Industry: Biotech Product: ModelRisk Application: Quantifying uncertainty in synthetic biology production processes
A synthetic-biology team had a design library of engineered production strains and an obvious favourite: design D6, with the highest intrinsic titer at 14.5 g/L on the bench. The production target was 8 g/L. On paper D6 cleared it with room to spare. After propagating the model across a full seed-to-harvest train, the mean realised titer for D6 was 2.7 g/L — and the probability of meeting the 8 g/L target had fallen to 11%. The highest-titer design on the bench was one of the worst designs at scale.
The culprit is genetic instability. A high-titer strain carries a heavy metabolic burden, which gives any spontaneous loss-of-function mutant — an "escaper" that stops producing but grows faster — a large fitness advantage. Over the 35 generations from seed to harvest, those escapers sweep the population, and the realised titer collapses far below the intrinsic value a bench measurement reports. A design's productivity at scale is a race between its titer and its stability, and the bench number only measures the first.
The deterministic approach measures intrinsic titer in a short bench run, compares it to the target, and ranks designs by that number. It implicitly assumes the producing strain stays the strain you inoculated. In a long fed-batch or a multi-stage seed train it does not: escaper frequency grows multiplicatively, so a tiny initial mutant load becomes a population-level takeover by harvest if the fitness advantage is high enough. Bench titer and genetic stability also trade off — the designs with the highest intrinsic titer carry the most burden and therefore the fastest escaper takeover — so ranking on titer alone systematically promotes the least stable designs.
The model treats each design as a population of fermentation runs. Intrinsic titer is LogNormal across runs (right-skewed). Escaper takeover follows logistic growth: an initial mutant load grows by a per-generation fitness advantage over 35 generations, and the producing fraction is what survives. A shared per-run quality factor (media lot, inoculum health, control loop) moves a run's titer ceiling and its escaper advantage together. A design is usable only if the realised titer meets 8 g/L and at least 85% of the population is still producing at harvest.
For design D6 the deterministic plug-in is 14.5 g/L — the bench intrinsic value. The realised-titer distribution at harvest tells a different story: a mean of just 2.7 g/L, with 89% of runs landing below the 8 g/L target. The distribution is dragged hard left and bimodal — runs where the strain held produce near the intrinsic value, runs where escapers won collapse toward zero — and the bench number sits far out in the right tail that almost never occurs at scale.
Decomposing each design's usability into its titer hurdle and its stability hurdle reveals an inverted-U in design space:
The best design is D3, with a 57% usability probability — against just 3% for the highest-titer D6, a 54-percentage-point gap. As intrinsic titer climbs from D3 upward, the stability pass-rate falls off a cliff (76% at D3, 41% at D4, 15% at D5, 3% at D6), and it drags usability down with it. The most productive strain at harvest is a mid-library design, not the bench champion.
Tracking the producing fraction generation by generation shows why. The stable best design D3 holds a P50 of 95% producing at harvest (P10 still 66%, just below the stability target). The high-burden D6 collapses to a P50 of 7% producing — P10 of 0% — well before the 35-generation harvest point. Two designs that look one rung apart on the bench are on opposite sides of the 85% stability line at the end of the run.
Around the chosen D3's baseline usability, the tornado ranks the levers: escaper growth advantage (circuit burden) swings usability from -39 to +26 points, initial mutant load (cell-bank QC) -28 to +19, generations to harvest -23 to +20, and intrinsic titer only -24 to +12. Stability and process levers dominate the realised outcome; raw intrinsic titer is the weakest of the four.
A 14.5 g/L bench strain that delivers 2.7 g/L at harvest is not a measurement error — it is a strain that lost the race against its own escapers. Simulating titer and stability together moved the lead design down the library, from a 3% usability champion on paper to a 57% usability winner at scale.