Increased sensitivity to strong perturbations in a whole-brain model of LSD

Recent neuroimaging studies indicate that LSD produces widespread changes in brain functional connectivity and an expanded repertoire of dynamical brain states, including increased BOLD variance and greater diversity of dynamic functional connectivity patterns. Despite these empirical observations, a mechanistic understanding of how whole-brain dynamics are altered by LSD is incomplete. Previous work using controlled perturbations (for example TMS or DBS) has shown the value of probing dynamical responses, but in vivo perturbations are limited in scope, duration and regional specificity. Jobst and colleagues set out to combine a whole-brain computational model with an in silico perturbational approach to explore how the brain’s dynamical stability differs under LSD versus placebo across three scanning conditions: rest, rest while listening to music, and rest after music. Using a Hopf-based whole-brain model fitted to fMRI data recorded under LSD and placebo, the investigators simulated region-specific perturbations and quantified recovery dynamics with the Perturbational Integration Latency Index (PILI). They hypothesised that LSD would produce stronger and longer-lasting responses to perturbation—consistent with dynamics closer to bifurcation or criticality—and expected music to amplify these effects.

Empirical data derived from an fMRI experiment in which 20 healthy participants underwent six scanning conditions: LSD and placebo (PCB) each in rest, rest while listening to music, and rest after music. Sessions were at least 14 days apart and order was balanced; participants were blind to condition. Eight participants were excluded (one early termination for anxiety, four for excessive head motion using a scrubbing threshold, three for technical problems with sound), leaving 12 participants for the analyses reported. The extracted text reports an intravenous LSD dose as "75 g" given ~70 minutes before scanning, but this value appears inconsistent with typical dosing and thus the exact dose is not clearly reported in the extracted text. fMRI acquisition parameters (TR/TE = 2000/35 ms, field of view = 220 mm, 64 × 64 matrix, 7:20 min per BOLD scan) and an extensive preprocessing pipeline are described, including motion scrubbing, spatial smoothing, band-pass filtering (0.01–0.08 Hz), nuisance regression and parcellation into 90 AAL regions (AAL90). Structural connectivity (SC) was obtained from probabilistic tractography applied to DTI data from an independent cohort of 16 healthy participants; the individual SC matrices were averaged to produce the anatomical coupling matrix used in the model. The whole-brain model comprised 90 coupled nodes whose local dynamics were given by the normal form of a supercritical Hopf bifurcation; nodes were coupled through the SC matrix scaled to a maximum of 0.2. Intrinsic node frequencies were estimated from the peak frequency of the filtered empirical BOLD time series for each region and condition. The global coupling parameter G was explored from 0 to 2 in steps of 0.01 and model fits to empirical functional connectivity (FC) were assessed using the Kolmogorov–Smirnov (KS) distance between empirical and simulated FC matrices; 50 simulations per G value were averaged to reduce noise. Functional connectivity matrices were computed as Pearson correlations of band-pass filtered (0.04–0.07 Hz) and z-scored regional BOLD time series, and group-level FC matrices were obtained via Fisher r-to-z transform and averaging. To test condition differences, surrogate datasets were constructed by random permutation of condition labels. A jackknife Gaussian classifier was implemented to test whether single-subject covariance structure predicted drug state: covariances were estimated on N-1 participants and log-likelihood used to assign the left-out subject; significance was evaluated against chance. Two in silico perturbation protocols were implemented by transiently changing the local bifurcation parameter a of a single node for 100 s: a = +0.6 to push a node into a more synchronous (oscillatory) regime and a = −0.6 to push it into a noisier regime. Each node was perturbed individually for 3000 trials. After perturbation offset the parameter was reset to a = 0. For each simulation the instantaneous phase of each regional time series was obtained via the Hilbert transform and a phase-locking matrix computed at each time point. Integration at time t was defined as the size of the largest connected component of the binarised phase-locking matrix averaged over 100 thresholds; integration curves were computed for 200 s in the basal state and following perturbation offset. The Perturbational Integration Latency Index (PILI) was defined as the area under the integration curve from perturbation offset until the integration reached the basal-state maximum (or recovery point), averaged over 3000 trials per node. Statistical comparisons between LSD and PCB PILI distributions used Mann–Whitney U tests (node-wise) with Bonferroni correction across 90 nodes, and RSN-level differences were summarised with Cohen’s d and tested with Mann–Whitney U tests corrected across 7 networks. Response variability was measured as the standard deviation of PILI values across nodes per trial and compared between conditions using a two-sided t-test.

After exclusions, 12 participants contributed to the primary analyses. Comparing empirical FC across conditions, a significant increase in mean FC under LSD versus placebo was observed in the music condition (LSD: 0.204 ± 0.179, PCB: 0.140 ± 0.197; p = 0.0297). Small non-significant increases in mean FC under LSD were reported for rest and rest after music. Fitting the Hopf model to empirical FC produced a shift of the optimal global coupling parameter G towards higher values under LSD in all three scanning conditions (rest: LSD 0.31 vs PCB 0.27; rest with music: LSD 0.35 vs PCB 0.25; rest after music: LSD 0.29 vs PCB 0.28), with the difference reaching significance for the music condition (p = 0.0099). The Gaussian jackknife classifier assigned single-subject covariance patterns to drug state with accuracy above chance in all conditions: 75% for rest, 79.17% for rest with music and 70.83% for rest after music; these performances exceeded the significance threshold given the sample size. Using the fitted models, simulated perturbations produced larger and longer-lasting deviations in the LSD state than in PCB when quantified by the global integration measure. Basal (unperturbed) integration was higher under LSD than PCB in all conditions and increased further in the music condition under LSD while music decreased basal integration in the PCB model. Between the two perturbation protocols, the synchronization protocol (a = +0.6) yielded larger LSD–PCB differences than the noise protocol, so subsequent analyses focused mainly on the synchronization perturbation. PILI values (node-wise, averaged over 3000 trials) were consistently higher under LSD than PCB across all three scanning conditions, with the largest effect in the music condition. Recovery latency analysis showed that for the synchronization protocol, 88/90 nodes had significantly higher latencies in LSD versus PCB in the first resting state, 90/90 nodes in the rest-with-music condition, and 62/90 nodes in the rest-after-music condition. Node-wise statistical testing with Bonferroni correction identified the most affected regions (top 20 reported in the paper), including the cingulate cortex, precuneus, medial orbitofrontal cortex, supplementary motor area, calcarine sulcus, olfactory sulcus, superior and medial frontal gyri, thalamus and hippocampus. At the network level, standardized differences (Cohen’s d) and statistical tests showed significant LSD–PCB PILI increases across all seven resting-state networks in the rest and rest-with-music conditions; in rest after music only two networks (limbic and default mode) survived correction. Across conditions the limbic, visual and default mode networks consistently exhibited the largest PILI differences, with the limbic network showing the highest effect in all three conditions. Finally, perturbation-response variability, measured as the standard deviation of PILI across nodes per trial, was significantly larger under LSD than PCB in all three conditions (p < 0.0001), with the largest effect in the music condition.

Jobst and colleagues interpret their findings as evidence that LSD shifts whole-brain dynamics toward a regime that is more highly connected, more responsive to perturbation and slower to recover — characteristics consistent with operation closer to a bifurcation or critical point. The observed increase in basal functional integration and the shift to higher optimal global coupling under LSD—most pronounced when participants listened to music—are taken to indicate enhanced propagation of activity and communication between regions in the psychedelic state. The larger and longer-lasting perturbation-elicited deviations (higher PILI and longer latencies) are framed as increased sensitivity and ‘‘critical slowing’’, which the authors link to prior work suggesting greater brain complexity under psychedelics. Locally, the limbic, visual and default mode networks were identified as especially sensitive to perturbation under LSD; the limbic system in particular showed strong effects, which the investigators relate to heightened emotional responsivity and the therapeutic relevance of emotional processing in psychedelic-assisted therapy. Increased perturbation-response variability across regions is interpreted as greater heterogeneity or independence of regional dynamics, consistent with a breakdown of typical hierarchical constraints and an expanded repertoire of brain states. Music appears to amplify these dynamical changes, consistent with reports that music intensifies emotional and imagery-related aspects of the psychedelic experience. The authors acknowledge several important limitations. The whole-brain model was homogeneous in local dynamics (all regions set to a = 0 in the basal state), meaning regional differences in results arise only from differences in connectivity rather than intrinsic regional parameters; introducing heterogeneity would require estimating many additional parameters. The model is also limited to slow BOLD frequencies and cannot speak to faster electrophysiological rhythms; the authors propose that MEG data would be informative for frequency-specific effects. The relatively small final sample (12 participants) is noted, and although single-subject classification exceeded chance, the authors state that larger samples would strengthen generalisability. Finally, pharmacokinetic timing may explain the reduced effects in the post-music scan, and the extracted text does not clearly report the precise LSD dose delivered. Overall, the investigators conclude that an in silico perturbational whole-brain modelling approach can reveal mechanistic features of the psychedelic state, showing increased global connectivity, regional sensitivity (notably limbic, visual and DMN), and heightened variability — findings the authors argue are consistent with theories that psychedelics move brain dynamics closer to criticality and that contextual factors such as music modulate these effects.