Extended Data Fig. 10: Hippocampal neurons encoded another complex temporal structure with longer paths.

(a) In a separate study, we tested five additional patients (seven recording sessions; see Methods). The procedure was the same as before, but we removed two edges from the pyramid graph during exposure. The resulting ‘diamond’ structure had the ‘shortest path’ distances of length one (D1; corresponding to the ‘direct’ category in the main study) or lengths two and three (D2 and D3, respectively, corresponding to the ‘indirect’ category in the main study). We recorded 221 neurons, of which 55 were located in the hippocampus (we did not have any recording sites in the entorhinal cortex in this additional study). Six hippocampal neurons responded preferentially during PRE to stimuli at the most distant nodes of the diamond (black circles) and were relational neurons according to the criteria described earlier (see Methods). (b) A relational neuron from the left hippocampus showing progressive tuning to the graph’s distances. Raster plots show individual spikes during each stimulus presentation in E1&2, E3&4, and E5&6. Line plots show the mean number of spikes ± s.e.m. (PRE included for reference). The left panel shows the stimuli and their locations on the graph. (c) During late-exposure phases, these hippocampal relational neurons responded more strongly to images located two edges away from their preferred stimulus than to images located three edges away (Wilcoxon signed-rank tests against E1&2 or zero; one-sided). Each circle corresponds to one neuron (the average D2 minus D3 difference in the 0.1 to 1.3 s time window, peak-normalized and baseline-corrected). For the definition of box plots, please see the legend of the Extended Data Fig. 1. (d) We replicated the main population decoding finding that neuronal representations of adjacent nodes progressively overlapped (see Fig. 3c). That is, the probability of decoding the actual stimulus as ‘actual’ was gradually decreasing, the probability of decoding the direct stimuli as ‘actual’ was increasing, and the probability of decoding indirect stimuli as ‘actual’ did not change significantly. This analysis was conducted on all hippocampal neurons from this additional study (n = 55) during all stimuli presentations. Please note that the indirect category combines D2 and D3. P values obtained from Kolmogorov-Smirnov tests comparing CDFs in PRE versus the subsequent study phases (one-sided). Exact P values for ‘actual’ E1&2&3: P = 1.06 × 10⁻⁷; E4&5&6: P = 1.94 × 10⁻⁷. (e) Next, we analysed CDFs for nodes separated by two versus three links away from the actual stimulus. This analysis was conducted only for trials where the actual stimulus was located at one of the ‘black nodes’ of the diamond (see panel a). We found that the probability of decoding D2 stimuli as ‘actual’ gradually increased during the study, while the probability of decoding D3 stimuli as ‘actual’ gradually decreased. P values obtained from Kolmogorov-Smirnov tests (two-sided). Exact P-value for ‘distance 3’ E4&5&6: P = 3.69 × 10⁻⁶.