Validation · Colocalization Benchmark Source
The tool recovers known colocalization to within half a percent
We ran the Colocalizer against the Colocalization Benchmark Source —
10 computer-simulated 1024×1024 images whose true colocalization is known exactly, from 0% to 90%.
Nothing was tuned to fit; the tool computed each coefficient blind and we compared to ground truth.
0.5%
mean error of Pearson r vs. known colocalization
0.9995
R² of Pearson r against ground truth
10 / 10
levels ranked in perfect order (Spearman 1.000)
Measured Pearson r vs. known colocalization
Each point is one benchmark image. If the tool were perfect, points sit on the dashed identity line. The background-excluded value at 0% dips marginally below zero — expected when there is no real correlation.
Pearson r (whole image)
Pearson r (background excluded)
perfect agreement
Bottom line. On known-truth data the whole-image Pearson coefficient landed on the identity
line — slope 0.999, intercept −0.001, mean error 0.5% and worst-case 1.5% across the full 0–90% range. That is
as good as this benchmark gets, and it means the numbers the tool reports can be trusted to reflect real colocalization.
Per-image results
| Image | Known % | Pearson r | Pearson (no bg) | MOC | |error| |
|---|
MOC = Manders' overlap coefficient. "Pearson (no bg)" excludes below-threshold background pixels (Costes).
What we learned about background removal
Reporting Pearson with and without a Costes background threshold is built in.
The benchmark showed something important about when each is right:
- Whole-image Pearson is the most accurate here (0.5% error). These simulations spread signal across the
whole frame, so "removing background" mostly removes real signal.
- Background-excluded Pearson still ranked every level perfectly (R² 0.9996) but read slightly lower,
because it drops ~8% of pixels as background.
- On real cells — a dark field with bright puncta — excluding background is often the more meaningful
reading, which is exactly why the tool shows both and lets you switch with the Ignore background checkbox.
One honest caveat. A colleague might ask about the Manders M1/M2 numbers: on these
particular synthetic images they read high even at 0%, because Manders is threshold-sensitive and these images have almost no
true "empty" background for the automatic threshold to latch onto. Manders still ranked the levels in perfect order. For a
single trustworthy number, Pearson r is the metric to lead with — standard advice, and the benchmark backs it up.
On “picking the channels for the best R.” Pearson's r is symmetric — swapping which
channel is red vs. green gives the identical value — so there is no assignment to optimize, and deliberately choosing whichever
pairing maximizes R would bias results upward. The legitimate way to get a cleaner R is background exclusion, above.