Transcriptomics-informed large-scale cortical model captures topography of pharmacological neuroimaging effects of LSD

The authors extended a previously validated biophysical whole-cortex model composed of 180 nodes, each representing a left-hemispheric cortical parcel from the HCP MMP1.0 parcellation. Each node contained coupled excitatory and inhibitory populations whose dynamics arise from recurrent synaptic interactions; nodes were interconnected by long-range excitatory projections whose strengths were given by a group-averaged diffusion-MRI structural connectivity (SC) matrix derived from 339 unrelated Human Connectome Project subjects. Simulated synaptic activity was transformed to BOLD-like signals using the Balloon–Windkessel haemodynamic model, enabling direct comparison with fMRI-derived measures. To simulate LSD, the investigators modelled modulation of neuronal gain — the non-linear mapping from input current to firing rate — as a fractional change in the gain parameter. Regional heterogeneity was introduced by scaling gain modulation in each node according to local HTR2A expression from the Allen Human Brain Atlas. The model allowed gain to be modulated separately on excitatory and inhibitory populations via two free parameters, thereby remaining agnostic about exact cell-type receptor distributions. The unperturbed (placebo) operating point was first calibrated by a one-dimensional grid search over a global coupling parameter to best match empirical placebo functional connectivity; subsequent simulations varied the two gain parameters over a defined grid ([0, 0.03] with step 0.003) to generate perturbed (LSD) model outputs. Empirical data came from a double-blind, placebo-controlled, within-subject fMRI experiment in 24 healthy participants who received placebo or 100 µg oral LSD; resting-state scans were collected 75 minutes after administration and the short 5D-ASC questionnaire was administered at 180 minutes. Global brain connectivity (GBC) maps were constructed both empirically and from model-derived BOLD FC matrices by averaging Fisher Z-transformed off-diagonal FC elements per parcel. Model–data similarity was quantified using a ‘‘loading’’ metric defined as the dot product between model and empirical ΔGBC vectors divided by the squared norm of the empirical ΔGBC map; this metric is sensitive to both topography and magnitude of perturbation. The authors also performed parcel-wise Spearman rank correlations to assess spatial correspondence. To test specificity to receptor topography, the grid search was repeated while modulating gain in proportion to alternative receptor gene maps (HTR1A, HTR2C, HTR7, DRD1, DRD2). Statistical significance of the HTR2A topography was assessed using spatial-autocorrelation-preserving surrogate brain maps generated with BrainSMASH, thereby controlling for the confounding influence of brain map spatial structure. Principal components analysis (PCA) characterised modes of individual ΔGBC variation in both empirical and model-derived maps. Finally, linear regression was used parcel-by-parcel to link individual ΔGBC to changes in each of the five 5D-ASC experiential dimensions, producing ‘‘experiential regression maps’’ that quantify spatial patterns predictive of subjective effects.

The calibrated unperturbed model reproduced placebo functional connectivity and produced a baseline model GBC map. When gain modulation was applied in proportion to regional HTR2A expression, the model generated a ΔGBC map that closely resembled the empirical LSD-minus-placebo ΔGBC map. A two-dimensional grid search over excitatory- and inhibitory-gain modulation parameters revealed a quasi-degenerate band of parameter values that maximised model–empirical loading. The best-fitting regime indicated preferential modulation of excitatory (pyramidal) populations. Analysing the model’s excitatory/inhibitory (E/I) firing-rate ratio showed that model–empirical loadings collapsed onto an approximately one-dimensional curve that peaked at positive shifts in E/I ratio, consistent with a net excitatory effect of the simulated perturbation. At the level of functional networks and parcels, the model recapitulated the empirically observed bidirectional pattern: LSD induced hyper-connectivity in sensory and somatomotor networks and hypo-connectivity in association networks. Parcel-wise spatial correspondence between simulated and empirical ΔGBC maps was strong (Spearman ρ = 0.73, p < 10^-5). The statistical association between model and data exceeded the direct association between the empirical ΔGBC map and the HTR2A expression map alone; the test for difference between dependent correlations yielded p = 0.011, indicating that heterogeneous physiological response in the model contributed beyond transcriptomic topography. Specificity analyses showed that modulating gain according to other tested receptor gene maps (HTR1A, HTR2C, HTR7, DRD1, DRD2) did not produce comparable model–data loadings. Using spatial-autocorrelation-preserving surrogate maps to build a null distribution, the HTR2A-based modulation produced a statistically significant loading (p = 0.012), while none of the alternative gene maps reached significance. The authors also examined the effect of preprocessing: model–empirical similarity and Spearman correlations were substantially reduced when global signal regression (GSR) was not applied to the empirical data, suggesting that GSR improved alignment between modelled neuronal effects and fMRI-derived maps. Turning to individual differences, PCA of empirical subject ΔGBC maps showed that the leading principal component (PC1) captured the dominant subject-level variation and correlated strongly with the group-averaged ΔGBC map (ρ = 0.72, p < 10^-5). Fitting the model separately for each subject (selecting gain parameters that maximised that subject’s model–empirical loading) yielded subject-specific model ΔGBC maps whose PCA demonstrated a model PC1 closely aligned with the group ΔGBC map (ρ = 0.87, p < 10^-5) and with empirical PC1 (ρ = 0.63, p < 10^-5). The model’s simulated ΔGBC variance was largely captured by two principal components (expected given two free parameters), whereas the empirical data exhibited approximately five significant modes; nonetheless, 67% of model variance lay within the five-dimensional empirical subspace, indicating that model variation occupied principal axes present in the data. Finally, experiential relevance was established by linking neural modes to 5D-ASC scores. Parcel-wise regression produced experiential maps for five dimensions of altered consciousness. Model PC1 correlated with the Meaning regression map (ρ = 0.38, p < 10^-5) and the Disembodiment map (ρ = 0.61, p < 10^-5) and most strongly resembled the dominant experiential regression map (ρ = 0.65, p < 10^-5). The authors report that the model fit the majority of subjects well, but note limited statistical power for individual-differences predictions and that several subjects exhibited idiosyncratic neural responses the model did not capture.

Burt and colleagues interpret their results as support for a mechanistic chain linking LSD’s agonism at 5-HT2A receptors to regionally heterogeneous modulation of neuronal gain, preferentially targeting pyramidal cells, which in turn produces the observed topography of GBC changes. They emphasise that the model’s ability to reproduce both group-level and individual-patterns of ΔGBC depended on using the HTR2A expression map to scale regional gain changes, and that this specificity held when controlling for spatial autocorrelation with surrogate-map testing. The authors situate their findings relative to prior work by noting conceptual convergence with other modelling studies that implicate gain modulation in LSD-induced changes, while highlighting that their transcriptomics-informed, spatially heterogeneous model uniquely captures cortical topographic effects. Evidence from single-cell RNA sequencing and prior anatomical studies is invoked to support the model result that HTR2A expression is enriched in excitatory cell types and localises to apical dendrites of pyramidal neurons, which may underlie LSD’s pronounced effects on perception and consciousness. Limitations acknowledged by the investigators include the model’s simplifications — excitatory and inhibitory populations are represented as statistical ensembles — and the study’s modest sample size for individual-differences analyses. They also note that LSD acts at multiple serotonergic and dopaminergic receptor subtypes (for example, 5-HT1A, which can exert inhibitory effects), and that their cortex-focused model does not capture potential contributions from subcortical structures or the claustrum. The authors discuss the methodological debate around global signal regression, reporting that GSR improved model–data correspondence in their analyses but recognising that no single pre-processing choice is universally correct. For future work, the study team recommends multi-scale and cell-type-resolved models that integrate bulk transcriptomics with single-cell sequencing, incorporation of additional receptor-rich regions such as the claustrum, and comparison to multiple imaging modalities. They also argue that transcriptomic atlases are a useful proxy for receptor topography when PET radioligands are unavailable, while acknowledging the limitations of this proxy. Overall, the study presents a proof-of-concept framework linking molecular targets through circuit mechanisms to empirically observable changes in human neuroimaging and experiential measures.