Luria-Delbrück Fluctuation Test (2D)

Plate map of independent cultures, the clone timeline of the one you click, and the jackpot distribution. Educational model.

Switch to 3D

Readouts

Presets

Experiment

Left: one dish per independent culture (Blender-rendered dish and colonies; at most 40 colonies drawn per dish, the numbers use the true counts); the ring colour shows the count class. Click a dish to see its clone timeline on the right: every bar is a mutation event that started a clone in that generation, and an early event gives a huge clone (a jackpot).

Independent numerical checks

The science

In 1943 Luria and Delbrück asked whether bacteria become resistant to a virus because of contact with it (induced) or because rare mutants already exist before contact (spontaneous). They grew many small independent cultures and plated each. Induced resistance gives counts clustered around a mean (Poisson, variance = mean). Spontaneous mutations at random times during growth give a very skewed distribution: a mutation early in a culture's growth leaves a large clone, so a few cultures are jackpots and many have none. The data matched the spontaneous picture.

Model. Each culture starts from one non-mutant cell. In every generation each of the W non-mutant cells divides, and with probability μ a division yields one mutant daughter (mutants breed true and never revert). With M mutants and X new mutant daughters, X ~ Binomial(W, μ) (a Poisson or normal draw when W is large), W′ = 2W − X, M′ = 2M + X, for G generations, so Nf = 2G. A mutant daughter born in generation g (0 to G−1) has 2G−1−g descendants at plating, so M = Σ Xg 2G−1−g: this is what the timeline draws. The expected number of events per culture is m = μ(Nf − 1) and the expected number in generation g is μ(2 − μ)g, so most events happen in the last generations but the rare early ones dominate the count.

Estimators. A culture has no mutants exactly when no event happened: P0 = (1 − μ)Nf − 1 ≈ e−m, so m̂ = −ln(fraction with no mutants) and μ̂ = m̂/Nf; it fails when no culture is empty. The Lea-Coulson median method solves r/m − ln m = 1.24 for the median count r (approximate, continuous growth). The sample mean converges very slowly (the expected mutants per culture, about m ln Nf/(2 ln 2), is dominated by jackpots), which is why P0 or the median is used.

Checks. Events per culture against the exact recursion; empty fraction against (1 − μ)Nf−1; exact enumeration of the whole Markov chain (G = 6, μ = 0.04) against the simulation code (total-variation distance) and against an explicit Bernoulli-per-cell version; the P0 estimate averaged over 400 repeated experiments; the Lea-Coulson estimate (wide tolerance). New for this view: the event-recording culture reproduces the 3D culture exactly with the same random draws and its clone sizes sum to M; and the events per generation in your experiment against μ(2 − μ)g (standard error taken from the spread across cultures). Rows say n/a instead of passing when the numbers are too small to compare.

Not modelled. Plating efficiency below 100 %, differential growth of mutants, back-mutation, cell death, lag phase, non-synchronous divisions, phenotypic delay and later refinements (Ma-Sandri-Sarkar estimators). Mutation rates 10−9 to 10−3 per division bracket typical spontaneous values (typical numbers from memory). Educational model only; harmless bacteria and plates are pictured and no laboratory protocol is given.

Dish and colony sprites: Blender (Cycles) renders of the 3D page's plates.

Share of cultures by mutant count (log bins): this experiment (blue) against Poisson with the same mean (yellow)
Running P0 estimate of m divided by the true m (1 is perfect)
Mutation events per generation: observed (bars) against expected mu(2-mu)^g C (line)
Made with 🟠 Blender