Extended Data Fig. 1: Behavioural performance.

(a) The participants responded more slowly and committed more errors in the ‘mirror task’ than in the ‘gender task’ (combined data from E1-E6 versus combined data from PRE and POST; P = 5.96 × 10⁻⁵ and P = 5.95 × 10⁻⁵, respectively; n = 21 sessions; Wilcoxon signed-rank tests; two-sided). (b) During exposure, the outer nodes could never immediately follow one another. If the participants extracted this rule, outer-after-outer trials (OaO) during POST should be unexpected and thus related to longer reaction times (RTs) than outer-after-inner trials (OaI). Indeed, we found that the difference between RTs in OaO minus OaI trials increased in POST compared to PRE (P = 0.0445; 10,000 permutations of PRE and POST labels). The plot shows means ± s.e.m. Circles correspond to datapoints from individual sessions (n = 21). The P-value was calculated as a number of permutations with a higher difference than the one actually detected, divided by the total number of permutations. (c) The above-mentioned behavioural conflict was especially pronounced among the patients who developed a robust hippocampal-entorhinal representation of the pyramid, calculated per participant as similarity between neuronal population responses and the successor template in POST (see Fig. 3g; ρ₁₉ = 0.53; P = 0.0077; Spearman correlation with 10,000 permutations of the session order). Another index of behavioural conflict—RTs in trials after ‘indirect’ transitions—was also positively correlated with the strength of the hippocampal-entorhinal successor representation (indirect trials: POST minus PRE; ρ₁₉ = 0.39; P = 0.04; Spearman correlation with 10,000 permutations of the session order). In contrast, RTs in trials without conflict (i.e., ‘direct’ and OaI transitions; POST minus PRE) did not significantly correlate with the strength of the hippocampal-entorhinal successor representation (ρ₁₉ = 0.32; P = 0.0814 and ρ₁₉ = 0.21; P = 0.173, respectively). All RT analyses were performed on correct trials only. Very short (<200 ms) and very long (> 5000 ms) RTs were discarded. The above P values were calculated as a number of permutations with a higher correlation coefficient than the one actually detected, divided by the total number of permutations. (d) Twenty-five healthy controls (see Methods) completed the same behavioural procedure as the patients. They were then asked an open question: ‘Have you noticed any pattern in the sequence of images?’ None of the participants reported noticing a graph-like organization of the sequence. Then, we informed them about the underlying structure and asked them to assign each image to a specific node (the ‘positions’ task). The pyramid has six variants (three rotations and two flips). The ‘positions accuracy’ was calculated as the maximum number of hits. For example, if someone’s highest score was three out of six for one variant and less than three hits for other variants, this person’s accuracy score was 50%. To calculate the ‘links accuracy,’ we checked whether each pair of images was linked directly or indirectly on the graph provided by each participant and compared it to the actual pyramid. Similarly, we calculated Spearman correlation coefficients (Fisher-transformed) between pairwise distances provided by each participant and the actual pairwise distances (‘distance similarity’ index). Finally, we checked how often the participants assigned the correct images to the inner versus outer nodes (‘inner-outer accuracy’). For each index, the chance level was estimated as the mean performance of 10,000 randomly generated ‘participants’ (red dashed line). To establish the ‘explicit benchmark’ (blue dashed line), we tested another eight control participants (see Methods). From the beginning, we informed them that the sequence of images during exposure phases will follow the pyramid graph, but we did not explain which image is located where on the graph. All other aspects of the procedure and analysis were the same as explained above. The mean performance of this additional group served as the explicit benchmark. We found that ‘positions accuracy’ did not significantly differ from chance level and was significantly below the explicit benchmark (P = 0.1584 and P = 8.54 × 10⁻⁶, respectively). Other indexes, arguably referring to less detailed knowledge of the graph, were significantly above chance level, but still below the explicit benchmark (‘links accuracy’ and ‘distance similarity’ versus chance: P = 0.007; ‘links accuracy’ and ‘distance similarity’ versus explicit: P = 7.04 × 10⁻⁶; ‘inner-outer accuracy’ versus chance: P = 0.0186; ‘inner-outer accuracy’ versus explicit: P = 0.0004). All the above P values are from the Wilcoxon signed-rank tests (two-sided). Together, these results suggest that the healthy control participants (and patients) did not have detailed explicit knowledge of the pyramid. The central marks of the box plots (panels a and d) indicate the medians. The bottom and top edges of the boxes indicate the 25th (Q1) and 75th (Q3) percentiles, respectively. The whiskers extend to the most extreme data points not considered outliers (1.5 × interquartile range above Q3 or below Q1).