Digital PCR Plate Map and Likelihood Curve (2D)

Every droplet is bright or dark; the Poisson likelihood of the sample concentration is drawn next to the plate. Educational model.

Switch to 3D

Readouts

Presets

Experiment

Left: the plate seen from above (Blender-rendered chip and droplets). After thermocycling, green droplets held at least one DNA copy; with "Reveal" the ones that held 2 or more turn orange. Right: ln L of the concentration given the bright count; its peak is the corrected estimate.

Independent numerical checks

The science

Digital PCR splits a sample into N partitions (about 1 nL each in a typical droplet system; a typical value quoted from memory). If molecules are far apart each partition receives 0, 1, 2, ... copies at random. After thermal cycling a partition holding at least one copy lights up; the instrument reports only how many are bright. Molecules are placed independently, so the count per partition is Poisson with mean λ: P(k) = λk e−λ/k!. A partition is dark with probability e−λ, so the expected bright fraction is p = 1 − e−λ and, inverting, λ̂ = −ln(1 − p̂). The naive fraction under-reports copies because droplets with two or more copies are counted once. Concentration is λ̂/V.

The likelihood curve (this view). With x bright droplets out of N the log-likelihood is ln L(λ) = x ln(1 − e−λ) − (N − x) λ. Its maximum is exactly the corrected estimate, its curvature there is N(N − x)/x (observed Fisher information), so sd(λ̂) = √((eλ − 1)/N), and the values of λ whose ln L is within 1.92 of the maximum form the 95 % likelihood-ratio interval. The Wilson-score interval (mapped through −ln(1 − p)) is drawn as a band for comparison; they agree closely except for very few bright or dark droplets. If every droplet is dark (x = 0) the curve only falls and the data give an upper bound; if all are bright it only rises and there is no finite estimate.

The relative error √(eλ − 1)/(λ√N) is smallest where λ eλ = 2(eλ − 1), λ ≈ 1.59.

Checks. Brute-force droplet filling against 1 − e−λ; the mean of the estimate over 400 repeats against the true λ (a small positive bias is included in the tolerance); the repeat-to-repeat spread against the delta-method formula; coverage of the Wilson interval; the optimal λ scan against a numerical root; and, new for this view, a golden-section search of ln L against the closed form, the numerical curvature at the peak against N(N − x)/x, and the likelihood-ratio interval width against the Wilson width.

Not modelled. False positives and negatives, droplets of unequal size, intermediate-brightness "rain", incomplete amplification, fragmentation and linked targets. Every molecule is amplified perfectly and all droplets are identical. A research instrument often uses about 20,000 droplets; the slider stops at 2,500 so that all can be drawn. Colours are for illustration. Educational model only; not for laboratory or clinical use and not medical advice.

Chip plate and droplet sprites: Blender (Cycles) renders of the 3D page's models.

Expected relative error of the estimate against λ for your N (dashed green: optimum 1.59)
Copies per droplet: this run (blue) against the Poisson prediction (yellow)
Made with 🟠 Blender