Rapid prototyping and design of cybergenetic single-cell controllers

Synthetic biology is increasingly able to realize complex genetic circuits, yet practical deployment is hindered by unanticipated cross-talk, cellular burden, stochastic intracellular environments, and long design-test cycles. Context-dependence further degrades circuit performance when moving between hosts or growth conditions. Robustness can be engineered through network architecture and, more generally, feedback regulation, but achieving perfect adaptation typically requires integral action. While several biomolecular integral control motifs have been proposed theoretically and a few demonstrated experimentally, mapping specifications to biological implementations remains challenging due to component availability, characterization limits, and stochasticity. Inspired by hardware-in-the-loop practices in engineering, the authors propose the Cyberloop: an in vivo/in silico single-cell framework where real-time cellular fluorescence measurements inform stochastic simulations of candidate biomolecular controllers running on a computer, which then compute control inputs delivered back to individual cells using optogenetics. This approach avoids assuming a precise model for the biological plant, shortens design cycles, and enables rapid testing under realistic cellular conditions. Using optogenetic activation of transcription and a live-cell nascent RNA reporter in Saccharomyces cerevisiae, the study investigates integral biomolecular controller motifs under realistic stochastic dynamics and non-idealities, deriving guidelines that inform future in vivo implementations.

Prior work has applied feedback control to synthetic circuits predominantly at the population level, with limited single-cell implementations; most used deterministic control schemes (e.g., proportional, PI, MPC) suitable for bioreactors but not directly for biomolecular controllers in single cells. Integral feedback enabling robust perfect adaptation has been analyzed in biological systems and proposed in synthetic contexts, with theoretical designs for autocatalytic and antithetic motifs. Experimental realizations of integral or quasi-integral control exist in bacteria, mammalian systems, and in vitro, yet achieving perfect adaptation is complicated by dilution and degradation of controller species. Theoretical studies have explored performance limitations, the impact of sequestration rates, and strategies to mitigate leaky integration. Hardware-in-the-loop is a mature engineering approach used where real-world testing is costly; analogous biology-in-the-loop strategies have been developed for in silico control of gene expression, including optogenetic frameworks enabling real-time control, but without explicit stochastic biomolecular controller architectures. This study builds on these strands by embedding stochastic biomolecular controllers in a single-cell cybergenetic loop, and by experimentally probing integral control motifs, including extensions (Rein and proportional-integral) that have been analyzed theoretically.

Cyberloop architecture: Cells are imaged periodically under an inverted microscope while a control computer performs automated segmentation, tracking, and fluorescence quantification. Each cell’s readout (nascent RNA spot intensity proportional to count) updates propensities of an in silico stochastic biomolecular controller (reaction network). Controller species counts are simulated using Gillespie’s SSA over the sampling interval, and the resulting control signal is converted to a blue light intensity delivered to that specific cell via a DMD-based projector at the next cycle. The sampling period was 2 minutes, set by computation time for image analysis and controller simulation between frames; approximately 75–100 cells were tracked per experiment over 4 hours. Biological system: A genetically engineered Saccharomyces cerevisiae strain expresses an optogenetic transcription factor (EL222-VP16) that dimerizes upon blue light and activates a promoter with EL-binding sites. The output is nascent RNA abundance from a reporter transcript bearing 24 PP7 stem loops; a PP7-mRuby3 fusion binds nascent transcripts, creating a nuclear fluorescence spot whose intensity reports nascent RNA count. Imaging and stimulation: Nikon Ti-Eclipse microscope, 40X oil objective, Hamamatsu ORCA-Flash4.0 camera. Brightfield and fluorescence (mRuby3) imaging with defined filters and exposure settings; five Z-stack fluorescence images per sampling time. The environment was maintained at 30 °C in an opaque enclosure. A custom DMD-based blue light projection system targeted individual cells; light intensities were constrained by projector limits and capped at a normalized maximum. Workflow per sampling interval: (1) Acquire images. (2) Segment cells (CellX) and track lineages. (3) Quantify nascent RNA spot intensity per cell. (4) Update controller reaction propensities with the measurement and simulate the controller via SSA for 2 minutes. (5) Compute per-cell light inputs and generate a projection mask corrected for optical aberrations. (6) Project the mask to deliver cell-specific blue light. Controller implementations tested: (i) Autocatalytic Integral Controller (single controller species V; reactions implementing integral action with autocatalysis and an actuation reaction proportional to V). Variants included V>0 constraint and addition of a basal leak production rate γ to remove the absorbing state. (ii) Antithetic Integral Controller (AIC) with controller species Z1 and Z2 produced proportional to reference and measured output, respectively, and sequestering each other with rate η; actuation via Z1-dependent light. Dilution effects modeled by adding degradation terms with rates drawn from the observed doubling time distribution (mean 77 min). Parameter sweeps explored η and controller production rates (θ, μ). (iii) Antithetic Integral Rein Controller (AIRC) adding a direct negative feedback from Z2 to the output; implemented by introducing an in silico output species Y driven by biological X1 and subject to Z2 repression to realize the extra feedback channel not physically available in the setup. (iv) Antithetic Proportional-Integral Controller (APIC) adding a proportional negative feedback term implemented as a nonnegative max-limited function of the output error. Data analysis: Population averages computed at each time point across tracked cells and filtered with a 5-sample moving average. Steady-state metrics computed over 60–202 minutes. Frequency-domain analysis of single-cell trajectories performed via DFT (MATLAB fft) to derive average amplitude spectra and extract peak frequency and amplitude. Controller species steady-state distributions computed from per-cell steady-state means. Strains and constructs: Yeast strains BY4741/BY4742 backgrounds; previously constructed optogenetic and reporter plasmids (pDB58, pDB96, pDB97) and strain DBY96 used. No new plasmids or strains constructed. Experimental constraints: Experiments limited to 4 hours due to growth in a 2D monolayer under agarose; imaging every 2 minutes. Code integrated with hardware in MATLAB; custom routines for segmentation, tracking, SSA, and light mask generation.

• Cyberloop enables real-time, single-cell, stochastic biomolecular control by closing the loop between fluorescence measurements and optogenetic actuation through in silico SSA simulations.
• Autocatalytic Integral Controller: Under stochastic dynamics, the controller species V reaches an absorbing state at V=0 with probability one, causing loss of control. Lowering the actuation gain k increases steady-state V abundance to sustain actuation and delays absorption, but does not eliminate it. Introducing a small basal production (leak) of V removes the absorbing state. Properly tuned leak maintains effective set-point tracking with minimal steady-state error, despite loss of perfect adaptation in theory.
• Antithetic Integral Controller (AIC): Achieves accurate set-point tracking across multiple reference levels and can follow time-varying references. Operates robustly at very low copy numbers (Z1, Z2 in single digits) without degrading steady-state tracking. Increasing sequestration rate η improves transient dynamics (faster settling), though benefits saturate when η is comparable to production rates. For the yeast system, adding physiologically realistic dilution (δ≈ln2/77 min⁻1; doubling time mean 77 min) to Z1 and Z2 had negligible impact on steady-state accuracy across a broad parameter range.
• Antithetic Integral Rein Controller (AIRC): Adding a direct negative feedback from Z2 to the output (implemented via an in silico output Y) preserves set-point tracking and improves transients—reduced overshoot, faster settling, and lower variance compared to AIC—while requiring lower Z2 abundance. When the output degradation rate is high, AIRC provides little additional benefit over AIC.
• Single-cell dynamics: AIC can induce oscillatory single-cell outputs despite stable population means; average single-cell amplitude spectra show a distinct peak at a nonzero frequency (~0.2 rad/sample) under certain parameters. Adding a proportional negative feedback term (APIC) quenches this spectral peak, mitigating oscillations, though higher proportional gains can increase overshoot and settling time.
• Through these experiments, the study derives actionable design guidelines: avoid absorbing states in single-species integral designs (e.g., via small basal production), select moderate-to-high sequestration rates for AIC, expect robustness to low copy numbers, and consider proportional or rein extensions to improve single-cell dynamics and transients.

By embedding stochastic biomolecular controllers within a real biological context, the Cyberloop bypasses model mismatch issues inherent to purely in silico design and accelerates iterative testing. The stark contrast between deterministic predictions for the autocatalytic integral motif and its stochastic failure due to an absorbing state underscores the importance of designing for stochasticity; a minimal basal production restores functionality in practice. Conversely, the antithetic integral controller benefits from intrinsic noise and ensures robust perfect adaptation at the population level while tolerating very low controller copy numbers. Practical non-idealities—particularly dilution—are often cited as threats to integral control, but for the transcriptional target network in yeast, dilution at physiological rates introduces negligible steady-state errors, suggesting that additional complexity to counter dilution may be unnecessary in similar contexts. Extensions such as the Rein and proportional actions provide levers to tune transients and single-cell dynamics: Rein decreases variance and settling time with low controller burden, while proportional action mitigates intrinsic oscillations at the single-cell level. Collectively, these findings validate the Cyberloop as an effective bridge between abstract controller designs and their eventual in vivo realizations, offering experimentally grounded guidelines for parameter regimes and architectural choices that yield reliable control in cells.

The study introduces the Cyberloop, a single-cell cybergenetic framework that enables rapid prototyping and optimization of stochastic biomolecular controllers directly in living cells. Applying the platform to yeast optogenetic transcription, the authors (i) revealed stochastic limitations of the autocatalytic integral motif and a practical remedy via basal production, (ii) confirmed robust set-point tracking of the antithetic integral controller under realistic noise and dilution, (iii) demonstrated performance gains from the Rein extension, and (iv) showed that adding proportional action suppresses oscillatory single-cell dynamics. These results provide concrete design guidelines and highlight when additional architectural elements are warranted. Future directions include extending Cyberloop-driven design to broader classes of synthetic circuits (e.g., cell-to-cell communication, pattern formation, division of labor), exploring different hosts and outputs, integrating additional sensing/actuation channels to implement richer feedback structures, and automating parameter exploration to further shorten design cycles.

The Cyberloop requires specialized integrated hardware-software infrastructure; the custom code is not standalone. Experiments were limited to 4-hour durations due to monolayer growth under agarose and hardware targeting constraints, potentially omitting longer-term behaviors. Sampling was constrained to 2-minute intervals by computation time, which may miss faster dynamics. Implementation of some controller extensions (e.g., Rein) required introducing an in silico output species (Y) to emulate an additional feedback channel not physically available, so results reflect a hybrid system. Controller architecture and parameter choices involve experimenter judgment; results are specific to the yeast optogenetic transcriptional system studied and may not generalize without adaptation. Perfect adaptation is theoretically compromised by basal leak introduced to avoid absorbing states, implying a trade-off between robustness to stochasticity and steady-state accuracy. High output degradation can negate benefits of the Rein extension. Although dilution had negligible effect here, its impact may be significant in other systems with different kinetics or faster growth.