Extended Data Fig. 6: Stability of neuronal encoding across animals, sessions and assay.

a, Normalized proportion of neuronal encoding competitive success, reward size, and relative rank across animals. b, An animal-dropping procedure revealed no difference in peak decoding performance for competitive success, reward size, and relative rank across animals (F_(6,1399)=1.27, p=0.20 for competitive success, F_(6,1399)=0.89, p=0.69 for reward size, and F_(6,1399)=0.44, p=0.98 for relative rank; one-way ANOVA). c, Population firing rates and proportion of neurons encoding competitive success were stable across trials within sessions (n=63 sessions across 7 animals; F_(1,116)=1.24, p=0.27 for firing rate; F_(1,125)=0.30, p=0.59 for proportion of encoding neurons; two-way RM-ANOVA). d, Population firing rates and proportion of neurons encoding relative rank were stable across trials within sessions (n=63 sessions across 7 animals; F_(1,116)=0.11, p=0.74 for firing rate; F_(1,125)=0.013, p=0.909 for proportion of encoding neurons; two-way RM-ANOVA). e, Population firing rates and proportion of neurons encoding competitive success were stable across sessions (n=63 sessions across 7 animals; F_(7,55)=0.67, p=0.76 for firing rate; F_(7,55)=0.758, p=0.679 for proportion of encoding neurons; two-way RM-ANOVA). f, Population firing rates and proportion of neurons encoding relative rank were stable across sessions (n=63 sessions across 7 animals; F_(7,55)=0.79, p=0.64 for firing rate; F_(7,55)=1.709, p=0.098 for proportion of encoding neurons; two-way RM-ANOVA). g, Top, Graphic showing the two-chamber arena in which the recorded animal was placed with cagemates following each group foraging session. Bottom, Heat map and trajectories of both animals during a representative trial. h, Animals spent similar amount of time investigating the other animal regardless of their hierarchical rank relative to the recorded animal (χ²_(5,304)=5.63, p=0.34; Kruskal-Wallis), but spent significantly less time investigating inanimate totems (Z=4.13, p=3.64x10^-5; Rank-sum). Dots represent trials (n=376) across n=58 sessions. Box-plot edges represent 25^th/75^th percentiles with centre=median and whiskers=1^st-99^th percentile range. i, PETH and spike raster plots of two representative neurons that displayed changes in firing rate based on the animals’ relative rank (high vs. low social rank in relation to the other animal). PETHs are aligned to the time point at which the recorded animal initiated an investigation of the other animal. Purple dots represent the end of an investigation. Shaded area denotes s.e.m. j, Top, Representative neuron during wireless recordings of the group foraging task (10 min after start of a session) compared to that of the two-chamber assay (3 h after start). Shaded area denotes s.e.m. Bottom, Venn diagram depicting the number of neurons encoding relative rank (p<0.01; Wilcoxon signed-rank) during the two-chamber assay and their overlap with those encoding relative rank during the group foraging task (n=115 overlap; χ²₍₁₎=42.17, p=8.35x10^-11; Chi-Square test). k, Correlation of normalized firing rates for neurons that encoded relative rank during the group foraging vs. two-chamber assay on a cell-by-cell basis (n=115 RR-encoding of n=174 total rank-encoding neurons; r=0.673, p=2.85x10^-24; Pearson correlation). Grey line depicts linear line of best fit. l, Decoding accuracy of relative rank during the two-chamber assay (n=942) was significantly higher than chance (p=0.0024, Permutation test) but lower than the peak decoding accuracy during the group foraging task (*p=0.013, Permutation test). m, SVM models trained on neuronal activity recorded during group foraging was used to decode validation data recorded during the two-chamber setup (i.e., switch model) or vice versa. Decoding accuracy for both switch models were significantly higher than expected from chance (p=0.0031 for group foraging model used to decode two-chamber data, p=0.021 for two-chamber model used to decode group foraging data; Permutation test). Overall, the group foraging model was better at decoding two-chamber rank rather than vice versa (p=0.026, Permutation test). Error bars denote mean±95%CI. N=500 bootstrapped samples for all decoding results.

Source data