Experimental overview

How the underlying experiment was run, and how it was turned into the analyses on this site.

The experiment

One field of live cells was imaged by brightfield time-lapse (live-cell imaging, "LCI") for 65 timepoints (Time00000–Time00064) at 3.25 µm/px (6.5 µm camera pixel through a 2.0x objective). Partway through the movie, the sample was taken off the scope, treated with fixative, and later returned for Cell Painting imaging.

Fixation boundary: Time59 → Time60. This was measured empirically, not assumed — total detected-cell count drops from 4,763 to 3,898 with a ~8px global shift at exactly that transition. Fixation isn't instantaneous: the slide leaves the scope, the fixative acts over roughly a few minutes, and the slide goes back on — so this single step reflects a real elapsed-time gap and a mechanical repositioning, not just one more frame of normal motion. Analyses of live motility are restricted to frame_index <= 59; the post-fixation frames (60–64) are used only to locate the terminal, fixed cell positions.

After fixation, the same field was stained and imaged for Cell Painting across 110 fields of view (.ims files) at a higher resolution, 0.3015 µm/px — roughly 10x finer than the LCI pixel scale. Six channels were acquired per FOV:

Channel	Stain	Target
0	405	DNA
1	637	Mitochondria
2	561	WGA / AGP (actin, Golgi, plasma membrane)
3	514	SYTO14 (RNA)
4	488	ConA (ER)
5	—	Combo channel, not used in analysis

This is the standard 5-stain Cell Painting panel (DNA / RNA / AGP / Mito / ER).

Turning images into cell tables

LCI segmentation. Every one of the 65 brightfield frames was segmented with Cellpose into per-cell masks, converted to polygons (one GeoDataFrame per frame, ~4,000–4,700 cells/frame) with a centroid per cell.
LCI tracking. Cells were linked frame-to-frame into motility trajectories: first by largest polygon overlap (one-to-one), then a shift-compensated nearest-centroid fallback for cells that moved without overlapping their previous selves, then a short gap-bridging pass for single-frame segmentation dropouts. ~12,000 trajectories result, median length 8 frames, ~29% reaching the final frame.
Cell Painting segmentation & features. CellProfiler segmented nuclei/cells in each of the 110 FOVs and measured ~499 features per cell (shape, intensity, texture, radial intensity distribution, granularity) across the 5 real stains.
Alignment. For each CP FOV, a similarity transform (scale + rotation/ reflection + translation) was fit mapping that FOV's local pixel coordinates into the LCI final-frame (Time64) coordinate space, via an interactive landmark-matching widget.
Linking. Every CP cell was matched to its nearest motility trajectory endpoint in that shared coordinate space (median match distance 2.5 px, 15 px threshold) — 2,130 linked 1:1 pairs, ~80% with usable live-motility metrics (the rest reach the final frame but have no earlier live-frame track).
Motility metrics. For each trajectory, net displacement, path length, straightness, mean per-frame speed, and heading were computed over the live frames only (<= Time59), explicitly excluding the fixation-transition frame.

What's on this site

The three "live, in your browser" pages are direct views of that pipeline's output: the motility scrubber plays back all 65 frames with tracked-cell paths; the linked view shows the final-frame motility panel side-by-side with the stitched Cell Painting mosaic, synced and click-to-highlight by the shared link_id; the clustergram page visualizes the ~499-dimensional CP feature space (via Celldega) and its correlation with motility.

The five NF pages go deeper on specific questions: is motion actually more directed than a random walk built from real observed steps (NF01); does live shape dynamics predict speed better than any fixed Cell Painting feature (NF02); does local crowding slow cells down, and do neighbors move in correlated directions (NF03); which Cell Painting ↔ motility correlations actually survive multiple-testing correction (NF04); and how often do cells physically touch, and does that redirect them (NF05).