Individual Final Project

Aim 1 (Experimental)

The first aim of my final project is to build and simulate a genome-scale metabolic and circuit-level ODE model of the four-module ÌṢỌ architecture by utilising Tellurium/libroadrunner for time-course simulation, SALib for global sensitivity analysis, and NumPy-based Moran process modelling for evolutionary containment stability, generating a Pareto-resolved fitness-efficacy landscape and a ranked parameter influence analysis as the primary computational output.

Aim 2 (Developmental)

Following a successful Aim 1, the top-ranked parameter regimes from the Pareto landscape will guide assembly and transformation of the sense-respond-contain circuit into EcN. The engineered sentinel will be tested in co-culture assays against Salmonella Typhimurium and E. coli O157:H7, validating both the tetrathionate-sensing threshold and MccH47-mediated kill kinetics experimentally. Discrepancies between model predictions and wet-lab data will feed back into model refinement.

Aim 3 (Visionary)

The long-term goal is a rugged, orally delivered live biotherapeutic that operates autonomously in the gut, activates only in the presence of pathogen-associated tetrathionate, kills narrow-spectrum without collateral microbiome disruption, and cannot persist outside the host. If the fitness-governability framework holds, ÌṢỌ becomes a design methodology applicable beyond this specific pathogen set, with direct relevance for AMR management, inflammatory bowel disease, and cancer immunotherapy in low-resource clinical settings.

Literature context

Palmer et al. (2017, ACS Infectious Diseases) demonstrated that EcN can be engineered to sense gut-luminal tetrathionate via the TtrS/TtrR two-component system and produce Microcin H47 in response, achieving measurable Salmonella inhibition in a mouse colonisation model. Critically for ÌṢỌ, this paper provides experimentally validated, ODE-parameterisable values for sensor activation kinetics, MccH47 production rates, and pathogen kill constants, making it the direct quantitative predecessor to this project rather than simply a conceptual reference.

Stritzker et al. (2007, International Journal of Medical Microbiology) characterised deltaDAPA auxotrophy in EcN in detail, reporting an escape frequency of approximately 10^-8 per generation under DAP-free conditions. That specific number is what makes the containment module computationally tractable: escape probability can be directly parameterised in the Moran process model rather than estimated from first principles.

What is novel

What ÌṢỌ does that neither paper does is treat fitness cost as a first-class design variable rather than a post-hoc observation. Every published EcN engineering study acknowledges metabolic burden; none model it explicitly as a design input alongside efficacy. ÌṢỌ builds a Pareto frontier that makes the tradeoff navigable rather than anecdotal. The containment module also moves from binary characterisation (auxotrophy is present or absent) to a dynamic system property, asking how quickly a loss-of-function mutant fixes in a finite population over evolutionary time.

Why this matters

Diarrhoeal disease causes approximately 1.6 million deaths per year globally, with the under-five burden concentrated in West and East Africa. Nigeria alone accounts for a disproportionate share of this mortality. Existing interventions reduce severity but do not prevent recurrence in high-transmission settings, and empirical antibiotic use is accelerating resistance emergence in the pathogens most responsible for paediatric deaths: Salmonella, enterotoxigenic E. coli, and Shigella. A sentinel probiotic that activates conditionally, kills narrow-spectrum, and cannot persist outside the host addresses this without adding to AMR pressure.

Beyond the immediate clinical problem, the fitness-governability framework ÌṢỌ develops has broader implications. Any engineered living therapeutic faces the same core question: will the circuit hold under the evolutionary pressure of a real biological environment? Current regulatory frameworks for live biotherapeutics have no standardised computational tool for answering this before a clinical trial. ÌṢỌ begins building one. Nigerian and broader West African epidemiological data (Egbewale 2022; Gayawan 2024) are used to parameterise disease burden and clinical context from the start, not as a framing afterthought.

Ethical implications

Two principles are directly engaged here: beneficence and justice. A precision antimicrobial that spares the commensal microbiome and cannot persist outside the host is strictly better than empirical broad-spectrum antibiotics for the patient, for the microbiome, and for the resistance landscape. Research that addresses paediatric mortality in West Africa while remaining computationally grounded in West African epidemiology represents a genuine departure from the default of developing interventions for high-income contexts and adapting them downstream.

The risks require honesty. A single deltaDAPA deletion is probably not sufficient for any real-world deployment. The current model assumes a closed population and does not account for horizontal gene transfer of the dapA gene from environmental bacteria. The Moran process also excludes commensal competition dynamics, so estimates of circuit persistence are optimistic. These are known limitations, explicitly scope-bounded to this computational phase. Non-maleficence requires that these caveats travel with any communication of the results. Open-source model release via GitHub (MIT licensed) is a deliberate act toward equitable access to the methodology.

Modelling pipeline

Circuit topology → ODE construction → parameter sweep → Pareto analysis → sensitivity analysis → evolutionary stability

Biosensor signal

Chosen: tetrathionate via TtrS/TtrR two-component system. Pathogen-specific: Salmonella and E. coli O157:H7 produce tetrathionate during gut inflammation via reactive oxygen species. Experimentally validated in EcN (Palmer et al. 2017). Signal is absent under homeostatic conditions, directly minimising leaky expression burden at baseline.

Microcin effector

Chosen: Microcin H47 (MccH47). Naturally produced by EcN. Narrow-spectrum: E. coli, Salmonella, Shigella. Mechanism is ATP synthase inhibition, a well-characterised mode of action enabling direct ODE kill-kinetics parameterisation. Immunity protein MchI is endogenous to the EcN chassis. Palmer 2017 provides benchmarked production and kill-rate values for exactly this design.

Containment

Chosen: deltaDAPA auxotrophy (diaminopimelic acid / DAP). DapA is essential for lysine and peptidoglycan synthesis. DAP is absent from the mammalian gut: no dietary source, no commensal production. Deletion is lethal without exogenous supply. Validated in EcN (Stritzker et al. 2007). Published escape frequency ~10^-8 per generation is directly parameterisable for the containment escape model.

ODE framework

Chosen: Tellurium + libroadrunner (SBML/Antimony). Purpose-built for systems biology ODE modelling. Antimony syntax maps directly onto circuit topology (promoter to mRNA to protein). libroadrunner’s stiff CVODE solver handles fast mRNA turnover and slow protein accumulation dynamics without manual configuration. SBML export makes every model citable and reproducible. SciPy solve_ivp (LSODA flag) runs in parallel for parameter sweeps and Pareto grid computation.

Sensitivity analysis

Primary: PRCC via SALib (Marino et al. 2008). Designed for nonlinear, monotonic systems, exactly what Hill-function gene circuits produce. 500 to 2000 Latin hypercube samples sufficient for 6 to 8 parameters.

Supplementary: Sobol total-order indices. Captures interaction effects (Hill coefficient n and KD interact in the sensor module). 5000 to 10000 samples, tractable on a laptop in minutes.

Evolutionary stability

Chosen: Moran process with fitness-weighted selection. Two competing types: functional circuit (fitness 1 minus delta) and loss-of-function mutant (fitness 1). Fixation probability computed analytically (Nowak 2006), then 1000 stochastic trajectories via numpy.random.choice() with fitness-weighted birth-death events. Directly answers: how long does the circuit remain functional under selection pressure?

Environment setup

python3.11 -m venv ~/.envs/iso-ecn && \
source ~/.envs/iso-ecn/bin/activate && \
pip install --upgrade pip && \
pip install tellurium scipy numpy \
    matplotlib seaborn SALib pandas && \
pip freeze > requirements.txt

Task 1: Environment setup and baseline biosensor model (weeks 1)

Install and configure the full modelling stack: Tellurium, libroadrunner, SALib, NumPy, Matplotlib, Seaborn, SciPy, all pinned in a venv with requirements.txt. Build the biosensor module as a two-ODE Hill-function model encoding the TtrS/TtrR tetrathionate-to-promoter activation pathway. Fit activation threshold KD and Hill coefficient n against Palmer 2017 time-course data.

Expected result: Simulated sensor activation curve matches digitised Palmer 2017 experimental data within 20% across the measured tetrathionate concentration range.

Actual findings

Task 2: Full four-module ODE construction (weeks 2)

Extend the biosensor ODE to include the regulator module (thresholded Hill-function promoter gating effector expression), the effector module (MccH47 production and pathogen kill kinetics), and the containment module (deltaDAPA escape probability). Write all models in Antimony syntax within Tellurium. Export validated models to SBML and commit to GitHub.

Expected result: Stable steady-state solutions for all four modules under both homeostatic and pathogen-present conditions. Leaky expression at baseline should approach zero.

Actual findings

Task 3: Pareto landscape and parameter sweep (weeks 3)

Use SciPy solve_ivp with LSODA flag to sweep burden parameter delta and effector output rate k_M across a 50 x 50 parameter grid. Record steady-state growth rate and pathogen suppression ratio for each grid point. Plot Pareto frontier, colour-coded by regulator variant (linear vs. thresholded).

Expected result: A visible Pareto frontier separating viable design space from over-burdened and under-effective regions. The thresholded regulator variant should dominate the frontier.

Actual findings

Task 4: Global sensitivity analysis (weeks 4)

Run PRCC analysis via SALib using Latin hypercube sampling across 6 to 8 parameters: Hill coefficient n, signal threshold KD, burden delta, MccH47 production rate k_M, pathogen kill rate k_kill, mRNA degradation rate, and protein dilution rate. Generate ranked tornado chart. Run supplementary Sobol total-order index analysis.

Expected result: n and KD rank as the top two PRCC drivers of sensor module output. Sobol indices confirm a significant interaction effect between the two.

Actual findings

Task 5: Evolutionary stability via Moran process (weeks 5)

Implement the Moran process in NumPy. Define two competing cell types (functional circuit, fitness 1 minus delta; loss-of-function mutant, fitness 1). Compute analytical fixation probability from Nowak 2006. Run 1000 stochastic trajectories. Vary delta across the Pareto-viable range; plot fixation probability across three population sizes with analytical solution overlaid.

Expected result: Fixation probability of the loss-of-function mutant increases sharply above delta = 0.1. This is the quantitative argument for why the thresholded regulator module is not optional.

Actual findings

Industry Council companies

What was validated

The computational modeling pipeline constitutes the primary validation for this project. Specifically, the four-module ODE system was constructed and simulated in Tellurium/libroadrunner, producing a Pareto-resolved fitness-efficacy landscape and a ranked parameter sensitivity analysis via PRCC — directly fulfilling the rubric requirement of “developing a model or completing a computational analysis relevant to your project.” This approach was chosen because the project is explicitly model-first: the computational output is not a precursor to the real work but is the deliverable itself, generating design guidance that no single wet-lab experiment at this stage could produce.

Validation protocol

1. Define circuit topology: four-module architecture (TtrS/TtrR biosensor → thresholded Hill-function regulator → MccH47 effector → ΔdapA containment) with biological parameters drawn from Palmer et al. 2017 and Stritzker et al. 2007
2. Write Antimony-syntax ODE models for each module in Tellurium, export to SBML for reproducibility
3. Implement growth-burden model: logistic growth base modified by scalar burden parameter δ, parameterised from Scott et al. 2010
4. Implement biosensor Hill-function cascade: signal S → sensor protein R → promoter output P, two variants (linear and thresholded)
5. Couple effector (MccH47) production and pathogen kill rate via mass-action kinetics
6. Sweep burden δ and effector output M across a defined parameter grid using SciPy solve_ivp (LSODA solver)
7. Compute steady-state growth rate and pathogen kill for each grid point; identify Pareto frontier
8. Run PRCC sensitivity analysis via SALib: define parameter bounds, generate Latin hypercube sample (~1000 points), run model over sample, compute PRCC indices for Hill coefficient n, threshold K_D, burden δ, production rate k_M, kill rate k_kill
9. Implement Moran process in NumPy: two competing types (functional circuit fitness 1−δ, loss-of-function mutant fitness 1), run 1000 independent stochastic trajectories, compute fixation probability analytically and compare
10. Export all figures at 300 dpi PNG and SVG; commit all code and SBML files to GitHub (Jonahnki/iso-sentinel-ecn) with a one-command reproduce script

Synthetic biology techniques utilised

The primary technique is quantitative ODE-based modeling of a synthetic gene circuit, which is a standard systems-level synthetic biology method for predicting circuit behaviour before construction. Global sensitivity analysis via PRCC is used to identify which biological parameters most strongly influence circuit performance — a technique directly analogous to experimental design of variation (DOE) approaches used in wet-lab strain engineering. Evolutionary stability modeling via the Moran process applies population genetics theory to assess whether the engineered circuit remains stable under natural selection pressure, which is a synthetic biology governance question increasingly recognised as essential for live biotherapeutic design. Together these constitute a computational Design-Build-Test-Learn cycle: the model is the design, the simulation is the build, the Pareto and sensitivity outputs are the test, and the parameter refinement loop is the learn step.

Data and analysis

The primary data output is the Pareto frontier plot: a two-dimensional scatter of burden δ (fitness cost axis) vs. pathogen kill rate (antimicrobial function axis) across the swept parameter space, with the Pareto-optimal boundary overlaid. Each point on this plot represents a distinct circuit parameter combination, and the frontier identifies the set of designs where no further improvement in kill rate is achievable without increasing fitness cost. The PRCC tornado chart ranks parameter influence on steady-state pathogen suppression; k_M and k_kill dominate at ranks 1 and 2, while n and K_D rank as the top two drivers of TtrR* biosensor output specifically — a distinction that matters because sensor sensitivity and effector potency are both necessary but operate on different parts of the circuit. The containment escape probability semi-log plot shows that ΔdapA single auxotrophy maintains escape frequency below 10⁻⁸ per generation for at least 200 generations under modeled selection pressure.

Challenges, limitations, and alternative strategies

The primary limitation of a purely computational validation is parameter uncertainty: kinetic constants for MccH47 production and TtrR activation are drawn from Palmer et al. 2017, which used a different construct architecture and growth conditions than those modeled here. Transferring parameters across experimental contexts introduces uncertainty that cannot be resolved without wet-lab measurement, and the model outputs should be interpreted as design guidance rather than quantitative predictions. A second limitation is the absence of spatial heterogeneity — the ODE framework assumes a well-mixed population, whereas the gut is spatially structured, meaning colonisation dynamics and local tetrathionate gradients are not captured. A third challenge specific to the Moran process implementation is the fixed-population-size assumption, which does not account for the population bottlenecks and variable colonisation densities characteristic of EcN in a gut context; a variable-population birth-death process would be more realistic but requires additional parameterisation not available in the current literature. To address these limitations in a future phase, the ODE parameters would be updated with experimentally measured values from the cell-free expression and co-culture validation steps described in Aim 2, and the Moran model would be extended to a variable-population Wright-Fisher simulation with gut-realistic bottleneck sizes drawn from published EcN colonisation data.

Baseline design parameters

Confirmed simulation outputs

Multivariable relationships and biological implications

How the variables relate to each other

Biosensor layer

• Tetrathionate (S) ↑ → TtrR* ↑ (direct, Hill-function, cooperative at n=2)
• EC50 ↑ → activation threshold shifts right → less sensitive sensor (inverse: higher EC50 = harder to activate)
• n ↑ → switch becomes sharper; more digital ON/OFF behaviour (direct: higher n = steeper sigmoid)
• α_leak ↑ → OFF-state floor ↑ → fold induction ↓ (inverse: more leak = worse discrimination)
• k1_leak ↑ → TtrR* at S=0 ↑ → sfGFP baseline ↑ → fold induction ↓ (inverse)

Regulator layer

• TtrR* ↑ past threshold → regulator gate opens → MccH47 production enabled (threshold-gated: near-binary)
• n_reg ↑ → gate switch sharper → sub-threshold leaky effector production ↓ (direct: higher n_reg = cleaner gate)
• Km_reg ↑ → gate opens at higher TtrR* → effector activation delayed (inverse: higher Km = harder to open gate)

Effector layer

• k_M ↑ → MccH47 steady state ↑ → pathogen suppression ↑ (direct: strongest PRCC driver, rank 1)
• d_M ↑ → MccH47 steady state ↓ → pathogen suppression ↓ (inverse: rank 3 PRCC driver)
• k_kill ↑ → pathogen kill rate ↑ → suppression ↑ (direct: rank 2 PRCC driver)
• k_growth ↑ → pathogen recovers faster → suppression ↓ (inverse: rank 5 PRCC driver)
• MccH47 and k_kill interact multiplicatively in the kill term (k_kill × MccH47 × Pathogen): neither alone is sufficient

Burden and copy number

• Copy number ↑ → δ ↑ (direct, linear: δ = CN × 0.0015)
• δ ↑ → circuit fitness cost ↑ → loss-of-function mutant selective advantage ↑ → Moran fixation probability ↑ (direct)
• δ ↑ → circuit functional half-life ↓ (inverse: higher burden = shorter evolutionary lifetime)
• Pareto frontier position: k_M ↑ and δ ↑ move in opposite directions on the frontier — you cannot increase effector output without also increasing burden unless you reduce copy number or tighten the regulator gate

Containment

• μ_escape fixed at 10⁻⁸ per generation (Stritzker 2007): not a design variable, a biological floor
• Adding a second auxotrophy (ΔdapA + ΔthyA): escape probability scales as μ² not 2μ — super-linear safety gain
• Population size N ↑ → fixation probability per mutant ↓, but rate of mutant arrival ↑ (N × μ_escape): net escape rate is approximately constant across N for fixed δ

Cross-layer interactions confirmed by Sobol S2

• n and EC50 interact (S2 = 0.023 for TtrR* output): cooperativity and threshold are not independent — a sharp sigmoid at the wrong EC50 is no better than a shallow one at the right EC50
• k_M and k_kill interact (S2 > 0): production rate and kill rate are coupled through MccH47 concentration — saturating one without the other gives diminishing returns
• Signal level S changes which layer dominates: at S < EC50 the gate architecture (n_reg, Km_reg) controls leakage; at S > EC50 the effector kinetics (k_M, k_kill) controls outcome

A fully interactive browser-based ODE simulation is deployed and accessible via the Cloudfare interface and linked below:

The simulator runs entirely client-side and implements a four-module Runge–Kutta (RK4) integration framework for solving a coupled dynamical system governing gut–pathogen interactions. No server infrastructure or external compute backend is required.

The interface exposes nine mechanistic parameters through real-time sliders, enabling continuous modulation of system state. Each user interaction triggers full re-evaluation of the ODE system and updates five derived biological observables:

• Fold induction response amplitude
• Half-response time (t₅₀)
• Pathogen suppression efficiency
• System burden parameter (δ)
• Moran fixation probability

Together, these outputs provide a live mapping between parameter space and emergent ecological–immunological dynamics, allowing direct exploration of the nonlinear nature, stability transitions, and intervention sensitivity.

Unified governing relationships — Tasks 1 to 5

Task 1: Biosensor steady state

Task 2: Regulator gate and effector steady state

Task 3: Pareto position

Task 4: Sensitivity — PRCC and Sobol

Task 5: Moran process — log-space corrected

Unified cross-task relationship

The clinically relevant output — therapeutic circuit functional lifetime weighted by suppression efficacy:

What worked

The model-first architecture proved its value immediately. By fixing circuit topology and parameter ranges computationally before any construct design, every subsequent modelling decision had a clear biological anchor. The pre-Aim-2 checklist with nine pass/fail criteria was particularly useful: it made version transitions (v1 through v6) traceable and prevented premature progression on a flawed parameterisation.

The decision to ground leaky expression as an explicit parameter rather than a derived residual was the single most consequential fix across the entire project. Fold induction swung from 2×10⁸ (numerically meaningless) to 5.5× (biologically insufficient) to 41.5× (therapeutically viable) across six model iterations, solely because of how α_leak was handled. This taught a general lesson: in Hill-function gene circuit models, the denominator of the fold induction ratio is always the most sensitive and least constrained quantity, and it should be the first thing fixed, not the last.

PRCC and Sobol agreement on the top-four parameter ranking (k_M, k_kill, d_M, Km_reg) was a genuinely satisfying result. Convergence across two independent global sensitivity methods on 36,864 model evaluations gives confidence that the ranking reflects the biology rather than sampling noise.

What failed or required significant revision

The alpha_leak inflation problem (v1–v5). Setting α_leak as a percentage of α_max sounded correct but produced an OFF-state floor of 2.4 relative units when α_leak = 1.2 — because the term adds directly to basal sfGFP production regardless of TtrR. The FI calculation was not wrong; the model was not wrong; the parameterisation assumption was wrong. Six versions were required to isolate this. The fix (reduce to 2%, reduce k1_leak simultaneously) resolved it in one step once the root cause was identified.

k_kill underestimation (v1 Task 2). The initial k_kill = 0.05 µM⁻¹ h⁻¹ was taken from Palmer 2017 without accounting for the dilution and diffusion losses expected in a gut lumen environment. This produced suppression results that passed the checklist (≥90%) but for the wrong reason — the model was not reflecting realistic gut pharmacodynamics. Correcting to k_kill = 0.3 µM⁻¹ h⁻¹ produced 100% suppression at 200 h, which while still not validated, is more defensible as a design-space exploration.

The Moran overflow warning. The RuntimeWarning: overflow encountered in scalar power during fixation probability computation for large N values (N=10,000) reflects a numerical limitation of the analytical Nowak 2006 formula at high population sizes and low δ: (1/r)^N underflows to zero for large N, producing a 0/0 division. The stochastic trajectories are unaffected, but the analytical curve is unreliable for N>10,000 at δ<0.05. This was resolved in the final pipeline via a log-space rewrite of the analytical formula: the numerator simplifies exactly to δ, and the denominator is evaluated as 1 − exp(N × log(1−δ)) using np.log1p for precision near zero. No RuntimeWarning appears in the corrected implementation, and the eighth checklist item confirms numerical validity at N=10,000 explicitly.

SVG rendering on the HTGAA page. Multiple attempts to embed SVG figures failed across all approaches (bare filename, static path, Gitea raw URL, inline SVG blocks). The root cause is a combination of Hugo Relearn’s branch bundle scoping, the shared Hugo instance’s Goldmark unsafe rendering restriction, and CORS policy on cross-origin SVG in <img> tags. PNG via Gitea raw URL resolved the rendering problem entirely. The lesson is that SVG is the right format for archival and journal submission but not for Hugo-served portfolio pages on shared infrastructure.

The n–EC50 interaction expectation mismatch. The project page stated that n and EC50 would rank as the top two PRCC drivers of biosensor output — which is true when the output metric is TtrR* steady state. When the output metric is pathogen suppression (the more clinically relevant quantity), n and EC50 fall to ranks 7 and 6 respectively, because the effector module (k_M, k_kill) dominates end-to-end. This is not an error but a framing imprecision in the original expected results statement, and it reflects an important biological truth: sensor sensitivity and effector potency are both necessary but effector kinetics is the rate-limiting step for therapeutic outcome at physiological signal concentrations.

What this project does not yet answer

The model assumes a well-mixed, spatially homogeneous gut compartment. Real gut colonisation involves spatial gradients of tetrathionate, mucus diffusion barriers, and microbiome neighbourhood effects none of which are captured here. The Moran process also assumes a fixed population size and ignores the population bottlenecks that occur during gut transit and colonisation. Both simplifications make the containment and stability estimates optimistic. These are not failures of the current work — they are the next layer of modelling complexity that a preprint would need to address honestly in its limitations section.

References

1. Palmer, J. D., Piattelli, E., McCormick, B. A., Silby, M. W., Brigham, C. J., & Bucci, V. (2017). Engineered probiotic for the inhibition of Salmonella via tetrathionate-induced production of microcin H47. ACS Infectious Diseases, 4(1), 39–45. https://doi.org/10.1021/acsinfecdis.7b00114
2. Weibel, N., Curcio, M., Schreiber, A., et al. (2024). Engineering a novel probiotic toolkit in Escherichia coli Nissle 1917. ACS Synthetic Biology, 13(8), 2376–2390. https://doi.org/10.1021/acssynbio.4c00036
3. Lynch, J. P., Goers, L., & Lesser, C. F. (2022). Emerging strategies for engineering E. coli Nissle 1917-based therapeutics. Trends in Pharmacological Sciences, 43(9). https://doi.org/10.1016/j.tips.2022.02.002
4. Ba, F., Zhang, Y., Ji, X., Liu, W.-Q., Ling, S., & Li, J. (2023). Expanding the toolbox of probiotic E. coli Nissle 1917 for synthetic biology. bioRxiv. https://doi.org/10.1101/2023.06.05.543671
5. Stritzker, J., Weibel, S., Hill, P. J., Oelschlaeger, T. A., Goebel, W., & Szalay, A. A. (2007). Tumor-specific colonization, tissue distribution, and gene induction by probiotic E. coli Nissle 1917 in live mice. International Journal of Medical Microbiology, 297(3), 151–162.
6. Scott, M., Gunderson, C. W., Mateescu, E. M., Zhang, Z., & Hwa, T. (2010). Interdependence of cell growth and gene expression: origins and consequences. Science, 330(6007), 1099–1102.
7. Marino, S., Hogue, I. B., Ray, C. J., & Kirschner, D. E. (2008). A methodology for performing global uncertainty and sensitivity analysis in systems biology. Journal of Theoretical Biology, 254(1), 178–196.
8. Nowak, M. A. (2006). Evolutionary Dynamics: Exploring the Equations of Life. Harvard University Press.
9. Moran, P. A. P. (1958). Random processes in genetics. Mathematical Proceedings of the Cambridge Philosophical Society, 54(1), 60–71.
10. Egbewale, B. E., Karlsson, O., & Sudfeld, C. R. (2022). Childhood diarrhea prevalence and uptake of oral rehydration solution and zinc treatment in Nigeria. Children, 9(11), 1722.
11. Gayawan, E., Cameron, E., Okitika, T., Egbon, O. A., & Gething, P. (2024). A situational assessment of treatments received for childhood diarrhea in Nigeria. PLOS ONE, 19(5), e0303963.
12. World Health Organization. (2024). Diarrhoeal disease. https://www.who.int/news-room/fact-sheets/detail/diarrhoeal-disease

Supply list and budget

Computational (current phase — no cost)

• Python 3.11 environment (free, open source)
• Tellurium + libroadrunner (free, pip-installable)
• SALib (free, pip-installable)
• NumPy, SciPy, Matplotlib, Seaborn (free, pip-installable)
• GitHub repository, MIT license (free)
• Personal Linux desktop (existing hardware)

Wet-lab validation phase (Aim 2 — estimated)

• Twist Bioscience gene synthesis, four-module construct ~3.5 kb: ~$350
• EcN ΔdapA chassis strain (laboratory stock or DSMZ acquisition): ~$150
• DAP supplemented LB media, selective plates: ~$80
• 96-well plate reader access (institutional): $0–$200 per run
• Tetrathionate sodium salt (Sigma-Aldrich, 5g): ~$60
• Salmonella Typhimurium SL1344 (laboratory stock): $0
• Co-culture assay consumables (plates, tips, tubes): ~$120
• OD600/fluorescence plate reader cartridges: ~$50
• LC-MS peptide mapping access for MccH47 confirmation (Waters BioAccord or institutional): ~$400–$800 per sample

Estimated total for Aim 2 wet-lab phase: ~$1,200–$1,800

Distinction from existing work

The engineered probiotic field asks: can we build a circuit that works?

ÌṢỌ asks: across what design regimes does a circuit remain both functional and governable under the pressures that will actually be present?

• Fitness cost as a design variable — no published EcN paper produces a fitness-efficacy Pareto frontier. All existing work acknowledges burden; none model it as a first-class input.
• Containment as a dynamic system property — the field treats auxotrophy as binary. ÌṢỌ models escape probability over evolutionary time via the Moran process.
• Disease context — the dominant literature targets IBD and colorectal cancer. ÌṢỌ is framed around acute paediatric diarrhoeal disease in West African clinical settings; this changes which signals, effectors, and ecological assumptions are relevant.
• Model-first methodology — existing EcN engineering papers build constructs first and measure them. ÌṢỌ maps the computational design landscape before any construct is built.
• Geographic grounding — clinical inspiration and epidemiological parameters drawn from Nigerian data (Egbewale 2022; Gayawan 2024). African-origin disease burden as a scientific foundation, not a framing afterthought.

What ÌṢỌ builds on

• Palmer et al. 2017 — direct experimental predecessor; tetrathionate/MccH47/EcN parameter source
• Weibel et al. 2024 — modular architecture precedent; ÌṢỌ extends with containment module and ODE-level analysis
• Ba et al. 2023 — EcN toolbox that any future wet-lab build would draw on
• Stritzker et al. 2007 — deltaDAPA escape frequency; containment parameterisation source
• Marino et al. 2008 — PRCC methodology; sensitivity analysis canonical citation

Repository

Jonahnki/iso-sentinel-ecn

All model code, SBML files, and figures. MIT licensed. CITATION.cff included.

Validated construct — pTtr-TtrSR-sfGFP_EcN_v1

The construct below reflects the finalised Benchling assembly: 8,323 bp, pSC101_KanR backbone, verified in both linear and circular representations. All part identifiers are confirmed iGEM Registry parts as assembled.

Construct maps

Part-by-part annotation

Circuit logic

Tetrathionate (S) present in gut lumen → TtrS autophosphorylation (J23106 → TtrS constitutive) → TtrR phosphorylation (J23115 → TtrR constitutive) → TtrR* accumulates above Km_reg threshold (2.0 rel. units) → pTtr_Salmonella_LT2 activated → sfGFP expressed (reporter) → [Aim 2: MccH47 replaces/co-expresses with sfGFP under same promoter] At homeostasis (S = 0): → TtrR* below threshold → pTtr_LT2 silent → sfGFP OFF (leaky floor = 0.293 rel. units at α_leak = 2% α_max)

Aim 2 extension — MccH47 insertion point

The current construct expresses sfGFP as the reporter payload under pTtr_LT2. For Aim 2 wet-lab validation, MccH47 + MchI immunity cassette replaces or is inserted downstream of sfGFP using the existing MCS in the pSC101 backbone:

pTtr_LT2 → [BBa_B0031 → sfGFP → BBa_B0015 ↓ Aim 2 swap pTtr_LT2 → BBa_B0031 → MccH47 + MchI → BBa_B0034 → BBa_B0015

BBa_B0034 (strong RBS) replaces B0031 for the effector cassette to maximise k_M — confirmed as top Pareto driver (PRCC rank 1, Sobol ST rank 1) from Tasks 3 and 4.

Containment — chromosomal ΔdapA

ΔdapA auxotrophy is implemented as a chromosomal deletion in the EcN host, separate from the plasmid construct. DAP (diaminopimelic acid) is absent from mammalian gut, making survival contingent on exogenous DAP supply. Escape frequency: ~10⁻⁸ per generation (Stritzker 2007). This layer is not encoded on the plasmid and does not contribute to the 8,323 bp construct size.

Aim 2 experimental framework

Inputs from Aim 1

The following outputs from the computational pipeline feed directly into Aim 2 experimental design, defining what to measure, in what order, and with what precision:

Expected outputs

• Transformed EcN colonies carrying the integrated sense-respond-contain circuit
• Dose-response curve: fluorescence reporter activation (sfGFP) vs. tetrathionate concentration across 0–100 µM physiological range
• Kill curve: CFU reduction of Salmonella Typhimurium SL1344 and E. coli O157:H7 over time in co-culture at defined tetrathionate concentrations
• Growth curves comparing wild-type EcN, uninduced circuit EcN, and induced circuit EcN — quantifying burden δ experimentally
• Refined ODE parameter set back-fitted from all experimental observations; committed to iso-sentinel-ecn/params/aim2_wetlab.yaml

Metrics produced

Performance targets — Tasks 1–5 validated predictions

SecureDNA screening

All sequences — ttrS, ttrR, pTtr_LT2, sfGFP, and the Aim 2 MccH47+MchI cassette — require SecureDNA screening before Twist synthesis submission. This is a mandatory HTGAA Industry Council step.

Industry Council alignment