NotebookEntry 06

NTA, simulated

A Monte Carlo of the whole nanoparticle tracking measurement: which vesicles the camera sees, how they jitter, which tracks survive, and how far the reported sizes drift from the truth.

  • Python
  • Monte Carlo
  • Script included
  • Made with Claude
  • 26 September 2026
Three-panel figure titled How faithfully does NTA recover an EV size distribution. A: two-second random walks for 70, 120 and 250 nm vesicles; the smallest wanders farthest. B: size histograms for the true population, the detected subset and the NTA-recovered sizes; the recovered curve is flatter with a long tail past 250 nm. C: recovered over true diameter against track length; tracks under about 20 frames scatter from a quarter to eight times the true size, and long tracks settle near one.
Fig. 1One run of the simulation, seed 2026. Select the figure to open it full size.
Download the script

Needs Python 3 with numpy, scipy and matplotlib. Runs in about a second.

The question

Nanoparticle tracking analysis sizes extracellular vesicles by watching them jitter: smaller particles diffuse faster. But the camera only sees particles that scatter enough light, the software drops short tracks, and every recorded position carries some noise. How far do the reported sizes drift from the truth?

What the model does

It builds a known population of 4,000 vesicles (75% near 90 nm, 25% near 190 nm) and decides which ones the camera detects. It simulates each detected particle's Brownian track frame by frame with localization error, drops tracks shorter than ten frames as NTA software does, fits a diffusion coefficient to every track that survives, and converts that back to a diameter with the Stokes-Einstein equation.

What it shows

The reported distribution comes out broader and shifted up: the median moves from 102 nm to 126 nm. Detection alone accounts for about a third of that shift, because the camera misses small, dim vesicles. The rest comes from short tracks. With only 10 to 14 frames, a track can put a vesicle at anywhere from half to four times its real diameter; past 50 frames the error mostly settles.

The model leaves out a few things real instruments add, such as motion blur and particles drifting out of focus, so a real measurement has more ways to go wrong than this one.

Claude wrote and ran the script. It pairs with Fig. 1 on the home page, which simulates the same measurement live, and with the NTA sizing I ran on EV preparations this summer.