Industry: Biotech Product: ModelRisk Application: Quantifying uncertainty in CRISPR-based gene editing outcomes
A team screening guide RNAs for a therapeutic CRISPR edit had a clear favourite. Sequence-prediction tools and a handful of replicates put guide g5 at a 72% mean on-target editing rate and a mean of 1.9 detected off-target events per cell. The acceptance criteria were a 70% on-target floor and no more than two off-target events. Both means clear the bar, so the spreadsheet said "clean — advance it." The Monte Carlo model said the probability that any single edited cell actually clears both criteria at once is 36%.
That gap is the whole problem with editing a living genome from a point estimate. On-target rate and off-target count are not fixed numbers; they are distributions that move together replicate to replicate, because the same things that make an editing run hot — a fresh Cas9 lot, an efficient transfection, deep sequencing — push both the productive edits and the spurious cuts up at the same time. A guide that looks clean on its averages can fail one criterion or the other far more often than either margin suggests.
The deterministic workflow evaluates a guide at its mean on-target rate and its mean off-target count, checks each against its threshold, and advances the guide if both pass. It silently assumes the two means describe the same cell. They do not. Because on-target activity and off-target activity are positively coupled within a run, the cells that edit best are also the cells most likely to over-cut — so the joint event "edits enough AND stays specific" is rarer than the product of the two marginal pass-rates, and far rarer than a two-margin check implies.
The model treats one guide as a population of editing outcomes. Each replicate draws a shared run-quality factor (Cas9 lot, transfection health, sequencing depth) that nudges on-target and off-target together. On-target rate is a Beta around a logistic mean in guide quality and run quality; off-target events are Poisson with a rate that falls with guide specificity but rises with editing activity and a permissive run. A "clean edit" requires the on-target rate to meet the target and the off-target count to stay at or below the limit, in the same cell.
For guide g5, the deterministic plug-in sits at its mean of (72%, 1.9 off-target events) — comfortably inside the acceptance box. But the cloud of replicate outcomes spreads across all four quadrants. The probability the on-target rate meets the 70% target is 60%; the probability the off-target count stays at or below 2 is 69%; and the probability both hold in the same cell — the acceptance region — is only 36%. The point estimate lands in the box; most of the distribution does not.
Screening the candidate library reframes the decision. Plotting each guide's on-target pass-rate, off-target pass-rate, and joint clean-edit probability shows that climbing the on-target axis is not what separates a usable guide from an unusable one.
The best guide is g6, with a 67% clean-edit probability — 30 percentage points above the team's favourite g5. The off-target hurdle, not the on-target hurdle, is what gates every guide: at each design level the off-target pass-rate is the binding constraint, and the guides that win do so by being more specific, not merely more active.
The off-target count for g5 is Poisson-shaped with a deterministic mean of 1.9 events, sitting just below the limit of 2. But the count is integer and right-tailed: the probability of landing above the safety limit is 31%, and the probability of staying at or below it is 69%. Reporting "1.9 off-target events on average" hides that nearly one cell in three exceeds the safety threshold — the single fact a safety reviewer most needs.
The tornado ranks the levers on the clean-edit probability around g5's 36% baseline: the on-target acceptance target swings it from -23 to +20 points, guide design quality from -19 to +21, the off-target safety limit from -14 to +11, and run-to-run variability only -5 to +5. The criteria themselves and the guide choice dominate; reagent consistency is a distant fourth.
A guide that edits 72% on target with 1.9 off-target events looks clean on its averages. Simulating the joint distribution shows a 36% clean-edit probability and a better candidate sitting 30 points higher — the difference between selecting a guide on its margins and selecting it on its odds.