Synergistic, multi-level understanding of psychedelics: three systematic reviews and meta-analyses of their pharmacology, neuroimaging and phenomenology

The first functional magnetic resonance imaging (fMRI) study of psychedelics, published in 2012 [20], was an exploratory analysis of psilocybin-induced changes in cerebral blood flow and BOLD activity in healthy human participants. Since then, dozens of fMRI studies have been performed on human participants under the influence of ayahausca, LSD, psilocybin and DMT, including several studies on depressed patients [21 –26].

The fMRI studies assess three different metrics of brain activity: BOLD activation, connectivity, and entropy. Studies on BOLD activation analyse changes in the trajectory of BOLD timeseries with a rapidly-acting psychedelic, such as intravenously-administered psilocybin. Connectivity studies can examine undirected, instantaneous correlations between different regions (a type of functional connectivity) or model the experimental factors that modulate directed connections between regions, using dynamic causal modelling (a tool for measuring effective connectivity). The entropy of spontaneous brain activity is best assessed with MEG and EEG, but efforts to compute entropy on the spontaneous BOLD signal, as well as the entropy rate (Lempel-Ziv complexity) and indices of criticality, have also been attempted (see Section S1.2 for a full list of citations). While studies on BOLD activation and functional connectivity are common for most subject areas within cognitive neuroscience, entropy is a rather unique feature of the psychedelic neuroimaging literature. The studies on entropy were motivated by Robin Carhart-Harris' influential Entropic Brain Hypothesis (EBH) [27, 28], which states that psychedelics alter consciousness by elevating the entropy of spontaneous brain activity across time.

We performed a quantitative meta-analysis of the pairwise functional connectivity data using a novel algorithm. We initially attempted to conduct a quantitative meta-analysis of the BOLD activation data with the GingerALE method, which examines common clusters of BOLD activity across studies. However, we found that the data was too heterogeneous for the results to be valid. From the outset, we decided to provide a qualitative discussion of the studies on entropy, due to the wide range of methods in that section of the literature.

We chose to perform a meta-analysis exclusively on the fMRI data and not on any of the other modalities used to measure brain activity, such as PET, SPECT, MEG, and EEG. There were too few primary PET (4), SPECT (4), or MEG (2) datasets to merit a meta-analysis; by comparison, 30 fMRI studies have collected original data. While there are many more primary EEG datasets (25) from the post-1960s era (and even more from the 1950s to 1960s era) than PET, SPECT, or MEG, most of the EEG studies do not spatially localise the brain activity recorded at EEG electrodes. Without the raw data, which may be difficult to access when some of the studies are nearly 20 years old, we cannot source-localise the EEG activity ourselves. On the other hand, most fMRI studies report the spatial coordinates of the brain regions that become more or less active or connected on psychedelics. In general, EEG has higher temporal resolution than fMRI [29], so a future meta-analysis that focuses on the temporal, rather than spatial, characteristics of brain activity on psychedelics could consider the EEG rather than fMRI data.

All classic psychedelics activate serotonin receptors. In particular, their interaction with the serotonin-2A (5-HT_2A) receptor is their primary mechanism of action [30, 31]. Multiple studies have demonstrated that blocking the 5-HT_2A receptor with antagonists such as ketanserin eliminates their subjective effects [32 –34]. 5-HT_2A expression is profuse in human cortex, with one study finding highest expression in the posterior cingulate cortex and V1 [35]. In both humans and monkeys, 86-100% of glutamatergic cells in layers II–V of cortex express mRNA encoding 5-HT_2A receptors [36]. 5-HT_2A has also been identified in many subcortical regions [2], though human PET imaging implies lower expression than in the cortex [35].

Although the consciousness-altering effects of psychedelics do primarily arise from their action at the 5-HT_2A receptor, different psychedelics display different binding profiles. For instance, unlike DMT and psilocybin, LSD has moderate affinity for dopamine receptors [37]. Differences in binding affinities could partially account for the distinctive phenomenology of each psychedelic. There is also some evidence that the 5-HT_1A and 5-HT_2C receptors may mediate the effects of psychedelics [38, 39].

However, binding affinity paints an incomplete picture of the pharmacology of psychedelics. Another key aspect is their functional activity at receptors. Serotonin and dopamine receptors tend to be G protein-coupled receptors (GPCRs) [40, 41], which perform essential roles in neurophysiological processes like smell, taste, light perception, and more [42]. GPCRs regulate these processes through G protein signalling pathways, which are initiated when agonists bind to GPCRs and activate G proteins [43]. Different families of G proteins are selective for particular signalling pathways [44]. In a phenomenon known as biased agonism, agonists can induce certain conformations in GPCRs, which then selectively activate certain pathways and not others [45]. 5-HT₂ receptors couple preferentially to a family of G proteins known as Gα_q/11, which activates the enzyme phospholipase C (PLC). This enzyme then catalyses the synthesis of the secondary messenger inositol triphosphate (IP₃, a species of inositol phosphate [IP]), which subsequently leads to the release or "mobilisation" of calcium from the endoplasmic reticulum [31]. Psychedelics are also known to recruit β-arrestin proteins [46 –49]. These proteins block the interaction between G proteins and the GPCR and remove the GPCR from the cell membrane, while also coupling GPCRs to signalling proteins without activating G proteins; thus, β-arrestin promotes alternative signalling pathways [50]. Functional GPCR assays on psychedelics have tended to focus on the three aspects of GPCR signalling described above: (1) IP formation, (2) calcium mobilisation, and (3) β-arrestin (specifically β-arrestin2) recruitment.

We performed a meta-analysis of the selective affinity of DMT, LSD, and psilocin for the 5-HT_2A, 5-HT_2C, and D₂ receptors, relative to the 5-HT_1A receptor. Additionally, we conducted a meta-analysis of their functional activity at the 5-HT_2A receptor, as measured by the three aforementioned assays.

We performed a meta-analysis on the 5D- and 11D-ASC scores of DMT, LSD, and psilocybin. Our literature search identified n = 44 phenomenology studies that met the inclusion criteria, including 5 studies on DMT (5D-ASC: n = 3; 11D-ASC: n = 3), 14 studies on LSD (5D-ASC: n = 9; 11D-ASC: n = 12), and 25 studies on psilocybin (5D-ASC: n = 12; 11D-ASC: n = 17). The eleven dimensions of the 11D-ASC are subscales of three dimensions in the 5D-ASC: oceanic boundlessness (OB), anxious ego dissolution (AED), and visionary restructuralisation (VR). Nevertheless, the 11D-ASC questionnaire contains fewer items than the 5D-ASC questionnaire, so 5D-ASC data cannot be directly compared to 11D-ASC data. Hence, we ran separate meta-analyses for the 5D-ASC and 11D-ASC data. In order to account for the measurement of multiple subjective dimensions, sometimes with more than one dose or drug, in individual studies, we performed a multilevel random-effects meta-analysis.

DMT was administered intravenously (IV) in the phenomenological studies, whereas LSD and psilocybin were administered orally (we excluded phenomenological studies that used IV LSD and psilocybin). IV administration yields different pharmacokinetics, which are known to influence the subjective experience of the drug [51, 52]. Therefore, while we conducted meta-analyses of all three drugs, we only examined significant between-drug differences for ASC ratings of LSD and psilocybin. Additionally, we only compared similar doses of LSD and psilocybin. We assumed 20 mg psilocybin is equivalent to 0.01 mg LSD [14], and we defined low, medium, and high doses based on the literature [53 –55].

Within-drug differences between the five subjective dimensions were quite similar. A clear separation can be seen between two groups of subjective categories: (1) OB, VR, RV, and (2) AA and AED. At medium doses, VR and OB scores were significantly higher than both AED and AA scores for both LSD and psilocybin. For DMT, there were no significant differences between dimensions at high doses (≥16 mg bolus injection, followed by continuous infusion of 1 mg/min), but for low doses (10–15 mg bolus + 0.6–0.99 mg/min continuous infusion), AED ranked significantly lower than both OB and VR. (Note that only one study measured AA and RV scores for DMT [52], so we did not perform a meta-analysis on these dimensions for DMT.)

The results of the 11D-ASC meta-analysis were not very consistent with those of the 5D-ASC meta-analysis (Fig.; Table). Note that we could not include DMT in the 11D-ASC meta-analysis because two of the three studies on DMT did not report standard errors, which are necessary for pooling the subjective data. At high doses, there were no significant differences between LSD and psilocybin in any of the dimensions, and standard errors were very large for psilocybin in some dimensions, particularly in some OB and VR subscales. At medium doses, where standard errors tended to be lower, we observed that two of the OB subscales (experience of unity and insightfulness), one of the AED subscales (impaired control and cognition), and three of the VR subscales (complex imagery, audio-visual synaesthesia, and changed meaning of percepts) ranked significantly higher for LSD than psilocybin. In line with the 5D-ASC results, we did not find any significant differences between LSD and psilocybin at low doses. However, unlike the 5D-ASC analysis, the 11D-ASC analysis showed that psilocybin rated higher than LSD in most of the dimensions; that being said, standard errors were very high for both LSD and psilocybin, so our estimates are not reliable. For medium doses, within-drug comparisons for the 11D-ASC analysis generally aligned with those of the 5D-ASC analysis. S3 S2

We did not find any significant relationship between the pooled 5D-ASC scores and three methodological covariates that were defined a priori: prior psychedelic use, the time that the questionnaire was administered relative to the drug, and the presence or absence of a task in the experiment. Additionally, there was significant residual heterogeneity in the pooled 5D- and 11D-ASC scores (p < 0.0001 in both cases). Unfortunately, we were unable to obtain reliable estimates of heterogeneity specifically attributable to within-study and between-study differences. For three of the 5D-ASC scales, we found evidence of significant publication bias with Egger's regression test, which determines whether small studies report disproportionately high effect sizes (OB: p = 0.0187, AED: p < 0.0001, VR: p = 0.6279, AA: p < 0.0001, RV: p = 0.6532). The results of our risk-of-bias assessments for individual studies are shown in Table S3. Due to the moderate-to-serious risk of bias, significant residual heterogeneity, high probability of publication bias, and indirectness (some studies measured subjective effects while participants were performing a specific task, as opposed to during resting-state), our certainty in the body of phenomenological evidence is low.

In order to directly compare the phenomenology of psychedelics to their neuroimaging and pharmacological profiles (Sectionand Section), we sought to determine the "phenomenology profiles" of DMT, LSD, and psilocybin based on the neural correlates of their subjective effects. We performed a separate literature search of studies that measured correlations between ASC ratings and fMRI activity. We determined the Yeo network that contained each brain region that exhibited a significant correlation. The phenomenology profiles were defined by a weighted combination of the mean correlations between each ASC scale and each Yeo network, in which the weights were the pooled ASC ratings of high doses of each psychedelic for the corresponding scale. 2.2 2.3

Finally, we also aimed to relate the phenomenology of different psychedelics by qualitatively synthesising the findings of papers that directly compared the subjective experiences of DMT/ayahuasca, LSD, and psilocybin. Sectioncontains a comprehensive summary of their major aims, analysis approaches and findings. S3

As stated above, the fMRI studies on psychedelics measure three different characteristics of brain activity: BOLD activation, entropy, and connectivity. We identified n = 17 studies on BOLD activation, including n = 3 studies on ayahuasca or DMT, n = 7 studies on LSD, and n = 4 studies on psilocybin. Initially, we used the GingerALE method to determine common clusters of BOLD activation across studies. We observed that psilocybin was associated with a cluster of activity in visual cortices, whereas LSD tended to affect more frontal areas. However, three of the four studies on psilocybin displayed visual stimuli to participants. Therefore, our results were likely confounded by the tasks that were used in the literature. We chose not to report the results of our GingerALE analysis because they are likely misleading.

The n = 12 fMRI studies on entropy and criticality in the psychedelic literature are very heterogeneous, employing a wide variety of measures on dissimilar variables of interest. A quantitative meta-analysis was extremely challenging, so we opted instead to perform a qualitative review of this literature, which is reported in Section S4.

The studies on connectivity were the only segment of the fMRI literature for which we determined that we could conduct a valid meta-analysis. These studies tended to use the same resting-state experimental procedure, and many of them employed similar methods for measuring connectivity. However, one major obstacle was the variety of parcellations in the literature. Since each study defined regions of interest (ROIs) in a different way, it was not possible to simply perform a weighted average of connectivity values across studies. (This weighted average method is essentially the core technique of the phenomenology and pharmacology meta-analyses.) An alternative approach is to "re-reference" the data to a coarse-grained parcellation: the Yeo networks, of which there are only seven in the brain [56]. That is, for every pair of functionally connected ROIs in the literature, we determine the corresponding Yeo networks that contain them. If the connection is positive (it increases under psychedelics), then the aggregate FC between those Yeo networks is incremented; otherwise, it is decremented. Our method for weighting each connection in the literature is described in Section S2.2.2. We determined significance by repeatedly applying our method to an independent resting-state fMRI dataset of sober, healthy individuals [57], which resulted in a null distribution (Fig. S4).

This method is capable of synthesising studies on pairwise FC, such as seed-to-seed, seed-to-voxel, and within- and between-network FC analyses using independent components analysis (see Table S5 for a full description of the inputs). The algorithm cannot accommodate global, structural metrics of connectivity such as global brain connectivity (GBC) or graph-theoretic modularity. Additionally, the algorithm does not capture any information about the direction of connectivity between regions; therefore, we did not incorporate studies of effective connectivity or directed FC. (However, the vast majority of FC studies measured undirected FC.)

Many FC studies are secondary analyses of the same primary dataset. For instance, the primary FC data from Carhart-Harris et al. [58] was re-analysed at least seven times with different methods in subsequent studies. To mitigate bias, we only inputted the most informative analysis of each unique dataset into the meta-analysis algorithm, such that each dataset was only represented once in the meta-analysis. Our method for selecting the most informative analysis is described in Section S2.2.2.

Figures–display the results of our meta-analysis on the FC of subcortical regions. Because several studies did not measure subcortical FC, most of the subcortical-to-subcortical and subcortical-to-cortical edges in the aggregate FC matrix are insignificant. Nevertheless, the significant edges indicate that psychedelics elevate connectivity from two regions in the subcortex – the anterior thalamus and cerebellum – to cortex, as well as from the anterior thalamus to some other subcortical structures. Compared to the other psychedelics, LSD vastly elevated the FC between the anterior thalamus and both the cortex and subcortex. S6 S8

Finally, there was no significant association between our results and any of the methodological covariates in the literature that we identified a priori: the width of the Gaussian kernel used to smooth the data, the use of FSL or SPM, the technique used to regress out white matter, the use of scrubbing to correct head motion, and route (oral or intravenous) and relative timing of drug administration. The results of our risk-of-bias assessment for individual studies are shown in Table. Because of the moderate-to-serious risk of bias, our certainty in the body of FC literature is low. S6

We performed two separate meta-analyses on the pharmacology of psychedelics. The first assessed the selective affinity of DMT, LSD, and psilocin for the 5-HT_2A, 5-HT_2C, and D₂ receptors, relative to 5-HT_1A, as well as selective affinity for the 5-HT_1A, 5-HT_2C, and D₂ receptors, relative to 5-HT_2A. The second examined the relative functional activity of the psychedelics at the 5-HT_2A receptor, as captured in three different assays of GPCR signalling.

For the phenomenology literature, we conducted a meta-analysis of both the 5-dimensional and 11-dimensional versions of the Altered States of Consciousness (ASC) scale, a common questionnaire for measuring changes in subjective experience on psychedelics. At both medium and high doses, LSD ranked significantly higher than psilocybin in "visionary restructuralisation," a dimension capturing the quality and intensity of visual hallucinations. Additionally, at medium doses, LSD was associated with significantly higher scores in the "oceanic boundlessness" dimension, which captures feelings of interconnectedness.

The whole-brain neuroimaging data consisted of three subsets of data: BOLD activation, entropy, and FC. We did not report a meta-analysis of the BOLD activation or entropy literature due to the heterogeneity of experimental procedures and analysis methods, respectively; however, we did conduct a qualitative review of the entropy literature (Section). To perform a meta-analysis on the FC data, we essentially re-parcellated the published data into the Yeo networks and then calculated a weighted sum of the connections between Yeo networks. LSD strongly elevated FC between the visual network and the other networks, whereas ayahuasca/DMT increased FC most between transmodal networks (specifically, the FPN and DMN) and other networks. S4

For pharmacology, we conducted two meta-analyses: one of selectivity and another of relative functional activity. There were no significant differences between psychedelics in selectivity for the 5-HT_2A, 5-HT_2C, or D₂ receptors relative to the 5-HT_1A receptor. Compared to both DMT and psilocin, LSD elicited significantly more relative activity in IP formation assays. We did not observe any significant between-drug differences for the calcium mobilisation and β-arrestin2 recruitment assays.

On a superficial level, there do appear to be some direct correspondences between the phenomenology, neuroimaging, and pharmacology. Phenomenologically, we only found one significant difference between psychedelics that was consistent across medium and high doses: LSD induces more visionary restructuralisation than psilocybin does. Neurobiologically, LSD enhanced connectivity between the visual network and all other networks more than the other psychedelics.

The relationship between the pharmacology results and the above findings is more difficult to establish. Our results show that LSD increases IP formation, whereas DMT and psilocin decrease it. IP formation occurs whenever Gα_q proteins are recruited to initiate cellular signalling. Furthermore, there is evidence that Gα_q activation is necessary for 5-HT_2A agonists to bring about hallucinogenic effects. When administered the psychedelic (±)1-(2,5-dimethoxy-4-iodophenyl)-2-aminopropane (DOI), mice that lack the gene for Gα_q proteins do not exhibit behaviours (head twitches) that typically indicate hallucinations [68]. More recently, activation of Gα_q pathways by 5-HT_2A agonists has been shown to predict head-twitch behaviour, whereas 5-HT_2A agonists that are biased for the β-arrestin2 pathway do not induce head twitches [69]. There appears to be a threshold level of Gα_q activation that must be surpassed in order for a 5-HT_2A agonist to induce hallucinogenic effects [69]. Intriguingly, the pro-psychotic properties of drugs can be predicted from the extent to which they increase Gα_q activation (relative to a different G-protein pathway) when these drugs interact with a complex formed between the 5-HT_2A receptor and a metabotropic glutamate receptor [70]. In other words, the more that a drug upsets the normal balance of Gα_q signalling, the more likely it is to engender hallucinations. However, these studies only establish the presence of hallucinations elicited by Gα_q signalling. They do not indicate whether Gα_q signalling induces visual hallucinations in particular. Thus, based on the available evidence, we cannot yet conclude that the pharmacology, neuroimaging, and phenomenology literature all convergently indicate that LSD has a uniquely strong effect on visual awareness, compared to the other psychedelics.

There do not appear to be any tools at the moment for relating the pharmacology of drugs to the precise nature of the subjective experiences that they produce. Methods for examining the relationship between neuroimaging and phenomenology are currently limited to correlations and linear regressions [58, 71], which do not capture causality. There is a need to develop techniques for elucidating the causal relationships between phenomenology, neuroimaging, and pharmacology.

We encourage researchers to develop tools for modelling the nonlinear relationships between the pharmacology, neuroimaging, and phenomenology of psychedelics, the three hierarchical levels of analysis that are at the heart of this meta-analysis. While some researchers have attempted to examine the associations between each level [15, 130], their techniques are model-free and only measure correlations between the effects of psychedelics, rather than elucidating the causal mechanisms that underpin them.

Recent whole-brain models have made significant advances in discovering these causal mechanisms (see [131] for a general review). Whole-brain models consist of nodes that approximate local or mesoscopic neuronal dynamics through mean field models of coupled excitatory-inhibitory interactions [132, 133] or Hopf models of bifurcations into and out of sustained oscillatory activity [134, 135]. Crucially, whole-brain models have enabled researchers to identify the specific regions that should be stimulated in order to force transitions between different states of consciousness, such as from sleep to wakefulness [136]. Whole-brain models have also been applied to psychedelics; in particular, the 5-HT_2A receptor density in each area of the brain was used to fit a mean field model of fMRI data on LSD, capturing the nonlinear interactions between connectivity and the distribution of 5-HT_2A receptors [137]. This model was recently extended to determine the associations between global brain connectivity on LSD and categories of subjective experience [71]. The model has also been applied to the relationship between neurotransmitters (namely, serotonin) and neuronal regions on psilocybin; in particular, Kringelbach et al. [67] coupled a mean field model of each neuronal region to the release-and-reuptake dynamics of serotonin concentration. Importantly, this coupled neuronal-neurotransmitter model exhibited a better fit to the empirical FC than a model of solely the neuronal activity. Overall, whole-brain models display great promise for linking neurotransmitter release, which is partly determined by pharmacology, to the brain regions that causally mediate, rather than merely correlate with, the effects of psychedelics. Most recently, whole-brain models have also been combined with another cutting-edge method, turbulence, to determine the susceptibility of the brain to external stimuli under psychedelics [138].

This meta-analysis was only performed on data from healthy participants, yet there is emerging research on the neuroimaging and phenomenology of psychedelics in clinical populations [21, 23, 139]. Future models that bridge the three levels of analysis should also be applied to the clinical data in order to shed light on the therapeutic mechanisms of psychedelics. It is essential to determine whether these models would be able to predict clinical outcomes. Recent research has succeeded in predicting treatment response to psilocybin for depression based on changes in the functional hierarchy of the brain [140]. These methods may be even more effective if they incorporated data on phenomenology and pharmacology. However, as mentioned in the Introduction, psychedelic treatments are always administered in conjunction with therapy or some form of "psychological support," which the models must also take into account.