Whole-brain multimodal neuroimaging model using serotonin receptor maps explains non-linear functional effects of LSD

Human brain function emerges from large-scale, non-linear interactions among interconnected neuronal populations. The authors note that whole-brain models based on structural and functional connectivity (for example, neural mass and mean-field models) have advanced understanding of resting-state networks, but such models often omit the spatially heterogeneous influence of neuromodulatory systems. Serotonergic neuromodulation, in particular, acts via regionally varying receptor densities that can change the gain and dynamics of local circuits and so may be necessary to explain some global functional effects observed in pharmacological manipulations. This study set out to test whether integrating an empirically derived map of serotonin 2A receptor (5-HT2A R) density into a biophysically informed whole-brain model can mechanistically explain the functional effects of the 5-HT2A agonist LSD in healthy participants. The authors introduced a global gain-scaling parameter (sE) that scales regional neuronal gain according to PET-derived 5-HT2A densities, fitted the model to placebo data, and asked whether adjusting sE (while keeping the placebo-derived coupling parameter fixed) could reproduce LSD-induced changes in functional connectivity dynamics. The work aims to provide a causal, mechanistic link between anatomy, receptor distribution, and functional dynamics induced by serotonergic stimulation.

The study combined multimodal neuroimaging and whole-brain computational modelling. Structural connectivity (SC) matrices were derived from diffusion MRI tractography averaged across 16 healthy young adults (5 women; mean age ≈ 24.8 years) to provide a 90-node network based on the Automated Anatomical Labeling (AAL) atlas. Functional MRI data came from a separate placebo-versus-LSD within-subject study run at Imperial College: participants attended two sessions (placebo and LSD, order balanced and participant-blind), received an intravenous LSD infusion (text reports 75 mg in the extraction) or placebo, and underwent eyes-closed resting-state BOLD scans (three 7-minute scans per session, the second including music). The extracted text does not clearly report the number of participants in the fMRI/LSD portion within this document. Neurotransmitter information used an in vivo 5-HT2A receptor density atlas obtained with PET ([11C]Cimbi-36), summarised to the same 90 AAL regions. The authors also referenced PET-derived maps for other serotonin receptors and the transporter for specificity tests. All imaging modalities were mapped to the AAL parcellation so that each node had structural, functional and receptor-density inputs. The computational core was a Dynamic Mean Field (DMF) whole-brain model in which each AAL node was represented by coupled excitatory and inhibitory population dynamics (a reduction of large spiking networks). Inter-regional coupling was given by the SC matrix and globally scaled by a single parameter G. Neuronal gain in each region was modulated by the empirical 5-HT2A density via an added global gain-scaling parameter sE: sE = 0 corresponds to the original uniform gain (placebo fit), and non-zero sE scales the regional gains proportionally to the PET-derived receptor density. The model thus effectively had two free parameters of interest: G (global coupling) and sE (strength of receptor-based gain modulation). Model fitting emphasised dynamical, time-resolved features: the authors computed Functional Connectivity Dynamics (FCD) by sliding-window FC (30 TR window, 2 TR steps) and comparing distributions of upper-triangular FCD values between empirical and simulated data using the Kolmogorov–Smirnov (KS) distance (smaller KS indicates a better fit). The pipeline was: fit the placebo FCD by adjusting G with sE = 0 to find an optimal G; then, keeping that G fixed, search sE values (scaling regional 5-HT2A densities) to fit the LSD FCD. Specificity tests included random shuffling of the 5-HT2A regional densities (200 permutations), use of alternative receptor maps (5-HT1A, 5-HT1B, 5-HT4, and 5-HTT), comparison with a uniform receptor distribution, and a 50% train/test participant split to probe generalisation. Tractography and preprocessing steps (dMRI and fMRI) followed common pipelines (probabilistic tractography, FSL MELODIC preprocessing, etc.) as outlined in the STAR Methods.

Fitting the DMF whole-brain model to the placebo condition (with uniform gain, sE = 0) identified an optimal global coupling parameter G at which the simulated FCD best matched the empirical placebo FCD; this optimum was reported as G = 2.1. Using that placebo-derived G and introducing regionally weighted gain modulation by the PET 5-HT2A map, the authors found a clear optimum in the sE parameter for fitting the LSD condition's FCD, with an optimal sE of approximately 0.2. In contrast, varying sE when fitting the placebo condition did not improve fit (the placebo FCD fit worsened monotonically when sE was varied), indicating that the receptor-weighted gain modulation specifically accounts for the LSD-induced changes. Model fit was assessed by KS distance between empirical and simulated FCD distributions; the authors report “excellent” FCD fitting at the identified parameter settings, and show that FCD is a far more informative fitting target than static grand-average FC. To test the importance of the precise spatial distribution of 5-HT2A density, the authors ran 200 simulations with the empirical 90 regional 5-HT2A values randomly reassigned across regions; these reshuffled maps produced significantly worse fits than the true map (Wilcoxon statistics; reported p < 0.0016). Using a uniform receptor-binding map produced the worst performance. Comparing receptor maps, the 5-HT2A map produced the best improvement in fit for the LSD condition, while other serotonin-related maps (5-HT1A, 5-HT1B, 5-HT4, and 5-HTT) performed significantly worse; 5-HT1A showed some intermediate explanatory power relative to the others. A 50% random train/test split (noting the authors regarded the sample as small for generalisation tests) produced similar training and testing performance, indicating some robustness. Finally, the FCD histograms for empirical LSD and placebo were distinct and the optimised model reproduced these distinct distributions, supporting the claim that modulation by the 5-HT2A density map can account for LSD-induced alterations in functional dynamics. The model uses only two principal parameters (G and sE) to capture these effects.

The authors interpret their findings as evidence that integrating empirically derived neurotransmitter receptor density maps into whole-brain dynamical models provides a causal mechanistic account of how neuromodulation shapes large-scale brain dynamics. Specifically, they argue that the non-linear functional changes induced by the 5-HT2A agonist LSD can be quantitatively explained by regionally heterogeneous gain modulation determined by the PET-derived 5-HT2A density map, interacting with the underlying anatomical connectivity. The results extend prior work emphasising gain modulation by showing that including receptor-binding maps yields additional explanatory power at the whole-brain level. They position their work relative to earlier research by noting that static functional connectivity is less informative than time-resolved measures (FCD) for constraining models of dynamical brain states, and that receptor-weighted gain modulation is necessary to account for the LSD condition. The specificity analyses (shuffling, alternative receptor maps, uniform map) are presented as supporting the principal role of 5-HT2A in mediating LSD effects, with a subsidiary role suggested for 5-HT1A. The authors acknowledge limitations and future directions. They note that the fMRI/LSD sample was not large (and that the model generalisation test used a small sample), and they recommend expanding models to include receptor distributions of other neuromodulators and to better characterise the distinct timescales of neuronal and neurotransmitter systems. They also highlight methodological choices (for example, use of the AAL parcellation) and justify these choices on grounds of precedent and computational tractability while recognising debates in parcellation research. The authors propose that the approach could be developed into a pipeline for modelling disease states and for rational drug-discovery, and that combining pharmacological manipulations with environmental/therapeutic interventions (for example, drug-assisted psychotherapy) might benefit from such mechanistic models. They emphasise that the present results provide a proof-of-principle that regional receptor maps can be integrated in a whole-brain model to yield new causal insights into anatomy–activity–neurotransmitter interactions.