NF05 · Collisions ("bumper cars"): does contact redirect cells?

Cellpose partitions the pixel grid, so adjacent cells' polygons genuinely touch (zero-area boundary contact) rather than needing an arbitrary distance threshold — the literal "bumper cars" collision event. This notebook counts every such contact across the live movie (all detected cells, not just tracked/linked ones), then asks the classic contact inhibition of locomotion question: does a cell change heading more sharply right after touching a neighbor than it does at a random moment?

1 · Collision count, all detected cells, all live frames

96,408 total collision (contact) events across 60 live frames

2 · Do cells turn more sharply right after a collision?

86,641 distinct (trajectory, frame) collision participations
collision moments: n=82,133, median |turn|=112.8 deg
non-collision moments: n=100,473, median |turn|=111.6 deg
Mann-Whitney U (collided > control): U=4177466310, p=2.26e-06

3 · Collision rate vs. time (density context) and per-cell collision count vs. morphology/CP

n_collisions vs. LCI shape-over-time features (all live trajectories):

	feature	rho	p
0	area_mean	0.184488	8.515860e-35
2	perimeter_mean	0.183974	1.310572e-34
8	equivalent_diameter_mean	0.183528	1.902668e-34
7	elongation_cv	0.167238	6.211724e-28
3	perimeter_cv	0.142216	1.425073e-20
4	solidity_mean	-0.136216	1.436563e-19
5	solidity_cv	0.134500	1.506579e-18
10	form_factor_mean	-0.098008	8.227222e-11
1	area_cv	0.098853	1.147190e-10
11	form_factor_cv	0.096145	3.639254e-10

n_collisions vs. CP morphology (2,130 linked cells):

	cp_feature	rho	p
60	Intensity_MeanIntensityEdge_OrigActin_Golgi_Me...	0.162542	4.438199e-14
85	Intensity_StdIntensityEdge_OrigActin_Golgi_Mem...	0.158044	2.202707e-13
50	Intensity_MaxIntensityEdge_OrigActin_Golgi_Mem...	0.153157	1.191898e-12
61	Intensity_MeanIntensityEdge_OrigDNA	0.147978	6.729138e-12
51	Intensity_MaxIntensityEdge_OrigDNA	0.145910	1.320938e-11
76	Intensity_MinIntensityEdge_OrigDNA	0.143134	3.218226e-11
143	RadialDistribution_RadialCV_OrigActin_Golgi_Me...	0.140812	6.689030e-11
81	Intensity_MinIntensity_OrigDNA	0.138635	1.313835e-10
26	Intensity_IntegratedIntensityEdge_OrigDNA	0.130497	1.494063e-09
35	Intensity_LowerQuartileIntensity_OrigActin_Gol...	0.108276	5.472850e-07

---

⭐ Key finding

(Corrected: an earlier pass accidentally included post-fixation frames in the turning-angle and shape/collision-count analyses -- now explicitly filtered to frame_index <= MOTILITY_LAST_FRAME. Also fixed a missing nan_policy="omit" -- single-frame trajectories give NaN coefficient-of-variation, which silently NaN'd whole Spearman correlations without it.)

96,408 genuine polygon-contact collisions across the 60 live frames (all detected cells, not just tracked/linked ones) — cells touch neighbors constantly in this dense monolayer.

Weak evidence for contact-inhibition-of-locomotion: across 82,133 collision moments vs. 100,473 non-collision moments, median |turning angle| is slightly higher right after a collision (112.8° vs. 111.6°, Mann-Whitney p=2.3e-06). Statistically robust given the large n, but the effect size is small against an already very noisy single-step baseline (median turn ~112° either way — frame-to-frame tracking is inherently jittery at this spatial scale, per NF01). So: collisions nudge cells to turn slightly more, but this is not a dramatic bounce-off effect.

Collision count correlates with irregular morphology: cells with more lifetime collisions have larger, more size-variable shape (area_mean ρ=+0.18, elongation_cv ρ=+0.17, both p<1e-27, n=4,375) and are less solid/convex (solidity_mean ρ=−0.14) — consistent with NF03's crowding-slows-speed finding and NF02's shape-variability-tracks-speed finding: frequently- colliding cells look like the same "protrusive/irregular" phenotype already implicated in slower, more crowded regions. On the CP side (2,130 linked cells), collision count correlates most with edge-intensity features (Intensity_Mean/MaxIntensityEdge_OrigActin_Golgi_Membrane, ρ≈0.15-0.16) — plausibly the same irregular-boundary signature showing up in the fixed CP image.

Caveat: collision detection depends on segmentation boundary precision -- boundary jitter (see NF01's noise-floor check) could create spurious touching/separating events between frames without any real change in physical contact, inflating or deflating collision counts independent of true "bumper cars" behavior. Not yet quantified here.