation

The autocorrelation functions were estimated by averaging the correlation functions of the individual cells, as it was more robust to outliers, and using the unbiased estimator. Similar functions were obtained by taking directly the autocorrelation of the population, but needed manual curation of the data to remove dead cells or filaments.

in which xi(t) is the production rate or the concentration (indicated in the figure caption) of the ith cell at time t. Averaging of the correlations functions of the cells was done taking into account the finite length of the time series (each cell has a different number of samples for a specific time lag). If Ai is the autocorrelation of cell i,

with , Δt the time between images, j the discrete delay index and Li the number of points in time trace i. The brackets are used to emphasize discrete sampling. The autocorrelations were cropped to a constantly decreasing envelope to keep only time lags with good estimates. This resulted in correlation functions very similar to the ones obtained by using the biased estimator, albeit with a slightly larger envelope.

The power spectrum was then estimated by taking the discrete Fourier transform (DFT) of the windowed autocorrelation function33,34:

in which DFTN is the N point DFT and a[m] is the windowed symmetric autocorrelation:

and w[m] is a window function. Then,

where X(ω) is the power spectrum of the signal and W(ω) the Fourier transform of the window function. We are therefore sampling the power spectrum of the signal convolved with W(ω). This is a consistent estimator of the power spectrum (it converges to the actual power spectrum as the amount of data goes to infinity)35. We used a triangular window function to avoid negative spectral leakage, and the length of the window function (2M) was chosen to maximize the resolution without introducing too much noise (50–225 frames, depending on the period of the oscillations). The approximate resolution loss was indicated by a red line of width 2π/M (1/(MΔt)) in the figures.

Period histograms and phase drift estimation

Peak-to-peak distances were evaluated by finding maxima using the findpeaks MATLAB function. The traces were first smoothed using a 3 or 5 points moving average and peaks were rejected if they were closer than 3 or 5 frames to avoid double counting, or smaller than the average of the trace. The peaks were then manually curated; this was especially useful for the noisy oscillators. Note that the average period was slighly shorter than the first maximum of the autocorrelation, most likely because longer periods have higher intensities and thus more weights in the correlation (but not in the period histogram).

The period histograms were made by using the peak-to-peak distance. The squared error on the nth period grew linearly with n, as expected for this type of oscillator undergoing a random walk in phase. We therefore used the coefficient of variation (CV, standard deviation divided by the mean) of the period as an indicator of phase drift; the normalization makes comparison between oscillators of different frequencies straightforward.

Most of the strains had a phase drift of 30–35% per period; except for the repressilator without degradation but with the titration sponge, where it was only 14%. Since the variance increased linearly, we can express the variance for n periods () as a function of the variance for one ():

Hence, it would take approximately 13 periods (179 generations) to obtain a standard deviation of half a period.

Another measure of the phase drift is the average time to reach half a period of phase drift, or the average first passage time. This could be calculated by drawing randomly directly from the period histogram until the first time the phase drift is reached, because subsequent periods were exceptionally well approximated as independent (Fig. 1f). This creates a distribution of first passage times, and after 105 iterations, we converge on an average first passage time of around 18 periods (240 generations), again for the repressilator without degradation with titration (LPT64).

IPTG synchronization and flask experiment

To synchronize the phase of the oscillators in the population, we diluted the strains in imaging medium supplemented with appropriate antibiotics and 1 mM IPTG so that they would be early exponential (OD600 = 0.2) 8 h later (∼1 × 10−6) at 37 °C. For the unsynchronized control, we did the same procedure but did not include IPTG. After, we diluted the cultures to OD600 0.05 every 50 min, while taking the fluorescent images of the (undiluted) flasks. The OD600 of the imaged culture varied slightly, but the effect was negligible, as can be seen on the unsynchronized control (Fig. 3b).

Photo acquisition

Photos were acquired using a digital camera setup equipped with emission filters and LEDs fluorescent excitation36. A custom-written software controls a Canon T3i digital single lens reflex (DSLR) camera with a Canon EF-S 60 mm USM lens, placed in front of a Starlight express filter wheel, and appropriate LEDs for excitation. A long exposition time of 10 s was used for the flask, enabling the use of small OD600, while the exposition time of the plates was 0.1–2 s.

Microscopy

Images for Fig. 3c were acquired using an Olympus MVX10 Macroview microscope equipped with a Zeiss AxioCam MRc camera. Flurophores were excited using a Lumen200 fluorescence illumination system (Prior Scientific) and we used the following Olympus filter sets: CFP (U-M40001XL), YFP (U-M49003XL) and mCherry (U-M49008XL).