An engineered multi-step differentiation program in Escherichia coli for self-organized spatial patterning

Reviewer #1 (Public review):

Summary:

This paper by Boni and colleagues presents the engineering of a multi-step differentiation program in Escherichia coli based on synthetic gene circuits. The motivation behind the study was to engineer a system capable of undergoing differentiation in a step-wise manner without the presence of external spatial cues and without inducers added during the differentiation process. To achieve this, the authors created several synthetic gene circuits, one being a toggle switch, and the others being quorum-sensing-mediated gene expression modules. The outputs of the differentiation process are fluorescent proteins, which allowed the authors to quantify the behavior of the system using fluorescence intensity measurements. The authors additionally built a multi-component mathematical model which is able to reproduce the experimental data and to make interesting predictions (which require future validation).
The data presented are convincing and support the claims, the work is well executed.

Strengths:

(1) The differentiation process proceeds autonomously after the initial step in liquid culture in the presence of external inducers.

(2) It is indeed a step-wise process.

(3) The mathematical model predicts the outcome (% of green, blue and red FP-expressing cells in the population) when changing the initial ratio of green:blue FP-expressing cells.

Comments on revised version:

The authors' replies to my comments were very satisfactory. I think the paper has been strengthened.

Reviewer #2 (Public review):

The revised manuscript by Boni et al. is substantially improved, and in my view the authors have responded constructively to the concerns raised during the first review. Most importantly, they now explicitly distinguish the bistable green/blue states generated by the toggle switch from the reversible red and yellow quorum-sensing outputs and acknowledge that the latter do not constitute irreversible differentiated states. This clarification improves the conceptual accuracy of the work.

The other revisions are also well justified. The authors clarify that Fig. 2d is derived quantitatively from flow-cytometry data and explain how the cross-sections in Fig. 2e were obtained; they introduce the nullcline interpretation of the toggle-switch landscape and distinguish stochastic single-cell fate from predictable population-level proportions. They also provide more information on pLux engineering, report the functional forms and parameters of fitted curves, quantify the different HSL induction ranges, explain the heuristic treatment of entry into stationary phase, and clarify image selection and sender-receiver distance calculations.

The principal strength remains the systematic engineering of a sequential multi-circuit programme in a single bacterial genetic circuitry (spread over two plasmids). The work combines bistable symmetry breaking, LuxI/LuxR-mediated communication and an orthogonal CinI/CinR layer to generate spatially organised colony-level behaviours. The DBTL engineering cycle is also convincing: substantial cross-talk in the initially tested Lux/Las quorum-sensing combination was experimentally identified and addressed by switching to the more orthogonal Lux/Cin architecture, while subsequent promoter leakiness was identified and mitigated by replacing pLux with pLuxLac.

Overall, I consider the revisions sufficient to address all the important concerns.

Reviewer #3 (Public review):

Summary:

This manuscript presents an engineered 3-step circuit in E. coli that combines toggle-switch-based symmetry breaking with quorum-sensing interactions to generate colony-scale spatial patterns. The work is interesting as a synthetic circuit integration study and as a demonstration of self-organized patterning across physically separated colonies. The authors provided a compelling demonstration of the characterization/tuning of parts to guide the overall system engineering. A notable strength is the demonstration that a single circuit can generate a range of self-organized spatial patterns across separate colonies.

Comments on revised version:

The authors have satisfactorily addressed my previously raised issues.

Author response:

The following is the authors’ response to the original reviews.

Reviewer #1 (Public review):

Summary:

This paper by Boni and colleagues presents the engineering of a multi-step differentiation program in Escherichia coli based on synthetic gene circuits. The motivation behind the study was to engineer a system capable of undergoing differentiation in a step-wise manner without the presence of external spatial cues and without inducers added during the differentiation process. To achieve this, the authors created several synthetic gene circuits, one being a toggle switch, and the others being quorum-sensing-mediated gene expression modules. The outputs of the differentiation process are fluorescent proteins, which allowed the authors to quantify the behavior of the system using fluorescence intensity measurements. The authors additionally built a multi-component mathematical model which is able to reproduce the experimental data.

The data presented are convincing and support the claims; the work is well executed.

Strengths:

(1) The differentiation process proceeds autonomously after the initial step in liquid culture in the presence of external inducers.

(2) It is indeed a step-wise process.

(3) The mathematical model predicts the outcome (% of green, blue and red FPexpressing cells in the population) when changing the initial ratio of green:blue FPexpressing cells.

We thank Reviewer #1 for the Summary and for highlighting the strengths of our work.

Weaknesses:

(1) No spatial pattern emerges. There are some isolated colonies that turn on the downstream FPs, but I do not see a pattern, really. Nonetheless, some colonies do differentiate (i.e. they turn on additional FPs).

The pattern does not arise within single colonies, but when looking at groups of colonies: a green sender is surrounded by a circle of blue-red receivers, which is in turn surrounded by blue-only receivers. This organization can be seen as a collective bullseye pattern (green centre, red annulus, blue background). When multiple green colonies are present in the plate, each gives rise to its own bullseye pattern. We have now clarified this detail in the text:

Lines 244-245 – “This motif is reminiscent of a bullseye pattern emerging not within a single colony, but in groups of colonies”

(2) The mathematical model appears somewhat superfluous. While it can clearly reproduce the data, it is not used to make interesting predictions, changing parameters (and not initial conditions) that guide further experimental implementations.

It is true that we presented the mathematical model primarily as being capable of recapitulating our experimental results. However, the model helped us identify the ideal set of initial conditions (i.e. inducers concentrations) for Figure 6, and to better understand the dynamics of 3O-C6-HSL and 3O-C14-HSL diffusion in our system (Supplementary Movie 6). While changing parameters would be theoretically possible (e.g. diffusion coefficients, protein production rates...) we have limited exploration in this direction, as fine-tuning a single molecular parameter, all other things being equal, is experimentally challenging.

Since several comments highlighted this weakness, we have now employed our mathematical model to predict patterns theoretically achievable with the sequential differentiation program under substantially different experimental conditions. Please find a detailed answer to this point below (see Reviewer #1 (Recommendations for the authors), point 8).

Future directions:

The utility of this differentiation process (e.g. in metabolic engineering or for the study of biofilm formation and antibiotic resistance) will become clearer once the FPs are substituted with functional proteins that exert an effect on the cells.

We agree with Reviewer #1 that, for the purposes of an application, we would have to replace the fluorescent proteins with functional proteins. In the last paragraph of the discussion, we outline several potential applications of our differentiation system, including division of labor, biocomputation, biosensing, and engineered living materials. However, expanding the system in this direction is beyond the scope of this manuscript.

Reviewer #1 (Recommendations for the authors):

(1) This sentence is confusing to me: ‘When cells were pre-cultured with 100 nM aTc and 1 mM IPTG, the resulting colonies were almost entirely homogeneously green and blue, respectively.’ It reads as if both inducers were added in the same culture. I would rewrite this as: ‘When cells were pre-cultured with either 100 nM aTc or 1 mM IPTG, the resulting colonies were almost entirely homogeneously green or blue, respectively.’

We adapted the text as suggested:

Lines 127-129 – “When cells were pre-cultured with either 100 nM aTc or 1 mM IPTG, the resulting colonies were almost entirely homogeneously green or blue, respectively.”

(2) It would be good to explain why the follow-up experiments are conducted with inducers at the beginning of the differentiation process, given that, in the absence of inducers, there is spontaneous symmetry breaking as noted by the authors: ‘When cells were pre-cultured in absence of inducers, both green and blue colonies grew in the plate (Figure 2b).’ Is it to obtain consistent results across biological replicates?

We used defined inducer concentrations to control the ratio of senders: receivers. In the absence of inducers, the population consisted of approximately equal proportions of senders and receivers. Under these conditions, nearly all blue colonies were close to a green colony and exhibited high expression of the red reporter. To enable spatial patterning, we therefore required a population strongly biased towards the receiver state. Accordingly, we used low concentrations of IPTG to shift the population to this state (as explained in lines 235-241 of the revised version). For testing and characterization of the system, we spotted senders alongside receivers (e.g. in Fig 7C and some supplementary figures). In these experiments, pre-culturing with high concentrations of aTc or IPTG ensured that the entire population was in the desired state. We have now revised the text to clarify that the choice of inducer concentrations ensured specific population ratios and the reproducibility of the differentiation assay:

Lines 235-241 – “We therefore selected a condition that would consistently generate a population strongly biased towards the receiver state, with only a few sparse senders to produce the diffusible signal. Having characterized the TS differentiation landscape (Figure 2), we decided to pre-culture cells starting from the green state in presence of 9, 12 and 18 µM IPTG, then plated at a cell density of 500-2000 colonies per plate, in absence of any positional information. This resulted in few sparse green senders densely surrounded by blue receivers, in a highly reproducible manner.”

(3) In the legend to Figure 4b, please indicate what ‘experimental points’ are (single colonies?)

We adapted the figure caption as suggested:

Figure 4b – “Hexagons, circles and triangles represent the average fluorescence intensity of single colonies.”

(4) This sentence was confusing to me: ‘The time-lapses supported our hypothesis that the differentiation key processes (colony growth, HSL production and detection, expression of the red reporter) happened simultaneously.’ I thought the whole idea behind the work was to have a step-wise differentiation process and that HSL production had to happen first, then its detection and finally the downstream activation of the red FP expression.

We acknowledge that this sentence was poorly phrased and may have caused confusion. Our multi-step differentiation program indeed operates in a sequential, stepwise manner at the single-cell level: a blue receiver cell can only transition to the red state after detecting 3O-C6-HSL, and red colonies can only transition to the yellow state after detecting 3O-C14-HSL. At the colony level, one might therefore expect colonies to first appear green or blue and only later acquire red fluorescence. However, our time-lapse experiments showed that colonies were already expressing the final fluorescent reporters by the time they became visible. Thus, our observations indicate that, at the colony level, HSL production and detection, induction of the red reporter, and colony growth occur concurrently over the same time period. Experiments at the single-cell level might make the step-wise progression blue - red - yellow more obvious than colony-level time-lapses. We have revised the text to clarify this point:

Lines 273-280 – “We therefore collected time-lapse videos of the plate assay, first inducing receiver cells with pure 3O-C6-HSL, then spotting sender cells alongside receivers, and finally performing the full differentiation assay (Supplementary Movies 1, 2, 3, 4). Whilst one might expect colonies to first appear green or blue and only later acquire red fluorescence, our time-lapse experiments showed that colonies were already expressing the final fluorescent reporters by the time they became visible. Thus, our observations indicate that, at the colony level, HSL production and detection, induction of the red reporter, and colony growth occur concurrently over the same time period.”

(5) The authors write: ‘The combination of spatial and temporal information allowed us to develop a mathematical model that could qualitatively recapitulate the patterning properties of the system’. It is strangely put in my opinion. One is ‘allowed’ to develop a mathematical model under other circumstances, too. I suggest rewriting. For example: ‘We developed a mathematical model combining spatial and temporal information to quantitatively ...’.

We adapted the text as suggested.

(6) In Figure 6b, the scale bar is missing

Thank you for noticing this, we have now added the scale bar.

(7) Figure 7c: Why is the green colony huge compared with the others? The formation of such huge colonies seems to occur sometimes, but not always (it is seen also in Supplementary Figure 12a and in Supplementary Figure 13, but nowhere else; by the way, in Supplementary Figure 12a there is a spatial pattern, with a clearly visible red ring in the otherwise black colony!).

The huge green colonies occur only in the experiments where 1 µl of green senders were inoculated at specific locations. In most of the differentiation assays, each colony arises from a single cell at unpredictable locations. For Figures 7c, S13a (previously 12a) and S14 (previously S13) we wanted a single green sender colony in the centre of the plate surrounded by blue receivers. To achieve this, as we briefly explained in the methods, we decided to spot 1 µl of culture at OD=1, therefore these green colonies originate from roughly 8×10⁵ cells, which justifies their larger size. We have now clarified this detail in the corresponding figure captions:

Figures 7c, S13a, S14 – “The sender colonies in this figure do not originate from one single cell, but from 1 µl of a culture of sender cells at OD=1, which explains their larger size (see Methods).”

Concerning the red rings in figure S13a (previously S12a), we hypothesise they are the effect of the leakiness of the pLux promoter. Observing expression of mCherry in the green sender colony was actually the main indication that the sequential differentiation circuit was not working as planned: if mCherry was being expressed in the sender state, then probably cinI was also being expressed, therefore the green sender colony was undesirably producing 3O-C14-HSL. We therefore replaced pLux with pLuxLac, resulting in tighter control over the mCherry-cinI operon in the green sender state. The improvement associated with this modification can be appreciated in Figure 7c, where the green sender colonies do not display any red fluorescence.

Regarding the specific pattern of this red signal, i.e. a ring, as opposed to an homogeneous signal in the entire colony, we have different hypotheses, but we have not investigated it thoroughly. Since the colony does not originate from a single cell but from a 1 µl inoculum, the effect might partially be ascribed to a ‘coffee ring stain’ phenomenon: upon absorption of the droplet, the outer ring could feature higher cell density compared to the centre, leading to higher red signal. In other cases, we have sometimes seen ring patterns arising in homogeneous colonies due to growth and temperature effects.

(8) It would be nice to see the model predictions being used to create a system with different properties than the actual one. Can the authors modify parameters such as diffusion of the quorum-sensing molecule or the gene expression response to it, and see how the output would change? Changing the type of quorum-sensing molecule experimentally is doable, as is the addition of, for instance, a delay module in the gene expression module. Nonetheless, I am aware that this would require some time to do, but it would, in my opinion, strengthen the paper.

We thank Reviewer #1 for the valuable suggestions. We have now employed the mathematical model to test the pattern resulting from various diffusion coefficients combinations and added Supplementary Figure S15 and a paragraph in the main text. Choosing diffusible signals with different diffusion coefficients would indeed allow to generate more complex patterns, including concentric rings where the expression of the red and yellow reporters does not overlap. While we agree that experimentally validating this point would be of considerable interest, we believe that changing diffusion rates is challenging and beyond the scope of a typical three-month revision, as it would most likely require substantial optimization and fine-tuning.

Concerning the delay module, experimentally it could be implemented by introducing an intermediate step with a transcription factor modulating the expression of the CinI synthase. While we agree it would technically be feasible, we did not implement it due to time constraints. We would expect the effect of a delay module to be the following: production of C14-HSL would be delayed, therefore the front wave of C14-HSL would be consistently lagging behind the front wave of C6-HSL. The distance between the two fronts would be proportional to the delay introduced by the module. We hypothesise this would generate patterns with small yellow circles inside larger red circles. Our simulations already capture these dynamics by varying diffusion coefficients, and given the absence of experimental data to constrain parameters, we decided not to include a hypothetical delay module in our mathematical model.

See Supplementary Figure S15

Lines 378-385 – “Our model suggests that varying the diffusion coefficients of the two diffusible signals could generate more complex patterns, particularly concentric rings where the expression of red and yellow reporters does not overlap (Supplementary Figure S15). Interestingly, the largest region of mCitrine reporter expression is predicted when the diffusion of the first signal is slow, while that of the second is fast. This suggests that sufficient local accumulation of the first signal is required to reach the threshold for production of the second signal, which can then spread further away to produce an outer blue-yellow region.”

Reviewer #2 (Public review):

In this manuscript, the authors implement a three-step genetic programme in E. coli that converts an initially homogeneous population into spatially structured sender, receiver, and ‘matured’ receiver colonies on agar without externally supplied positional information. They combine a TetR/LacI toggle switch for symmetry breaking, LuxI/LuxR quorum sensing for a paracrine signalling step, and CinI/CinR for an autocrine signalling-like maturation step, and complement the experiments with a mathematical model that qualitatively reproduces pattern formation over a range of initial conditions.

While the article has many strengths such as a clear conceptual framing using Waddington landscapes, a modular and carefully optimised circuit design, thorough experimental characterisation of the toggle and quorum-sensing modules, integration of spatial modelling with experiments, and generally clear writing and figures, I think it will benefit the article to clarify the definition and stability of ‘differentiated’ states, clarify several quantitative and modelling aspects, better explain how fitted curves and promoter engineering were done, and improve some figure design and wording to avoid ambiguity.

We thank Reviewer #2 for the summary and for highlighting the strengths of our work. Please find below our detailed responses addressing the weaknesses and gaps you identified.

Detailed comments below:

(1) P5-8 / and more generally: A major concern is that producing a reporter output is not, by itself, differentiation. For a state to be credibly called ‘differentiated’, it should be stable (self-maintained) over relevant timescales, ideally in the absence of the inducing context. As written, the manuscript sometimes seems to equate cell type with reporter expression. I strongly suggest adding a short subsection explicitly defining state versus output, and for each claimed state, stating whether it is stable/bistable or unstable/reversible, with evidence. Concretely, the authors should enumerate: a) Togglederived sender versus receiver: stable? under what conditions (inducer ranges, hysteresis window)? b) Paracrine-induced ‘red’ receivers: is this a stable differentiated state, or a context-dependent induction requiring proximity to senders? c) ‘Mature’ (yellow) state: does it persist after removal from the spatial signal field? If not, it should be described as an induced output programme rather than a mature lineage state.

At present, later sections (and the ‘maturation’ language) risk over-stating what is demonstrated.

We acknowledge Reviewer #2’s concern regarding the stability and irreversibility of our states and agree this is a limitation in our work. The hysteresis property of the toggle switch is well documented in previous studies (Litcofsky et. al., 2012; Barbier et. al., 2020), so we did not formally re-quantify it. However, the fact that pre-culturing cells without inducer yielded predictable green: blue ratios and very few colonies exhibiting both states indicates that the toggle switch is both irreversible and stable under our experimental conditions. For the second and third steps, the expression of the fluorescent proteins is maintained throughout the duration of the experiment and for several hours to days thereafter. It is indeed established that cells in stationary phase have protein half-lives on the scale of tens of hours (Wiechecki et. al., 2017; Gervais et. al., 2025). However, we agree that re-streaking these cells in the absence of sender cells would eventually lead to the cessation of fluorescent protein expression, making the quorum sensing-mediated differentiation steps reversible.

You can find further details on this point in our reply to Reviewer #3 (Public review), point 1.

We have now adjusted the wording throughout the manuscript, and in particular we modified the abstract to highlight that our system ‘mimics’ cell differentiation. We have also added sections to explain how red and yellow are reversible cellular programs and not stably differentiated cellular states:

See Abstract

Lines 92-104 – “In the present work, we aimed to address this gap by creating an autonomous multi-step program recapitulating cell differentiation in Escherichia coli. [...] Next, we activated a new molecular program in a subset of receivers that are in close proximity to a sender through intercellular communication, implemented via the quorum sensing (QS) system LuxI-LuxR. Finally, the newly emerged population underwent maturation by producing an autocrine signal via the quorum sensing system CinI-CinR.”

Lines 171-173 – “Taken together, these observations showed that, owing to the TS bistability, a group of initially undifferentiated cells bifurcated into one of the two possible states, which were stably maintained thanks to the TS hysteresis (Litcofsky et. al., 2012).”

Lines 255-266 – “It is crucial to distinguish that, while the blue/green identity associated with the toggle switch represents a true bistable state (Gardner et. al., 2000; Litcofsky et. al., 2012; Barbier et. al., 2020), the expression of the red reporter is reversible and dependent on proximity to the 3O-C6-HSL source. Although the QS ONOFF transition is considerably slower than the OFF-ON activation, reversibility remains a fundamental characteristic of QS (Abraham et. al., 2024). Furthermore, protein half-lives in stationary-phase cells extend over tens of hours (Wiechecki et. al., 2017; Gervais et. al., 2025), allowing for the stable detection of mCherry fluorescence over days. Consequently, mCherry serves as a robust readout for the molecular program’s output; however, a truly differentiated cell state would necessitate irreversible modifications to the gene expression profile. Therefore, in our landscape analogy we represented the receiver valley as a continuum where red intensity decreases as the distance from the sender valley increases, without local minima (Figure 1, 3rd row).”

Lines 337-342 – “Thanks to the production of an ‘autocrine signal’ that affects only the red cells, this population drifts apart from the blue receiver state, increasing the distance between the two cell types in the differentiation landscape (Figure 1, IV° row). Similarly to the 2nd step, the 3rd step recapitulates a differentiation trajectory via expression of a fluorescent reporter, which in this system is reversible and does not introduce irreversible modifications of the cell state.”

Lines 428-430 – “In this work, we successfully engineered a multistep program mimicking the differentiation of an initially homogeneous population into three distinct cell types without any external cues, while still achieving fine-tuning of the populations’ ratios through pre-culture conditions.”

(2) Figure 2d: It is unclear whether this panel is intended to be qualitative (schematic/ illustrative) or generated from quantitative data. The legend should explicitly state the origin (e.g., representative image, averaged data, simulation output, schematic) and, if quantitative, what was measured, how many replicates, and how the visualisation was constructed.

We acknowledge the potential confusion generated from this image. While its purpose is to visually illustrate the toggle switch differentiation landscape in 3D, the figure is derived from quantitative data, specifically from the same flow cytometry data used for Figure 2c. Basically, it is the density plot of the (GFP, mCerulean) events recorded for a single replicate (initial state: mixed, induced with 0.003 mM IPTG), but the density is represented with a third dimension (depth) instead of colour intensity (as we did for Figure 2e). We have now adjusted the figure caption to clarify the figure purpose and how it was generated:

Figure 2d – “Representative 3D energy landscape: valleys represent the green and blue stable states, red line represents the separatrix. The landscape was generated from quantitative flow cytometry data from a single replicate in c (initial state: mixed, induced with 0.003 mM IPTG). We measured GFP and mCerulean fluorescence intensity from 50,000 single cells (see Methods). The depth of each (x,y) point in the landscape corresponds to the number of recorded cells with a specific (GFP, mCerulean) intensity. This specific condition was chosen for illustrative reasons, to show the landscape in an almost symmetrical condition.”

(3) Figure 2e: The cross-sectional line is described as meant to be comparable, yet the leftmost plot appears to have a different slope from the others. The authors should explain whether this reflects a different scaling/normalisation, a different underlying dataset/condition, or simply a plotting artefact. If these are fitted trends, report the fit function (see also the comment on fitted lines below).

The observation is correct, the cross-sectional lines do not have the same slope. They are obtained by connecting the local minima of the green and blue population, i.e. the two points with highest density in the density plots. As the mean fluorescence intensity of the two population is not the same across conditions, the resulting lines have different slopes. We chose this strategy instead of using fixed-slope cross-sectional lines to capture the maximum depth of each valley. We have now adjusted the figure caption to clarify how the cross-sectional lines were generated:

Figure 2e – “Flow cytometry density plots of 5 representative conditions from c (initial state: mixed, inducer concentration indicated below each plot). For each density plot, we computed the coordinates of the local minima in the green and blue populations, then generated an orthogonal plane crossing them. Black lines represent the projections of these planes on the (GFP, mCerulean) plane. [...]”

(4) Around P7-8: (saddle/separatrix description): When describing the saddle or separatrix between the two valleys, it would be helpful to briefly connect this more directly to a quantitative dynamical-systems perspective: for instance, the intersection of nullclines and how nullcline geometry changes under IPTG/aTc induction. This will make the landscape picture more complete for readers familiar with the original genetic toggle switch work (Garder et al., 2000).

We thank Reviewer #2 for the suggestion. We have now included the concept of nullclines in our description of the differentiation landscape in the main text:

Lines 153-159 – “Mathematically, the profile of the toggle switch landscape corresponds to the number of intersections between the nullclines (the curve where the derivative of a given species over time equals zero), which can be one or three (Gardner et. al., 2000). If they intersect only once, there is only one minimum on the potential curve, which corresponds to a single stable state. If they intersect three times, there are two minima and one maximum, which corresponds to two stable states, and the maximum represents the separatrix.”

Lines 168-171 – “From a quantitative dynamical-systems perspective, the addition of aTc or IPTG modified the geometric shapes of the nullclines, shifting the three solutions. If the inducer concentration is large enough, the nullclines intersect only once, producing a single stable steady state (Gardner et. al., 2000).”

(5) P9, lines 157-159: The current phrasing (‘in absence of noise, the system would be fully deterministic... in living cells, however, stochastic bursts... change the trajectory’) risks conflating predicting population-level percentages with predicting colony-level trajectories. It would help to clearly separate (i) the ability to predict the overall fraction of ON/OFF (green/blue) colonies from inducer conditions (which is largely deterministic at the population level) from (ii) the intrinsically stochastic choice of state made by any given founder cell and its colony.

We acknowledge that this wording has generated confusion, but we are not completely sure we understood the suggestion made by Reviewer #2. If we understood correctly, the concern is that single-cell trajectories and population-level ratios are substantially different and should not be confused, and should be investigated and modelled differently.

To clarify, our initial sentence referred to the toggle switch differentiation landscape, not to the likelihood of a cell or a colony to be in a given state. In absence of noise, the TS landscape is deterministic: one of the two states is always the stronger attractor, for the same initial conditions cells would fall 100% of the times in that valley. In living cells, however, there are sources of noise, which make it possible for two cells starting from the same initial conditions to end up in two different valleys. The stochastic bursts in gene expression can ‘push cells back up’ and beyond the separatrix. Very large noise would allow cells to end up in the green and blue valley regardless of the TS previous state or the inducer concentration.

The differentiation induced by the toggle switch when the system is initiated close enough to the separatrix is stochastic. The probability of observing a certain ratio green: blue at the population level reflects exactly the likelihood of a single cell falling into one of the two stable states. Our mathematical model does not take into account molecular details (synthesis and degradation/dilution of proteins, affinity of transcription factors for their cognate promoter...) to mimic the stochastic choice at the cell level. Instead, it estimates the TS differentiation landscape from the population-level percentages based on thousands of single cells.

To avoid confusion, we have removed the initial sentence and replaced it with an explanation of the link between individual cell trajectories and the resulting population-level ratios:

Lines 159-164 – “When the system is close to the separatrix, cells can progress towards both the green and the blue destiny. The differentiation trajectory of a single cell is largely stochastic and is influenced by bursts of gene expression. Once the TS is locked in one state, the progeny of that cell maintains a memory of that state, hence a colony has the same state as the founder cell. At the population level, the ratio of green and blue colonies reflects the probability of each cell to fall into the green or the blue valley.”

(6) P11, lines 193-195 (promoter engineering): The main text currently only refers to screening variants and choosing pLux76; I suggest briefly stating in the main text (not only in the supplement) what was changed (for example, promoter box variants, core promoter strength modifications) and what design criteria were used (reduced leakiness, increased dynamic range).

We adapted the text as suggested:

Lines 206-211 – we carried out a screening of pLux promoter variants, testing combinations of Lux boxes (G1 [iGem part BBa K1216007] and pLux76 (Grant et. al., 2016) and promoters with different strengths (100% and 54%), in order to identify variants with minimal leakiness and a high fold-change. We identified pLux76 (Grant et. al., 2016) as the regulatory region with the highest fold-change and sufficiently low leakiness among our candidates (Supplementary Figure S4)”

(7) Use of fitted lines (Figures 2, 4, 5, 7): Wherever fitted curves are overlaid on data, the authors should indicate in the figure legend the explicit form of the fit as well as the fit equation/ parameters. As a reader, it is difficult to interpret what is empirical smoothing versus what is a mechanistic functional form.

In most cases, with the exception of Figure 2c, the fitted curves have an illustrative purpose and result from empirical smoothing of the data points. Unless otherwise stated, the resulting parameters were not implemented in our mathematical model. For Figure 2c, instead, the experimental data is fitted with a mechanistic functional equation and the calculated parameters were implemented in our mathematical model.

We have now added the equations and parameters of the lines fitting the experimental points, either directly in the figure caption or in Supplementary Information Tables (Tables VII to XII). In the latter case, the exact table is referenced in the corresponding figure caption.

(8) P13, lines 232-235: The comparison between induction directly with C6-HSL and induction from sender colonies is qualitative (‘significantly smaller range’). The authors should provide distances (for example, in mm) for the induction range in each case and, if possible, approximate total HSL amounts or concentrations, so that the reader can appreciate the magnitude of the difference.

We have now calculated the induction range as the distance where half-maximal induction is observed, which allowed us to compare the induction range across conditions. We adapted the text accordingly. We also provide an estimate of the amount of 3O-C6-HSL produced by a green sender colony, based on simulations of our mathematical model, but we highlight that the two conditions are substantially different and care should be used when comparing them: in one case, a fixed amount of C6-HSL is present from t=0 and simply diffuses outwards; in the second case, a source continuously produces C6-HSL that progresses as a wavefront, therefore the concentration profiles and diffusion dynamics are different.

Lines 220-226 – “The induction range obtained with 100 picomoles of pure 3O-C6-HSL was approximately 6-8 mm (50% of maximal induction at 3.74 mm). The induction range around sender colonies was significantly smaller (50% of maximal induction at 1.8 mm, Figure 4b). Mathematical simulations yielded a similar slope when using 0.25 picomoles of 3O-C6-HSL (data not shown), even though care should be used when comparing diffusion of a fixed amount of inducer with continuous production from a growing source.”

(9) P13, lines 259-262: The authors model the transition to the stationary phase via a monotonically decreasing sigmoid in time for biosynthetic capacity. What is the rationale or literature basis for this approach to model entry into the stationary phase? The authors should cite prior work and clarify why this form is appropriate here, versus alternatives (nutrient diffusion limitation, logistic growth with resource depletion, etc.).

In most mathematical models involving pattern formation, entry of cells in stationary phase is simply treated as a step-wise function (Zwietering et al., 1990): the entire population is active (activity = 1) until it suddenly becomes inactive (activity = 0). However, experimental evidence suggests that the decay in activity is better represented by a smooth decreasing function (Gefen et al., 2014). While several frameworks and mathematical equations exist to describe loss of cell viability (for example under scenarios of heat inactivation (Mafart et al., 2002; Van Boekel et al., 2002), or nutrient limitation), we could not find previous works modeling the loss of metabolic activity upon entry in stationary phase. Again, experimental evidence suggest that switches in metabolic state are rather heterogeneous and occur stochastically at the single cell level (Nikolic et al., 2013; Kiviet et al., 2014; Van Heerden et al., 2014). We therefore decided to implement a simple heuristic approximation that captured well our experimental data. We have now added a paragraph in the section Mathematical modelling - Bacterial activity, supported by the appropriate references, to justify our reasoning:

Supplementary Material, section A. “Mathematical modelling, subsection 4. Bacterial activity – In our system, production of diffusible molecules and fluorescent reporters happens on a timescale of several hours, therefore we decided to take into account the entry of cells in stationary phase. Traditionally, bacterial activity has been modelled as a step-wise inactivation function (Zwietering et al., 1990). However, experimental evidence has shown that bacteria support a low constant rate of protein expression even while growth-arrested, suggesting a low decay in bacterial activity (Gefen et al., 2014). Even when considering cell death upon heat inactivation, a Weibull frequency distribution model is preferred over a step-wise viability function (Mafart et al., 2002; Van Boekel et al., 2002). While we could not find mathematical descriptions specifically for entry in stationary phase, several single-cell metabolism studies support gradual, asynchronous state transitions (Nikolic et al., 2013; Kiviet et al., 2014; Van Heerden et al., 2014). We therefore decided to describe the loss of activity function in our system via a monotonically decreasing sigmoid function, that captured well our experimental data.”

(10) Figure 6c: Are the areas of the plate shown in each column the same field of view across conditions/time, or are these simply representative regions selected per condition (possibly from different plates)? The caption/legend should clarify whether these are matched locations and how images were chosen.

They are indeed representative regions per each condition. Each condition is a different plate, as the differentiation assays requires plating the culture homogeneously on a fresh plate without inducers. The plates used for imaging were the same used for the quantification in Figure 6b, four images per plate were collected, one representative image was chosen for Figure 6c. We have now adapted the figure caption to clarify how images were collected and chosen:

Figure 6c – “Representative microscopy images of the spatial patterns generated by cells harbouring the 2-step system, pre-cultured with different inducer concentrations (indicated at the bottom of each column). Each column displays one representative image (from four locations imaged) of the seven plates in b. Rows (from top to bottom): GFP channel, CFP channel, mCherry channel, composite image.”

(11) Figure 7a: The combination of solid, dashed, and dash-dot arrows/lines is visually hard to read. I suggest replacing the dash-dot line with a fully dotted line or using different colours (if consistent with journal style) to improve readability.

Thank you for noticing this. We have now replaced the dash-dot line with a simple dotted line in Figures 3c, 7a, S12a (previously S11a) and S13b (previously S12b) to improve readability.

(12) Figure 7e and similar analyses: The authors should explain in the Methods and/or captions how ‘distance from sender colonies’ is computed when multiple senders exist. Is the distance always measured to the nearest sender, and how are cases handled where a receiver is in the overlapping influence of several senders? This clarification is important for interpreting the fitted curves.

The calculation of the distance in 4b, 5b and 7e was performed only in cases where sender colonies were sufficiently sparse (i.e., more than 1 cm away) to assume each receiver was under the influence of a single sender. When collecting the microscopy images, we carefully avoided to image fields of view with green senders just outside the edges of the images. Images with multiple senders (e.g., 7c) would require a non-trivial calculation to compute the relative contribution of each sender to the red intensity of each receiver. We have now clarified this detail in the Methods:

Lines 558-560 – “When collecting microscopy images with senders and receivers, we carefully avoided to image fields of view with green senders just outside the edges of the images.”

Lines 590-593 – “Calculation of the distance between receiver and sender colonies was performed only for images collected from plates with few sparse (i.e., more than 1 cm away) senders, ensuring that each receiver only sensed the 3O-C6-HSL produced by a single sender colony.”

Reviewer #2 (Recommendations for the authors):

(1) P8, ‘upon transformation’: This phrasing is ambiguous and can be misread as referring to DNA transformation. I recommend changing ‘upon transformation’ to ‘after transition’ (or similar) to avoid confusion.

The phrasing indeed refers to the DNA transformation of circuits into the cells. The two plasmids were co-transformed into MG1655, cells recovered for approximately 1 h, then the bacteria were plated on solid medium in absence of inducers. The resulting colonies, which we refer to as ‘upon transformation’, were a mixture of green and blue (both expressed at low intensity, as highlighted in Supplementary Figure 2). We only used these colonies for Figure 2c, middle row. For all other experiments, we selected colonies that had been pre-differentiated in one of the two states via chemical inducers, and showed strong expression of the respective reporter (Supplementary Figure 2). We have now adapted the figure caption to make this detail more explicit:

Figure 2c – “For the initial state green and blue, cells were taken from colonies that were homogeneously green or blue, while for the initial state mixed, cells were taken from a colony obtained immediately after transformation of the circuit plasmids into cells.”

(2) More generally, consider disambiguating terminology around ‘cell type’, ‘state’, and ‘output’, since the current wording occasionally implies stable fate commitment where the data (as presented) may instead support reversible, context-driven induction.

We went carefully through the text and adapted it appropriately, highlighting which states are stable and which are reversible. See our detailed reply to your point 1 (Public review) above.

Reviewer #3 (Public review):

This manuscript presents an engineered 3-step circuit in E. coli that combines toggleswitch-based symmetry breaking with quorum-sensing interactions to generate colonyscale spatial patterns. The work is interesting as a synthetic circuit integration study and as a demonstration of self-organized patterning across physically separated colonies. The authors provided a compelling demonstration of the characterization/tuning of parts to guide the overall system engineering. A notable strength is the demonstration that a single circuit can generate a range of self-organized spatial patterns across separate colonies.

However, I think the paper needs to tone down the extent to which the system demonstrates multi-step differentiation or morphogenesis, which is not critical for making the paper valuable. Only the first step of their circuit design (Figure 1), the toggle switch, generates stable alternative states. The latter steps are mainly signal-dependent reporter activation states layered on top of the blue receiver state, rather than true fate transitions. The authors explicitly state that red expression is added without replacing the blue identity, and they also acknowledge that red cells lose their identity upon restreaking unless they remain near sender cells. That substantially weakens the differentiation analogy and makes the Waddington framing too strong.

We acknowledge that our sequential program does not recapitulate all the features of a differentiation trajectory, and in particular we do recognize the only two stable states are the identities associated with the toggle switch, while red and yellow are outputs indicating activation of a new molecular program. A complete differentiation program would require irreversible cell fate determination, as we mention in the discussion. For the same reason, in our illustrative Waddington landscape (Figure 1) we only represent two valleys, or minima, corresponding to the green and blue states, while the activation of the red and yellow programs does not result in further valleys. We modified the text at various points to clarify this important distinction, and the fact that our system mimics some key steps happening during multicellular differentiation, without claiming that expression of a fluorescent reporter is a stably differentiated state. Notably, we modified the abstract to highlight that our system ‘mimics’ cell differentiation.

We believe the differentiation analogy and the Waddington framing remain valuable in our work, considering the physical constraints and timescale of our system. While it is true HSL-induced molecular program would eventually deactivate in absence of signal, this would require a significantly longer amount of time than the one we use to observe patterns. Also, upon storage of Petri dishes at 4 °C, the pattern (including red and yellow) remains stable for a few weeks. Finally, our main purpose was to investigate the capacity of a sequential program to generate autonomous spatial patterns, without the need for human intervention. The removal of a colony from the local signal field, e.g. to re-streak it on a fresh plate, represents a strong disturbance of the pattern, not dissimilar from early developmental biology experiments where transplant of tissue portions would sometimes result in fate reprogramming.

You can find further details on this point in our reply to Reviewer #2 (Public review), point 1.

See Abstract

Lines 92-104 – “In the present work, we aimed to address this gap by creating an autonomous multi-step program recapitulating cell differentiation in Escherichia coli. [...] Next, we activated a new molecular program in a subset of receivers that are in close proximity to a sender through intercellular communication, implemented via the quorum sensing (QS) system LuxI-LuxR. Finally, the newly emerged population underwent maturation by producing an autocrine signal via the quorum sensing system CinI-CinR.”

Lines 171-173 – “Taken together, these observations showed that, owing to the TS bistability, a group of initially undifferentiated cells bifurcated into one of the two possible states, which were stably maintained thanks to the TS hysteresis (Litcofsky et. al., 2012).”

Lines 255-266 – “It is crucial to distinguish that, while the blue/green identity associated with the toggle switch represents a true bistable state (Gardner et. al., 2000; Litcofsky et. al., 2012; Barbier et. al., 2020), the expression of the red reporter is reversible and dependent on proximity to the 3O-C6-HSL source. Although the QS ON-OFF transition is considerably slower than the OFF-ON activation, reversibility remains a fundamental characteristic of QS (Abraham et. al., 2024). Furthermore, protein half-lives in stationary-phase cells extend over tens of hours (Wiechecki et. al., 2017; Gervais et. al., 2025), allowing for the stable detection of mCherry fluorescence over days. Consequently, mCherry serves as a robust readout for the molecular program’s output; however, a truly differentiated cell state would necessitate irreversible modifications to the gene expression profile. Therefore, in our landscape analogy we represented the receiver valley as a continuum where red intensity decreases as the distance from the sender valley increases, without local minima (Figure 1, 3rd row).”

Lines 337-342 – “Thanks to the production of an ‘autocrine signal’ that affects only the red cells, this population drifts apart from the blue receiver state, increasing the distance between the two cell types in the differentiation landscape (Figure 1, IV° row). Similarly to the 2nd step, the 3rd step recapitulates a differentiation trajectory via expression of a fluorescent reporter, which in this system is reversible and does not introduce irreversible modifications of the cell state.”

Lines 428-430 – “In this work, we successfully engineered a multistep program mimicking the differentiation of an initially homogeneous population into three distinct cell types without any external cues, while still achieving fine-tuning of the populations’ ratios through pre-culture conditions.”

A related concern is that the 3rd step does not introduce a new spatial organizing rule. The authors show that the second signal remains confined to cells already receiving the first signal, and explicitly conclude that it functions only as an autocrine cue rather than a second paracrine layer. As a result, the 3-step system seems more like an added local readout or maturation layer. Overall, the main 2-step outcome is sparse green sender colonies surrounded by red-expressing blue receivers, with distant receivers remaining blue. That is a valid engineered pattern, but it is still a local, threshold-response circuit architecture.

The comment is correct, we indeed refer to the 3rd step as ‘maturation’ in the manuscript, as the newly emerged red population activates a new molecular program, which cannot be activated in blue receivers that have not been exposed to C6-HSL. While we had initially expected that production of a second diffusible signal could result in signal propagation, therefore generating an outer yellow ring surrounding the red ring, our experiments showed C14-HSL only acted locally, and our mathematical simulations confirmed that the front wave of the second signal was always lagging behind the first one. We have now employed our mathematical model to explore which conditions would support a new patterning rule, and added Supplementary Figure S15. If the diffusion coefficients of C6-HSL and C14-HSL were 10-100 times smaller and 2-5 times larger respectively, a yellow-only area could appear around the red area.

Regarding the last sentence, we are unsure about the concern raised and which alternatives Reviewer #3 would recommend. We fully agree our genetic program leverages a local, threshold-response circuit architecture, it is indeed a reaction-diffusion system based on quorum sensing signalling. Many natural patterning and morphogenesis systems do rely on local, rather than global, interactions, yet they are capable of forming complex and hierarchical structures.

The autonomy claim should be toned down and stated more precisely. The plate patterning occurs without externally imposed spatial gradients, which is a strength. However, by design, the overall system behavior depends strongly on pre-culture inducer conditions that set the sender:receiver ratio, and this externally imposed history is central to the final pattern. This property is tied to how the circuit is designed where steps 2 and 3 largely respond to symmetry breaking introduced in step 1, which is dependent on both history and initialization on the plate. In particular, currently the pattern formation process is quite variable (e.g. figure 5), depending on how different colonies flip the toggle switch, and consequently, how many become senders and how many become receivers. It would have been fascinating if they could also demonstrate the differentiation within individual colonies, leading to intra-colony patterns. This aspect should at least be discussed.

We would like to clarify that, by ‘autonomous’ we refer to the reaction-diffusion system that starts functioning upon the seeding of the cells on the plate. All previous steps (including the cell culturing, dilution and plating) correspond to setting the initial conditions. We agree with the description of Reviewer #3 concerning the features of our system, but we argue that autonomy and dependence on the initial conditions are two different properties. We claim our system is both autonomous and sensitive to initial conditions. Upon initialization of the system on the plate, there is no further intervention, or nudging (e.g. time-dependent light stimuli, addition or removal of inducers at specific times...), therefore the system is autonomous. It is also dependent on the initial conditions, which are set homogeneously for all the cells (no positional information, each cell experiences the exact same conditions). We believe this is a strength of our system, which allows to generate a variety of diverse yet reproducible spatial patterns. Independency from initial conditions would consistently generate the same pattern, regardless of the initial inducer concentration, which might be interesting for certain applications (e.g., ensuring a fixed population ratio, provide robustness to environmental variability...), but it is not universally superior.

In order to clarify that we only refer to the reaction-diffusion system on the plate as the autonomous component, we have now added a sentence in the main text:

Lines 132-134 – “In our differentiation assay, cell culturing, dilution and plating correspond to setting the initial conditions for the system. Upon initialisation on the plate, no further intervention was performed, therefore the system evolved in an autonomous fashion.”

Concerning the intra-colony patterns, that was admittedly our initial interest. We explored conditions that would consistently generate colonies with sectors, for example by inducing cells in one of the two states and then growing them on agar supplemented with the opposite inducer. However, we immediately observed that, due to the close proximity of sender and receiver bacteria, and the rapid diffusion relative to the timescale of gene expression, all blue sectors were also red. This outcome effectively eliminated the distance-dependent nature of the quorum sensing response. In order to take full advantage of the diffusible system, we focused instead on well-separated homogeneous colonies. We have now added Supplementary Figure 5, the corresponding figure caption, and we briefly discuss in the main text the occurrence of intra-colony patterns and why we did not investigate them further:

See Supplementary Figure S5

Lines 228-231 – “We then tested the potential of the 2-step differentiation system to generate self-organized spatial patterns. We observed rare motifs arising within the sporadic colonies showing both green and blue sectors. Due to the close proximity of the sender and receiver bacteria within a colony, the blue sectors always showed strong red signal (Supplementary Figure S5).”

The mathematical model is useful in guiding both the characterization of parts, modules and the overall system. However, the claims around its quantitative predictive power should also be made narrower. The simulations are built from multiple fitted and partly hand-tuned components, including toggle-switch response curves, colony-growth rules, diffusion, reporter-response functions, and activity decline. This supports a calibrated qualitative reconstruction of the observed patterns, but not a strong predictive or mechanistic validation.

We accepted the suggestion and replaced ‘predicted’ with ‘recapitulated’ or ‘simulated’ at various locations in the main text:

Lines 105-107 – “Throughout the work, experimental results were used to develop a mathematical model, providing us with insights that guided further experimental efforts.”

Lines 292-293 – “We developed a mathematical model combining spatial and temporal information to qualitatively recapitulate the patterning properties of the system.”

Lines 316-318 – “We therefore employed the mathematical model to simulate the patterns generated from the 2-step differentiation system for different initial blue: green ratios. The model suggested a variety of outcomes [...]”

Other specific points:

(1) Given the topic of the work, the authors should cite closely relevant studies in programming pattern formation, including: Cao et al, Cell 2016 Collective space-sensing coordinates pattern scaling in engineered bacteria Rajasekaran et al, Cell 2024 A programmable reaction-diffusion system for spatiotemporal cell signaling circuit design Lu et al, BioRxiv 2024 Discovery of interpretable patterning rules by integrating mechanistic modeling and deep learning

We have added these references to the introduction and discussion.

(2) The model assumes identical diffusion coefficients for C6-HSL and C14-HSL despite their substantially different molecular sizes and hydrophobicities. This assumption could distort kinetic lag with differential diffusion in explaining the autocrine confinement of the third step. Its impact should at least be explored in the simulations.

The comment is very appropriate. To the best of our knowledge, no exact values have been published for C6-HSL and C14-HSL diffusion in water, let alone for diffusion in agar. However, estimates in water range between 3 ·10⁻⁶ cm/s and 5 · 10⁻⁶ cm/s at 25°C, depending on the chain length (i.e. a factor of 1.7 at most). In response to your concern, we have now computationally explored the effect of using signals with different diffusion coefficients and their impact on the resulting spatial pattern (Supplementary Figure S15). Varying the diffusion coefficient of the second diffusible signal does not substantially change the pattern: lower diffusion leads to the yellow region being slightly more concentrated around the green sender and more intense. Higher diffusion has the opposite effect, with yellow areas being wider but less intense. Conversely, changing the diffusion coefficient of the first signal leads to significant changes in the final pattern, suggesting that the kinetics of C6-HSL accumulation and dispersal have a strong influence on the production of C14-HSL and therefore of the yellow signal.

(3) The mCherry response parameters change significantly between the 2-step and 3-step systems. The authors acknowledged this change but did not provide a clear explanation.

We believe that the elements contributing to the different mCherry response in the 2-step and 3-step systems are: a) the replacement of pLux76 with pLuxLac, which leads to tighter regulation and overall reduced expression of the downstream genes; b) the addition of the cinI synthase and a RBS in a monocistronic unit upstream of mCherry, which might justify reduced expression of mCherry; c) circuit-induced cell burden and reduced growth rate for the 3-step system, which leads to a slower response and lower reporter expression. We have now added a paragraph in the section Mathematical modelling Bacterial activity that lists those differences. We have also added Supplementary Figure S20, containing the experimental data that was fitted to obtain the parameters listed in the second row of Table II. Finally, we highlight that for the implementation of our mathematical model we were only interested in the response function to varying 3O-C6-HSL concentrations and not in the exact values of individual parameters.

Supplementary Material, section A. “Mathematical modelling, subsection 3. mCherry production and 30-C6-HSL sensing – The parameters fitted for mCherry production in the 2-step and 3-step systems are quite different. We hypothesize that the following elements contribute to the difference: a) the replacement of pLux76 with pLuxLac, which leads to tighter regulation and overall reduced expression of the downstream genes; b) the addition of the cinI synthase and a RBS in a monocistronic unit upstream of mCherry, which might justify reduced expression of mCherry; c) circuit-induced cell burden and reduced growth rate for the 3-step system, which leads to a slower response and lower reporter expression.”

Supplementary Material, section A. “Mathematical modelling, subsection 4. Bacterial activity – For the implementation of our mathematical model, we were not interested in analysing individual parameters values, but rather in the resulting response function to varying 3O-C6-HSL concentrations. Further analysis would be needed to determine whether these parameters are statistically significant, but this is beyond the scope of this present manuscript.”

(4) The 3-step system is evaluated at only a single condition with no simulation comparison, in contrast to the systematic 11-condition validation of the 2-step system.

The rationale for not repeating the 11-condition assay with the 3-step system is that the patterns would be substantially the same as Figure 6c (red colonies would also be yellow). Furthermore, the Nikon SMZ25 stereo microscope we used to take images with a large field of view (ideal for the differentiation assay in Figure 6c) would not allow us to discriminate well between green and yellow colonies, making the interpretation of the patterns difficult. However, following up on your comment, we have now used our mathematical model to produce a representative image corresponding to Figure 6d, that we included as Supplementary Figure S16, exploring the effect of varying the green: blue ratio with the 3-step system.

Reviewer #3 (Recommendations for the authors):

Please see above. My comments are largely about improving the rigor and clarity of the writing, particularly those related to conceptual claims, as well as some modeling analysis to strengthen their conclusions.

We thank Reviewer #3 for the suggestions. See our detailed replies to your points above.