Fig. 4: Extending NP-Cloud and RR-NP Cloud to 3D drift correction.

a Zoom-in view of a small region of the simulated 3D-SMLM dataset of 66,000 frames with features of 100 clusters/µm², without drift (left), after the addition of 3D-drift (center), and after drift correction by RR-NP Cloud-3D (right). Color presents the z values. b Comparison of the ground-truth z drift curve (black) with those calculated from the simulated 3D-SMLM dataset using NP-Cloud-3D, RR-NP Cloud-3D, and AIM-3D at the same segment size of 50 frames/segment (green). c Distributions of residual z drifts of n = 5000 randomly sampled frames from the drift correction results, calculated as the distance between the estimated drift and the ground truth, for the three methods at different segment sizes. Dots: residuals of individual frames; box boundaries: the first and third quartiles; center lines: medians; small squares: means; whiskers: SD. Note that AIM-3D failed for the 30 frames/segment condition and did not generate meaningful drift correction. d RR-NP Cloud-3D calculated z drift for the experimental 3D-STORM data of dye-labeled antibody adsorbed to a coverslip, in which z was stabilized by the microscope focus lock. e Histograms: Distribution of z values for single-molecule localizations before (left) and after (right) 3D drift correction. Red curves: fits to normal distribution, yielding FWHM of 57 and 55 nm, respectively. f, g Similar to d, e but with the focus lock disabled, and the focal plane was continuously adjusted manually during imaging. Fits to normal distribution yield FWHM of 75 and 55 nm before and after RR-NP Cloud-3D drift correction. h RR-NP Cloud-3D calculated z drift for the experimental 3D-STORM data of immunolabeled mitochondrial outer membrane (TOM20) in a fixed COS-7 cell. When collecting this dataset, the focus lock was disabled, and the sample was allowed to continuously drift in z by ~200 nm each time before the focus was manually adjusted to bring the sample back. i Virtual vertical (xz) cross sections for a region of the data (marked in Supplementary Fig. 12), without drift correction (left), after only in-plane (xy) drift correction through RR-NP Cloud (middle), and after 3D drift correction through RR-NP Cloud-3D (right).