NF03 · Local crowding & collective motion¶

An earlier exploratory draft (DV-motility_metrics_sanity_check.ipynb) tested "isolated vs. crowded" neighbor-density groups against many columns, but only on the 2,130 cells that link to Cell Painting. Here we compute local density for every detected cell in every live frame (~300k cell-frame rows, no tracking/linking/final-frame restriction) via a per-frame KD-tree, and ask: does local crowding predict instantaneous speed? Bonus: is there evidence of coordinated ("flocking") motion between nearby cells?

1 · Local density, all detected cells, all live frames¶

282,178 detected cells across 60 live frames
count    282178.000000
mean          4.303425
std           2.855956
min           0.000000
25%           2.000000
50%           4.000000
75%           6.000000
max          22.000000
Name: neighbor_count, dtype: float64

2 · Density vs. instantaneous speed, all tracked cells (continuous, not just linked-to-CP)¶

186,840 tracked cell-frame steps with a matching density value
Spearman(neighbor_count, step_speed_um) = -0.091, p = 0, n = 186,840
No description has been provided for this image

3 · Isolated vs. crowded (generalized): Welch's t-test + BH-FDR¶

Reproduces the earlier draft's isolated (0 neighbors) vs. crowded (>=5 neighbors) comparison, but on the full live-trajectory population instead of just the 2,130 CP-linked cells, and against LCI-only features (motility metrics + the shape-over-time summary from NF02) since most cells here have no CP match.

isolated (mean nbrs <1): 222 | crowded (mean nbrs >=5): 1,425
13 / 26 features significant after BH-FDR
feature t p isolated_mean crowded_mean p_adj significant
22 equivalent_diameter_mean 7.805724 1.017015e-13 16.934110 15.485750 1.322120e-12 True
14 area_mean 7.820011 9.645215e-14 233.585522 196.304885 1.322120e-12 True
16 perimeter_mean 6.905334 2.868788e-11 59.916580 55.212229 2.486283e-10 True
12 path_length_um 4.126386 4.801246e-05 223.731975 190.446957 1.929361e-04 True
13 mean_step_um 4.115748 5.194434e-05 6.393794 5.070380 1.929361e-04 True
9 mean_step 4.115748 5.194434e-05 1.967321 1.560117 1.929361e-04 True
7 path_length 4.126386 4.801246e-05 68.840608 58.599064 1.929361e-04 True
25 form_factor_cv 3.618377 3.542329e-04 0.089441 0.075009 1.151257e-03 True
1 start_y 3.423717 7.233240e-04 1616.658481 1380.241328 1.880642e-03 True
3 end_y 3.439238 6.848829e-04 1617.471433 1380.223256 1.880642e-03 True
20 elongation_mean -2.713283 6.997272e-03 1.300884 1.337798 1.653901e-02 True
23 equivalent_diameter_cv 2.557181 1.107642e-02 0.118313 0.107121 2.399891e-02 True

Bonus · Collective motion: velocity correlation vs. distance¶

If nearby cells move in correlated directions (beyond what crowding-vs-speed alone shows), that's evidence of coordinated/"flocking" migration rather than independent movement.

No description has been provided for this image
bin_left_um bin_right_um mean_cos_similarity n_pairs
0 0 50 0.090447 20740
1 50 100 0.020828 106979
2 100 150 0.009791 167841
3 150 200 0.007869 230767
4 200 300 0.006245 641812
5 300 400 0.003431 901101
6 400 600 0.005177 2499409

⭐ Key finding¶

(Corrected: an earlier pass accidentally included post-fixation frames -- now explicitly filtered to frame_index <= MOTILITY_LAST_FRAME. This also exposed a second bug: once capped, frame_last becomes ~constant, and a zero-variance t-test's NaN silently corrupted the whole BH-FDR vector -- both are now excluded/guarded against.)

Crowded cells move measurably slower. Across 186,840 tracked cell-frame steps (all live trajectories, not just CP-linked ones), local neighbor count (150 µm radius) correlates negatively with instantaneous step speed: Spearman ρ = −0.091 (p ≈ 0) -- this number is unaffected by the fix since the density join already implicitly excluded post-fixation steps. The generalized isolated-vs-crowded comparison (222 isolated / 1,425 crowded trajectories, by mean neighbor count) confirms this at the trajectory level: isolated cells average 6.39 µm/frame vs. 5.07 µm/frame for crowded cells (both mean_step_um and path_length_um, BH-FDR significant). 13 of 26 LCI features tested survive BH-FDR -- isolated cells are also consistently larger (area_mean 234 vs. 196 px², equivalent_diameter_mean 16.9 vs. 15.5 px), suggesting crowding may restrict both spreading and movement (contact inhibition-like behavior), not just an incidental correlation.

Bonus — collective motion: velocity direction correlation between cell pairs is consistently positive and decays with distance (mean cos-similarity 0.090 at 0-50 µm down to ~0.003-0.005 beyond 300 µm, n up to 2.5M pairs at the largest bin) — weak but real local coordination in migration direction, on top of (and distinct from) the crowding-slows-speed effect above.

Caveat: this notebook's step-level signal-to-noise ratio is low (see NF01's caveats section -- post-fixation "should be motionless" steps measure ~95% of the live median step size), and the tracker's distance-fallback/gap-bridging links are more likely to mislink in dense regions. If mislinking is more common in crowded areas and produces artificially smaller apparent displacement, that alone could partly explain "crowded cells look slower" without real biology behind it. Not yet ruled out -- see NF01 for the proposed overlap-only-vs-fallback comparison.