LSD flattens the functional hierarchy of the human brain

Classic serotonergic psychedelics such as psilocybin and lysergic acid diethylamide (LSD) are under renewed investigation for therapeutic applications and are known to profoundly alter conscious experience via 5-HT2A receptor agonism. Earlier neuroimaging work using fMRI, MEG and EEG has characterised psychedelic effects on undirected functional connectivity and on some directed measures (for example, Granger causality and travelling waves). A recent theoretical framework, RElaxed Beliefs Under pSychedelics (REBUS), predicts that psychedelics 'flatten' hierarchical information processing by weakening top-down priors encoded in higher-order associative regions and thereby increasing the relative influence of bottom-up sensory signals. Testing REBUS directly is challenging, but one accessible implication is that psychedelics should reduce top-down and/or increase bottom-up directed information flow in the brain. Shinozuka and colleagues set out to operationalise hierarchy as directional asymmetry in interregional signalling and to test whether LSD alters that hierarchy. They applied a measure of temporal irreversibility (INSIDEOUT), together with complementary metrics—hierarchical coherence and hierarchical inhomogeneity, defined from incoming versus outgoing lagged connectivity—to previously acquired MEG recordings in healthy participants who received 75 μg intravenous LSD or placebo. The authors also examined the dynamics of these hierarchy metrics over time, compared their discriminatory power to conventional undirected measures using machine learning, and tested whether regional changes related to serotonin receptor expression maps.

The analysis used MEG data from a prior study in which each participant underwent two sessions (placebo and 75 μg i.v. LSD). Recordings occurred ~165 minutes after drug administration and included four seven-minute conditions: eyes-open rest (Open), eyes-closed rest (Closed), listening to music with eyes closed (Music), and watching a silent nature video (Video). From an original cohort of 20, three participants were excluded for incomplete sessions or excessive movement and one lacked the Music condition, leaving 16 participants for analysis. Data were sampled at 600 Hz, high-/low-pass filtered at 1 and 100 Hz, epoched into 2-second segments, downsampled to 200 Hz and cleaned (manual inspection, ICA and automatic algorithms). Source reconstruction used an LCMV beamformer to obtain amplitude envelopes for 90 AAL regions; symmetric orthogonalisation was applied to remove spurious zero-lag correlations and analytic envelopes were extracted via the Hilbert transform. Broadband analyses were primary; frequency-specific analyses used theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz) and gamma (30–48 Hz) bands. Irreversibility (INSIDEOUT) was computed as the squared difference between cross-correlations of forward and time-reversed amplitude-envelope time series for every pair of regions across a range of time-lags t, yielding epochwise 90 × 90 irreversibility matrices. Regional irreversibility values were derived and, to reduce noise, matrices were thresholded at the 95th percentile (the authors report rerunning analyses without thresholding with consistent results). Three strategies were used to select a representative t: integrate across all t (total irreversibility), choose the t maximising mean irreversibility across all data (orthogonal contrast), and a stratified cross-validation with permutation testing to select the t that maximised placebo-versus-LSD differences; the latter (toptimal = 0.225 s) is reported in the main text. Hierarchical coherence and inhomogeneity were defined from the incoming–outgoing asymmetry at each region: the magnitude difference between a region’s strongest incoming and strongest outgoing lagged connections. The mean of these regional asymmetries is hierarchical coherence and the standard deviation is hierarchical inhomogeneity. To probe temporal structure, a sliding-window analysis (1 s windows, 80% overlap) produced timeseries for each hierarchy metric. Recurrence network analysis was then applied: time-delay embedding parameters (tRNoptimal and moptimal) were chosen by mutual information and False Nearest Neighbours criteria, a phase-space distance threshold was set using Eroglu’s method, and recurrence rate (edge density) quantified how often similar states recurred. Statistical testing included cluster-based permutation testing over t, two-way repeated-measures ANOVAs (drug × condition) and three-way ANOVAs when region was included; normality of irreversibility distributions was evaluated and noted to be imperfect for LSD in some cases. For comparison to more conventional FC measures, undirected FC was taken as Pearson correlations of amplitude envelopes and forward cross-correlations were computed at the selected t. A random forest classifier (bagged decision trees; default 100 trees) was trained 100 times with 80/20 train/test splits to discriminate LSD versus placebo using features drawn from the four conditions; performance was assessed by area under the ROC curve and differences between classifiers were tested with Wilcoxon signed-rank tests. Finally, spatial correlations between regional irreversibility and serotonin receptor expression (5-HT2A, 5-HT1A, 5-HT1B) used existing PET maps averaged into AAL regions; significance was assessed against 1,000 spatial-autocorrelation-preserving shuffles (variogram-matching).

Across analyses, LSD reduced measures of directional asymmetry and hierarchy relative to placebo. When irreversibility was computed across all time-lags t, placebo showed greater irreversibility than LSD at nearly all t (significant in all four conditions; pOpen = 0.0002, pClosed = 0.0001, pMusic = 0.0001, pVideo = 0.001). Integrating irreversibility across t produced a 'total irreversibility' that was significantly decreased by LSD in every condition (pOpen < 0.0001, pClosed < 0.0001, pMusic < 0.0001, pVideo = 0.0023). The t that maximised irreversibility was lower on LSD in all conditions (p-values range 0.0009–0.0214) and showed greater between-subject variance under LSD. Using the orthogonal-contrast t (0.12 s) or the cross-validation-selected t (0.225 s), the LSD-related decrease in irreversibility remained highly significant across conditions (e.g. at t = 0.225 s: pOpen < 0.0001, pClosed < 0.0001, pMusic < 0.0001, pVideo = 0.0032). Regional analyses revealed that occipital cortex regions consistently exhibited the highest irreversibility across conditions, even in eyes-closed states. Under Open, Closed and Music conditions, ≥85 of the 90 regions showed significantly reduced irreversibility on LSD compared to placebo; in the Video condition 74 regions were significantly reduced (insignificant regions were largely parietal). Frequency-specific analyses showed that alpha and beta band results aligned with broadband effects, but theta and gamma bands diverged: in theta LSD sometimes increased irreversibility for certain t ranges (p-values indicate increased theta irreversibility under LSD in all four conditions), whereas gamma showed increased irreversibility in three conditions but not Video. Partial time-delayed correlations (which regress out other regions) still showed significant LSD-related reductions in irreversibility across conditions. Measures derived from incoming–outgoing asymmetry paralleled the irreversibility findings. LSD significantly decreased regional incoming–outgoing asymmetry and, when summarised, reduced hierarchical coherence (mean asymmetry) and hierarchical inhomogeneity (standard deviation) in all four conditions (hierarchical coherence: pOpen < 0.0001, pClosed = 0.0002, pMusic < 0.0001, pVideo = 0.0062; hierarchical inhomogeneity: pOpen < 0.0001, pClosed = 0.0139, pMusic = 0.0003, pVideo = 0.0121). Under both drug states, the eyes-closed conditions (Music and Closed) had higher coherence and inhomogeneity than eyes-open conditions, with some condition differences significant only on LSD. Dynamics analyses showed that the temporal evolution of each hierarchy metric was more recurrent (higher recurrence rate or edge density) under LSD, indicating the brain revisits a narrower set of hierarchical states on LSD. Recurrence rates for irreversibility were significantly higher under LSD across conditions (pOpen = 0.0001, pClosed = 0.0185, pMusic = 0.0032, pVideo = 0.0416) and similar significant increases were found for hierarchical coherence and inhomogeneity across conditions. Comparing metrics, undirected FC and forward cross-correlations were also reduced by LSD and shared similar spatial topographies (lower in frontal, higher in occipital/temporal regions). Nevertheless, a random forest classifier trained on regional irreversibility outperformed classifiers trained on forward cross-correlations, undirected FC and the other hierarchy metrics in discriminating LSD from placebo (statistically significant differences in AUC reported by the authors). Finally, regional irreversibility did not correlate with PET-derived maps of serotonin receptor expression: no significant spatial correlations were found with 5-HT2A (p-values 0.8100–0.9560 across conditions) nor with 5-HT1A or 5-HT1B maps (all non-significant).

Shinozuka and colleagues interpret their primary finding—that LSD reduces directional asymmetry (irreversibility) and decreases hierarchical coherence and inhomogeneity across experimental conditions—as empirical support for the REBUS hypothesis that psychedelics flatten functional hierarchy. They note the concordance of their results with earlier MEG and EEG findings showing reduced measures of directed connectivity under psychedelics and with EEG results indicating altered travelling-wave dynamics. The authors argue that irreversibility may be a more specific biomarker of psychedelic state than conventional undirected FC measures because classifiers based on irreversibility distinguished LSD and placebo more accurately. The discussion emphasises conceptual points about hierarchy: asymmetry in directed information flow is a necessary but not sufficient signature of hierarchy. To illustrate this, the authors present toy network models in which irreversibility can be high without coherent hierarchical organisation; they therefore considered both the mean (coherence) and the dispersion (inhomogeneity) of incoming–outgoing asymmetry as joint indicators of hierarchy and report LSD-related reductions in both. They further compare their functional hierarchy measures to graph-theoretic constructs such as trophic coherence and a recently proposed 'democracy coefficient' framework, noting methodological differences (for example, those other measures typically assume non-negative networks). The authors also discuss how the topography of their results relates to anatomical visual hierarchies: visual cortical regions showed the largest asymmetries and tended to act as net transmitters in their measures, an outcome that the analysis treats agnostically with respect to whether transmitters represent higher or lower hierarchical levels. On dynamics, the investigators found that hierarchy metrics evolve over time and that LSD constrains this evolution—recurrence analyses showed the brain occupies a smaller repertoire of hierarchical dynamics under LSD. They place this alongside prior literature on increased entropy under psychedelics, clarifying the distinction between entropy (state variability) and entropy production (a measure related to irreversibility). The authors note that increased entropy under psychedelics does not contradict their finding of reduced irreversibility, because entropy and entropy production can behave differently and are not equivalent; they invoke the Crooks fluctuation theorem to motivate links between irreversibility and entropy production under Markovian assumptions and propose future application of Ornstein–Uhlenbeck modelling to derive entropy production directly. Key limitations are acknowledged: causality between altered hierarchy and subjective or therapeutic effects is not established; the hierarchical metrics were computed across all brain regions although a smaller latent subset may mediate directed effects; dimensionality-reduction choices (e.g. averaging top 5% values) could be improved with more principled approaches; and perturbations of horizontal versus vertical connections in networks may have different implications for true hierarchical strength than captured here. The authors present their principal conclusion as evidence that LSD weakens directional asymmetry and the imbalance between top-down and bottom-up information flow, consistent with a flattened functional hierarchy, while noting outstanding uncertainties and directions for further work.

Shinozuka and colleagues conclude that LSD reduces the directional asymmetry (irreversibility) of brain activity and decreases the imbalance between incoming and outgoing connectivity, which they interpret as a reduction in functional hierarchy and in entropy production. These results are presented as empirical corroboration of the REBUS hypothesis that psychedelics flatten hierarchical information processing, and the authors suggest irreversibility-based metrics may be informative biomarkers for psychedelic states and for future mechanistic work on entropy production in the brain.