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

In this neuroimaging study (n=14) data from participants who were given 75μg of intravenous LSD or placebo and brain activity was assessed using fMRI and a novel whole-brain computer model (in silico) approach. The largest deviations from normal brain function were found in the limbic network, the visual network and the default mode network (DMN). It was found that the computer model used allows for the exploration of changes in brain dynamics that can be challenging to observe via experiments in living subjects (in vivo).

Abstract

“Lysergic acid diethylamide (LSD) is a potent psychedelic drug, which has seen a revival in clinical and pharmacological research within recent years. Human neuroimaging studies have shown fundamental changes in brain-wide functional connectivity and an expansion of dynamical brain states, thus raising the question about a mechanistic explanation of the dynamics underlying these alterations. Here, we applied a novel perturbational approach based on a whole-brain computational model, which opens up the possibility to externally perturb different brain regions in silico and investigate differences in dynamical stability of different brain states, i.e. the dynamical response of a certain brain region to an external perturbation. After adjusting the whole-brain model parameters to reflect the dynamics of functional magnetic resonance imaging (fMRI) BOLD signals recorded under the influence of LSD or placebo, perturbations of different brain areas were simulated by either promoting or disrupting synchronization in the regarding brain region. After perturbation offset, we quantified the recovery characteristics of the brain area to its basal dynamical state with the Perturbational Integration Latency Index (PILI) and used this measure to distinguish between the two brain states. We found significant changes in dynamical complexity with consistently higher PILI values after LSD intake on a global level, which indicates a shift of the brain’s global working point further away from a stable equilibrium as compared to normal conditions. On a local level, we found that the largest differences were measured within the limbic network, the visual network and the default mode network. Additionally, we found a higher variability of PILI values across different brain regions after LSD intake, indicating higher response diversity under LSD after an external perturbation. Our results provide important new insights into the brain-wide dynamical changes underlying the psychedelic state – here provoked by LSD intake – and underline possible future clinical applications of psychedelic drugs in particular psychiatric disorders.”

Authors: Beatrice M. Jobst, Selena Atasoy, Adrián Ponce-Alvarez, Ana Sanjuán, Leor Roseman, Mendel Kaelen, Robin Carhart-Harris, Morten L. Kringelbach & Gustavo Deco

Notes

Author Highlights

• Novel offline perturbational method applied on functional magnetic resonance imaging (fMRI) data under the effect of lysergic acid diethylamide (LSD)

• Shift of brain’s global working point to more complex dynamics after LSD intake

• Consistently longer recovery time after model perturbation under LSD influence

• Strongest effects in resting state networks relevant for psychedelic experience

• Higher response diversity across brain regions under LSD influence after an external in silico perturbation

Despite a large amount of research, the exact mechanisms through which psychedelics exert their effects on the brain remain unknown. Understanding these mechanisms is imperative to realizing the therapeutic potential of psychedelic’s and make these drugs accessible to people with mental disorders. Neuroimaging studies are helping to illuminate these mechanisms of action and therefore, it is important we continue to explore all possible techniques.

In the present study, researchers investigated the effects of LSD on the brain using data from 14 healthy participants using fMRI (from Carhart-Harris et al., 2016) and a novel silico perturbation approach. In general, silico models are logical extensions of controlled experiments in living participants or cells (in vivo) that are created by a computer that models the pharmacological process.

The main findings

• LSD intake led to a shift of the brains global working point further away from a stable equilibrium when compared to normal conditions.
• The largest deviations from normal brain function were found in the limbic system, the visual system and the default mode network (DMN).
• The novel perturbation approach allowed researchers to better understand how the changes in brain function induced by LSD interact with the connectome (a complete map of structural connectivity in the brain) to produce different network dynamics.
• The model may be useful for futures research as it allows for the exploration of characteristic changes in whole-brain dynamics in ways that are extremely challenging to do via in vivo experiments.

The authors acknowledge that limitations exist surrounding their novel technique. Mainly, the model assumes that all brain regions have the same intrinsic dynamics and it has a limited frequency range. Nonetheless, the present study has enriched our understanding as to how psychedelics may exert their effect on the brain all while providing a novel technique that may be useful for future research.

Summary

1. Introduction

LSD is a potent psychoactive drug that was first synthesized in 1938 and whose potent psychological effects were discovered in 1943. It was made illegal in the late 1960s, but has since undergone a renaissance in clinical and brain research.

Human neuroimaging studies have identified several neural correlates of the psychedelic state provoked by hallucinogenic drugs, including an increase in visual cortex blood flow and an expanded visual cortex functional connectivity.

In psychedelic states, the brain shows a variety of functional alterations, including an increased variance of the Blood-Oxygen Level Dependent (BOLD) signal and a higher diversity of dynamic functional connectivity states. We use a computational model to simulate external perturbations of any brain region.

In the last 15 years, several studies have explored the dynamical responses of different brain regions to controlled artificial external perturbations. These studies have been limited to transcranial magnetic stimulation (TMS) in healthy human subjects and to deep brain stimulation (DBS) in patients.

We apply a novel in silico model perturbation approach to study the perturbation-elicited changes in global and local brain activity under the influence of LSD in three different scanning conditions. We used a computational whole-brain model to simulate the dynamics in each brain area with the normal form of a supercritical Hopf bifurcation. This model allows systematical perturbation of each brain region in silico without needing to perturb the brain activity explicitly.

Brain activity without the model based perturbation is consistent with more complex and less stable dynamics, as well as brain dynamics closer to bifurcation or critical regime. LSD has been shown to re-organize brain dynamics at the edge of criticality, to take longer to go back to its original state after perturbation, and to propagate to other brain regions beyond the original stimulation site in an awake resting state as opposed to deep sleep.

2.1. Functional magnetic resonance imaging (fMRI) data

20 healthy participants were scanned in 6 different conditions: LSD resting state, placebo resting state, LSD and PCB resting state while listening to music, LSD and PCB resting state after listening to music. All participants gave informed consent. Eight out of 20 subjects were excluded from further analyses, four were excluded due to high levels of head movement, and three were excluded due to technical problems with the sound delivery in the music condition. Participants received either 75 g of LSD (intravenous, I.V.) or saline/placebo (I.V.) 70 min prior to MRI scanning. The MRI scans lasted for about 60 min, and participants performed subjective ratings inside the scanner via a response box. Carhart-Harris et al. (2016 b) performed the following pre-processing steps: 1) removal of the first three volumes; 2) de-spiking; 3) slice time correction; 4) motion correction; 5) brain extraction; 6) rigid body registration to anatomical scans; 7) non-linear registration to 2 mm MNI brain; 8) scrubbing.

The BOLD signals were averaged over 90 regions of interest (ROIs) in the automated anatomical labeling (AAL) atlas, comprising 45 regions in each hemisphere. The AAL parcellation seems to be particularly well fitted for studying the spatiotemporal dynamics on a whole brain level.

2.2. Anatomical connectivity

The anatomical connections between the different brain areas were obtained from Diffusion Tensor Imaging (DTI) data of 16 healthy right-handed participants. The data was averaged across participants and used to construct structural connectivity maps.

2.3. Hopf computational whole-brain model

The brain activity in each brain region was simulated with a computational whole-brain model, which was previously described in various publications. The model is based on the 90 coupled brain regions, comprising cortical and subcortical areas, retrieved from the AAL parcellation explained above. The dynamics of a given uncoupled node can be described by the normal form of a supercritical Hopf bifurcation, which can be used to simulate electroencephalography, magnetoencephalography and fMRI dynamics.

The normal form of the system possesses a supercritical Hopf bifurcation at a = 0. The system settles into a stable limit cycle, producing self-sustained oscillations with frequency f j = j 2 , and noise-induced oscillations are observed.

This model can be interpreted as an extension of the Kuramoto model with amplitude variations, and a global coupling strength scaling equally the total input in each brain node.

2.4. Functional connectivity estimation

The BOLD signal of each AAL region was detrended, demeaned and band-pass filtered within the range of 0.04 – 0.07 Hz following Glerean et al. (2012 ). The functional connectivity (FC) matrices were then calculated for each participant in each condition.

We generated 6 final FC matrices, one for each condition, and compared them across subjects individually for each condition. We then generated 100 surrogate datasets to test the significance of the differences between drug-induced conditions.

To ensure that there would be no difference between FC matrices between the group of participants who received PCB in their first session and the group of participants who received PCB in their second session, we performed a statistical significance analysis.

We further analyzed the differences between the LSD and PCB states between the two groups, those who received PCB in their first session ( “First “) and those who received PCB in their second session ( “Second “), and tested for statistical significance.

2.5. Drug state classification with Gaussian classifier

We classified the drug state based on the covariance of fMRI signals using a jackknife cross-validation approach, and then associated the data of the remaining subject to a drug state by selecting the zero-mean multivariate Gaussian process which maximised the log-likelihood of the test data given the trained model.

We computed the probability of getting k correct classifications by chance and found that the performance exceeded the 95th percentile.

2.6. Fitting the model to experimental data

We explored the parameter space of the whole-brain computational model by varying the global coupling strength parameter G from 0 to 2 in steps of 0.01. The coupling parameter values where the fitting curves were minimal were then used for the following analysis steps.

2.7. Model perturbation protocols

We used the locally defined bifurcation parameter a of the Hopf model to simulate two kinds of off-line perturbation protocols. We perturbed each node individually, repeated the perturbation procedure 3000 times and performed statistical analyses using the error of the distribution averaged over the 3000 trials.

2.8. Integration measure

We measured the level of integration across all brain regions for each time point using fMRI data.

The phase locking matrix P was calculated for two brain regions p and q using the Hilbert transform.

We binarized the phase locking matrix P for 100 evenly spaced thresholds between 0 and 1, and calculated the level of integration at time t as the integral of the curve of the largest connected component as a function of the thresholds.

2.9. Perturbative integration latency index (PILI)

We calculated the Perturbative Integration Latency Index to characterize the return of the brain dynamics to the basal state after a model perturbation.

We calculated the PILI for each brain area by integrating the 200 s basal state over 3000 trials and then perturbing the system following the procedure described above. The statistical significance tests were performed across the 3000 trials applying a Mann-Whitney U test.

2.10. Region-wise and resting state network analysis

We computed PILIs for 90 brain areas and compared LSD and PCB in three scanning conditions. Bonferroni correction was applied to correct for multiple comparisons.

We evaluated the differences between PILI values in seven commonly observed resting state networks (RSNs) by calculating Cohen’s d-values, and then ordered the RSNs from highest to lowest Cohen’s d-value, where the higher the d-value, the larger the response to a model perturbation under the influence of LSD in one particular RSN.

The significance between the LSD and PCB state models in each condition was tested using the Mann Whitney U test.

2.11. Response variability

We calculated the variability of the PILI values over different brain regions and evaluated statistically significant differences between the LSD and PCB induced brain states.

3. Results

We applied a previously published offline perturbational approach based on a whole-brain model to investigate the differences between LSD and PCB brain states in three different scanning conditions.

3.1. Functional connectivity and optimal working point

We investigated the differences in functional connectivity between the LSD and PCB brain states in all three scanning conditions. We found a significant difference in the mean FC values between LSD and PCB in the music condition, but not in the resting conditions.

We fitted the Hopf whole-brain model to the fMRI data in each condition and found that the optimal global coupling parameter G shifted towards higher values under the influence of LSD in all three scanning conditions, with a significant difference in the music condition.

3.2. Drug state classification with Gaussian classifier

We assessed how specific the functional connectivity is to the drug state (LSD or PCB) using jackknife cross-validation. We found that the drug states were predicted with an accuracy exceeding the significance level for all 3 scanning conditions.

3.3. Global differences in integration

We simulated two kinds of model perturbation protocols for each brain state, and compared the brain states using the global integration measure.

We adjusted the whole-brain model to the fMRI data and simulated two perturbation protocols for LSD and PCB states. We quantified the perturbation-caused changes in brain-wide signal interactions over time.

In the LSD state, the level of BOLD signal connectedness was higher without perturbation than in the PCB state, and the basal integration increased under the influence of LSD while listening to music, whereas in the PCB state, the basal integration decreased with music. We found that the deviations from the basal activity were stronger and longer-lasting under the influence of LSD in comparison with PCB after being exposed to the same kind of perturbation.

3.4. Global and local differences in perturbative integration latency index

We computed the Perturbative Integration Latency Index (PILI) for each node to capture the strength of deviation from the basal state and duration of the recovery after a model perturbation.

We found that the LSD induced brain state model showed higher PILI values than the PCB induced brain state model in all three scanning conditions, where the effect was strongest for the music condition.

We calculated the time for the perturbed signals to come back to the basal state in the first resting state, in the Rest with Music condition, and in the Rest after Music condition when compared to PCB. The longer lasting perturbation effect is most prominent in the Music condition.

We looked at the PILI values on a node-to-node basis to gain further insights into local processes. The results show that the 20 brain areas with the highest PILI differences are shown in order from smallest to largest p-value with their according effect sizes.

The results reveal that the dynamical responses of the brain as a whole to an external model perturbation are stronger and longer lasting under the influence of LSD when compared to PCB. Furthermore, this effect is amplified in the model estimated from data in which participants listen to music.

3.5. Relationship of PILI to resting state networks

We computed Cohen’s d values to assess differences in PILI values between LSD and PCB state models in seven reference RSNs.

In all three scanning conditions, the limbic, visual and default mode networks showed the highest PILI differences between the LSD and PCB state models. The somato-motor network came fourth to the first three RSNs by Cohen’s d values, whereas the ventral attention network gained more importance in the music condition.

Three resting state networks, limbic, visual and default mode, show increased sensitivity under the influence of LSD, in line with previous studies.

3.6. Increased perturbation response variability in LSD condition

We found that the perturbation response variability was higher under the influence of LSD than for PCB in all three scanning conditions and both drug states for the synchronization protocol.

4. Discussion

We used a novel in silico model-based perturbational approach to analyze the perturbation-elicited changes in global and local brain activity under the influence of LSD compared with PCB in three consecutive scanning conditions. We found that the limbic network, the visual network, and the default mode network were most sensitive to these changes.

We found that the functional connectivity between the thalamus and the association cortices was higher on average in the LSD condition compared with the PCB condition, and this difference was especially pronounced in the music condition, where the effects of LSD seem to be amplified. Similar results have been reported for other psychedelic drugs such as psilocybin, which increases the variance of the Blood-Oxygen Level Dependent (BOLD) signal measured with fMRI and the diversity of dynamic functional connectivity states.

We assessed how specific the functional connectivity is to the drug state and found that the characteristics of the single participants are reflected in the group-level results. This suggests that even a small number of participants can be seen as a representative sample.

We applied a whole-brain model to simulate fMRI BOLD responses and found that the global working region of brain dynamics shifts to higher global coupling parameters in the LSD state when compared with PCB, and that this shift is accentuated under conditions of significant emotional evocation.

Increased brain criticality is consistent with the entropic brain hypothesis.

We used a model-based perturbation approach to evaluate the responses to strong offline model perturbations in each scanning condition, and found that the new methodology allowed for stronger, longer lasting and brain node-specific perturbations in ways not possible experimentally.

We found that the global integration was increased under LSD in contrast to PCB, which was again amplified in the music condition. This indicates that the communication and interaction between distinct brain areas is enhanced under the influence of LSD.

Music increased basal integration in the LSD condition, while it decreased basal integration in the PCB condition. This could be related to an accentuated psychological response to music under psychedelics. Music could be characterized as a type of intrinsic perturbation under LSD, but less so under placebo, where it is more likely to be witnessed more as an external object.

Almost every node revealed a marked difference in PILI values under LSD versus placebo, and this was evident across all three scanning conditions. This suggests that the brain exhibits a diminished ability to homeostatically ‘right itself’ after perturbation under LSD.

A simulated perturbation to LSD fMRI data was found to increase sensitivity and the stronger reaction to a model perturbation. This finding is consistent with the entropic brain hypothesis and the finding of elevated brain complexity under psychedelics.

We found that the limbic network showed the highest perturbation-elicited differences between the LSD and PCB state models, and that the cingulate cortex showed a remarkably large sensitivity. The cingulate cortex and the limbic system are both implicated in emotional processing and the brain action of psychedelics, and are also associated with producing transient dreamlike states with visual hallucinations upon electrical depth stimulation. The findings of enhanced sensitivity to a model perturbation of the limbic network support the idea that heightened sensitivity of the limbic circuitry in particular is implicated in the heightened emotional responsivity that has been found in relation to psychedelic therapy.

Two other networks, the visual network and the default mode network, were strongly altered by LSD, consistent with previous studies reporting changes in the functioning of visual areas and in the functional properties of the DMN under LSD.

We found that the response variability was higher in all three scanning conditions under LSD than PCB, which indicates an enhanced diversity in brain dynamics, as also previously suggested for the LSD state. This is consistent with a breakdown in the usual hierarchical constraints governing global brain function. The relationship between hierarchical organization in the brain and criticality was the focus of a recent major review.

The present study suggests that the brain is more sensitive to perturbation under the influence of psychedelic drugs, which is consistent with evidence-based assumptions about increased emotional sensitivity to environmental and other contextual factors (such as music) under psychedelics.

The present model allows us to understand how the global changes induced by LSD interact with the connectome and produce different network dynamics. Its main limitation is its homogeneity, but it could be extended by introducing heterogeneity in local dynamics.

We used a novel in silico perturbational approach to explore the underlying mechanistic properties of the whole-brain dynamics in the LSD state, and have provided important new insights into global brain function underlying a possible altered state of consciousness.

Author contributions

The authors designed the study, performed data analyses and numerical simulations, and wrote the first version of the manuscript.

Credit author statement

BMJ, GD, SA, AP-A, AS, LR, MK, RC-H and MLK designed the study, performed data analyses and numerical simulations, and contributed significantly to the writing of the manuscript.