Fig. 2.

Testing NuRIM’s theoretical accuracy using simulated data. (A) Point-like NPCs (green solid dots) and reference fiduciary markers (red open dots) are distributed on the NE, itself modeled as a sphere. Variable levels of confounding fluorescence are introduced inside and outside the NE (blue open dots). (B) Convolution with a realistic point spread function (PSF) yields 3D diffraction-limited stacks (a 3D rendering of the NE is shown). (C and D) Single Z slices sampled from the simulated image volumes. Small transverse shifts in the distributions of NPCs away from the NE lead to measurable subpixel shifts in the simulated images (red and green dashed circles). (E) Over 10 million simulated images were generated under variable background conditions. The plot shows the error made by NuRIM when recovering the ground-truth shift for an NE signal level corresponding to nucleoporins. Black dots represent average shift error for batches of 64 simulated images. The fitting surface was obtained using neural network training, thus yielding estimates of bias error in any background conditions. These small bias errors are subsequently subtracted from experimental values to obtain adjusted positions (SI Appendix and SI Appendix, Fig. S1). (Scale bars, 1 μm.)