Multicellular Feedback Control Strategies in Synthetic Microbial Consortia: From Embedded to Distributed Control

Multicellular F eedbac k Con trol Strategies in Syn thetic Microbial Consortia: F rom Em b edded to Distributed Con trol Mario di Bernardo Departmen t of Electrical Engineering and Information T ec hnology Univ ersity of Naples F ederico I I, 80125 Naples, Italy Scuola Sup eriore Meridionale, Naples, Italy mario.dibernardo@unina.it Abstract Living organisms rely on endogenous feedback mec hanisms to maintain homeostasis in the presence of uncertain ty and en vironmental fluctuations. An emerging c hallenge at the in- terface of con trol systems engineering and syn thetic biology is the design of reliable feedbac k strategies to regulate cellular b eha vior and collectiv e biological functions. In this article, w e review recent adv ances in multicellular feedbac k control, where sensing, computation, and actuation are distributed across different cell p opulations within syn thetic microbial consor- tia, giving rise to biological multiagen t con trol systems gov erned b y molecular communica- tion. F rom a control-theoretic p ersp ectiv e, these consortia can b e in terpreted as distributed biomolecular con trol systems, where co ordination among populations replace embedded reg- ulation. W e survey theoretical frameworks, control architectures, and modeling approac hes, ranging from aggregate p opulation-lev el dynamics to spatially a w are agen t-based sim ula- tions, and discuss exp erimen tal demonstrations in engineered Escherichia c oli consortia. W e highlight ho w distributing control functions across p opulations can reduce metab olic burden, mitigate retroactivit y , impro ve robustness to uncertaint y , and enable modular reuse of control components. Beyond regulation of gene expression, we discuss the emerging prob- lem of p opulation comp osition control, where co ordination among growing and comp eting cell p opulations b ecomes an integral part of the control ob jective. Finally , we outline key op en challenges that m ust b e addressed b efore m ulticellular con trol strategies can b e de- plo y ed in real-w orld applications suc h as biomanufacturing, environmen tal remediation, and therap eutic systems. These c hallenges span mo deling and simulati on, experimental platform dev elopmen t, coordination and comp osition control, and long-term ev olutionary stability . Keyw ords: cyb ergenetics, syn thetic biology , multicellular feedbac k control, microbial consor- tia, biomolecular con trollers, quorum sensing, agent-based modeling 1 In tro duction The engineering of living systems represents one of the most comp elling fron tiers where control theory meets biological implementation [1–4]. While traditional control applications often b en- efit from well-c haracterized ph ysical mo dels, biological systems presen t distinct challenges as they are inherently sto c hastic, op erate at molecular scales where deterministic approximations b ecome questionable, gro w and evolv e ov er time, and m ust function robustly despite parame- ter uncertaint y and environmen tal p erturbations that substantially exceed t ypical engineering scenarios. These characteristics demand b oth new theoretical framew orks and nov el imple- men tation strategies that extend classical control paradigms while resp ecting the constraints imp osed b y living matter. 1 Y et these same challenges rev eal a profound opp ortunity . Living organisms themselves are a pro of that robust con trol at molecular scales is ac hiev able, as they can main tain homeostasis through sophisticated feedback mec hanisms that regulate in ternal states despite p erturbations and uncertain ties. F rom hormone secretion and signaling path w ays in m ulticellular organisms to bacterial chemotaxis, naturally ev olved negative feedback lo ops accomplish what app ears daunt- ing from an engineering persp ective achieving reliable regulation amid extreme sto chasticit y and parameter v ariation. This natural precedent has inspired syn thetic biologists to engineer con- trol systems within living cells [4–7], op ening new p ossibilities for applications ranging from optimized biopro duction [8] to targeted therap eutic deliv ery [9].The cen tral challenge lies in translating natural regulatory principles into rationally designed implemen tations that are b oth exp erimen tally tractable and suitable for deplo yment in applications. Con trol strategies for synthetically engineered biological systems go from fully external to fully embedded implemen tations. External control employs computer algorithms interfaced with cells through sensors (e.g. fluorescence microscopy) and actuators (e.g. c hemical inducers, op- togenetics), providing excellen t p erformance and flexibilit y for lab oratory applications [10–12]. Ho wev er, this approach requires sp ecialized hardware, contin uous monitoring, and precise envi- ronmen tal control, making it unsuitable for autonomous applications such as in vivo therap eu- tics or industrial bioreactors where precise real-time monitoring and external interv entions are impractical [13]. Em b edded biomolecular controllers implemen t complete feedback lo ops within cells using syn thetic gene regulatory net works, achieving autonomous op eration without external hardw are. Notable ac hiev ements include an tithetic in tegral feedback controllers, now a common motif in the field, that guaran tee robust p erfect adaptation through molecular sequestration [14, 15], biomolecular PID con trollers implementing prop ortional, in tegral, and deriv ative actions [16, 17], and v arious circuits demonstrating precise gene expression regulation in bacteria and mammalian cells [18, 19]. While single-cell embedded con trollers ha v e demonstrated impressive capabilities, they face three fundamen tal arc hitectural limitations arising from cen tralizing all con trol functions within individual cells. First, metab olic burden emerges when synthetic circuits consume cellular re- sources (rib osomes, RNA p olymerase, A TP , and metab olic in termediates) comp eting with en- dogenous pro cesses [20, 21]. This burden scales with circuit complexity , slowing growth and creating selective pressure for loss-of-function m utations that disrupt the controller but con- fer growth adv an tages. Second, retroactivity o ccurs when biological mo dules are connected, with downstream comp onents affecting upstream behavior through loading effects suc h as ribo- some or polymerase comp etition [22, 23]. This violates the mo dularit y assumption essen tial for plug-and-pla y design, making it difficult to predict assem bled circuit p erformance from the c har- acterisation of its individual components. Third, po or mo dularity emerges because once cells are engineered with embedded con trollers, any c hange in con trol strategy requires a complete system re-engineering, limiting adaptability and reuse of existing designs. More broadly , there is curren tly no systematic metho dology to translate abstract control laws into implementable bio c hemical reactions, making biomolecular controller design a largely ad-ho c process regardless of the c hosen architecture. T o address both the architectural and ecological limitations of single-cell implementations a promising strategy is distributing regulatory and control functionalities across distinct cell p opulations within microbial consortia. In [24], w e introduced the concept of multic el lular fe e db ack c ontr ol , in whic h sensing, computation, and actuation are distributed across dedi- cated p opulations. This approach treats syn thetic consortia as biological m ultiagent systems, a p ersp ective we develop formally in Section 2, where control functions are distributed across p opulations rather than concentrated within individual cells. P opulations communicate via dif- fusible molecular signals and collectiv ely implement closed-lo op regulation of a target pro cess. This architecture reduces metab olic burden p er cell, mitigates retroactivity betw een controller 2 and plant comp onen ts, and enables mo dular reuse of controller p opulations across different applications [24–26]. Ho wev er, distributing con trol across populations introduces orc hestration challenges absent in single-cell implementations. Ho w can appropriate population comp ositions b e maintained o ver time despite differential gro wth rates arising from metab olic burden and resource comp e- tition [27–34]? Ho w can comm unication proto cols based on quorum sensing remain effective in the presence of spatial gradien ts, diffusion dela ys, and signal degradation [35]? What con- trol arc hitectures can ensure reliable collectiv e b eha viour despite cell-to-cell v ariability and the absence of explicit error correction in molecular signalling [36, 37]? These challenges arise b ecause microbial consortia constitute a distinct class of multiagen t systems, characterised by sto chastic dynamics, con tinuous p opulation gro wth, and communica- tion through analog molecular diffusion. These prop erties differ substantially from traditional m ultiagent systems in rob otics or distributed computing [38]. Addressing them requires con trol strategies that op erate across molecular, cellular, and p opulation scales sim ultaneously . This review synthesizes theoretical foundations, design principles, computational mo deling approac hes, and exp erimental implementation strategies for multicellular feedback control in syn thetic bacterial consortia, with emphasis on co ordination c hallenges unique to biological m ultiagent systems. W e examine ho w distributing sensing, computation, and actuation func- tions across p opulations addresses architectural limitations while in tro ducing new orchestration problems requiring explicit management. Drawing extensively on recent computational [24] and exp erimental [39] demonstrations as illustrative case studies, w e identify critical op en c hal- lenges in ev olutionary stability , scalabilit y , and biological realizability that m ust b e addressed to translate lab oratory demonstrations in to industrial and therap eutic applications. The re- view is organized as follo ws. Section 2 presen ts design principles for distributed multicellular arc hitectures, including quorum sensing communication channels and the fundamen tal t w o- p opulation controller framework, with explicit comparison to traditional multiagen t systems. Section 3 examines computational v alidation approaches spanning aggregate population mo dels and agent-based simulations, follow ed by exp erimental v alidation in engineered E. c oli consortia. Section 4 extends the basic arc hitecture to the m ulticellular PID con troller family . Section 5 addresses the challenge of maintaining stable p opulation ratios. Finally , Section 6 identifies outstanding researc h challenges and concludes with persp ectives on future directions. 2 Multicellular Control Arc hitectures The design of a m ulticellular feedback control system requires engineering three interdependent comp onen ts (see Fig. 1a): the gene regulatory netw orks (GRNs) embedded within each pop- ulation enco ding the desired functions, the comm unication c hannels that enable information exc hange b et ween p opulations, and the consortium comp osition that determines the relative abundance of eac h p opulation. The GRNs define the internal dynamics of each cell t yp e, encoding sensing, computation, or actuation functions dep ending on the p opulation’s role in the con trol arc hitecture. In bacterial systems, comm unication betw een populations is t ypically mediated b y diffusible quorum sensing (QS) molecules, which act as molecular signals carrying information ab out cellular states across the consortium. Finally , the composition of the consortium, that is, the ratio of controller to target cells, directly affects closed-lo op p erformance, as it determines the effectiveness of b oth the actuation and feedbac k pathw ays. These three components are tigh tly coupled. The GRNs pro duce and resp ond to QS signals, the communication dynamics dep end on p opulation densities, and the comp osition evolv es ov er time due to differential gro wth rates that may themselv es dep end on the me tab olic burden imp osed by the syn thetic circuits. A successful multicellular con trol design must therefore address all three asp ects in an in tegrated manner. In the follo wing, we first present the tw o- 3 Figure 1: Multicellular control conceptual framework. (A) Key ingredients of a cellular consortium. (B) Tw o-p opulation arc hitecture in which a Contr ol ler cell p opulation regulates the expression of a gene within a T ar get p opulation through quorum sensing (QS)–mediated comm unication, forming a closed feedbac k loop (repro duced from [25]). (C) F our-population arc hitecture implemen ting a distributed PID con trol, with separate cell p opulations resp onsible for prop ortional, integral, and deriv ative actions whose com bined signals regulate the target population, enabling robust and tunable con trol of collective behavior (repro duced from [26]). p opulation architecture that forms the basis of multicellular feedbac k control, then examine the quorum sensing systems that provide the communication substrate, and finally derive the mathematical framew ork describing the coupled dynamics of the consortium. 2.1 Tw o-Population con trol arc hitecture The simplest multicellular control architecture consists of tw o in teracting p opulations, c on- tr ol lers and tar gets , whic h collectively implemen t closed-lo op regulation of a biological pro cess of interest [24, 39]. This arc hitecture, illustrated in Fig. 1b, establishes the basic framework up on whic h more sophisticated control strategies can be built. Con troller cells p erform three k ey functions in the distributed control lo op: they sense the output of target cells, compute the error b etw een a desired reference and the measured output, and pro duce an actuation signal prop ortional to the computed control action. The reference signal is typically provided as an external inducer, while sensing and actuation are mediated b y orthogonal quorum sensing (QS) channels, describ ed in detail in Section 2.2. An error computation mec hanism is implemen ted through molecular mec hanisms suc h as comp eting transcriptional regulators or molecular titration circuits as first prop osed in [40] and discussed in [41]. 4 T arget cells execute the desired biological function, representing the plan t in con trol termi- nology , while simultaneously rep orting their state bac k to the con trollers. The target genetic circuit resp onds to the con trol signal b y mo dulating expression of the gene or pathw ay of in ter- est. Concurren tly , target cells synthesize a feedback signaling molecule whose pro duction rate reflects the curren t state of the controlled output, thereb y closing the feedback lo op. This arc hitecture directly addresses the limitations of single-cell em b edded control discussed in Section 1. Metab olic burden is reduced b ecause eac h p opulation carries only a subset of the full circuit. Retroactivit y b etw een controller and plant is mitigated since their molecular comp onen ts never interact within the same cellular compartment. Modularity is enhanced b ecause the same controller p opulation can, in principle, regulate different target p opulations b y establishing appropriate communication c hannels. The realization of this arc hitecture requires a reliable molecular comm unication system capable of transmitting information bidirectionally b et w een p opulations. 2.2 Quorum Sensing as Comm unication Channels Cell-to-cell comm unication through quorum sensing (QS) provides the molecular substrate en- abling distributed control across p opulations. This naturally evolv ed bacterial comm unication system consists of matched pairs of signaling molecules and transcriptional regulators, and has b een extensiv ely characterized for synthetic biology applications [35]. Sender cells produce dif- fusible signaling molecules, such as acyl-homoserine lactones (AHLs), that cross cell membranes in to the extracellular environmen t follo wing a concen tration gradien t. Receiv er cells express cog- nate receptors that bind these molecules and mo dulate transcription of target genes, creating a comm unication channel betw een spatially separated cells. The dynamics of QS-mediated communication can b e captured through coupled differen- tial equations describing intracellular pro duction, cross-membrane exchange, and extracellular diffusion [24]. W e present here an aggregate p opulation mo del in which each p opulation is treated as homogeneous, represented by a single effectiv e cell whose state v ariables corresp ond to p opulation-av eraged concen trations. F or a signalling molecule in sender cells, the intracellular concen tration Q s ev olves as dQ s dt = f prod − γ Q Q s − η s ( Q s − Q e ) , (1) where f prod denotes the pro duction rate, γ Q is the intracellular dilution and degradation rate, and the final term captures diffusion across the cell membrane with rate η s . Here Q e represen ts the extracellular concentration of the QS molecule. In receiver cells, whic h lac k pro duction capabilit y , the intracellular dynamics reduce to dQ r dt = − γ Q Q r − η r ( Q r − Q e ) , (2) where for simplicit y we assume the same dilution and degradation rate in both p opulations. The extracellular concen tration Q e ( x , t ), whic h in general depends on b oth spatial p osition x and time t , ev olves according to: ∂ Q e ∂ t = η s ( Q s i − Q e ) N s + η r ( Q r i − Q e ) N r − γ e Q e + Θ ∇ 2 Q e (3) where N s and N r are the densities of sender and receiver p opulations resp ectively , γ e is the extracellular degradation rate, and Θ is the spatial diffusion co efficient. The term Θ ∇ 2 Q e accoun ts for spatial diffusion in structured en vironments or under limited mixing; it can b e ne- glected when diffusion is fast relative to other dynamics, as in well-mixed microfluidic c hambers or bioreactors. 5 These equations reveal k ey design considerations for multicellular con trol. Mem brane ex- c hange rates η must b e sufficiently fast relativ e to cellular dynamics to enable timely com- m unication, as slo w diffusion introduces dela ys that migh t degrade control p erformance. The extracellular degradation rate γ e determines ho w rapidly signals decay in the en vironment. F ast degradation can limit the spatial range o v er whic h p opulations can co ordinate but improv es tem- p oral resolution b y preven ting signal accum ulation. In structured en vironmen ts such as biofilms or microcolonies, the diffusion co efficien t Θ b ecomes critical, as spatial gradien ts can create pro- nounced concen tration differences that disrupt co ordination b etw een p opulations [42], [43]. Sev eral properties of QS systems are particularly relev ant for their use in feedback control arc hitectures. First, the av ailability of orthogonal systems that op erate in parallel without crosstalk enables m ultiple indep endent comm unication c hannels within the same consortium. W ell-characterized orthogonal pairs include the Lux system (3-oxo-C6-HSL) and Las system (3- o xo-C12-HSL), which exhibit minimal cross-activ ation when prop erly configured [44]. Second, QS systems are tunable. Pro duction rates can b e adjusted through promoter strength and rib o- some binding site engineering, while degradation rates can b e mo dulated b y adding degradation tags to the QS synthase proteins. Third, QS signaling is inherently analog, with output levels v arying contin uously as a function of signal concen tration according to Hill-t yp e dose-resp onse curv es. Ho wev er, the effective dynamic range is constrained by saturation of regulatory pro- teins at high inducer concentrations and by basal expression at lo w concentrations, limiting the amplitude av ailable for control. These c haracteristics, orthogonality , tunabilit y , and analog re- sp onse with b ounded dynamic range, define the design space within which multicellular con trol arc hitectures must operate. 2.3 Mathematical Mo del W e now deriv e a mathematical description of the tw o-p opulation arc hitecture illustrated in Fig. 1b. The mo del captures the dynamics of three coupled subsystems: the gene regulatory net work within controller cells, the bidirectional QS comm unication channels, and the gene regulatory netw ork within target cells. W e adopt an aggregate p opulation formulation in which eac h p opulation is treated as homogeneous and well-mixed; extensions to spatially explicit agen t-based mo dels are discussed in Section 3. 2.3.1 Con troller P opulation Dynamics The controller p opulation implements error computation through an antithetic motif [14], in whic h t w o molecular sp ecies Z 1 and Z 2 m utually sequester one another. Sp ecies Z 1 is pro duced at a rate prop ortional to the reference signal Y d ( t ), provided as an external inducer, while sp ecies Z 2 is pro duced in resp onse to the feedback QS molecule Q x receiv ed from the target p opulation. The sequestration reaction Z 1 + Z 2 → ∅ implements a comparison b etw een reference and feedback signals: the net concen tration of free Z 1 enco des the control error. The dynamics are giv en by d Z 1 dt = µY d − γ z Z 1 Z 2 − γ Z 1 (4) d Z 2 dt = θ Q i x − γ z Z 1 Z 2 − γ Z 2 (5) where µ is the pro duction rate constant for Z 1 , θ gov erns the activ ation of Z 2 b y the in tracellular feedbac k signal Q i x , γ z is the sequestration rate, and γ is the dilution rate due to cell gro wth. The con trol signal is enco ded in the QS molecule Q u , whose intracellular concen tration Q i u in con troller cells evolv es as dQ i u dt = β u Z 1 − γ Q i u + η ( Q e u − Q i u ) (6) 6 where β u is the production rate prop ortional to the free Z 1 concen tration, and η is the mem brane diffusion rate gov erning exchange with the extracellular po ol Q e u . F or simplicity , we assume a common dilution rate γ for all intracellular sp ecies; this assumption can b e relaxed when sp ecies- sp ecific degradation is significan t. 2.3.2 T arget P opulation Dynamics T arget cells respond to the control signal Q u b y modulating expression of the controlled output X c , and sim ultaneously pro duce the feedbac k signal Q x to report their state to the controllers. The con trolled sp ecies evolv es according to dX c dt = f ( Q t u ) − γ X c (7) where Q t u denotes the in tracellular concen tration of the con trol signal in target cells and f ( · ) is a Hill-t yp e activ ation function f ( Q t u ) = α 0 + α max − α 0 1 + ( K u /Q t u ) n u (8) with basal expression α 0 , maximal expression α max , disso ciation constan t K u , and Hill c o efficien t n u . The feedbac k QS molecule Q x is produced in prop ortion to the controlled output X c . Its in tracellular concentration in target cells ev olves as dQ t x dt = β x X c − γ Q t x + η ( Q e x − Q t x ) (9) where β x is the pro duction rate constan t. 2.3.3 Quorum Sensing Communication Channels The t w o p opulations exchange information through tw o orthogonal QS channels: the con trol c hannel carrying Q u from con trollers to targets, and the feedback channel carrying Q x from targets to controllers. In each case, the in tracellular concen tration in receiver cells (whic h lack pro duction capabilit y) equilibrates with the extracellular p o ol through passive diffusion. F or the control signal in target cells: dQ t u dt = − γ Q t u + η ( Q e u − Q t u ) (10) F or the feedback signal in con troller cells: dQ i x dt = − γ Q i x + η ( Q e x − Q i x ) (11) The extracellular concentrations dep end on the balance b et w een production by sender cells, uptak e by receiver cells, and degradation in the environmen t. Assuming a well-mixed regime where spatial diffusion is fast relative to other dynamics, the extracellular concen tration of the con trol signal evolv es as dQ e u dt = N c η ( Q i u − Q e u ) + N t η ( Q t u − Q e u ) − γ e Q e u (12) where N c and N t are the densities of con troller and target populations resp ectiv ely , and γ e is the extracellular degradation rate. Similarly , for the feedback signal: dQ e x dt = N t η ( Q t x − Q e x ) + N c η ( Q i x − Q e x ) − γ e Q e x (13) 7 2.3.4 Closed-Lo op System Equations (4)–(13) constitute a coupled dynamical system describing the closed-lo op b ehaviour of the multicellular controller. The reference signal Y d ( t ) acts as the external input, the con- trolled output X c (or equiv alently the feedback signal Q x ) represen ts the regulated v ariable, and the p opulation densities N c and N t en ter as parameters that affect the effective gains of b oth actuation and feedbac k pathw ays. Under appropriate conditions, this arc hitecture implements an in tegral feedback controller. Sp ecifically , when the sequestration rate γ z is sufficiently high that the sequestration flux dom- inates dilution (i.e., γ z Z 1 Z 2 ≫ γ Z i for i = 1 , 2), the system achiev es Z 1 ≈ Z 2 at steady state, whic h implies µY d ≈ θ Q i x . This enforces a fixed relationship b etw een reference and output that is largely insensitive to parameter v ariations in the target p opulation [24]. This prop erty , kno wn as robust p erfect adaptation, is cen tral to the regulatory p erformance of the m ulticel- lular arc hitecture, whic h implemen ts a distributed version of the embedded an tithetic feedbac k con trol strategy prop osed in [45]. When dilution is not negligible relative to sequestration, the in tegrator b ecomes “leaky” and p erfect adaptation is compromised; achieving the required separation of timescales is therefore an imp ortant consideration in the design of both single-cell and m ulticellular antithetic con trollers [46]. 2.3.5 Implemen tation Considerations Implemen ting m ulticellular feedbac k con trol in tro duces tuning challenges that go b ey ond those encoun tered in single-cell synthetic circuits, as parameters in physically separated p opulations m ust b e co ordinated to ac hieve desired closed-loop b ehavior. A first consideration concerns the effective loop gain, whic h in the multicellular arc hitecture dep ends on parameters distributed across b oth p opulations and the communication c hannels connecting them. In a single-cell controller, gain tuning inv olv es adjusting promoter strengths and rib osome binding sites within one genetic con text. In the multicellular setting, the o v erall gain is the pro duct of contributions from controller gene expression, QS signal pro duction and transmiss ion, and target resp onse characteristics. This distributed nature complicates systematic tuning: mo difying a promoter in the con troller p opulation affects only part of the lo op, and the resulting change in closed-lo op b ehavior dep ends on the (potentially uncertain) c haracteristics of the target p opulation and communication c hannels. Libraries of characterized genetic parts [40] can facilitate exploration of this design space, but predicting closed-lo op p erformance from individual comp onen t characterization remains c hallenging. A second consideration is the population ratio. Unlike parameters in ternal to cells, the rel- ativ e abundance of controllers and targets is influenced b y growth dynamics and ma y drift ov er time. Exp erimental studies ha v e demonstrated that functional regulation is maintained across con troller-to-target ratios ranging from approximately 1:5 to 5:1 [39], pro viding substan tial op erating margin. This robustness arises b ecause the feedback mechanism partially comp en- sates for comp osition c hanges: an increase in controller abundance strengthens actuation but also increases consumption of feedback signal, attenuating the net effect on closed-lo op gain. Nev ertheless, this self-comp ensation has limits, and sev ere comp osition imbalances degrade p er- formance; a c hallenge we address in Section 5. Finally , the communication channels imp ose constraints absent in single-cell implementa- tions. The dynamic range of QS-mediated signaling is bounded by promoter leakiness at low signal concentrations and receptor saturation at high concentrations, limiting the effective op- erating region of the controller. Diffusion delays betw een p opulations introduce phase lag that can destabilize the feedback lo op if gain is set to o high. These constraints couple the achiev able bandwidth and stability margins to physical parameters of the growth environmen t, including cell densit y , mixing regime, and cham b er geometry in microfluidic implemen tations. 8 2.4 Multicellular Consortia as Biological Multiagent Systems Multicellular bacterial consortia represen t a unique class of m ultiagen t systems that differ fun- damen tally from traditional engineered m ultiagent systems in rob otics, distributed computing, or co op erativ e control. These differences, summarized in T able 1, highlight b oth c hallenges and opp ortunities sp ecific to biological implementations, necessitating no vel con trol design ap- proac hes that account for the unique ph ysics and biology of microbial systems. T able 1: Comparison of biological and traditional multiagen t systems Prop ert y T raditional Systems Biological Consortia Agen t dynamics Deterministic or near- deterministic Sto c hastic (molecular noise) Agen t p opulation Fixed or con trollable Dynamic (growth and divi- sion) Comm unication Digital, error-corrected, high- bandwidth Analog molecular diffusion, no error correction Comm unication timescale Milliseconds to seconds Min utes to hours Agen t identit y P ermanent, fixed function Can c hange via differentiation or sto c hastic switching P arameter uncer- tain ty T ypically single-digit p ercent- ages Significan t (20-30% or more across cells) Con trol bandwidth High (millisecond resp onse) Limited (minute-to-hour re- sp onse) F ailure mo des Hardw are/softw are faults (rare, detectable) Mutations (con tinuous, often undetectable) Ev olutionary pres- sure None Con tinuous selection against metab olic burden Spatial effects Often negligible or well- con trolled Diffusion-limited gradien ts, biofilm structure T raditional multiagen t systems maintain agen t p opulations with fixed or con trollable size and exhibit deterministic or near-deterministic dynamics. Communication is t ypically digi- tal, error-corrected, and op erates on millisecond-to-second timescales. Biological consortia, by con trast, feature p opulations that grow and divide contin uously (following logistic dynamics when resources are limited), exhibit intrinsically sto c hastic gene expression with substantial cell-to-cell v ariability , and communicate through analog diffusion of quorum sensing molecules o ver minute-to-hour timescales. The bandwidth limitations and spatial gradients in molecular comm unication fundamentally constrain the complexit y of achiev able co ordination compared to electronic systems, while the absence of error correction mechanisms demands robust enco ding of con trol signals. P arameter uncertain t y presents p erhaps the most significan t difference from traditional sys- tems. Engineered m ultiagent systems t ypically op erate with single-digit percentage uncertaint y , w ell within the capabilities of standard robust con trol tec hniques. Biological systems exhibit substan tially greater parameter v ariation across cells, growth conditions, and exp erimental con- texts, arising from v ariable promoter strengths, fluctuating plasmid cop y n umbers, and con text- dep enden t transcription rates. In our computational studies, robustness is typically assessed under parameter v ariations of 20% or more [24, 26], and the con trol arc hitectures m ust tolerate this lev el of uncertaint y while maintaining stable regulation. Unique to living systems are evolutionary dynamics that ha ve no parallel in traditional mul- tiagen t control. Mutations that reduce metabolic burden b y disabling controller genes confer 9 immediate gro wth adv antages: disabled con trollers reduce cellular resource consumption, allow- ing faster division and competitive displacemen t of functional strains. This creates ev olutionary pressure to eliminate precisely the control functions that maintain desired system b eha vior—a failure mo de entirely absent from engineered systems where comp onen ts do not comp ete for surviv al. This challenge b ecomes particularly significan t in long-horizon applications suc h as con tinuous bioreactors or en vironmental deploymen t where many generations of gro wth o ccur o ver extended operation. Despite these challenges, biological multiagen t systems offer unique capabilities una v ailable in traditional implemen tations: autonomous op eration without external infrastructure, self- repair through p opulation turnov er, the abilit y to function in unstructured environmen ts, and p oten tial for adaptation through ev olution when prop erly harnessed. Understanding these fun- damen tal differences enables the design of con trol strategies that work with rather than against the constraints of living matter, translating classical m ultiagent con trol theory in to forms ap- propriate for the sto c hastic, growing, ev olving systems that microbial consortia represent. 3 V alidation of Multicellular Control: F rom Sim ulation to Ex- p erimen t T ranslating the t wo-population con trol arc hitecture from mathematical framew ork to functional biological system requires systematic v alidation through computational mo deling follow ed b y exp erimen tal implementation. This workflo w enables rapid design iterations in silic o b efore in vesting in time-consuming w et-lab exp eriments. 3.1 In Silico V alidation The mathematical framew ork introduced in Section 2.3 enables systematic v alidation of mul- ticellular feedback con trol through b oth aggregate p opulation mo dels and agent-based sim ula- tions [24, 47, 48]. T ogether, these complementary approaches allow rapid exploration of design trade-offs under idealized conditions, follo w ed b y realistic assessment of spatial effects, stochastic gene expression, and single-cell heterogeneit y prior to exp erimental implemen tation. 3.1.1 Aggregate P opulation Mo dels As shown in Section 2.3, aggregate mo dels such as Equations (4)–(13) treat each population as homogeneous, enabling efficien t simulation and systematic exploration of the design space. This form ulation supp orts extensiv e parameter sw eeps o v er promoter strengths, Hill co efficients, degradation rates, and QS pro duction rates, as well as linearisation around equilibrium p oin ts to assess stabilit y margins. As shown in Fig. 2a, numerical simulations of Equations (4)–(13) confirm that the mul- ticellular architecture in tro duced in Sec. 2.1 achiev es accurate tracking of different reference signals with negligible steady-state error, limited ov ersho ot, mo derate phase lag, and settling times on the order of 3–5 hours [24], which is compatible with bacterial gro wth dynamics under standard conditions. The aggregate framework also enables assessment of robustness through Mon te Carlo simulations with random parameter p erturbations, t ypically within ± 20% of nom- inal v alues. Regulation p erformance remains satisfactory when parameters in b oth p opulations are sim ultaneously p erturb ed, and degrades only mo derately when uncertaint y is confined to the target p opulation [24]. Imp ortantly , v ariations in target dynamics hav e limited impact on closed-lo op b eha vior, supp orting the notion of functional mo dularity as a single controller design can regulate targets with differen t intrinsic parameters without retuning. T o ev aluate the impact of spatial separation and diffusion-limited signalling, we v aried the distance b et ween controller and target p opulations. As sho wn in Fig. 2c, increasing this distance 10 Figure 2: In Silic o v alidation of the m ulticellular feedbac k con trol strategy using aggregate and agen t-based models. (a) Aggregate population model: output of the T arget p opulation under set-p oint and time-v arying reference signals, sho wing accurate regulation, limited ov ersho ot, and stable steady-state b ehav- ior despite nonlinear dynamics and diffusiv e comm unication. (b) Agen t-based sim ulations in a microfluidic-like spatial domain: p opulation-av eraged T arget output trac king trapezoidal and sin usoidal reference signals with neg- ligible phase delay and low cell-to-cell v ariability . (c) Robustness analysis: effect of increasing spatial separation b et w een Controller and T arget populations and of parameter p erturbations ( ± 20%) on regulation p erformance, sho wing preserv ed stability and reduced but nonv anishing dynamic range under severe comm unication attenu- ation. (d) Comp osition indep endence and heterogeneity: regulation p erformance under v arying Controller-to- T arget p opulation ratios and single-cell parameter v ariability , demonstrating that accurate control is maintained ev en when Con trollers represent a small fraction of the consortium and when biological parameters v ary across cells. T ogether, the results indicate that the prop osed distributed feedback architecture provides reliable regula- tion across spatial scales, biological noise, and p opulation composition. All panels are repro duced from [24]. 11 atten uates both the effective control input and the measured output, reducing the ac hiev able dy- namic range. Nevertheless, closed-lo op stability and conv ergence to the desired op erating point are preserved ov er distances exceeding those t ypically encoun tered in microfluidic or colon y-scale exp erimen tal platforms. 3.1.2 Agen t-Based Sim ulations While aggregate mo dels enable efficien t exploration of system-lev el behaviour, they neglect key biological features suc h as stochastic gene expression, cell growth and division, spatial cro wding, and heterogeneous cell–cell in teractions. Agen t-based sim ulations using frameworks such as BSim [47, 48] address these limitations by explicitly mo delling individual cells with sto chastic in tracellular dynamics, ph ysical interactions, cell growth, motility , and spatially resolved QS diffusion fields (see Box 1 for details). Each cell evolv es according to the same regulatory equations used in the aggregate mo del, but with parameters drawn from realistic distributions to capture cell-to-cell v ariability . Communication occurs through discretised reaction-diffusion pro cesses in realistic microfluidic geometries. Agen t-based results (Fig. 2b) confirm that multicellular feedbac k regulation remains effectiv e under biologically plausible spatial and sto c hastic conditions. T rac king of dynamic reference signals remains highly accurate, with p opulation-a veraged output closely following trap ezoidal and sin usoidal tra jectories. Conv ergence times remain comparable to aggregate mo del predic- tions, albeit with mo dest increases in transient ov ersho ot due to the discrete nature of cell-to-cell comm unication. The agent-based framew ork further reveals ho w p erformance dep ends on p opulation density and comp osition (Fig. 2d). Regulation remains effective across a wide range of total densities and con troller-to-target ratios, including cases in which con trollers represen t only a small fraction of the consortium. Significant degradation emerges only at v ery low densities, where diffusion- mediated comm unication b ecomes insufficient to sustain coordination. T ogether, the aggregate and agent-based v alidation studies provide strong computational evidence that multicellular feedback architectures can achiev e reliable regulation and tracking across spatial scales, sto chastic fluctuations, and p opulation v ariabilit y . Once this pro of-of- concept is established, the abstract arc hitecture presen ted in Fig. 1b can serve as a blueprin t for biological implementation using appropriate biomolecular parts, as we describ e in the following section. 3.2 Exp erimen tal Implementation and Results The computational v alidation presen ted in the previous section established the theoretical feasi- bilit y of multicellular feedback control under idealized and realistic conditions. T ranslating these predictions into functional biological systems requires addressing implementation challenges that are difficult to capture fully in silico, including promoter leakiness, context-dependent gene expression, and the difficult y of predicting closed-lo op b ehavior from component-lev el c harac- terization (see Bo x 2 for a discussion of the biological implementation pipeline). The first successful exp erimen tal demonstration of m ulticellular feedback con trol employ ed t wo engineered E. c oli p opulations implemen ting distributed regulation through molecular titra- tion mechanisms [39]. As illustrated in Fig. 3a, controller cells compute the regulation error b y comparing a reference signal pro vided via an external inducer (IPTG) with a feedback signal pro duced by target cells in the form of the QS molecule 3-O-C6-HSL. Error computation is implemen ted through a mo dule first presented in [40] which is based on molecular titration, a mec hanism in whic h t w o molecular sp ecies bind and m utually sequester one another, effectively computing the difference b etw een their concentrations. Sp ecifically , the architecture emplo ys σ /anti- σ factor pairs [49]: σ factors are bacterial transcription initiation proteins, and an ti- σ factors are cognate inhibitors that bind and inactiv ate them. The reference signal induces 12 Figure 3: V alidation of t wo-population feedback con trol. (A) Schematic of tw o-p opulation arc hitec- ture with bidirectional QS communication and molecular titration for error computation. (B) Co efficient of v ariation across biological replicates for closed-lo op (blue) v ersus open-lo op (orange) con trol at differen t IPTG concen trations, showing sixfold reduction at 3 µ M. (C) Normalized target fluorescence versus consortium comp o- sition (target p ercentage), demonstrating comp osition-independent output for closed-loop (flat, p > 0.05) versus comp osition-dependent op en-lo op (negative slop e, p < 0.01). (D) Input-output relationships showing improv ed linearit y for closed-lo op ( R 2 = 0 . 91) v ersus op en-lo op ( R 2 = 0 . 67) con trol. All panels are repro duced from [39]. 13 σ factor pro duction while the feedbac k signal induces anti- σ pro duction; since these sp ecies sequester each other in a one-to-one stoichiometry , their relative abundances determine the effectiv e amount of σ av ailable to activ ate downstream control genes, thereby enco ding the magnitude of the error signal. The resulting control signal is transmitted via a second quorum sensing channel (3-O-C12- HSL) to the target p opulation, whic h mo dulates GFP expression accordingly while co-pro ducing the feedback molecule 3-O-C6-HSL, thereby closing the lo op. This design instantiates the mathematical framew ork introduced in Section 2.3 using w ell-characterized genetic parts and orthogonal Lux/Las quorum sensing systems. Figure 3b–d summarises the experimental comparison betw een closed-loop feedbac k control and op en-lo op configurations in whic h con trollers do not receive information about target out- put. Closed-lo op op eration substantially improv ed regulation reliability , reducing v ariability across biological replicates by approximately sixfold at intermediate reference lev els (Fig. 3b). Input–output linearity also increased mark edly under feedback (Fig. 3c). The co efficien t of determination improv es from R 2 = 0 . 67 in op en lo op to R 2 = 0 . 91 under closed-lo op con trol, yielding predictable dose-resp onse b ehaviour in whic h the output v aries prop ortionally with the reference input. This prop erty is essential for practical deplo ymen t where precise titration of the con trolled v ariable is required. Output lev els were largely indep endent of controller-to- target population ratios o ver the range 1:1 to 15:1 (Fig. 3d), demonstrating the comp osition- indep enden t regulation predicted by the mathematical analysis. Increases in con troller abun- dance raise actuation, which elev ates target output and hence feedback signal, suppressing further con troller activity and stabilising the collectiv e resp onse despite comp ositional fluctua- tions. Settling times remained on the order of three hours for b oth configurations. Although maxim um induction levels w ere low er under feedback than in op en lo op, the improv ed precision and robustness are more relev ant for applications requiring consistent and reliable performance. Bey ond v alidating theoretical predictions, the exp erimen ts revealed implemen tation-level constrain ts not fully captured by computational mo dels. Promoter leakiness limited the achiev- able dynamic range at lo w induction lev els and required the use of additional op erator sites to sharp en regulatory resp onses. Separately , the inclusion of degradation tags on key molecular sp ecies prov ed critical for tw o reasons: accelerating the dynamics to enable resp onsiv e feed- bac k on timescales compatible with bacterial gro wth, and reducing sensitivity to p opulation im balances by prev en ting accumulation of signaling molecules. F urthermore, comp onen t-lev el c haracterization using proxy fluorescent rep orters did not quantitativ ely predict closed-lo op b eha vior, underscoring the necessity of testing complete feedback systems rather than extrap- olating from isolated parts. 3.3 Syn thesis and Practical Design Insigh ts The com bined computational and exp erimental results highlight sev eral design principles that are cen tral to the feasibilit y of m ulticellular feedback control. First, physical separation b etw een con troller and target p opulations effectiv ely reduces metab olic burden and mitigates retroac- tivit y , enabling regulatory architectures whose complexity w ould be difficult to sustain within single cells. This separation allows controller p opulations to implemen t sensing and computa- tion without directly perturbing the regulated bio chemical pro c esses, thereby preserving both stabilit y and mo dularity . Second, orthogonal quorum sensing systems provide reliable bidirec- tional comm unication c hannels b et w een populations. When appropriately selected and tuned, Lux- and Las-based signaling exhibits sufficien tly lo w crosstalk to support independent feedback path wa ys, ev en in dense cultures. Third, molecular mechanisms suc h as σ /anti- σ factor titration offer practical bio chemical realizations of abstract con trol operations, including error computa- tion, allo wing classical feedback concepts to b e mapp ed on to genetic circuit implemen tations. Finally , p opulation-lev el redundancy introduces an inheren t form of fault tolerance: feedback regulation main tains stable collective b eha vior despite substantial sto chastic v ariabilit y at the 14 single-cell lev el. A t the same time, v alidation studies highlight limitations that remain critical for deploy- men t b ey ond con trolled lab oratory settings. In large batch cultures or applications requiring extended autonomous op eration, maintenance of stable p opulation ratios may become challeng- ing, as even mo derate differences in metab olic burden can lead to comp ositional drift through comp etitiv e exclusion; bioreactor en vironments with con tinuous dilution are less susceptible to this issue, but long-term batch processes or environmen tal applications may require additional safeguards. Reactor-scale op eration in tro duces further constrain ts, since mixing times and transp ort dela ys ma y limit the effectiv e bandwidth of quorum-sensing communication, particu- larly in large-v olume or spatially structured environmen ts. Evolutionary stability also represen ts a p ersistent c hallenge, as mutations that disable con troller functions can confer growth adv an- tages, generating selection pressures that gradually ero de regulatory p erformance ov er man y generations. Moreo v er, spatial organization in biofilms or p o orly mixed systems creates con- cen tration gradients that weak en coupling b etw een p opulations, suggesting the need for either enhanced mixing, increased signaling capacity , or con trol arc hitectures that explicitly accoun t for spatial dynamics. Despite these c hallenges, the qualitativ e agreemen t b etw een computational predictions and exp erimen tal observ ations supports the view that m ulticellular feedbac k con trol is not merely a theoretical construct but a practically realizable engineering strategy . The mo deling frame- w ork captures the key mec hanisms underlying composition-indep endent regulation and pro vides a useful to ol for guiding circuit design, while experimental v alidation clarifies the op erational regimes in which distributed control can b e reliably implemen ted. T ogether, these results pro- vide a foundation for extending multicellular control architectures tow ard more complex con- trollers and tow ard application domains where robustness, adaptabilit y , and long-term stability are essen tial requirements. 4 Adv anced Con trollers: The Multicellular PID F amily Building on the basic tw o-p opulation feedback architecture, subsequen t studies hav e demon- strated that increasingly sophisticated biomolecular control structures inspired b y classical prop ortional-in tegral-deriv ative (PID) con trollers can b e implemen ted in a mo dular m ulticellu- lar fashion. While single-cell in tegral feedback has been demonstrated through antithetic control mec hanisms (guaranteeing robust p erfect adaptation) [15] and an embedded PID controller w as analyzed in-silico in [17], the multicellular approac h offers complemen tary adv antages. Namely , eac h control action can b e realized b y a functionally distinct p opulation, distributing the ge- netic burden and enabling indep endent tuning of controller parameters, p oten tially enabling plug-and-pla y synthesis of differen t controllers in the PID family [16, 26]. The resulting architecture, illustrated in Fig. 1c, assigns prop ortional, integral, and deriv a- tiv e actions to separate cell p opulations. Each p opulation receiv es the error signal (encoded as a quorum-sensing molecule) and pro duces a contribution to the ov erall control input correspond- ing to its designated action. These con tributions com bine in the shared extracellular medium b efore reaching the target p opulation, effectively pro viding the equiv alent biomolecular P , I, and D terms found in the classical parallel PID structure. The key distinction from classic PID implemen tations lies in how eac h action is realized biomolecularly . Briefly , prop ortional action arises from direct transcriptional activ ation, integral action from the accumu lating im balance in an antithetic motif, and deriv ativ e action from an incoherent feedforward lo op that responds to the rate of c hange of its input. It is important to note that these biomolec ular implementations are inheren tly nonlinear, go verned by Hill-t yp e response functions and saturation effects, and resem ble their classical linear counterparts only within certain operating regions where the un- derlying bio chemical reactions approximate the desired input-output relationships. A detailed analysis of these op erating regimes and the conditions under which the analogy to classical PID 15 con trol holds can be found in [16, 25, 26]. Despite these nonlinearities, the distributed realiza- tion enables systematic controller design while preserving mo dularity and reuse of controller p opulations. The foundational m ulticellular con troller implemen ts in tegral action through an an tithetic motif distributed across t wo p opulations, as describ ed in Section 2.1. Multicellular PI con- trol [25] extends this b y adding a prop ortional branc h to improv e transien t p erformance. While integral action alone can exhibit slo w conv ergence and significant o vershoot, the pro- p ortional term accelerates the resp onse and provides additional damping. Theoretical analysis further sho ws that proportional-integral arc hitectures can reduce output v ariability arising from sto c hastic gene expression [50]. Numerical and agent-based v alidations confirm that PI architec- tures ac hiev e robust set-p oint regulation ev en under significant parameter v ariability , sto chastic gene expression, and p opulation heterogeneit y , as w ell as under v ariations in con troller-to-target p opulation ratios [25]. Multicellular PD con trol [51] is primarily aimed at impro ving transient response. By intro- ducing a deriv ative branch that responds to the rate of c hange of the error signal, the closed-lo op system exhibits increased damping and reduced o vershoot ov er broad regions of the gain space. Agen t-based simulations confirm that PD arc hitectures can significan tly atten uate oscillatory transien ts and shorten settling times compared to proportional control, while maintaining com- parable steady-state accuracy . How ev er, as in classical con trol, deriv ative action alone does not remo ve steady-state error and remains sensitiv e to noise, limiting its standalone applicability for precise regulation in biological systems. A unified analysis of the multicellular P , PI, PD, and PID families shows that the b ene- fits of in tegral and deriv ative actions are largely complemen tary [26]. In tegral action primarily determines robustness of steady-state regulation to biological uncertain ty and parameter disp er- sion, whereas deriv ativ e action expands the region of stabilizing controller gains and improv es transien t shaping. In com bined PID arc hitectures, agen t-based sim ulations indicate impro ved tolerance to cell-to-cell v ariability and faster conv ergence compared to PI con trol alone, while preserving near-zero steady-state error. Across all architectures, regulation remains effective under realistic communication constrain ts, including diffusion-limited signaling and spatial sep- aration b et w een p opulations, as w ell as mo derate heterogeneity in cellular parameters, con- firming that feedbac k comp ensation is not compromised b y the ph ysical distribution of con trol functions. These results supp ort the view that multicellular PID con trollers constitute a family of mo dular arc hitectures rather than a single fixed design. Dep ending on application requiremen ts, designers ma y prioritize robustness of steady-state regulation (fav oring PI or PID structures) or improv ed transient p erformance and damping (fav oring PD or PID), while retaining the adv antages of distributed implementation, reduced p er-cell genetic burden, and comp osition indep endence that motiv ate multicellular control strategies. How ever, increasing the n umber of con troller populations in tro duces additional challenges: main taining stable co existence among three or four distinct p opulations b ecomes progressively more difficult as differential growth rates accumulate. The choice of con troller complexity m ust therefore balance the b enefits of impro ved dynamic p erformance against the practical difficulty of sustaining m ulti-p opulation consortia o ver extended operation, a problem we address in the follo wing section. 5 P opulation Comp osition Con trol P opulation comp osition con trol (also referred to as ratiometric control) is a fundamen tal chal- lenge unique to multicellular systems. It concerns the main tenance of stable co existence of functionally distinct cellular subp opulations in the presence of differential gro wth rates arising from unequal metab olic costs, resource comp etition, and en vironmental v ariability [23, 54]. Because engineered functions typically imp ose unequal metab olic loads, ev en mo dest growth- 16 Figure 4: Approac hes to ratiometric con trol. (A) Sc hematic of a comp osition control scheme (repro duced from [52]). (B) Representation of a dual c ham b er bioreactor architecture for comp osition con trol (left), and exp erimen tal data of biomass and comp osition regulation with a switc hing and learning-based con troller (repro- duced from [53]). (C) Composition con trol in microfluidics exploiting a reversible memory machanism. On the righ t panel the microfluidics platform is represen ted. On the right panel a representativ e example of regulation with a switching con troller is sho wn (repro duced from [52]). 17 rate mismatches can generate sustained drift in p opulation ratios o ver exp erimen tally relev an t timescales. F rom a con trol p ersp ective, comp osition control therefore constitutes an additional con trol la y er acting on p opulation-level states. This lay er is coupled to, but conceptually distinct from, the design of intracellular and in tercellular feedback dynamics discussed in previous sec- tions, and determines whether the assumptions under whic h molecular con trollers are designed remain v alid ov er time. It is useful to distinguish t wo orthogonal design dimensions in p opulation comp osition con- trol. The first concerns the lo cus of fe e db ack implementation , whic h can b e external (reactor- or computer-mediated orchestration) or embedded within genetic circuits. The second concerns the c ontr ol le d variable , which may b e the relativ e abundance of distinct strains or the frac- tion of cells o ccup ying different phenot ypic states within a single strain. Figure 4 illustrates three representativ e approac hes within this design space: external feedbac k regulation of strain ratios, ratiometric con trol via rev ersible phenotypic switching, and reactor-lev el orc hestration using dual-c hamber architectures. 5.1 External orchestration of p opulation ratios One class of approaches treats consortium composition as an externally regulated pro cess v ari- able, manipulated through environmen tal or reactor-level inputs. In a single chemostat, the dilution rate can act as a control input to regulate the ratio of t wo p opulations when their gro wth-rate curv es in tersect, enabling selectiv e adv an tage across differen t op erating regions. In this setting, feedbac k strategies based on gain scheduling or switc hing con trol ha ve b een sho wn to stabilize co existence and regulate p opulation ratios without requiring genetic mo dification of the strains [31, 33, 55, 56]. This formulation provides a clear control-theoretic interpretation of co existence as a reac hable op erating p oin t under suitable actuation (Fig. 4A). This p ersp ectiv e also reveals a fundamental limitation. When sp ecies are non-complementary (meaning that one p opulation outgro ws the other across the relev an t op erating range), the comp etitiv e exclusion principle implies that co existence cannot b e guaran teed at steady state using dilution-rate actuation alone [27], [57]. T o ov ercome this structural limitation, dual- c hamber bioreactor platforms hav e b een prop osed [32], in which the slow er-growing strain is cultiv ated in a reservoir and p erio dically injected into a mixing cham b er hosting the co-culture (Fig. 4B). This arc hitecture introduces an indep endent actuation c hannel through the inter- c hamber transfer rate, allowing population ratios to b e regulated indep endently of in-cham b er gro wth comp etition. Mo del-based analyses show that switc hing feedbac k strategies can robustly regulate b oth total biomass and comp osition for strain pairs that w ould otherwise b e incom- patible in a single-reactor setting [32]. Subsequen t w ork has extended this framew ork to w ard exp erimen tally grounded and data-efficient control, combining mo del-based and learning-based con trollers trained via sim-to-real pip elines, with in vivo v alidation demonstrating reference trac king, recov ery from p erturbations, and robustness to parameter uncertaint y [58]. These results illustrate how reactor-lev el orchestration can provide a stable exp erimental substrate for testing m ulticellular molecular controllers under con trolled p opulation ratios. Related external orc hestration strategies hav e b een developed within cyb ergenetic control framew orks, where optical or environmen tal actuation is combined with real-time feedback. Optogenetic regulation of differen tiation has been used to dynamically con trol the composition of yeast consortia, enabling reversible and tunable allo cation of cells to distinct functional roles under light-driv en feedback [59]. Other strategies emplo y externally driven switching inputs to toggle genetic circuits asso ciated with different phenotypes in a consortium (Fig. 4C) [52]. While these approaches regulate comp osition through phenotypic switching rather than direct strain comp etition, they share the ratiometric ob jectiv e of con trolling functional allo cation. Similarly , optogenetic mo dulation of an tibiotic resistance has b e en used to regulate effective gro wth rates and enforce desired p opulation ratios in bacterial co-cultures under closed-lo op con trol [33]. These approaches typically assume complementary strains, but demonstrate how 18 external feedbac k can comp ensate for biological v ariability and enforce co ordination ob jectives that are difficult to ac hieve through em b edded circuits alone. 5.2 Em b edded comp osition control In embedded comp osition control, feedback regulation of p opulation structure is implemented directly within genetic circuits, without requiring contin uous external in terv ention. A prominent mec hanism in this class is phenotypic switc hing, whereby functional roles are regulated within a genetically homogeneous p opulation through reversible transitions b et ween cellular states [60]. In this formulation, the con trol ob jectiv e is the regulation of the fraction of cells o ccupying eac h phenot yp e rather than the abundance of ph ysically distinct strains. Other studies hav e demonstrated fully embedded mechanisms in whic h p opulation comp o- sition emerges from genetic feedbac k alone. A notable example is pro vided by differen tiation circuits that couple cellular s tate to growth or metab olic capacity . In a recent exp erimental study , a multi-stage differen tiation program in E. c oli w as sho wn to main tain a stable mix- ture of gro wth and pro duction phenot yp es while suppressing tak eo ver by fast-growing m utan ts, thereb y enhancing evolutionary stability [61]. In this setting, differentiation effectively imple- men ts a p opulation-level negative feedback mec hanism that regulates functional heterogeneity and mitigates comp etitiv e exclusion. Theoretical w ork has similarly sho wn that state-dependent phenotypic switching can stabi- lize coexistence even when static ecological equilibria predict dominance of a single phenot yp e. Mo dels incorp orating resp onsive switching dynamics predict robust main tenance of mixed p opu- lations and increased resilience to p erturbations, suggesting that con trollable switc hing can sub- stitute for explicit p opulation-level actuation [62]. Related analyses of bistable and co-repressiv e circuits further indicate that embedded multistabilit y and delay ed feedback can generate stable ratios or oscillatory p opulation structures in microbial consortia [63]. Another class of embedded strategies relies on self-limiting gro wth implemented through quorum-sensing–regulated killing or growth inhibition. Early w ork demonstrated that density- dep enden t toxin expression can regulate total p opulation size through programmed cell death [64]. This concept was later extended to multi-strain systems using kill–rescue or predator–prey ar- c hitectures, in which recipro cal signalling mo dulates surviv al and leads to stable co existence or sustained oscillations [27, 37]. More recently , orthogonal quorum-sensing–controlled lysis circuits hav e b een used to stabilise otherwise comp etitive co-cultures b y selectively p enalising fast-gro wing p opulations, enabling long-term coexistence without enforced metab olic dep enden- cies [28]. A closely related but conceptually distinct line of w ork focuses on embedded regulation of p opulation gro wth rates through genetic feedback, providing a foundational actuation mech- anism for comp osition control. F usco et al. [65] prop osed an architecture in whic h cells self- regulate their growth rate using a tunable expression system coupled with quorum sensing. By mo dulating the pro duction of a growth-inhibitor protein as a function of p opulation density , the circuit implements a fully em b edded negativ e feedback lo op that stabilises biomass densit y and allows the reference set-p oint to be adjusted through genetic or slowly v arying external tuning. Although demonstrated in a single-strain setting, this approac h highligh ts ho w em b ed- ded regulation of effective gro wth rates can serve as an inner con trol lay er to supp ort long-term co existence and robustness in multicellular consortia, either as a standalone density con troller or in com bination with higher-level composition or differentiation mec hanisms. F rom a con trol-theoretic standp oint, em b edded comp osition regulation can b e interpreted as feedbac k acting on population growth, death, or switc hing probabilities rather than directly on molecular outputs. F ormal arc hitectures hav e b een prop osed in which multiple feedback lo ops regulate b oth total p opulation size and relative abundance through differentiation, death, or phenotypic transitions [66]. These form ulations highligh t strong analogies with co ordination con trol in m ulti-agen t systems, with actuation channels intrinsically tied to cellular ph ysiology . 19 Ov erall, embedded strategies based on phenot ypic switc hing and self-regulation provide a complemen tary route to comp osition control that reduces reliance on external actuation and reactor-lev el interv en tion. While these approaches shift design complexity tow ard genetic cir- cuit stability and ev olutionary robustness, they offer the prosp ect of long-term op eration and tigh ter integration b etw een molecular ob jectives and p opulation-level organization. As suc h, they complement external orchestration and multi-strain architectural approac hes within the broader design space of m ulticellular feedback con trol. 6 Conclusions and F uture Directions in Multicellular Con trol Multicellular feedback control in syn thetic microbial consortia provides a principled wa y to distribute sensing, computation, and actuation across interacting populations, ov ercoming k ey limitations of single-cell control strategies related to metab olic burden, retroactivit y , and lack of mo dularity . The results review ed in this pap er demonstrate that feedback regulation can b e ac hiev ed robustly despite sto chastic gene expression, spatial comm unication constraints, and heterogeneous p opulation comp ositions, establishing multicellular control as a viable engineering paradigm rather than a purely conceptual proposal. A t the mo deling level, an imp ortan t op en c hallenge is bridging the gap b et ween current agen t-based simulations and the physical environmen ts relev an t for large-scale cultiv ation and field deplo ymen t. While existing mo dels capture sto chastic intracellular dynamics and diffusion- mediated communication, they typically neglect realistic transp ort phenomena suc h as adv ec- tion, mixing heterogeneity , and shear effects. Dev eloping reduced-order or m ultiscale mo dels that in tegrate p opulation dynamics with represen tations of fluid transp ort will b e essential for predicting p erformance and stability of m ulticellular con trollers beyond microfluidic and small- batc h settings. Exp erimen tal infrastructure also remains a limiting factor. Most bioreactors are designed for mono cultures and provide limited support for independent monitoring and actuation of m ultiple in teracting p opulations. Scaling m ulticellular con trol to liter-scale or larger systems in tro duces fundamen tal constraints on signaling bandwidth and resp onse times, as well as increased risks of con tamination and p opulation im balance. Reactor architectures and sensing tec hnologies sp ecifically tailored to consortia, together with automated con trol of dilution and feeding, will b e critical to mo v e m ulticellular feedback strategies from lab oratory demonstrations tow ard sustained long-term op eration. F rom a con trol design p ersp ective, curren t m ulticellular con trollers largely implement low- order architectures inspired by classical PID control, though only the antithetic integral motif has been v alidated experimentally to date; the prop ortional and deriv ative extensions remain computationally demonstrated but a w ait in vivo implementation. Extending these concepts to incorp orate adaptive, sto chastic, or predictive con trol mechanisms could substan tially im- pro ve p erformance under uncertain and time-v arying gro wth conditions. How ever, realising suc h strategies requires solving the biological realizabilit y problem of identifying molecular mecha- nisms and comm unication patterns capable of implementing abstract con trol la ws under severe constrain ts on reaction kinetics, noise, and resource a v ailability . Communication dela ys, satura- tion effects, and crosstalk further imp os e performance limits that must b e explicitly accounted for in con troller synthesis. A distinctiv e and unresolved asp ect of multicellular control is long-term ecological and ev o- lutionary stabilit y . Differences in growth rates, metab olic costs, and mutation-driv en loss of function can destabilize carefully tuned architectures ov er extended time horizons. Ac hieving p ersisten t regulation ma y require in tegrating feedback control with ecological design principles, including m utualistic dep endencies, density-dependent gro wth regulation, or spatial structuring that aligns evolutionary fitness with control ob jectiv es. Developing control architectures that remain functional under evolutionary pressure represents a fundamen tal challenge that has no 20 direct analogue in traditional engineered systems. Despite these challenges, the range of p otential applications for multicellular con trol is broad, spanning bioman ufacturing, environmen tal remediation, and therap eutic interv entions [67–69]. In all these domains, the ability to co ordinate specialized populations through feedbac k offers opp ortunities to ac hiev e levels of robustness and functional complexity that are difficult to attain within single-cell designs. Con tin ued progress will dep end on tighter in tegration b etw een con trol-theoretic analysis, circuit-lev el synthetic biology , and scalable exp erimental platforms designed explicitly for consortium-based op eration. In summary , multicellular con trol shifts the fo cus of biological regulation from isolated en- gineered cells to co ordinated p opulations acting as distributed con trol systems. By treating microbial consortia as controllable multiagen t systems, this paradigm op ens new directions for b oth control theory and syn thetic biology , while p osing uniquely in terdisciplinary challenges that will require sustained collab oration across traditionally separate researc h communities. Ac kno wledgemen ts The researc h program on multicellular feedbac k control described in this review represents the outcome of a decade-long collab oration b et w een the Univ ersit y of Naples F ederico I I in Italy and the Univ ersity of Bristol in the U.K. I am deeply grateful to m y colleagues Nigel Sa very (Sc ho ol of Biochemistry , Bristol), Claire Grierson (Sc ho ol of Biological Sciences, Bristol), and Lucia Marucci (Department of Engineering Mathematics, Bristol) for their sustained partner- ship throughout this endea v or. The work w ould not ha v e b een p ossible without the dedicated con tributions of the PhD students and p ostdo ctoral researchers in Bristol and Naples who tack- led differen t asp ects of the pro ject o ver the years: Thomas Goro cho wski, Gianfranco Fiore, F abio Annunziata, Antoni Matyjaszkiewicz, Criseida Zamora-Chimal, Barbara Shannon, Da- vide Fiore, Da vide Salzano, Sara Brancato and Vittoria Martinelli. This researc h w as supp orted o ver the years b y funding from the UK Biotechnology and Biological Sciences Researc h Council (BBSR C), the UK Engineering and Physical Sciences Research Council (EPSR C), the Ital- ian Ministry of Universit y and Researc h (MUR), and the Bristol Centre for Synthetic Biology (BrisSynBio). The author wishes to thank Dr. Davide Salzano (Universit y of Naples F ederico I I, Italy) for his though tful con tributions to the b oxed sidebars, for meticulous pro ofreading of the man uscript, and for pro viding useful commen ts. The use of AI-based writing assistance to ols for language editing and stylistic refinement of p ortions of the text is also ac knowledged. All tec hnical conten t, interpretations, and conclusions remain the sole resp onsibility of the author. References [1] Domitilla Del V ecc hio and Richard M. Murra y . Biomole cular F e e db ack Systems . Princeton Univ ersity Press, Princeton, NJ, USA, 2014. [2] Domitilla Del V ecchio, Aaron J Dy , and Yili Qian. Con trol theory meets synthetic biology . Journal of The R oyal So ciety Interfac e , 13(120):20160380, 2016. [3] Victoria Hsiao, Anandh Sw aminathan, and Richard M Murra y . Control theory for syn- thetic biology: recen t adv ances in system characterization, con trol design, and controller implemen tation for syn thetic biology . IEEE Contr ol Systems Magazine , 38(3):32–62, 2018. [4] M. H. Khammash. Cyb ergenetics: Theory and Applications of Genetic Con trol Systems. Pr o c e e dings of the IEEE , 110:631, 2022. [5] D. E. Cameron, C. J. Bashor, and J. J. Collins. A brief history of synthetic biology . Natur e R eviews Micr obiolo gy , 12(5):381–390, 2014. 21 [6] M. Filo, C.-H. Chang, and M. Khammash. Biomolecular feedback controllers: from theory to applications. Curr ent Opinion in Biote chnolo gy , 79:102882, 2023. [7] I. Ruolo, S. Nap olitano, D. Salzano, M. Di Bernardo, and D. Di Bernardo. Con trol engineer- ing meets syn thetic biology: F oundations and applications. Curr ent Opinion in Systems Biolo gy , 28:100397, 2021. [8] Sebasti´ an Sosa-Carrillo, Henri Galez, Sara Nap olitano, F ran¸ cois Bertaux, and Gregory Batt. Maximizing protein pro duction by k eeping cells at optimal secretory stress levels using real-time con trol approaches. Natur e Communic ations , 14(1):3028, 2023. [9] M Omar Din, T al Danino, Arth ur Prindle, Matt Sk alak, Jangir Selimkhano v, Kaitlin Allen, Ellixis Julio, Eta Atolia, Lev S Tsimring, Sangeeta N Bhatia, et al. Sync hronized cycles of bacterial lysis for in viv o delivery . Natur e , 536(7614):81–85, 2016. [10] Andreas Milias-Argeitis, Marc Rullan, Stephanie K Aoki, Peter Buchmann, and Mustafa Khammash. Automated optogenetic feedbac k control for precise and robust regulation of gene expression and cell gro wth. Natur e c ommunic ations , 7(1):1–11, 2016. [11] Jannis Uhlendorf, Agn` es Miermont, Thierry Dela veau, Gilles Charvin, F ran¸ cois F ages, Sam uel Bottani, Gregory Batt, and Pascal Hersen. Long-term mo del predictiv e con trol of gene expression at the p opulation and single-cell lev els. Pr o c e e dings of the National A c ademy of Scienc es , 109(35):14271–14276, 2012. [12] Filipp o Menolascina, Gianfranco Fiore, Erica Orabona, Luigi De Stefano, Mik e F erry , Jeff Hast y , Mario di Bernardo, and Diego di Bernardo. In-viv o real-time control of protein expression from endogenous and syn thetic gene netw orks. PL oS c omputational biolo gy , 10(5):e1003625, 2014. [13] G. Perrino, A. Hadjimitsis, R. Ledesma-Amaro, and G.-B. Stan. Con trol engineering and syn thetic biology: w orking in synergy for the analysis and control of microbial systems. Curr ent Opinion in Micr obiolo gy , 62:68, 2021. [14] C. Briat, A. Gupta, and M. Khammash. An tithetic Integral F eedback Ensures Robust P erfect Adaptation in Noisy Biomolecular Netw orks. Cel l Systems , 2:15, 2016. [15] S. K. Aoki, G. Lillacci, A. Gupta, A. Baumsc hlager, D. Sc hw eingrub er, and M. Khammash. A univ ersal biomolecular in tegral feedback controller for robust p erfect adaptation. Natur e , 570:533, 2019. [16] M. Filo, S. Kumar, and M. Khammash. A hierarch y of biomolecular prop ortional-integral- deriv ative feedbac k con trollers for robust p erfect adaptation and dynamic p erformance. Natur e Communic ations , 13:2119, 2022. [17] M. Chev alier, M. G´ omez-Sc hia von, A. H. Ng, and H. El-Samad. Design and Analysis of a Prop ortional-In tegral-Deriv ative Con troller with Biological Molecules. Cel l Systems , 9:338, 2019. [18] T. F rei, C.-H. Chang, M. Filo, A. Arampatzis, and M. Khammash. A genetic mam- malian prop ortional–integral feedbac k control circuit for robust and precise gene regula- tion. Pr o c e e dings of the National A c ademy of Scienc es of the Unite d States of Americ a , 119:e2122132119, 2022. [19] Christian Cuba Samaniego and Elisa F ranco. Ultrasensitive molecular con trollers for quasi- in tegral feedback. Cel l Systems , 12(3):272–288, 2021. 22 [20] Cameron D McBride and Domitilla Del V ecchio. Predicting comp osition of genetic circuits with resource comp etition: demand and sensitivity . ACS Synthetic Biolo gy , 10(12):3330– 3342, 2021. [21] F rancesca Ceroni, Alice Bo o, Simone F urini, Thomas E Goro cho wski, Olivier Bork owski, Y aseen N Ladak, Ali R Awan, Charlie Gilbert, Guy-Bart Stan, and T om Ellis. Burden- driv en feedback con trol of gene expression. Natur e metho ds , 15(5):387–393, 2018. [22] Domitilla Del V ecchio, Alexander J Ninfa, and Eduardo D Sontag. Mo dular cell biology: retroactivit y and insulation. Mole cular Systems Biolo gy , 4(1):161, 2008. [23] F rancesca Ceroni, Rhys Algar, Guy-Bart Stan, and T om Ellis. Quantifying cellular capacit y iden tifies gene expression designs with reduced burden. Natur e Metho ds , 12(5):415–418, 2015. [24] Gianfranco Fiore, Antoni Mat yjaszkiewicz, F abio Ann unziata, Claire Grierson, Nigel J Sa very , Lucia Marucci, and Mario di Bernardo. In-silico analysis and implementation of a m ulticellular feedback control strategy in a synthetic bacterial consortium. ACS Synthetic Biolo gy , 6(3):507–517, 2017. [25] Vittoria Martinelli, Davide Salzano, Da vide Fiore, and Mario di Bernardo. Multicellular pi con trol for gene regulation in microbial consortia. IEEE Contr ol Systems L etters , 6:3373– 3378, 2022. [26] Vittoria Martinelli, Da vide Fiore, Da vide Salzano, and Mario di Bernardo. Multicellular pid con trol for robust regulation of biological pro cesses. Journal of the R oyal So ciety Interfac e , 22(222):20240583, 2025. [27] F rederick K Balagadd´ e, Hao Song, Jun Ozaki, Cyn thia H. Collins, Matthew Barnet, F rances H. Arnold, Stephen R. Quak e, and Lingchong Y ou. A synthetic esc herichia coli predator–prey ecosystem. Mole cular Systems Biolo gy , 4:187, 2008. [28] Sp encer R. Scott, M. Omar Din, Philip Bittihn, Liyang Xiong, Lev S. Tsimring, and Jeff Hast y . A stabilized microbial ecosystem of self-limiting bacteria using synthetic quorum- regulated lysis. Natur e Micr obiolo gy , 2:17083, 2017. [29] Kristina Stephens, Maria Pozo, Chen-Y u Tsao, Pricila Hauk, and William E. Bentley . Bacterial co-culture with cell signaling translator and growth controller mo dules for au- tonomously regulated culture comp osition. Natur e Communic ations , 10:4129, 2019. [30] Alex J. H. F edorec, Behzad D. Kark aria, Michael Sulu, and Chris P . Barnes. Single strain con trol of microbial consortia. Natur e Communic ations , 12:1977, 2021. [31] Da vide Fiore, Davide Salzano, Giancarlo Russo, and Mario di Bernardo. F eedback ratio- metric control of tw o microbial p opulations in a single chemostat. IEEE Contr ol Systems L etters , 5(2):555–560, 2021. [32] Sara M. Brancato, Da vide Salzano, Davide Fiore, Giancarlo Russo, and Mario di Bernardo. Ratiometric control of tw o microbial p opulations via a dual c hamber bioreactor. IEEE Contr ol Systems L etters , 8:1301–1306, 2024. [33] Joaqu ´ ın Guti´ errez Mena, Sant Kumar, and Mustafa Khammash. Dynamic cyb ergenetic con trol of bacterial co-culture comp osition via optogenetic feedback. Natur e Communic a- tions , 13(1):4808, 2022. 23 [34] Nicolas E Grandel, Amanda M Alexander, Xiao P e ng, Caroline Palamoun tain, Razan N Alnahhas, Andrew J Hirning, Kre ˇ simir Josi´ c, and Matthew R Bennett. Long-term home- ostasis in microbial consortia via auxotrophic cross-feeding. Natur e Communic ations , 16(1):8573, 2025. [35] Alice Bo o, Ro drigo Ledesma Amaro, and Guy-Bart Stan. Quorum sensing in synthetic biology: A review. Curr ent Opinion in Systems Biolo gy , 28:100378, 2021. [36] T al Danino, Octa vio Mondrag´ on-P alomino, Lev Tsimring, and Jeff Hast y . A sync hronized quorum of genetic clo c ks. Natur e , 463(7279):326–330, 2010. [37] Y e Chen, Jae Ky oung Kim, Andrew J Hirning, Kre ˇ simir Josi´ c, and Matthew R Bennett. Emergen t ge netic oscillations in a synthetic microbial consortium. Scienc e , 349(6251):986– 989, 2015. [38] Nicolas E Grandel, Kiara Reyes Gamas, and Matthew R Bennett. Control of synthetic microbial consortia in time, space, and comp osition. T r ends in Micr obiolo gy , 29(12):1095– 1105, 2021. [39] Da vide Salzano, Barbara Shannon, Claire Grierson, Lucia Marucci, Nigel J. Sav ery , and Mario di Bernardo. In Vivo Multicellular F eedback Con trol in Syn thetic Microbial Con- sortia. ACS Synthetic Biolo gy , 14(7):2537–2547, July 2025. Publisher: American Chemical So ciet y . [40] F abio Ann unziata, Antoni Mat yjaszkiewicz, Gianfranco Fiore, Claire S Grierson, Lucia Marucci, Mario di Bernardo, and Nigel J Sa very . An orthogonal m ulti-input in tegration system to con trol gene expression in esc herichia coli. A CS synthetic biolo gy , 6(10):1816– 1824, 2017. [41] Christian Cuba Samaniego and Elisa F ranco. Ultrasensitive molecular con trollers for quasi- in tegral feedback. Cel l Systems , 12(3):272–288.e3, 2021. [42] Alvin T amsir, Jeffrey J T ab or, and Christopher A V oigt. Robust multicellular computing using genetically enco ded nor gates and chemical ‘wires’. Natur e , 469(7329):212–215, 2011. [43] Jae Kyoung Kim, Y e Chen, Andrew J Hirning, Razan N Alnahhas, Kre ˇ simir Josi´ c, and Matthew R Bennett. Long-range temp oral co ordination of gene expression in synthetic microbial consortia. Natur e chemic al biolo gy , 15(11):1102–1109, 2019. [44] Nicolas Kylilis, Zoltan A T uza, Guy-Bart Stan, and Karen M Polizzi. T o ols for engineering co ordinated system b ehaviour in synthetic microbial consortia. Natur e c ommunic ations , 9(1):2677, 2018. [45] Coren tin Briat, Christoph Zechner, and Mustafa Khammash. Design of a synthetic inte- gral feedbac k circuit: dynamic analysis and dna implemen tation. A CS synthetic biolo gy , 5(10):1108–1116, 2016. [46] Yili Qian and Domitilla Del V ecchio. Realizing ‘integral control’ in living cells: ho w to o vercome leaky in tegration due to dilution? Journal of The R oyal So ciety Interfac e , 15(139):20170902, 2018. [47] Thomas E. Goro cho wski, Antoni W. Matyjaszkiewicz, Thomas T o dd, Neera j Oak, Kira Ko walsk a, Stephen Reid, Krasimira T. Tsanev a-Atanaso v a, Nigel J. Sav ery , Claire S. Grier- son, and Mario di Bernardo. Bsim: An agen t-based to ol for mo deling bacterial p opulations in systems and syn thetic biology . PL oS ONE , 7(8):e42790, 2012. 24 [48] An toni Matyjaszkiewicz, Gianfranco Fiore, F abio Ann unziata, Claire S Grierson, Nigel J Sa very , Lucia Marucci, and Mario Di Bernardo. Bsim 2.0: an adv anced agen t-based cell sim ulator. ACS synthetic biolo gy , 6(10):1969–1972, 2017. [49] Virgil A. Rho dius, Thomas H. Segall-Shapiro, Brian D. Sharon, Amar Gho dasara, Ek a- terina Orlo v a, Hannah T abakh, David H. Burkhardt, Kevin Clancy , T o dd C. Peterson, Carol A. Gross, and Christopher A. V oigt. Design of orthogonal genetic switc hes based on a crosstalk map of σ s, anti- σ s, and promoters. Mole cular Systems Biolo gy , 9:702, 2013. [50] Coren tin Briat and Mustafa Khammash. In-silico prop ortional-in tegral moment control of sto c hastic gene expression. IEEE T r ansactions on Automatic Contr ol , 66(7):3007–3019, 2020. [51] Vittoria Martinelli, Da vide Salzano, Da vide Fiore, and Mario di Bernardo. Multicellular p d con trol in microbial consortia. IEEE Contr ol Systems L etters , 7:2641–2646, 2023. [52] Da vide Salzano, Da vide Fiore, and Mario di Bernardo. Ratiometric control of cell phenot yp es in monostrain microbial consortia. Journal of the R oyal So ciety Interfac e , 19(192):20220335, 2022. [53] Sara Maria Brancato, Davide Salzano, Da vide Fiore, F rancesco De Lellis, Gio v anni Russo, and Mario di Bernardo. A bioreactor-based architecture for in vivo mo del-based and sim-to-real learning control of microbial consortium comp osition. arXiv pr eprint arXiv:2511.08554 , 2025. [54] Yili Qian, Hsin-Ho Huang, Jos´ e I. Jim´ enez, and Domitilla Del V ecchio. Resource com- p etition shap es the resp onse of genetic circuits. ACS Synthetic Biolo gy , 6(7):1263–1272, 2017. [55] Ting An Lee, Jan Morlock, John Allan, and Harrison Steel. Directing microbial co-culture comp osition using cyb ernetic con trol. Cel l R ep orts Metho ds , 5(3), 2025. [56] F ran¸ cois Bertaux, Sebasti´ an Sosa-Carrillo, Viktoriia Gross, Achille F raisse, Chetan Adity a, Mariela F urstenheim, and Gregory Batt. Enhancing bioreactor arrays for automated mea- suremen ts and reactive con trol with reacsight. Natur e c ommunic ations , 13(1):3363, 2022. [57] Stefano Cardinale and Adam P aul Arkin. Contextualizing context for syn thetic biology– iden tifying causes of failure of syn thetic biological systems. Biote chnolo gy Journal , 7(7):856–866, 2012. [58] Sara M. Brancato, Davide Salzano, F rancesco De Lellis, Davide Fiore, Giancarlo Russo, and Mario di Bernardo. In viv o learning-based con trol of microbial populations density in bioreactors. In Pr o c e e dings of the 6th Annual L e arning for Dynamics and Contr ol Con- fer enc e (L4DC) , v olume 242 of Pr o c e e dings of Machine L e arning R ese ar ch , pages 941–953, 2024. [59] Chetan Adit ya, F ran¸ cois Bertaux, Gregory Batt, and Jakob Ruess. A light tunable differ- en tiation system for the creation and control of consortia in y east. Natur e c ommunic ations , 12(1):5829, 2021. [60] Elise W eill, Virgile Andreani, Chetan Adity a, Pierre Martinon, Jak ob Ruess, Gr ´ egory Batt, and F r ´ ed ´ eric Bonnans. Optimal control of an artificial microbial differentiation system for protein biopro duction. In 2019 18th Eur op e an Contr ol Confer enc e (ECC) , pages 2663–2668. IEEE, 2019. 25 [61] Da vid S. Glass, Assaf Bren, Iftach V aisb ourd, Avi May o, and Uri Alon. A synthetic differen tiation circuit in esc herichia coli for suppressing mutan t takeo v er. Cel l , 187(4):931– 944.e12, 2024. [62] Pierre A. Haas, Maria A. Gutierrez, Nuno M. Oliv eira, and Raymond E. Goldstein. Sta- bilization of microbial comm unities b y resp onsiv e phenotypic switc hing. Physic al R eview R ese ar ch , 4(3):033224, 2022. [63] Mahdi Sadeghp our, Alan V eliz-Cuba, G´ ab or Orosz, Kre ˇ simir Josi´ c, and Matthew R. Ben- nett. Bistability and oscillations in co-repressiv e syn thetic microbial consortia. Quantitative Biolo gy , 5(1):55–66, 2017. [64] Lingc hong Y ou, Rob ert S Cox II I, Ron W eiss, and F rances H Arnold. Programmed p opu- lation control b y cell–cell communication and regulated killing. Natur e , 428:868–871, 2004. [65] Virginia F usco, Davide Salzano, Da vide Fiore, and Mario di Bernardo. Embedded control of cell gro wth using tunable genetic systems. International Journal of R obust and Nonline ar Contr ol , 33(9):4893–4907, 2023. [66] Xin ying Ren, Ania-Ariadna Baetica, Anandh Sw aminathan, and Richard M Murra y . Pop- ulation regulation in microbial consortia using dual feedbac k control. In 2017 IEEE 56th A nnual Confer enc e on De cision and Contr ol (CDC) , pages 5341–5347. IEEE, 2017. [67] In Y oung Hw ang and Matthew W o ok Chang. Engineering commensal bacteria to rewire host–microbiome in teractions. Curr ent opinion in biote chnolo gy , 62:116–122, 2020. [68] Inessa Arif, Maria Bato ol, and P eer M Schenk. Plant microbiome engineering: exp ected b enefits for improv ed crop gro wth and resilience. T r ends in Biote chnolo gy , 38(12):1385– 1396, 2020. [69] Jasmine Shong, Man uel Rafael Jimenez Diaz, and Cynthia H. Collins. T ow ards syn thetic microbial consortia for biopro cessing. Curr ent Opinion in Biote chnolo gy , 23(5):798–802, 2012. [70] M. E. Guti´ errez, P . Gregorio-Go do y , G. P´ erez del Pulgar, L. E. Mu ˜ noz, L. S´ aez, and A. Ro dr ´ ıguez-P at´ on. A new improv ed and extended version of the multicell bacterial sim ulator gro. ACS Synthetic Biolo gy , 6(10):1496–1508, 2017. [71] CellMo deller Developmen t T eam. Cellmo deller: Python based framework for m ulticellular agen t-based simulation. https://cellmodeller.github.io/CellModeller/ , 2025. [72] Mic hael W. Sneddon, James R. F aeder, and Thierry Emonet. Efficien t modeling, sim ulation and coarse-graining of biological complexit y with rule-based methods. PL oS Computational Biolo gy , 7(7):e1002266, 2011. [73] Stev en Thomas, Nathaniel D Maynard, and John Gill. Dna library construction using Gibson assem bly . Natur e Metho ds , 12(11):i–ii, 2015. [74] Carola Engler and Sylv estre Marillonnet. Golden gate cloning. In DNA cloning and as- sembly metho ds , pages 119–131. Springer, 2013. [75] Kevin D Litcofsky , Raffi B Afey an, Russell J Krom, Ahmad S Khalil, and James J Collins. Iterativ e plug-and-pla y methodology for constructing and mo difying synthetic gene net- w orks. Natur e metho ds , 9(11):1077–1080, 2012. 26 [76] Elisa P edone, Lorena Postiglione, F rancesco Aulicino, Dan L Rocca, Sandra Mon tes-Oliv as, Mahmoud Khazim, Diego di Bernardo, Maria Pia Cosma, and Lucia Marucci. A tunable dual-input system for on-demand dynamic gene expression regulation. Natur e c ommuni- c ations , 10(1):4481, 2019. 27 BSim F ramework for Agen t-Based Mo deling Figure 5: a Agen t-based sim ulation explicitly mo dels intracellular dynamics, together with cell growth, division, mec hanical interactions, and extracellular signaling. b Com- parison of representativ e agent-based simulators for microbial p opulations. Both panels are repro duced from [48]. Accurate analysis of m ulticellular feedback con trollers requires mo dels that capture not only intracellular gene regulation but also the spatial and ph ysical interactions among cells and their environmen t. Agent-based mo deling addresses this need b y representing eac h cell individually and explicitly simulating growth, division, mo vemen t, mec hanical con tact, and diffusion of signaling molecules. As illustrated in Fig. 5, this approac h naturally links single-cell dynamics to emergen t p opulation-level behaviors. BSim is an op en-source, Ja v a-based framew ork for agen t-based sim ulation of bacterial p opulations, first introduced in 2012 [47] and extended in version 2.0 [48]. Each cell car- ries its o wn gene regulatory net work, which can be describ ed using ordinary , delay ed, or sto c hastic differen tial equations. The simulator also mo dels realistic geometries such as microfluidic cham b ers and chemostats, together with diffusion, degradation, and trans- p ort of extracellular sp ecies. Compared with well-mixed p opulation mo dels, agent-based sim ulators offer tw o key ad- v antages. First, they reveal emergen t collective phenomena—including quorum sensing, spatial pattern formation, and co existence dynamics—that dep end on lo cal in teractions. Second, they capture cell-to-cell v ariability and rare sto c hastic even ts that are typically a veraged out in deterministic descriptions. In practice, such in silic o exp eriments pro vide a v aluable design to ol for m ulticellular con trol. Candidate arc hitectures can be tested, tuned, and stress-tested under realistic v ariability b efore exp erimental implemen tation, reducing b oth developmen t time and lab oratory cost. The latest release of BSim, together with documentation and example models, is av ail- able at https://github.com/bsim- bristol/bsim . Alternativ e pac k ages compared with BSim in Figure 5b are describ ed in [70–72]. 28 S Exp erimen tal Implemen tation Guide: Design and Characterization Figure 6: Pip eline for implementing a multicellular feedback con trol scheme in vivo . Eac h strain is developed through a build–test–learn cycle b efore integration in to the final consortium. Implemen ting a multicellular feedbac k controller requires the co-design of t wo in teracting bacterial strains that play the roles of Contr ol ler and T ar get . As illustrated in Fig. 6, de- v elopment follows an iterativ e build–test–learn workflo w applied first to each p opulation individually and only later to the in tegrated consortium. F or eac h strain, suitable biological components must b e selected to implement the desired con trol functionality . F or example, the Controller requires a comparator mo dule that measures the deviation b et w een the T arget’s gene expression and a reference signal. Suc h behavior can b e realized using an annihilation (an tithetic) motif, here implemen ted through σ /anti- σ factor pairs [40]. This motif pro vides a biomolecular analogue of integral action and enables robust regulation of gene expression. Once the functional mo dules are identified, the gene net work is partitioned across com- patible plasmids. This design balances construct size, mo dularity , and genetic stabilit y while resp ecting replication-origin constrain ts of the host chassis. Mo dular partitioning also facilitates indep enden t testing and debugging of subsystems. Eac h construct is assembled using standard molecular biology techniques [73, 74] and exp erimen tally characterized. Promoter strengths, rib osome binding sites, and degra- dation rates are tuned to match mo del predictions. Measured input–output resp onses (e.g., from batch cultures or flo w cytometry) are used to up date mathematical mo dels and guide subsequen t redesign steps, closing the build–test–learn lo op [75]. This mo del-informed w orkflo w allo ws multicellular con trol arc hitectures to b e v alidated in silic o and exp erimentally at the single-strain level before moving to full consortium in tegration, significantly reducing dev elopment time and cost. 29 Exp erimen tal Implementation Guide (con t.): Consortium In tegration and V alidation After individual c haracterization, the Controller and T arget strains are in tegrated in to a consortium that p erforms the desired collectiv e task. At this stage, the primary de- sign challenge b ecomes establishing reliable communication and v alidating closed-lo op b eha vior at the p opulation lev el. Bidirectional information exchange is typically achiev ed through quorum-sensing molecules that diffuse across the medium. Because multiple signaling channels may in- tro duce cross-talk, orthogonal acyl-homoserine lactone (AHL) pairs must b e selected to minimize unintended coupling [44]. In our implemen tation, 3-O-C6 and 3-O-C12 HSL w ere chosen for their minimal in terference and suitable op erating ranges. System-lev el v alidation follows a control-orien ted proto col. Op en-lo op exp eriments first v erify that the Controller p opulation can influence T arget gene expression. The feedback path wa y is then enabled to close the lo op, and the complete architecture is ev aluated using classical control metrics, including set-p oin t tracking, transient resp onse, steady-state accuracy , and robustness to disturbances and parameter v ariability . These experiments establish whether the consortium b eha ves as a reliable m ulticellular feedback con troller. Sev eral exp erimental platforms can be emplo y ed. Batc h or c hemostat cultures coupled with flow cytometry provide scalable and cost-effective measurements of p opulation- a veraged behavior [56]. In con trast, microfluidic devices combined with time-lapse mi- croscop y enable contin uous, single-cell monitoring and rev eal spatial effects and hetero- geneit y [76]. While microfluidics offers higher temp oral and spatial resolution, it requires more complex exp erimental setups and careful strategies to main tain long-term co exis- tence of m ultiple strains. In practice, com bining bulk and microfluidic exp eriments provides complementary insight in to b oth p opulation lev el p erformance and single-cell v ariabilit y , enabling comprehensiv e v alidation of multicellular con trol strategies. 30