Extended Data Fig. 5: Behaviour state model fits animals across models and treatments.

(a) Cartoon of unfolded transition matrices between six spontaneous behaviours (top), which are k-means clustered in a 36-dimensional space. Model centroids are described by 36 dimensional points (unfolded transition matrices) in that state space (middle). Behavioural trajectories can be plotted in this state space, and portions of this trajectory are classified as belonging to a state by calculating the nearest centroid (Euclidean distance, bottom). (b) 100-fold cross-validated models with k = 6 clusters maximize silhouette score and represent the “elbow” of the sum of squared distances. (c) Fit across full model and control models (left; permutation tests: p < 0.0001 for each condition compared to n = 100 shuffles) and fraction of matching classification across control models (right; permutation tests: p < 0.0001 for each condition compared to n = 100 shuffles) in the pain model animals used as training data. (d-f) Fit across full model and control models in uninjured (e), capsaicin (f), and formalin (g) mice across morphine doses (Permutation tests: p < 0.0001 for each condition and treatment compared to n = 100 shuffles). (g,h) Fit across full model and control models in mice expressing either MORp-YFP (left) or MORp-hM4Di (right) before and after spared nerve injury (SNI) with DCZ administration (Permutation tests: p < 0.0001 for each condition compared to n = 100 shuffles). (i) Log-transformed transition matrices between states. (j,k) Bouts (j) or mean duration (k) per session of each state in capsaicin (cyan), formalin (magenta), and SNI (dark red) animals (One-way RM ANOVA, Tukey correction, pState, bout <0.0001, pState, duration <0.0001). ⋆ = p < 0.05. Dots are mean; error bars are SEM. See Table S1 66–74 for statistics.