Cortisol administration impairs PI performance irrespective of incoming distance and availability of spatial cues

The analysis regarding the success of cortisol administration revealed elevated cortisol concentrations under cortisol compared to placebo treatment for the entire PI task (S2 Table, see also Materials and methods). In the PI task, we assessed measures for incoming distance, PI performance and navigational pattern during the experimental task (see Materials and methods). Incoming distance refers to the Euclidean distance between retrieval and goal location (Fig 1C). Overall PI performance was assessed by the drop error, i.e., the Euclidean distance between response location and goal location (Fig 1D). For the navigational pattern analysis, to investigate the specific role of the landmark under cortisol, we measured the distance between the goal and the spatial cue (goal-to-landmark distance), and the mean Euclidean distance of the moving participant from the landmark across all time points of the incoming phase (movement-to-landmark distance).

To investigate whether cortisol influences PI performance, we built a linear mixed model, in which we investigated effects of subtask (two levels: Pure PI versus Landmark PI), incoming distance, and treatment (two levels: cortisol versus placebo) on overall PI performance (i.e., drop error) on the level of single trials (see S3 Table for extended analyses). This analysis revealed similar findings as in our previous work [22,31,32]: We observed main effects of subtask (F_(1,4945) = 432.88, p < .001, η_p² = .080), and of incoming distance (F_(1,4945) = 364.99, p < .001, η_p² = .069), indicating higher errors in Pure PI than Landmark PI and for longer incoming distances, respectively. Besides, we found an interaction effect between subtask and incoming distance (F_(1,4949) = 44.99, p < .001, η_p² = .009; Fig 3A, left), indicating a stronger relationship between incoming distance and drop error in Pure PI as compared to Landmark PI (t₍₄₉₄₉₎ = 6.71, p < .001, d = 0.095). Moreover, we found a main effect of age (F_(1,36) = 12.23, p = .001, η_p² = .254; Fig 3A, right), reflecting higher drop errors in older age. No effects of testing day (F_(1,4945) = 0.39, p = .532, η_p² < .001) or sequence (F_(1,36) = 2.18, p = .148, η_p² = .057) emerged, indicating the absence of an order effect. Importantly, we found a main effect of treatment on drop error (F_(1,4945) = 10.46, p = .001, η_p² = .002; Fig 3B, left), indicating that cortisol impaired overall PI performance. Cortisol did not interact with other variables, suggesting a general effect independent of spatial cues or incoming distance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Predictors of PI performance and strategy use.

(A) The effect of longer incoming distances leading to higher drop errors was more pronounced when no spatial cues were available (left), and older age led to higher drop errors (right). (B) CORT led to overall higher drop errors (left), but navigational pattern during landmark PI was preserved (right). (C) In Landmark PI, the movement distance to the landmark predicted performance, irrespective of treatment. Significant slopes are indicated via asterisks. Plots show estimated marginal means derived from linear mixed-effects models (see S2 Fig for extended plots). Error bars and confidence bands represent SEM. Pure PI: pure path integration, Landmark PI: landmark-supported path integration, CORT: cortisol, PLA: placebo, vm: virtual meters, n. s.: not significant, ***p < .001, **p < .01, *p < .05. The data underlying panel 3B can be found at https://osf.io/c8u57.

https://doi.org/10.1371/journal.pbio.3003661.g003

In a second step, we built linear mixed models only including Landmark PI to examine the navigational pattern in the presence of the landmark under heightened cortisol levels. These models aimed at investigating the role of goal-to-landmark distance and its interaction with cortisol on PI performance, and at investigating whether cortisol affected the navigational pattern (irrespective of performance), respectively. In the first model, we found that a longer distance between goal location and landmark predicted worse PI performance, and cortisol did not moderate this effect (F_(1,2453) = 348.20, p<.001, η_p² = .124). Like in the main models, older age again significantly reduced PI performance (F_(1,36) = 6.30, p = .017, η_p² = .149). In the second model, we found no effect of treatment on movement distance to the landmark (F_(1,2455) = 1.37, p=.243, η_p² < .001; Fig 3B, right), showing that cortisol treatment had no effect on how closely participants navigate to the landmark.

The behavioral results so far have demonstrated that cortisol impairs PI performance, while the navigational pattern with respect to landmark proximity was preserved (if available). Generally, these variables may be connected, and we conducted a further analysis examining the effect of movement distance to the landmark on the drop error. Indeed, navigating further away from the landmark was associated with worse performance (F_(1,2453) = 106.37, p < .001, η_p² = .042; Fig 3C, left), and this effect was equally observed under placebo and under cortisol as evidenced by a lack of an interaction effect (F_(1,2459) = 0.16, p = .689, η_p² < .001; Fig 3C, right).

Cortisol administration enhances activation in the right caudate nucleus in the presence of landmarks

We tested the contrasts “Landmark PI> Pure PI”, “CORT> PLA” and their interaction “(Landmark PI> Pure PI)_CORT> (Landmark PI> Pure PI)_PLA”—as well as the respective reverse contrasts—during the outgoing and incoming phases. We performed exploratory whole-brain analyses and region of interest (ROI) analyses in left and right hippocampus, left and right caudate nucleus, bilateral posterior cingulate, and left and right posterior-medial entorhinal cortex for each contrast. For the whole-brain analysis, statistical parametric maps were initially thresholded at a family-wise error (FWE)-corrected α level of p < 0.05 across the whole brain and clusters were considered significant at p < 0.05, FWE-corrected (extent threshold of five voxels). For ROI analyses, the significance threshold was set to p < 0.05 on voxel level, corrected for multiple testing within each ROI using FWE-correction with the small volume correction option of SPM12.

The exploratory whole-brain analyses revealed significantly higher activation in left superior parietal lobule, bilateral precuneus (adjacent to posterior cingulate) as well as in visual areas including left superior and right middle occipital gyri during Landmark PI compared to Pure PI, presumably reflecting the higher amount of visual information (see S3 Fig and S4 Table). No significant clusters were found on whole-brain level for the other contrasts.

The ROI analyses (see S4 Fig for overview of masks) showed higher activation of posterior cingulate (t₍₃₅₎ = 6.33, p_FWE < .001) and right caudate nucleus (t₍₃₅₎ = 4.45, p_FWE = .010) in Landmark PI compared to Pure PI (Landmark PI > Pure PI; Fig 4A), while no significant differences were found for the other ROIs. Furthermore, no ROI showed differential activation in the contrast between cortisol and placebo (CORT > PLA). Importantly, however, we observed higher activation in the right posterior-medial entorhinal cortex (t₍₃₅₎ = 2.78, p_FWE = .045), right caudate nucleus (t₍₃₅₎ = 4.62, p_FWE = .006), and left caudate nucleus (t₍₃₅₎ = 3.90, p_FWE = .035) in the interaction contrast of Landmark PI versus Pure PI after cortisol compared to placebo administration ([Landmark PI > Pure PI]_CORT> [Landmark PI > Pure PI]_PLA).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Univariate fMRI results.

(A) Significantly higher activation in posterior cingulate and right caudate nucleus during Landmark PI as compared to Pure PI. (B) Significantly higher differentiation between subtasks under different treatments or between treatments in different subtasks in right caudate nucleus, left caudate nucleus and right posterior-medial entorhinal cortex. Activation shown for peak voxels of significant ROIs only. Error bars represent SEM. PI: path integration, CORT: cortisol, PLA: placebo, n.s.: not significant, ** p < .01, * p < .05, ^† p < .10. The data underlying this figure can be found at https://osf.io/c8u57.

https://doi.org/10.1371/journal.pbio.3003661.g004

Pairwise post-hoc tests showed that these interactions reflected relatively higher activation of the right caudate nucleus (i) under cortisol compared to placebo during Landmark PI (t₍₃₅₎ = 3.43, p_Bonferroni = .006, d = 0.550), and (ii) during Pure PI compared to Landmark PI under placebo (t₍₃₅₎ = 2.82, p_Bonferroni = .032, d = 0.291), as well as a trend for a higher activation of the right posterior-medial entorhinal cortex during Pure PI than during Landmark PI under placebo (t₍₃₅₎ = 2.45, p_Bonferroni = .077, d = 0.366). No significant pairwise comparisons emerged for the left caudate nucleus (all t ≤ 2.29, all p_Bonferroni ≥ .112, all d ≤ 0.360; Fig 4B). These findings indicate a reduced recruitment of the right caudate nucleus during Landmark PI compared to Pure PI under placebo, but enhanced recruitment during Landmark PI under cortisol compared to placebo, suggesting a relatively increased relevance of this region in the presence of landmarks under cortisol. Conversely, the findings suggest that the presence of landmarks can reduce recruitment of the right posterior-medial entorhinal cortex. No other ROI exhibited higher activation for this contrast, nor for any of the reverse contrasts.

On an exploratory level, we analyzed whether the neural differences between the subtasks were confounded by difficulty (S5 Fig). In this analysis, the contrast between Landmark PI and Pure PI was only significant in the posterior cingulate (t₍₃₅₎ = 4.44, p_FWE = .029), while the right caudate was not significant any longer (t₍₃₅₎ = 1.97, p_FWE = .705). The contribution of the posterior cingulate to landmark processing thus persists after controlling for total difficulty. On the other hand, right caudate nucleus appears to not be generally involved in landmark processing. However, when looking into the interaction with treatment, we observed a higher differentiation between the subtasks under different treatments for the caudate nucleus (both left: t₍₃₅₎ = 3.96, p_FWE = .026, and right: t₍₃₅₎ = 3.86, p_FWE = .034), resembling the results in the total sample. Moreover, the left hippocampus (t₍₃₅₎ = 4.80, p_FWE = .004) also showed higher activation for the contrast between Landmark PI and Pure PI for different treatments, while the right posterior-medial entorhinal cortex did not any longer. This lack of an effect in the posterior-medial entorhinal cortex in the subsample could be related to reduced power due to the halved number of trials, particularly because this is a very small brain region.

Lastly, in an attempt to replicate the findings reported in Bierbrauer and colleagues [22], we conducted two additional analyses examining representations of integrated path and goal proximity (see S1 Text). These analyses partially replicated the original results (see S6 and S7 Figs).

Cortisol administration impairs GLRs, which are associated with PI performance depending on subtask and treatment

To assess GLRs, we used a multivariate approach and focused on the right entorhinal cortex, consistent with previous studies [22,36]. Accordingly, due to their 6-fold rotational symmetry, grid cells should show more similar activity during movements along directions that differ by 60° (aligned movements) as compared to those differing by 30° (misaligned movements; see also Materials and methods).

We first aimed to confirm the presence of GLRs by comparing activity during aligned versus misaligned movements (Fig 5A) and then tested whether they were affected by cortisol. Indeed, we found higher similarities between aligned movements (mod(α,60°) = 0°) than misaligned movements (mod(α,60°) = 30°) across all days and treatments, demonstrating the presence of GLRs in our study (t₍₇₁₎ = 1.75, p = .042, Fig 5B, left). The permutation test further strengthened the robustness of this finding, as the obtained z-scores were significantly different from zero on group level (t₍₇₁₎ = 1.69, p = .048). Subsequent analyses revealed no main effect of treatment on GLRs (F_(1,67) = 0.61, p = .437, η_p² = .009), but we observed a significant treatment by day interaction (F_(1,67) = 5.99, p = .017, η_p² = .082; Fig 5B, right). Due to the crossover design, the interaction cannot easily be interpreted, as it may also include order effects. We thus decided to examine differences between groups on day one [38]. When doing this, post-hoc tests revealed a significant effect of treatment (t₍₆₇₎ = 2.28, p = .026, d = 0.761), indicating higher GLRs during placebo than during cortisol (whereas analogous univariate contrast analyses restricted to day one, both whole-brain and ROI-based, did not reveal any brain region with significantly higher activation). Indeed, separate analyses confirmed the presence of GLRs under placebo (t₍₁₈₎ = 2.16, p_Šídák = .044) but not under cortisol treatment (t₍₁₆₎ = −0.97, p_Šídák = .826). Control analyses of GLRs from day one under placebo (Fig 5C, top) confirmed that pattern similarity differed between aligned and misaligned movements only for 6-fold rotational symmetry (t₍₁₈₎ = 2.16, p = .022), but not for 4-fold, 5-fold, 7-fold, or 8-fold rotational symmetries (all t ≤ 1.81, all p_Bonferroni ≥ .175), and removing movements during same heading directions (±15° from 0°) preserved (by trend) 6-fold rotational symmetry and thus GLRs (t₍₁₈₎ = 1.37, p = .093). In contrast, under cortisol, the control analyses showed no evidence for 6-fold symmetry (Fig 5C, bottom). Although potentially confounded by order effects and thus difficult to interpret, GLRs were notably not found on day two under either treatment.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Cortisol effects on GLRs and their behavioral relevance.

(A) Inner numbers represent angular differences in 360° space, whereas outer numbers show differences in 60° space (left panel). Higher pattern similarity was expected for angular differences of mod(α,60°) = 0° (rose; α as angular difference between two movement directions) than for angular differences of mod(α,60°) = 30° (gray). The same result was expected when discarding pattern similarities of movements for the same heading direction (dark rose). Two exemplary fMRI volumes representing different movement directions (blue arrows; right panel). In the first volume, the participant navigates at an angle of 0° with respect to the (arbitrary) reference axis, while navigating at an angle of 90° in the second volume. This results in an angular difference of 90° in 360° space, corresponding to 30° in 60° space and thus mod(α,60°) = 30° (compare blue arrows on the left panel). (B) Average pattern similarity over the entire time series of all voxels in the right entorhinal cortex for all directions of aligned (rose) and misaligned (gray) fast runs (left panel). Differences in GLRs (i.e., pattern similarity difference) depending on treatment and day (right panel). (C) Average pattern similarity (left panels) as well as control analyses (right panels) for treatment groups on day one. (D) Relevance of GLRs for PI considering day one. Error bars and confidence bands represent SEM. GLRs: grid-like representations, PI: path integration, CORT: cortisol, PLA: placebo, vm: virtual meters, n.s.: not significant, ** p < .01, * p < .05, ^† p < .10. The data underlying panels 5B and 5C can be found at https://osf.io/c8u57.

https://doi.org/10.1371/journal.pbio.3003661.g005

Then, we investigated whether GLRs were behaviorally relevant using a linear mixed model incorporating effects of subtask, incoming distance, and GLRs. Because of the interaction effect between day and treatment, we again only focused on day one. We found similar effects as in the purely behavioral model (main effects of subtask, incoming distance, age, and the subtask by incoming distance interaction) with the addition of an interaction between subtask, treatment and incoming distance (F_(1,2264) = 4.33, p = .037, η_p² = .002), and an interaction between subtask, treatment, and GLRs (F_(1,2256) = 5.48, p = .019, η_p² = .002; Fig 5D). For the former effect, post-hoc analyses revealed that the relationship between incoming distance and drop error in pure PI is stronger under placebo than under cortisol (t₍₂₂₆₃₎ = 3.05, p_Šídák = .014), indicating a higher importance of incoming distance for performance under placebo. In contrast, post-hoc analyses for the latter interaction failed to reach significance, but a visual inspection of this finding suggested that GLRs are generally positively associated with PI performance, except under cortisol in the Pure PI subtask, which showed the inverse relationship.

Lastly, we examined whether the separate analysis of GLRs in the Pure PI and Landmark PI conditions revealed further insights. This analysis showed a largely similar pattern as in the general analysis (see S8 Fig), and further suggested that GLRs may be more relevant under Pure PI than under Landmark PI.

Exploratory parametric modulation analysis of PI performance

We further conducted a parametric modulation analysis using a general linear model (GLM), including trial-wise inverted drop error values as regressor for the combined outgoing and incoming phases. We used separate parametric regressors for Landmark PI and Pure PI trials, estimating each regressor’s average effect against baseline across treatments as well as the contrast between treatments. In addition, we computed the contrast between LPI and PPI trials across and between treatments. These effects were tested exploratorily across the whole brain as well as in ROI analyses.

The whole-brain analyses revealed one cluster showing modulation across treatments on Landmark PI trials as well as three clusters showing modulation across treatments on Pure PI trials, while no differences were found between treatments (S9 Fig and S5 Table). For the contrast between subtasks (Landmark PI versus Pure PI), one cluster showed significantly stronger positive modulation on Landmark PI compared to Pure PI trials across treatments, while there were again no differences between treatments (S9 Fig and S5 Table). The ROI analyses did not reveal any significant modulation of the BOLD signal for any of the contrasts (all p_FDR > .983; S9 Fig).

Conclusions

Our current findings provide essential information about the underlying mechanisms of stress effects on PI, supporting the hypothesis that they are strongly mediated by effects of cortisol on the entorhinal cortex. Cortisol impaired overall PI performance in young healthy men, while the navigational pattern with respect to landmark proximity was preserved. The fMRI data further indicated that cortisol administration increased activation in the right caudate nucleus in the presence of landmarks. Lastly, we demonstrated GLRs in the right entorhinal cortex, which were associated with PI performance in interaction with the spatial environment and the administration of cortisol.

Ethics statement

All participants were informed about the study procedures and provided written informed consent before the study began. The research was conducted in accordance with the Declaration of Helsinki and approved by the local ethics committee (Ethics Committee of the Medical Faculty, Ruhr University Bochum, approval no. 18-6368).

Participants

We used a two-day within-subject crossover design and recruited 42 healthy men, two of which had to be excluded due to technical failures of the joystick and another one due to unsuccessful experimental manipulation (no cortisol increase), leaving a final sample size of n = 39, aged 19–34 years (24.18 ± 4.55 years; mean ± SD) and with a body mass index ranging between 18 and 29 kg/m² (23.95 ± 2.54 kg/m²; mean ± SD). We recruited participants through online advertisements in social media networks, mailing lists and in university classes at Ruhr University Bochum. Exclusion criteria comprised an acute or history of disease (i.e., neurological, psychiatric, cardiovascular, immunologic), current or history of medical or psychological treatment, drug use, female sex, previous experience with the PI paradigm (see Experimental task), or any contraindication for participation in MRI studies. We only included men because of the well-known influence of sex hormones on the secretion of stress hormones [60] and the related changes in neural functioning [61]. All participants had normal or corrected-to-normal vision and received a compensation of 10€/hour (60–70€ in total) or course credits.

Experimental design

Sample size was based on previous studies investigating PI [17], the functional relevance of GLRs [21,22], and effects of cortisol administration [62]. In addition, we conducted a power analysis for a repeated measures analysis of variance (rANOVA) with two conditions (placebo versus cortisol), α = 0.05, power (1 − β) = 0.8, and a medium effect size (f = 0.25). This analysis suggested a minimum of 34 participants. While this analysis does not exactly match our linear mixed-effects modeling approach, it provides a reasonable benchmark.

Participants received two tablets of 10 mg cortisol (hydrocortisone, Hoechst) on one testing day, and two visually identical placebos on the other testing day in a double-blind two-day within-subject crossover design. Order of conditions was counterbalanced across participants. The dosage of 20 mg was chosen in accordance with previous studies demonstrating effects of cortisol on behavioral and neuronal responses with similar dosages [63–67]. We assessed salivary cortisol using Salivettes (Sarstedt, Nümbrecht, Germany) collected at several time-points (see Fig 2). Saliva samples were stored at −20 °C until assayed. Salivary cortisol concentrations were extracted from the samples using a time-resolved fluorescence immunoassay (IBL, Hamburg, Germany) at the local biochemical laboratory of Ruhr University Bochum and are reported in nanomole per liter (nmol/l). Intra- and inter-assay coefficients of variations were below 6.1%.

Experimental task

The Apple Game is a virtual PI task that was implemented by Bierbrauer and colleagues [22] via Unreal Engine 4 (Epic Games, version 4.11; see also Fig 1 and S1 Fig). In this task, PI is considered as the ability to integrate across several paths and to calculate home-coming vectors based on visual cues only. A circular arena formed by an endless grassy plain with a blue sky rendered at infinity, with a diameter of 13,576 virtual meters (vm), built the environment. Briefly, each trial was composed of three phases. In the “start phase”, participants first moved to a basket and memorized its location (goal location). Then, the “outgoing phase” followed, in which they navigated to a variable number of trees (1–5), which appeared consecutively in different locations, until a tree containing a red apple (retrieval location) was reached. The variable number of trees allowed manipulating the path of the outgoing phase and thus difficulty. Basket and trees disappeared upon arrival at the respective locations. During the “incoming phase”, participants lastly were asked to take the shortest path from the retrieval location back to the goal location. They pressed a button when arriving at the presumed location (response location) and received visual feedback via zero to three stars corresponding to the Euclidean distance between response location and goal location (drop error; three stars for <1,600 vm, two stars for <3,200 vm, one star for <6,400 vm). To investigate the influence of spatial cues on PI, the original version of the task contained three subtasks, of which we used two in this study. The “Pure PI” subtask was composed only of a grassy plain, forcing participants to solely rely on visual flow, while the “landmark-supported PI” (Landmark PI) subtask contained a central lighthouse as spatial cue. Participants first played up to eight training trials to get familiar with the task, during which a structural MRI scan (~6 min) was obtained. Then, a total of 64 experimental trials (32 per subtask) divided into four blocks of 16 trials were presented, during which functional MRI scans were assessed. The outgoing phase of the 32 trials of each subtask was composed of six trials with 1, 2, 4, and 5 trees, respectively, and eight trials with three trees in randomized order. Before every new trial, a fixation cross was presented with a variable duration 5–7.5 s (randomly distributed).

Experimental procedure

Testing sessions were performed after 1 PM to avoid interference with the cortisol awakening response [68]. On day one, upon arrival, participants read study information and gave written informed consent, before undergoing a cognitive test battery, after which the pharmacological intervention and a series of questionnaires followed (see S6 Table). About 30 min after pharmacological intervention, participants were prepared for the fMRI session, during which the experimental PI task was conducted. Lastly, participants answered a final set of questionnaires. The testing session on day two followed one week later, and the procedure was essentially the same, except for the absence of a cognitive test battery and the respective other pharmacological intervention. Finally, participants were debriefed and compensated (see Fig 2).

Behavioral data

To extract Apple Game data from computer-generated log-files we used MATLAB (2021b, The MathWorks, Massachusetts, US), including the Parallel Computing Toolbox (v6.12) and the CircStat Toolbox [69]. Statistical analyses were conducted in R [70] using the lme4 [71], lmerTest [72], and emmeans [73] packages.

Behavioral data analysis

One participant did not show any cortisol increase after cortisol administration and thus was excluded from statistical analyses, leading to a sample size of n = 39 for behavioral analyses. We assessed several measures for path distance, PI performance and navigational pattern during the experimental task. For path distance, two parametrical measures were assessed: Outgoing distance represents the accumulated distance from the goal to the retrieval location, while incoming distance refers to the Euclidean distance between retrieval and goal location (Fig 1C). These measures represent different subcomponents of PI: Outgoing distance is relevant for keeping track of the traveled path in relation to the starting point (i.e., the later goal location), and incoming distance for calculating a direct vector to this goal location. Because previous studies indicated incoming distance to be more closely related to entorhinal cortex activation [52,74], and the results of our previous work supported this view [31,32], we again concentrated on this measure as a proxy of path distance. Moreover, overall PI performance is represented by the drop error, i.e., the Euclidean distance between response location and goal location (Fig 1D). Lastly, to investigate the specific role of the landmark under heightened cortisol levels, we assessed two variables only applicable in Landmark PI. The first of these reflects the distance between the goal and the spatial cue (goal-to-landmark distance), while the other represents the mean Euclidean distance of the moving participant from the landmark across all time points of the incoming phase (movement-to-landmark distance).

MRI data

We acquired MRI data during the whole experimental task at the Bergmannsheil hospital in Bochum using a 3T Philips Achieva Scanner (X-Series, the Netherlands) with a 32-channel head coil. High-resolution whole-brain structural brain scans were acquired using a T1-weighted sequence with a 1 mm isotropic resolution, an FOV of 240 mm × 240 mm, and 220 transversally oriented slices during a total acquisition time (TA) of six minutes and three seconds. Blood oxygenation level-dependent (BOLD) contrast images were registered with a T2*-weighted gradient echoplanar imaging sequence with 2.5 mm isotropic resolution, TR = 2,500 ms, TE = 30 ms, FA = 85°, FOV = 96 mm × 96 mm, 45 transversal slices in ascending order without slice gap, and TA = 17.32 ± 2.01 mins (mean ± SD), corresponding to 415.67 ± 48.27 (mean ± SD) volumes. Variation in TA emerged because the task was self-paced, leading to different durations of runs. In addition to three dummy scans preceding data acquisition, the first five images of each session were discarded to allow for signal steady-state transition. We further used a shim box of 60 mm × 60 mm × 60 mm around the medial temporal lobe to increase temporal signal-to-noise ratio. The virtual environment was presented to participants via MR-compatible liquid crystal display goggles (VisuaStim Digital, Resonance Technology, Northridge, CA, USA) with a resolution of 800 × 600 pixels, and they navigated within the virtual environment using an MR-compatible joystick (Nata Technologies, Coquitlam, Canada).

fMRI data were preprocessed and subsequently analyzed using SPM12 implemented in MATLAB (2021a, The MathWorks, Massachusetts, US) and nilearn for Python (Nilearn contributors, 2025). Preprocessing included slice time correction and spatial realignment. For whole-brain and ROI analysis, fMRI scans were further normalized to MNI space using parameters from the normalization procedure of the segmented structural T1 image. Further, we applied spatial smoothing with a 5 mm isotropic Gaussian kernel to the normalized fMRI data.

We used masks of the left and right hippocampus, left and right caudate nucleus, and bilateral posterior cingulate (maximum probability masks; probability threshold: 0.25; Harvard-Oxford Cortical and Subcortical Structural Atlases, Harvard Center for Morphometric Analysis; https://cma.mgh.harvard.edu/). Because the medial entorhinal cortex is especially related to grid-cell activity in rodents and the human entorhinal cortex consists of structurally and functionally distinct subparts, where the anterior-lateral entorhinal cortex and the posterior-medial entorhinal cortex represent the homolog of rodent lateral and medial entorhinal cortex, respectively, we focused our univariate analysis on left and right posterior-medial entorhinal cortex using masks created by Maass and colleagues [75]. For the multivariate analyses, we performed automated segmentation of the entire right entorhinal cortex using the Automatic Segmentation of Hippocampal Subfields (ASHS) software with the ASHS-PMC-T1 Atlas [76,77].

fMRI data analysis

We excluded three participants, whose entorhinal cortices were only partly covered, most likely due to susceptibility artifacts and other distortions, which are very strong in the medial temporal lobe [78], leaving a final sample size of n = 36 for the fMRI data analysis. In the first GLM, we modeled start phase, outgoing phase, incoming phase, and feedback separately for each of the two subtasks and each of the two treatments (16 regressors in total). The duration of these regressors corresponded to the duration of the respective phases in each trial. We tested three contrasts (plus their respective reverse contrasts): “Landmark PI> Pure PI”, “CORT> PLA” and their interaction “(Landmark PI> Pure PI)_CORT> (Landmark PI> Pure PI)_PLA”, specifically during the outgoing and incoming phases, which represent the actual PI task. All regressors were convolved with the hemodynamic response function before entering the GLM. As nuisance regressors, we included motion parameters as estimated in the realignment procedure and a high-pass filter (time constant = 128 s) was implemented. We performed whole-brain analyses on each contrast. Contrast images from the first-level analysis of each participant were entered into a GLM. Statistical parametric maps were initially thresholded at a FWE-corrected α level of p < 0.05 across the whole brain. We considered clusters significant at p < 0.05, FWE-corrected (extent threshold of five voxels). For all significant clusters, we provide maximum probability tissue labels with MNI coordinates derived from the Neuromorphometrics atlas as implemented in SPM12 (www.oasis-brains.org/; http://neuromorphometrics.com/). For ROI analyses, the significance threshold was set to p < 0.05 on voxel level, corrected for multiple testing within each ROI (FWE-corrected; using the small volume correction option of SPM12).

On an exploratory level, we re-ran the first GLM with matched task difficulty between subtasks. Disentangling task difficulty from the effects of spatial cues is inherently challenging, as difficulty varies across multiple levels and partially overlaps with cue availability: spatial cues facilitate navigation by enabling additional cognitive strategies, while difficulty can also be manipulated via path distance or cumulative angular rotation. To approximate matching for difficulty, we compared trials with high incoming distances in Landmark PI to trials with low incoming distances in Pure PI, matching overall performance levels rather than specific subprocesses. Consequently, residual differences in difficulty between subtasks cannot be fully excluded. To address this, univariate analyses were repeated in a subsample of trials, in which, within each participant, trials were ranked by incoming distance and the upper half of Pure PI trials and the lower half of Landmark PI trials were excluded. This substantially reduced performance differences between subtasks, although a small difference remained, which was accepted to preserve statistical power.

In a second GLM, we conducted an exploratory parametric modulation analysis of the drop error, which for each of the two subtasks included an unmodulated trial regressor as well as a parametric modulation regressor for trial-wise drop error modeling the combined outgoing and incoming phases of each trial (four regressors in total). The onsets of these regressors corresponded to the onset of the outgoing phase and the modeled durations corresponded to the added total durations of the outgoing and incoming phases of each trial. Drop error values were symmetrically inverted so that higher values of the parametric regressor would reflect better task performance. In addition, drop error values were normalized to range between 0 and 1 and mean-centered. All regressors were convolved with the hemodynamic response function before entering the model. We included motion parameters resulting from realignment of functional images during preprocessing as nuisance regressors. Additionally, we used a cosine model with a high-pass frequency of 0.01 Hz to account for slow temporal drifts in the signal. For each participant and session, beta maps were computed for the contrast of each subtask’s drop-error parametric modulation against baseline, and for the contrast between subtasks’ parametric modulations (Landmark PI versus Pure PI). We performed whole-brain analyses on the first-level drop error parametric modulation contrast images, computing two different second-level contrasts for each while accounting for subject-specific effects: The average first-level effect across treatments as well as the difference in first-level effects between cortisol and placebo. Resulting statistical parametric maps were thresholded at a false discovery rate (FDR)-corrected α level of p < 0.01 across the whole brain. Clusters were considered significant at p < 0.01, FDR-corrected and extending a minimum cluster size of 10 voxels. For all significant clusters, maximum probability tissue labels are reported based on the peak voxel’s MNI coordinates and segmentations derived from the Harvard-Oxford atlas as implemented in nilearn. For the ROI analysis of parametric modulation by drop error, p values were corrected for the number of repeated tests (number of ROIs and number of contrasts) using FDR-correction at α < 0.01.

To assess GLRs during fMRI, we performed a representational similarity analysis (RSA), using an approach analogous to previous studies [22,36]. This method assumes that grid cells in the entorhinal cortex, due to their 6-fold rotational symmetry, show similar activity patterns during movements that differ by 60° in angular space. Activity patterns should thus show higher similarity for movements with an offset of n*60°, where n = {0, 1,..., 6}, and lower similarity for movements with an offset of n*60° + 30°, where n = {0, 1,..., 5}. We refer to the first condition as “mod(α,60°) = 0°” and to the second condition as “mod(α,60°) = 30°” (where α represents the angular difference between movement directions, and “mod” denotes the modulo operator; see Fig 5A). For the analysis we used the preprocessed fMRI data. To account for the hemodynamic response, BOLD time series were shifted by two TRs (2.5 s each) to align the peak response (∼5 s) with behavioral events. The RSA involved the following steps: (i) extracting the BOLD signal within the right entorhinal cortex ROI from the preprocessed fMRI volumes without normalization or smoothing; (ii) calculating the mean orientation and speed for each fMRI volume based on the trajectory data; (iii) keeping only the volumes with movement speed in the top third; (iv) averaging volumes into bins of 5° orientation based on the mean orientation; (v) calculating the angular difference of the movement orientation between the volume bins; and (vi) calculating Fisher z-transformed Pearson correlations between the volume bins. We then compared the pattern similarity of pairs of volume bins offset by ±15° from the mod(α,60°) = 0° condition (aligned movements) with those offset by ±15° from the mod(α,60°) = 30 condition (misaligned movements). Pattern similarity was expected to be higher for aligned movements than for misaligned movements, thus reflecting GLRs. Two control analyses were performed. First, to ensure that the increase in pattern similarity was specific to 6-fold symmetry, the same analysis was performed for other types of rotational symmetry (4-, 5-, 7-, and 8-fold). Second, the analysis was also performed after excluding correlations that were ±15° from 0°, to ensure that increases in pattern similarity were not primarily driven by movements in the same direction, which could reflect head direction signals.

To investigate whether the magnitude of GLRs is modulated by the spatial environment (i.e., with or without spatial cues), we reconducted the GLR procedure in a secondary analysis, this time separating the data to calculate two distinct metrics for each testing day: one for Landmark PI and one for Pure PI. This isolated assessment enabled a comparison of GLR magnitude across the two navigational contexts.

Statistical analysis

Our first statistical analysis examined the success of pharmacological intervention by comparing cortisol concentrations using an rANOVA with time and pharmacological intervention as within-subject factors. Because cortisol concentrations typically exhibit a right-skewed distribution, we conducted a natural log (ln) transformation to obtain normally distributed data. In case of violation of the sphericity assumption, we used Greenhouse-Geisser adjustment (and for the sake of presentation, we rounded the corrected degrees of freedom to the nearest whole number). Post-hoc pairwise comparisons were performed using Bonferroni-corrected t-tests (or Welch’s t-tests in case of unequal variances).

To test the hypothesis postulating differential effects of cortisol depending on subtask and incoming distance on PI performance, we then built a linear mixed model with PI performance (i.e., drop error) on the level of single trials as criterion, and subtask (two levels: Pure PI versus Landmark PI), incoming distance, and treatment (two levels: cortisol versus placebo) as within-subject predictors (Model 1; S1 Table).

To test the hypothesis concerning the effects of cortisol on the role of the landmark, we built three further linear mixed models (Models 2A–C; S1 Table). The first of these aimed at investigating the role of goal-to-landmark distance and its interaction with cortisol on PI performance and thus included PI performance on the level of single trials as criterion, goal-to-landmark distance (only in Landmark PI), and treatment as within-subject predictors. The second of these models aimed at investigating whether cortisol affected navigational pattern (irrespective of performance) and included movement-to-landmark distance (only in Landmark PI) on the level of single trials as criterion, and treatment as within-subject predictor. The third of these models aimed at investigating whether the navigational pattern (in interaction with cortisol) affected PI performance and thus included PI performance on the level of single trials as criterion, movement-to-landmark distance (only in Landmark PI), and treatment as within-subject predictors.

To evaluate the overall existence of GLRs, we conducted a one-tailed one-sample t test comparing activation of entorhinal cortex during aligned versus misaligned movements against zero (Model 3A; S1 Table). To further test the robustness of our findings, we performed a non-parametric permutation test. For each participant, we constructed a null distribution of similarity differences by randomly shuffling the angular labels and recomputing the similarity measure for each of 10,000 permutations. The position of the observed difference within this null distribution yielded a p-value, which we transformed into a signed z-statistic (positive and negative values allowed). At the group level, these z-statistics were tested against zero using a one-sample t test. To test whether GLRs were related to treatment, we conducted a linear model with GLRs as criterion, treatment and day as within-subject predictors, and age as covariate (Model 3B; S1 Table). For the secondary analysis, we conducted a linear model with GLRs as criterion, and treatment, day, and subtask as within-subject predictors, as well as age as covariate (Model 3C; S1 Table). We conducted separate t-tests for the control analyses (other symmetries or correction for heading direction). In the next step, to examine the behavioral relevance of GLRs, we built linear models with PI performance on the level of single trials as criterion, and subtask, incoming distance, and GLRs as within-subject predictors (Models 4A–B; S1 Table).

In all linear mixed models, “subject” was added as random factor, and age, testing day, and sequence (to control for order effects) as covariates. For all analyses, we centered within-subject parametric predictors (incoming distance, goal-to-landmark distance) to the participant’s mean and age to the grand mean of all participants [79]. For analysis of fixed effects, we always used type III sum of squares. In case of interactions between parametric predictors, we discretized all but one parametric predictor and calculated estimated marginal means (“emmeans” or “adjusted means”) or estimated marginal means of linear trends (“emtrends” or “conditional regression equations”) based on the minimum, mean, and maximum values of these discretized predictors. For post-hoc tests against zero or pairwise comparisons using Fisher’s tests, we used Šídák-adjustment correcting for number of discretized predictors (3), number of subtasks (2), number of treatments (2), or a combination of those. For an approximation of degrees of freedom, we used the Kenward-Roger method (and for the sake of presentation, we rounded them to the nearest whole number). To estimate effect sizes, we used Partial Eta-Squared (η_p²) for F-tests and Cohen’s d for t-tests. Multicollinearity between predictors was not problematic (all variance inflation factors <5). Assumptions of statistical tests were verified before application, and all tests were conducted two-tailed (except for t-tests on GLRs) at a significance level of α = 0.05.