Psilocybin modulation of time-varying functional connectivity is associated with plasma psilocin and subjective effects

Psilocybin, via its active metabolite psilocin and agonism at the serotonin 2A receptor (5-HT2AR), produces acute alterations of consciousness and rapid, sometimes lasting, changes in mood and cognition. Previous resting-state fMRI studies report broad alterations in thalamic and whole-brain connectivity, reduced segregation among canonical networks, and increased signal complexity, but most work has used “static” functional connectivity estimated across an entire scan and thus may miss rapidly evolving neural dynamics that accompany the psychedelic experience. This study set out to characterise acute psilocybin effects on time-varying functional connectivity (tvFC) during the extended oral psilocybin experience, and to relate those dynamics to measured plasma psilocin level (PPL) and subjective drug intensity (SDI). Olsen and colleagues applied Leading Eigenvector Dynamics Analysis (LEiDA) to instantaneous phase-coherence estimates from BOLD rs-fMRI and introduced diametrical clustering (a method that respects the antipodal spherical geometry of normalised eigenvectors) to define discrete brain states. They also modelled state dwell time using survival analysis (Cox frailty models) rather than simple averages, and explored robustness across a range of cluster numbers (k = 2–20). The primary goal was to map brain-state fractional occurrence and dwell time onto PPL and SDI during the psilocybin session.

Fifteen healthy adults (mean age 34.3 ± 9.8 years, six female) with no or limited prior psychedelic experience participated in a brain imaging study that included a single oral psilocybin intervention. Participants were screened for neurological, psychiatric and somatic illness, gave written informed consent, and were prepared and supported by two psychologists. Psilocybin was administered in body-weight-calibrated doses (mean total dose 0.24 ± 0.04 mg/kg) as part of a single-blind cross-over protocol (only the psilocybin arm is reported here). Resting-state fMRI (10 minutes, TR = 2 s, 300 volumes per session) was acquired once before drug intake and at approximately 40, 80, 130–140 and 300 minutes after administration; 74 scan sessions were acquired in total and two sessions (from one participant) were excluded for motion, leaving data that correspond to 21,600 volumes across sessions. Immediately after each rs-fMRI session participants rated subjective drug intensity (SDI) on a 0–10 Likert scale and a blood sample was drawn to quantify unconjugated psilocin (PPL). MRI acquisition used a 3T Siemens Prisma and a 90-region AAL parcellation (excluding cerebellum). Preprocessing (performed per session) included slice-timing correction, realignment and field-unwarping, co-registration and segmentation of T1-weighted images, normalization of the AAL atlas, smoothing (4 mm FWHM), nuisance regression (motion parameters and derivatives, aCompCor components and derivatives for white matter and CSF), band-pass filtering (0.008–0.09 Hz) and motion-based session exclusion using ART thresholds. LEiDA was implemented by computing the instantaneous phase of each regional BOLD time series via the Hilbert transform, building phase-coherence matrices (cosine of pairwise phase differences) for every time point, and retaining the leading eigenvector for each matrix as a P-dimensional unit vector representing the dominant instantaneous coherence pattern (P = 90 regions). Because leading eigenvectors are unit-length and sign-ambiguous (antipodal symmetry), the investigators applied diametrical clustering — a k-means-style algorithm appropriate for points on the unit hypersphere that uses squared Pearson correlation (squared cosine similarity) as the similarity metric and acknowledges sign ambiguity. Clustering was performed for k = 2…20, with multiple initialisations (k-means++ style) and many replications; the best replication per k was retained and additional stability analyses ran thousands of repetitions. Outputs were hard assignments of each time point to a brain-state centroid. Two principal brain-state descriptors were extracted per scan session: fractional occurrence (FO), the proportion of time points assigned to a given state, and dwell time, the duration of continuous visits to a state. FO models used linear mixed-effects models with a random intercept for subject and either PPL or SDI as fixed effects; significance was tested via likelihood-ratio tests and family-wise error control used permutation-based max-T correction (100,000 permutations) within each k. Dwell time was modelled using Cox proportional hazards models with a subject-level frailty term to account for inter-subject heterogeneity; hazard ratios (HRs) express the change in instantaneous switching hazard per unit increase in PPL or SDI. Because no suitable permutation for frailty Cox models was available, multiple-comparison correction for dwell-time tests used Bonferroni–Holm within each k. Centroid matching across different k values used template centroids to track analogous states, and diametrical clustering outputs were compared with Euclidean k-means (with centroid normalisation) to assess methodological differences.

Across all sessions, 21,600 leading eigenvectors were clustered into k brain states for k = 2…20. Centroid locations were generally stable across contiguous k values and a ‘‘global’’ or ‘‘fully connected’’ state (all centroid elements sharing the same sign) appeared for k ≥ 4. Other states exhibited bipolar patterns of coherence (positive and negative loadings), with two particular centroids resembling frontoparietal or central-executive network configurations (labelled here as frontoparietal state 1 and frontoparietal state 2). Visual and salience-related regions loaded strongly in some other centroids. The principal tvFC findings related brain-state fractional occurrence to both plasma psilocin and subjective intensity. For frontoparietal state 1 (template taken at k = 7), FO was negatively associated with PPL (k = 7 slope = -0.0064 FO per μg/L PPL; 95% CIunadj = [-0.0091, -0.0036]; pFWER-maxT < 0.001) and with SDI (k = 7 slope = -0.013 FO per SDI rating; 95% CIunadj = [-0.017, -0.009]; pFWER-maxT < 0.001). This negative FO–PPL association held for k ∈ {4,…,8,10}, and the FO–SDI association was significant for nearly all k ≥ 4 (except k = 17). For frontoparietal state 2 (template at k = 11) FO was likewise negatively associated with PPL (k = 11 slope = -0.0045 FO per μg/L PPL; 95% CIunadj = [-0.0064, -0.0026]; pFWER-maxT < 0.001) and SDI (k = 11 slope = -0.0089 FO per SDI rating; 95% CIunadj = [-0.011, -0.0065]; pFWER-maxT < 0.001), with significant PPL effects across many k ≥ 8. In contrast, a fully connected state showed a positive FO association with SDI for all k ≥ 5 (example k = 7 slope = 0.0078 FO per SDI rating; 95% CIunadj = [0.0027, 0.013]; pFWER-maxT = 0.026) but not with PPL (k = 7 pFWER-maxT = 0.284). Dwell-time analyses produced complementary results. For the frontoparietal states, higher PPL and SDI yielded larger hazard ratios (i.e. an increased hazard of switching away from the state and thus shorter continuous dwell time); the PPL–dwell-time association was statistically significant for k ∈ {4,…,9}, and the SDI–dwell-time association for k ∈ {4,…,11,13}. Dwell time for the fully connected state was significantly positively associated with both PPL and SDI in some k configurations, suggesting longer uninterrupted visits to that state when subjective intensity was higher. The authors note that the numerical magnitudes of hazard ratios across the three highlighted states were similar. Stability analyses showed the three template states could be tracked across k, although state centroids shifted modestly as k changed. Some associations, notably for the fully connected state and, to a lesser extent, the frontoparietal centroids, showed relationships with head motion for certain k values, indicating a potential motion-related contribution to that state’s FO. Finally, the investigators report that about 58% of variance in phase-coherence matrices was explained on average by the first eigenvector, and that clustering centroids were generally robust to initialization and to choice of k.

Olsen and colleagues interpret their results as evidence that acute psilocybin effects map onto changes in time-varying whole-brain connectivity: higher plasma psilocin levels and greater subjective intensity corresponded to reduced prevalence and shorter dwell time of two frontoparietal-like brain states, alongside an increase in a ‘‘fully connected’’ state. They propose that the reduction in frontoparietal state prevalence constitutes a systems-level neural correlate that scales with psilocin availability and subjective experience magnitude; psilocin has previously been linked to 5-HT2AR occupancy and subjective intensity, which the authors cite to connect pharmacology and functional dynamics. The findings converge with a prior LEiDA study of intravenous psilocybin that reported decreased occurrence of a frontoparietal state, even though the present work differs methodologically (oral dosing, multiple post-dose scans, measurement of plasma psilocin, and use of diametrical clustering). The investigators highlight that the affected states predominantly involve higher-order ‘‘transmodal’’ networks thought to support abstract cognition and that their decreased prevalence may relate to altered thought processes and stimulus interpretation during the psychedelic state. Methodological contributions are emphasised: diametrical clustering more appropriately handles unit-length, sign-ambiguous eigenvectors than Euclidean k-means and obviates arbitrary sign-flip procedures; survival (Cox) modelling of dwell time captures the conditional dependence of state persistence and complements FO analysis. The authors acknowledge trade-offs, such as increased numbers of statistical tests when exploring many k and the hard-assignment nature of k-means–type methods. They suggest future methodological directions including Watson or Bingham mixture models, directional archetypal analysis, and consideration of modelling beyond the first eigenvector since the phase-coherence matrix is rank two and on average the first eigenvector explained 58% of the variance. Key limitations reported by the investigators include the absence of a placebo condition (the study compares pre- and post-drug scans without a control arm), lack of pulse and respiratory recordings to regress physiological noise directly, and greater head motion during post-psilocybin scans which could confound some state estimates. The extracted text ends partway through a sentence describing motion-related issues, so some details about how motion was handled or the extent of its impact are not fully available in the provided text. Despite these caveats, the authors conclude that their results provide a novel, pharmacologically grounded mapping between psilocybin exposure, subjective effects, and brain-state dynamics, and they recommend diametrical clustering as a preferred approach for LEiDA-based tvFC analyses.