A Virtual Clinical Trial of Psychedelics to Treat Patients With Disorders of Consciousness

L. and colleagues situate their study within the view that conscious experience depends on spatiotemporally complex brain dynamics supported by structural connectivity (white matter, from diffusion-weighted imaging) and functional dynamics (BOLD fMRI). Healthy brain dynamics are argued to operate near a critical point — a transition between order and disorder — which affords high sensitivity to perturbation, long-range correlations and a balance of integration and differentiation thought to underpin rich conscious experience. Disorders of consciousness (DoC) are described as states of pathologically reduced complexity and stability (sub‑critical), whereas psychedelic states show increased complexity and sensitivity to perturbation. Because administering psychedelic drugs to DoC patients raises practical and ethical challenges, the authors propose using whole‑brain computational models — digital twins — to explore mechanistic effects and perform a "phase zero" virtual clinical trial simulating drug administration. The study sets out to (1) create individualized whole‑brain models for patients with DoC (based on each patient's DWI and fMRI), (2) validate a model‑based perturbational biomarker, the perturbative integration latency index (PILI), for discriminating states of consciousness, and (3) virtually simulate administration of LSD and psilocybin to each patient model to evaluate whether psychedelics shift DoC brain dynamics toward criticality. The authors also aim to identify baseline structural or functional biomarkers that predict the simulated treatment effect, thereby assessing personalised treatment potential with in silico methods.

L. and colleagues combined six previously published datasets to build group‑level and individual whole‑brain models. Key samples cited include a DoC dataset with 35 healthy controls (CNT), 20 unresponsive wakefulness syndrome (UWS) patients and 26 minimally conscious state (MCS) patients; anesthetic datasets (propofol N = 13, dexmedetomidine N = 11); and psychedelic datasets (LSD N = 12, psilocybin N = 15, plus an auxiliary psilocybin dataset). For DoC subjects, diffusion MRI (DWI) and resting‑state fMRI from the same individuals were used; models for healthy drug datasets used an average structural connectome derived from the DoC healthy controls because individual SC was not available for those datasets. Ethical approvals and consent procedures are reported for all source studies. Preprocessing: All fMRI data were reprocessed with a common pipeline (FSL/MELODIC). Steps included discarding initial volumes, motion correction, ICA denoising with manual component inspection, band‑pass filtering (0.01–0.08 Hz; modelling analyses used 0.04–0.07 Hz for signal extraction), registration to T1, and extraction of regional BOLD time series from the 90 AAL atlas regions (cortical and subcortical, cerebellum excluded). Pairwise Pearson correlations produced 90×90 functional connectivity (FC) matrices (Fisher z‑transformed for group averaging). Structural connectivity: Patient structural connectomes were built from single‑subject tractography (MRtrix3 pipeline). Preprocessing included denoising, eddy and susceptibility correction, multi‑tissue constrained spherical deconvolution, probabilistic tractography with 20 million streamlines per subject, SIFT2 weighting, and mapping to the AAL parcellation to produce weighted 90×90 SC matrices. Fractional anisotropy (FA) and graph‑theory measures (graph strength, global/local efficiency, characteristic path length, betweenness centrality) were computed. Computational model and fitting: Each brain region was modelled as a Stuart‑Landau (normal form of a supercritical Hopf) oscillator coupled according to SC and scaled by a global coupling parameter G. Local bifurcation parameters a j determined node dynamics (a < 0 noisy fixed point; a = 0 at bifurcation; a > 0 oscillatory). For group models where individual SC was unavailable, the average healthy SC from the DoC control group was used. Model fitting proceeded in two steps: (1) optimise global coupling G by minimising the Kolmogorov–Smirnov (KS) distance between empirical and simulated FC across G values (500 simulations per G), and (2) fit local a parameters constrained to a nine‑dimensional space defined by nine resting‑state networks (derived from ICA of the DoC healthy controls). A genetic algorithm was used to find the optimal nine network a values (range −0.2 to 0.2), and each region's a j was the linear combination of the networks it belongs to. Perturbation and PILI: The perturbation protocol switched a single region's bifurcation parameter to a synchronous regime (a = 0.6) for 100 s, repeating this for each of the 90 regions and over 100 trials per region. Integration at each time point was calculated from instantaneous phases (Hilbert transform) by binarising phase‑locking across thresholds and measuring the size of the largest connected component; integration is the integral across thresholds. The perturbative integration latency index (PILI) was computed as the area between the perturbed integration curve and the maximal basal integration until return to baseline, capturing both amplitude and latency of the return to baseline. Regional PILI values were averaged across trials; a global mean PILI was computed as the average across regions. Virtual pharmacology: To simulate LSD and psilocybin administration in patient models, the authors derived parameter shifts (changes in G and the nine network a parameters) from group‑level drug versus placebo models (healthy subjects). These extracted shifts were applied to each patient model to produce a simulated drug state. The method was validated with two cases of empirical fMRI from patients scanned before and during a light propofol dose: parameter shifts derived from group models of light propofol were applied to patient models and compared with empirical propofol‑state models. Statistics: At the group model level, Cohen's d was used to compare PILI distributions. For individual models, conventional frequentist tests were used: one‑way ANOVA and post‑hoc Tukey HSD for three‑group comparisons, Mann–Whitney U tests, Wilcoxon signed‑rank tests for paired comparisons, paired t‑tests for region‑wise changes (Bonferroni correction across 90 regions), and correlation analyses with Bonferroni correction where indicated. The authors note using each simulation as a datapoint for group‑level models would artificially inflate sample size, hence reliance on effect sizes there.

PILI distinguishes states of consciousness: Using group‑level Hopf models fitted to FC for multiple states, PILI discriminated reduced‑consciousness states from richer phenomenological states. DoC groups and anaesthesia showed reduced PILI compared with their controls, while psychedelic conditions showed increased PILI relative to placebo. Reported effect sizes (Cohen's d) included UWS versus CNT = −0.52 and MCS versus CNT = −0.48; propofol versus wake = −0.54 and dexmedetomidine versus wake = −0.32. Psychedelic comparisons yielded smaller positive effects (LSD Cohen's d = 0.11; psilocybin Cohen's d = 0.28). Network‑wise PILI changes were tabulated in the Supporting Information. Individualised models and group differences: Individualised patient models (CNT N = 35, UWS N = 20, MCS N = 26) showed a significant main effect of diagnostic group on mean regional PILI (one‑way ANOVA F(2,78) = 8.84). Post‑hoc Tukey tests indicated higher PILI in CNT versus UWS (mean difference = 6.80, 95% CI [2.94, 10.66], p < 0.001) and higher PILI in MCS versus UWS (mean difference = 4.11, 95% CI [0.09, 8.12], p = 0.044). No significant difference was found between CNT and MCS (mean difference = 2.69, 95% CI [−0.92, 6.30], p = 0.182). All p‑values reported had multiple‑comparison correction applied. Virtual clinical trial—simulated psychedelics: Applying parameter shifts derived from healthy drug datasets to individual DoC models, simulation of LSD and psilocybin increased PILI across patients with DoC, with larger effects in MCS than UWS. For LSD, increases were greater in MCS (z = 4.13, p < 0.001) than UWS (z = 2.8, p < 0.001); for psilocybin, UWS z = 3.71, p < 0.001 and MCS z = 3.92, p < 0.001. Region‑wise analyses showed that, in MCS patients, the largest network‑level PILI increases after simulated LSD and psilocybin occurred in the default mode network (DMN), attention and frontoparietal networks. In UWS patients, the sensorimotor network exhibited the highest PILI increase for both drugs. Region‑wise PILI did not show significant correlations with regional structural connectivity (SC) or functional connectivity (FC) after Bonferroni correction (reported uncorrected correlations: e.g. LSD MCS r = 0.21, p = 0.04; others non‑significant). Validation with propofol: As a validation of the virtual pharmacology method, parameter shifts for a light propofol dose were applied to two patient models with empirical pre/post propofol scans. The mean differences in residual per‑region PILI between simulation‑derived and empirical‑derived models were reported as ~0.0002 in one patient and ~1 in another; one patient with low baseline PILI showed a small simulated effect, while a patient with higher baseline PILI showed a larger decrease after propofol. The authors interpret this as the method capturing differential individual effects, with further discussion in Supporting Information. Baseline predictors of simulated treatment response: The simulated treatment effect was defined as change in PILI from baseline to simulated drug. Correlational analyses revealed that, in UWS patients, mean SC correlated with simulated treatment effect for both LSD (r = 0.66, p = 0.001) and psilocybin (r = 0.66, p < 0.002), surviving Bonferroni correction. In contrast, in MCS patients, mean FC correlated with simulated treatment effect (LSD r = 0.65, p < 0.001; psilocybin r = 0.57, p = 0.002), and network‑specific FC (notably DMN FC) also correlated with treatment effect for both drugs; structural measures did not significantly correlate with outcome in MCS. Representative patient examples illustrated that high baseline mean FC predicted strong simulated response in MCS, whereas in UWS high mean SC distinguished those with larger simulated responses. Other graph metrics showed group differences (e.g. characteristic path length lower in UWS than MCS; global efficiency and graph strength differed between controls and UWS), but mean SC was the dominant predictor in UWS.

L. and colleagues interpret their findings as demonstrating that PILI, a model‑based perturbational biomarker, differentiates states of consciousness: diminished states (DoC, anaesthesia) had lower PILI consistent with dynamical rigidity and sub‑critical regimes, while psychedelic states showed higher PILI consistent with increased sensitivity and proximity to criticality. Using personalised whole‑brain models (digital twins), simulated administration of LSD and psilocybin increased the brain's perturbational response in patients with DoC, shifting dynamics closer to criticality; this effect was larger in MCS than UWS. The authors emphasise that increased PILI indicates greater spatially extensive and sustained integration after perturbation — properties associated with complexity — but caution that a change in PILI need not directly equate to improved behavioural consciousness or beneficial phenomenology. The authors note dissociations in baseline predictors: in UWS patients structural integrity (mean SC) correlated with simulated response, implying that when structural scaffolding is severely compromised it limits capacity for psychedelic‑induced complexity; in MCS patients, functional connectivity — particularly within the DMN, attention and frontoparietal networks — predicted response, suggesting preserved functional organisation is necessary to support dynamical changes. They present this as a hierarchical relationship where white‑matter architecture provides the scaffold on which functional dynamics and complexity emerge. The authors discuss practical and conceptual implications: virtual pharmacology and digital twins could enable mechanistic investigations, drug discovery and personalised treatment testing without exposing vulnerable patients to risk. They acknowledge important limitations: PILI shows substantial within‑subject variability and currently lacks single‑subject diagnostic precision comparable to measures like the perturbational complexity index (PCI); cross‑dataset differences in acquisition complicate comparisons; the Hopf model is phenomenological and not mechanistic at a biophysical level; and the virtual pharmacology relies on a superposition assumption that parameter shifts observed in healthy subjects generalise linearly to patients with DoC. The authors also recognise reliance on parameter estimates from single datasets as a source of potential bias. They stress that empirical studies in DoC patients would be required to validate simulated predictions and refine parameter sets. Overall, the authors present their virtual clinical trial as a computational proof‑of‑principle that psychedelics can increase modelled brain complexity in DoC and as a step towards computational personalised medicine and drug discovery, while explicitly noting many open questions about phenomenological and behavioural consequences and the need for empirical validation.