Fig. 5: Transition to and from the blanched state.

a, Chromatophore-space trajectories from camouflages (disruptive (D), background N29; mottled (M), background N13) to blanched (B) in response to approaching visual stimuli (sepia219 is shown throughout, except in c, k and l; other animals are shown in Extended Data Figs. 8–10). The solid and dashed lines show the motion to and from the blanched state, respectively. b, The outwards and return paths of all trajectories (n = 17, sepia219). Note the slower returns. The colour shows the speed of pattern change. c, Blanching (blanch) trajectories are straighter (lower mean curvature, 75-PC space) than camouflage (camo.) transitions (Student’s t-test, camo. versus blanch, 3 animals each, P = 0.0017). d, Whole-mantle, ‘disruptive’ camouflage. Scale bar, 10 mm. e, Chromatophore segmentation in a cropped region (indicated by the yellow box in d) at the trial start, at maximum blanching and at the trial end, with corresponding marginal histograms (red). D and M, camouflage at trial onset as in a. Note that traces of the start and end patterns can be seen at the blanch timepoint. f, Hierarchical clustering of whole-mantle patterns during (top) and after (bottom) blanching reveals conserved subtrees (colours). The open circles show the trials in e (D (black); M (blue); n = 17 trials, cophenetic correlation = 0.26; P = 0.015, Mantel test). g, Chromatophore size over time (single trial), ordered by the time of recruitment during return from blanching. Scale bar, 4 s. h, Chromatophores coloured by recruitment-time rank (as in g), suggesting a non-random, compartmentalized sequence. i, Leiden clustering (17 trials) reveals six components (colours). j, The density distribution of the chromatophore mean rank over trials for each component (on the basis of i), showing a reliable sequence (Kruskal–Wallis, H = 67.3, P = 3.7 × 10⁻¹³), all pairs being significantly different (post hoc multiple hierarchical permutation tests, P = 0.001). The shading shows the binned s.d. k, The explained variance after reduction to components derived from the 200-PC baseline (unclustered), the individual trajectories (cluster by trial), the whole dataset (cluster all trials) and a shuffled dataset (shuffle). n = 3 animals per dataset. l, The ratio of explained variance between cluster by trial and cluster by all trials. Blanching components are more generalizable across trials (two-tailed permutation t-test, camouflage versus blanching, 3 animals each, P = 0.0475).