Decoupling of neurophysiological activity from structure mirrors global microarchitectural and neuromodulatory trends

Abstract

The brain’s functional activity is shaped by the complex architecture of its fibers. Yet, the lack of a direct one-to-one mapping between functional and structural connections makes this relationship elusive. To date, most studies on structure–function coupling (SFC) have conceptualized function in terms of resting-state functional Magnetic Resonance Imaging (fMRI) connectivity. Here, we extend this framework to neurophysiological data by examining how magnetoencephalography (MEG) activity relates to the structural connectome, leveraging its rich spectral content and direct sensitivity to neuronal population dynamics. We show that the decoupling of MEG activity from structure is strongly associated with the expression levels of synaptic plasticity markers, pointing to a link between flexible functional reconfiguration and the molecular mechanisms of plasticity. Moreover, regions with greater decoupling exhibit higher neurotransmitter receptor diversity, underscoring neuromodulatory heterogeneity as a substrate for functional flexibility. This association is especially pronounced for slow-acting metabotropic receptors, whose diffuse and prolonged signaling may facilitate functional reorganization atop the structural connectome.

Analysis of neurophysiological (MEG) structure–function coupling sheds light on its association with molecular plasticity markers and diversity of neuromodulatory receptors in the human brain.

Introduction

The brain’s structural scaffolding, or structural connectivity (SC), is considered critical for explaining the complex, multifaceted functional interactions among neural ensembles^1,2. In line with this, considerable effort has recently been devoted to understanding how functional connectivity (FC) arises from the brain’s structural connectome^2–6. A progressive decoupling of FC from structural wiring has been reported along the sensory-fugal axis of brain organization^5,7–9. In other words, moving from sensory areas to transmodal cortices, the predictability of synchronized neural activity based on SC decreases^10–13. Importantly, graded principles of cortical microstructural variation have been used to address the regional heterogeneity of structure–function coupling (SFC). For example, SFC has been shown to be higher in brain areas with higher myelin content and predominantly short-range structural pathways^11,13,14. Thus far, most studies on the brain’s SFC have defined functional connectivity based on correlations between the Blood-Oxygenation-Level-Dependent (BOLD) signal profiles of pairs of nodes (see Fotiadis et al.³ for a review), as assessed by functional Magnetic Resonance Imaging (fMRI). More recently, the relationship between structural and functional organization has also begun to be explored using electrophysiological data from EEG and MEG, which provide measurements of neuronal population dynamics that offer a complementary perspective to fMRI-based approaches^15,16. This emerging body of work has started to elucidate how structural architecture shapes fast-scale functional interactions across frequency bands. EEG-based investigations have shown that functional synchronization partly reflects both direct and indirect white-matter pathways, with stronger dependence on anatomical connectivity for higher oscillatory frequencies¹⁷. Computational modeling informed by diffusion tractography has further demonstrated that structural networks can explain a substantial fraction of the variance observed in EEG coupling patterns¹⁸. Other multimodal approaches combining fMRI, MEG, and DWI have revealed that structure–function mapping is modality-dependent and shaped by both direct and polysynaptic anatomical walks¹⁹. In animal models, both phase- and envelope-based intrinsic coupling modes have been shown to depend on the underlying white-matter architecture in a frequency-specific manner²⁰. Finally, recent multimodal studies confirmed that coupling strength follows a unimodal-to-transmodal gradient²¹, albeit frequency-specific patterns have also been described⁹. However, due to the limited number of studies on neurophysiological SFC, our understanding remains incomplete—particularly regarding its potential microarchitectural and neuromodulatory underpinnings.

Here, we sought to bridge this gap by investigating patterns of structural–functional (de)coupling in the neocortex using MEG and DWI data from the Human Connectome Project (HCP)^22,23. To elucidate the biological underpinnings of SFC, we integrated multiple open-source resources—including ex vivo gene transcription data from Allen Human Brain Atlas (AHBA)²⁴ and in vivo PET maps—through a combination of univariate and multivariate analytical approaches. Our investigation is guided by two interrelated hypotheses that link macroscale functional flexibility to underlying neurobiological specializations.

We propose that functional flexibility relative to structural constraints reflects a continuous, experience-dependent need for dynamic remapping. This remapping may be particularly prominent in transmodal cortical regions²⁵, where we consequently expect high-level cognitive and emotional processes to show the strongest associations with functional-structural decoupling^13,26. Furthermore, from a neurobiological perspective, we expect molecular signatures of plasticity to co-vary with decoupling. Operationally, we hypothesize that biological factors promoting synaptic plasticity, as opposed to those associated with synaptic stabilization or inhibition of plasticity, spatially co-localize with decoupling. Here, we focus on two markers of synaptic plasticity—regional expression of GAP43 (Growth-associated protein 43) and BDNF (Brain-Derived Neurotrophic Factor) genes—both of which play pivotal roles in experience-driven neural remodeling^27–30. For markers associated with synaptic and network stability, we examine the regional expression of the MBP (Myelin Basic Protein) gene, encoding myelin basic protein, and PVALB, reflecting parvalbumin expression. Both myelin and parvalbumin have been shown to help end critical periods and to limit Long-Term Potentiation (LTP)^31–33.

We further hypothesize that functional flexibility is underpinned by the regional diversity of neuromodulatory profiles. In particular, we propose that regions exhibiting pronounced structure–function decoupling are characterized by a richer repertoire of neurotransmitter systems, thereby enabling a broader dynamic range of functional configurations³⁴. Beyond receptor diversity, we further examine whether the temporal dynamics of neuromodulatory signaling shape the relationship between structure and function^34–36. We work under the hypothesis that sensory areas, whose activity is largely hardwired, benefit from fast-acting ionotropic neurotransmission to promptly respond to external environments. Conversely, association cortices, which have a higher density of metabotropic receptors—exerting diffuse, long-lasting modulatory effects—would be less constrained by the timing of external stimuli and characterized by more flexible, less hardwired functional activity^2,25,37.

Taken collectively, these hypotheses posit that structure–function decoupling reflects a macroscale expression of cortical flexibility, arising from molecular mechanisms that regulate plasticity and neuromodulation along the cortical hierarchy. In this framework, regions positioned higher in the hierarchy—characterized by lower myelination, greater receptor diversity, and increased metabolic cost—exhibit weaker structural constraints yet greater capacity for dynamic, context-dependent reconfiguration. This integrated view provides a mechanistic bridge between structural architecture, molecular specialization, and cognitive adaptability.

Results

We performed source localization on resting-state MEG data from 89 participants of the Human Connectome Project (HCP). Subject-specific FC matrices were constructed by computing the amplitude-envelope correlation of broadband activity³⁸. Source time courses were parcellated according to the Schaefer 200 regions 7 networks atlas³⁹ prior to FC computation. This approach provides an R² coefficient for each node, quantifying how well the FC profile of a node recapitulates its SC profile. We then obtained a nodal measure of structure–function decoupling as 1 − R². The decoupling index thus naturally ranges from 0 to 1, indicating perfect alignment and maximal decoupling, respectively. We aggregated the results across subjects to generate a group consensus structure–function decoupling map (Fig. 1A). The highest decoupling values were observed in the orbitofrontal, ventromedial, and dorsolateral prefrontal cortices. Given the availability of three separate resting-state MEG sessions in the HCP dataset, we divided the data accordingly and generated session-specific decoupling maps (Fig. 1B). Similarity between session-specific maps was assessed using Lin’s Concordance Correlation Coefficient (CCC), a stringent measure of agreement⁴⁰. The minimum Lin’s CCC was 0.86, indicating high reproducibility across sessions. Interestingly, the regions showing the highest decoupling were also those that displayed the greatest across-subject variability, as indicated by the standard deviation of regional decoupling (Fig. 1C). We further analyzed the distribution of decoupling values across resting-state networks (RSNs)⁴¹, calculating the average decoupling value for each RSN (Fig. 1D). Statistical significance of these mean values was assessed using a spin test to account for spatial autocorrelation (N = 10,000) under a bidirectional hypothesis. Decoupling was significantly higher than chance in the Default Mode Network (DMN, p_spin < 0.05) and Frontoparietal Network (FPN, p_spin < 0.05), and significantly lower than chance in the Visual network (p_spin < 0.05). To further confirm the unimodal-transmodal nature of the decoupling map, we correlated it to the first gradient of FC⁴². The decoupling index showed a significant positive correlation with the principal functional connectivity gradient (r = 0.54, p_spin = 0.0001), indicating that structure–function decoupling increases along the sensory–fugal axis.

Fig. 1. Characterization of structural–functional decoupling patterns.

A Group-level template map of the decoupling index, quantifying the proportion of functional connectivity variance which cannot be explained by structural connectivity. B Session-specific decoupling maps from HCP resting-state MEG data, which demonstrate high reproducibility (minimum Lin’s concordance correlation coefficient [CCC] = 0.86 across sessions). C Inter-subject variability of decoupling values, revealing that structurally decoupled regions exhibit greater between-subject differences. D Decoupling distribution across Yeo’s resting-state networks. The Default Mode Network (DMN) and Frontoparietal Network (FPN) exhibited significantly elevated decoupling (spin test, N = 10,000 spins, p_spin < 0.05), while the Visual Network (Vis) showed reduced decoupling.

Multivariate meta-analytic decoding links decoupling to high-level cognitive processes

We further validated the transmodal nature of decoupling by analyzing its spatial associations with the Neurosynth dataset⁴³ and a curated subset of terms from the Cognitive Atlas⁴⁴. This enabled us to associate our index with a proxy for cognitive functions. The subset of 124 terms was drawn from Hansen et al. ⁴⁵ to minimize selection bias. To examine the relationship between decoupling and cognition, we employed Partial Least Squares (PLS) correlation, augmented with a spin test (N = 10,000) to account for spatial autocorrelation. PLS aims to maximize the covariance between latent structures of datasets—in this case, subject-specific regional decoupling and meta-analytic activation maps from Neurosynth. We evaluated the reliability of the weights using a bootstrap resampling procedure (N = 10,000 bootstraps), considering a feature as reliable if its 95% confidence interval did not include zero. The first component identified through PLS (PLS-1) explained 72% of the covariance between decoupling and cognition (p_spin = 0.003; Fig. 2A). We additionally implemented a distance-dependent cross-validation procedure⁴⁵ to evaluate model generalizability. Cross-validation confirmed the generalizability of this relationship (median out-of-sample Spearman’s ρ = 0.58), supporting the robustness of the PLS model (Supplementary Fig. S2). The structure–function decoupling map and the Neurosynth score relative to PLS-1 were strongly and positively correlated (r = 0.69, p_spin = 0.009; Fig. 2B, C). Higher-level cognitive and emotional processes, such as “cognitive control”, “decision”, “emotion”, and “emotion regulation”, showed positive loadings in the decoupling–cognition relationship. In contrast, perceptual processes, including “visual attention”, “gaze”, and “object recognition”, exhibited negative loadings (Fig. 2D).

Fig. 2. Multivariate cognitive decoding of structure–function decoupling using the Neurosynth dataset.

A Explained covariance for the first four PLS components. The first component accounted for a significantly greater proportion of covariance (72%) than expected by chance (spin test; gray boxes show null model distributions). B Pearson correlation between the PLS-1-derived Neurosynth score (weighted sum of Neurosynth maps) and the structural–functional decoupling index. C Spatial distribution of the PLS-1 Neurosynth score across the cortex. D Top 10% of Cognitive Atlas terms showing either negative (left) or positive (right) loadings in the decoupling–cognition relationship. Error bars represent 95% confidence intervals.

Cortical microarchitecture tracks regional variation of decoupling

Building on our hypothesis that molecular signatures of plasticity co-vary with decoupling patterns, we analyzed gene expression data from the AHBA (Fig. 3). Our analyses focused on two sets of genes: (1) markers of cortical stability – myelin basic protein (MBP) and parvalbumin (PVALB); and (2) plasticity-related genes—GAP43 and BDNF. As predicted, decoupling showed strong negative correlations with stability markers—PVALB (r = –0.59, p_spin = 0.02) and MBP (r = –0.66, p_spin = 0.009)—and significant positive correlations with plasticity markers—GAP43 (r = 0.66, p_spin = 0.0001) and BDNF (r = 0.47, p_spin = 0.04). These findings align closely with our operational hypotheses, indicating that regions with higher expression of plasticity-related genes exhibit greater decoupling from structural constraints, whereas areas enriched in stability markers maintain tighter structure–function coupling.

Fig. 3. Relationship between structural–functional decoupling and gene expression from the Allen Human Brain Atlas.

As hypothesized, markers of cortical stability (MBP for myelin basic protein and PVALB for parvalbumin) showed negative correlations with decoupling, whereas plasticity markers (GAP43 and BDNF for brain-derived neurotrophic factor) exhibited positive correlations. Statistical significance was assessed using spin test permutations (N = 10,000 spins).

Multivariate contribution of neurotransmitters to decoupling patterns

As a final step, we investigated which neurotransmitters might be more strongly implicated in explaining the cortical patterning of structural–functional decoupling. We considered a curated set of PET receptor/transporter atlases⁴⁶ and employed a PLS correlation approach. The first PLS component (PLS-1) was statistically significant (explained covariance = 80%, p_spin = 0.001) and was therefore retained for further analysis. Structural–functional decoupling and the neurotransmitter score derived from PLS-1 showed a strong positive correlation (r = 0.7, p_spin = 0.002; Fig. 4A). We further confirmed the generalizability of this relationship using a distance-dependent cross-validation procedure⁴⁵, which showed robust out-of-sample performance (median Spearman’s ρ = 0.45, Supplementary Fig. S2). Bootstrap resampling (N = 10,000) revealed that the vast majority of neurotransmitters contributed reliably and positively to the decoupling–neuromodulation relationship, with the exception of GABA_A and the noradrenaline transporter (NAT), which exhibited reliable negative contributions (Fig. 4C). Collectively, brain regions that are more strongly decoupled from structure display a high diversity of neurotransmitter systems. We further observed that slow-acting metabotropic receptors contributed most strongly to the decoupling—neurotransmitter relationship. These decoupling patterns align with one of the principal axes of neuromodulation in the brain: the ionotropic vs. metabotropic axis³⁴. This suggests that increased activity of G-protein-coupled receptors may facilitate functional flexibility beyond structural constraints.

Fig. 4. Neurotransmitter correlates of structure–function coupling (SFC) revealed by PLS correlation analysis.

A Pearson correlation between the PLS-1-derived neuromodulation score (weighted sum of receptor/transporter maps) and the structural–functional decoupling index. B Cortical distribution of the PLS-1 neuromodulation score. C Neurotransmitter loadings from bootstrap resampling (N = 10,000). Bars represent loading values; error bars indicate 95% confidence intervals. Only neurotransmitters/receptors with confidence intervals excluding zero were considered reliable contributors to the decoupling–neuromodulation relationship.

Robustness analysis

To verify that our findings were robust to methodological variations, we conducted comprehensive sensitivity analyses. First, we evaluated whether our results depended on the specific parcellation scheme by repeating the analyses using an alternative brain atlas (Supplementary Fig. S1A). Second, we assessed the robustness of our findings to source localization methods by comparing our beamformer approach with Minimum-Norm Estimation (MNE)⁴⁷, which yielded highly consistent spatial patterns (Supplementary Fig. S1B). Finally, to assess the impact of different SFC estimation methods, we implemented the nodal multilinear regression approach developed by Liu et al.⁹, which incorporates graph-theoretical properties of the structural connectome. The resulting decoupling map showed strong spatial correspondence with our primary eigenmode-mapping results (Supplementary Fig. S1C), confirming the robustness of our findings across analytical pipelines. Moreover, we verified that the spatial pattern remained consistent when using phase-based functional connectivity instead of amplitude envelope correlations (Supplementary Fig. S3) and when omitting the orthogonalization step from the amplitude envelope correction procedure (Supplementary Fig. S5).

Discussion

In this work, we combined resting-state MEG with DWI data to map how neurophysiological activity recapitulates the pathways of SC. We developed a local index of structural–functional decoupling, defined as the fraction of variance in functional connectivity not explained by the brain’s physical wiring. We then related this index to cognitive processes using the Neurosynth dataset⁴³, as well as to microchemical factors derived from ex vivo gene transcription and in vivo molecular neuroimaging. By combining MEG-based estimates of functional connectivity with molecular and transcriptomic data, we aimed to uncover the neurobiological substrates of structure–function decoupling, linking electrophysiological flexibility to underlying mechanisms of plasticity and neuromodulation. Importantly, by employing MEG—a direct measure of neuronal population activity—we extend previous fMRI-based findings to the electrophysiological domain. This demonstrates that large-scale decoupling between structure and function is not limited to hemodynamic coupling, but represents a general property of cortical dynamics that mirrors underlying molecular specializations. The convergence between MEG-derived decoupling and neuromodulatory receptor distributions underscores the tight interplay between electrophysiological flexibility and the brain’s chemical architecture.

The spatial distribution of structural–functional decoupling we described bears similarity with the sensory-fugal gradient of brain organization, which spans from primary sensory and motor regions to high-level integrative areas^42,48. This pattern may stem from the rapid evolutionary expansion of the neocortex, which likely uncoupled functional connectivity in transmodal regions from the rigid, bottom-up hierarchical constraints of unimodal cortices^2,5,13. Regarding the functional consequences of structure–function decoupling, influential theoretical accounts suggest that it may enable high-level flexible behaviors^12,13. Consistent with this idea, the multivariate cognitive signature of decoupling derived from the Neurosynth dataset showed alignment with high-level cognitive and emotional processes but a negative correlation with sensory and perceptual processes. Altogether, these findings suggest that structure–function decoupling indexes a gradient of cortical flexibility, whereby regions at higher levels of the hierarchy can transiently disengage from fixed anatomical constraints to support abstract, integrative functions.

We next examined which neurochemical factors contribute to local structure–function decoupling. Based on the idea that functional flexibility reflects dynamic, experience-dependent remapping, we hypothesized that decoupling would co-localize with markers of synaptic plasticity. Supporting this view, decoupling showed strong positive associations with two plasticity-related genes, GAP43 and BDNF, which promote neurite outgrowth and synaptic remodeling, and are central to experience-dependent plasticity^27–30,49. In contrast, we observed robust negative associations with MBP and PVALB, markers of myelin and parvalbumin-positive (PV⁺) interneurons, respectively—both associated with cortical stability³¹. MBP expression was especially anti-correlated with decoupling, consistent with myelin’s role in stabilizing evolutionarily conserved, low-complexity circuits^31,32 and limiting synaptic remodeling^50,51. Similarly, PVALB, indexing inhibitory PV⁺ interneurons, also showed strong negative associations with decoupling, aligning with recent findings linking lower E/I ratios to stronger SFC¹¹. Together, these findings highlight a neurobiological gradient in which regions enriched in plasticity-related genes support functional reorganization and flexible network dynamics, while stability—prompted by myelin and PV⁺ interneurons—anchors functional activity to anatomy, balancing flexibility with architectural constraints.

We further hypothesized that decoupling is supported by diverse neuromodulatory profiles, as a richer neurotransmitter repertoire may enable broader functional configurations³⁴. To test this, we used a curated set of PET receptor and transporter maps, previously applied to study the brain’s functional and metabolic architecture^35,46. The analysis revealed a robust link between neuromodulation and decoupling, with many neurotransmitter systems showing positive loadings, suggesting that overall neurotransmitter density and diversity shape large-scale functional–structural decoupling and promote neural flexibility, especially in transmodal cortices³⁴. Furthermore, we observed that metabotropic receptors loaded more strongly than ionotropic receptors in the PLS-derived relationship, highlighting a temporal asymmetry in neurotransmission: fast ionotropic signaling mediates rapid synaptic responses, while slow metabotropic signaling enables prolonged modulation³⁴. In this regard, it has been shown that during naturalistic stimulation such as movie watching, unimodal cortices track sensory changes over seconds, while transmodal regions integrate inputs across minutes⁵². We therefore propose that fast ionotropic transmission may underpin rigid functional–structural alignment in unimodal areas to swiftly respond to external stimuli, while slow metabotropic signaling promotes dynamic network reconfiguration in transmodal regions.

Collectively, our results demonstrate robust associations between structure–function decoupling and regional neurochemical profiles across the human neocortex. Using direct neurophysiological measures, we show that the brain’s neuromodulatory diversity—and its associated temporal asymmetry—likely contributes to the decoupling of function from structure that supports dynamic reconfiguration of activity atop the structural connectome. Taken together, these findings point toward a hierarchical model in which molecular mechanisms of plasticity and neuromodulatory control jointly shape the emergence of functional flexibility beyond structural constraints. This framework unifies anatomical hierarchy, neurochemical diversity, and cognitive adaptability within a single continuum of cortical organization.

Our findings should be interpreted in light of several methodological considerations. Although DWI-based tractography remains the most commonly used technique for in vivo reconstruction of the structural connectome, it presents known limitations, including biases in tracking both long-range fibers and short U-fibers, as well as sensitivity to processing pipeline parameters^53,54. In this study, we employed advanced fiber tracking and tractogram filtering techniques, validated against realistic connectome benchmarks⁵⁵. Secondly, our SFC framework builds on the connectome eigenmodes perspective^6,56–58. It is well established that inter-regional distance strongly shapes both structural and functional connectivity^59,60. Our approach, based on the eigenmodes of the structural connectome, captures topological rather than geometric relationships and thus does not explicitly control for the effects of physical proximity. Future work could disentangle the respective contributions of topology and spatial embedding to SFC. Beyond distance and topology, several alternative frameworks have been proposed to model structure–function relationships, ranging from simple correlations to communication models based on graph-theoretical properties^2,3. Furthermore, the geometry eigenmodes framework has recently gained momentum as an alternative to the connectome-based approach⁶¹. As neurophysiology-based SFC research is still emerging, the optimal analytic strategy remains to be established.

For functional connectivity estimation, we employed orthogonalized power envelope correlations³⁸, which effectively mitigate volume conduction and offer high test–retest reliability and fingerprinting capability^62,63. Importantly, we also observed that the same spatial pattern of structure–function decoupling emerges when FC is estimated from phase-based interactions, further supporting the robustness of our findings. Future work could extend this approach to explore coupling between structural connectivity and EEG/MEG-derived effective connectivity, or to examine how structure–function relationships dynamically evolve over time (see for example Facca et al.¹⁴ and Liu et al.⁹). In the present study, we focused on broadband activity; however, frequency-specific decoupling maps derived from band-limited MEG connectivity are provided in the Supplementary Materials (Supplementary Fig. S4). These maps largely preserve the broadband hierarchy, although certain frequency bands show marked deviations, pointing to frequency-dependent variations that may differentially relate to regional microarchitectural and chemoarchitectural features—an interesting direction for future research. The strength and spatial distribution of structure–function coupling are also likely influenced by data quality and the signal-to-noise characteristics of MEG recordings. Advances in sensor technology—particularly on-scalp and optically pumped magnetometer (OPM) MEG systems—promise higher spatial precision and sensitivity, which should enhance SNR and improve the fidelity of structure–function mapping⁶⁴. Finally, our analyses focused exclusively on spontaneous brain activity. Whether perturbations—via active tasks, neuromodulation, or psychoactive agents—alter structure–function coupling remains a promising avenue for future investigations.

Methods

Human Connectome Project (HCP) data

Resting-state MEG and DWI data were drawn from the HCP 1200 subject data release^22,23. We considered the 89 HCP subjects from whom MEG data had been collected. Subjects were scanned on a Siemens “Connectome” Skyra 3 T scanner and underwent an extensive imaging protocol including T1-weighted imaging (3D multi-echo magnetization-prepared rapid gradient echo [MEMPRAGE] sequence; voxel size: 0.7 mm isotropic; repetition time [TR]: 2400 ms; echo time [TE]: 2.14 ms), T2-weighted imaging ([SPACE] sequence; voxel size: 0.7 mm isotropic; TR: 3200 ms; TE: 565 ms), and diffusion-weighted imaging (spin-echo planar imaging sequence; voxel size: 1.25 mm isotropic; TR: 5520 ms; TE: 89.5 ms; maximum b-value: 3000 s/mm²; 270 non-collinear diffusion directions; 18 B0 acquisitions). MEG data were acquired using a whole-head MAGNES 3600 system (4D Neuroimaging, San Diego, CA), consisting of 248 magnetometers and 23 reference sensors. Informed consent was obtained from all participants, and the full protocol was approved by the Washington University Institutional Review Board. All ethical regulations relevant to human research participants were followed.

Source localization of MEG data and functional connectivity analysis

The data were downloaded pre-processed at the sensor level²². Briefly, three runs of 6 min each were recorded per subject, with a sampling frequency of 2 kHz. Each recording underwent a state-of-the-art pre-processing pipeline, which included bandpass filtering (1.3–150 Hz), notch filtering (59–61 Hz and 119–121 Hz), and ICA-based removal of non-brain components. All available resting-state recordings for each subject were included in the analyses. The HCP consortium also provides subject-specific pre-computed head models based on a single-shell approach, along with pre-computed surface-based source models in the fsLR space with 4000 vertices per hemisphere. Source reconstruction was performed using custom MATLAB scripts based on the FieldTrip Toolbox⁶⁵ and the megconnectome Toolbox (https://github.com/Washington-University/megconnectome). As an additional preprocessing step, prior to source localization, sensor-level data were low-pass filtered to a frequency range of 1.3–57 Hz using a zero-phase, fourth-order Butterworth filter. Data were then projected into source space via a Linear Constraint Minimum Variance (LCMV) Beamformer. Source time courses were parcellated according to the Schaefer 200 regions 7 networks atlas³⁹. Each ROI time course was reconstructed via Singular Value Decomposition (SVD) of the time courses of its constituent vertices. Broadband functional connectivity was computed as the correlation between power envelopes after multivariate orthogonalization of the signals³⁸. Connectivity analyses utilized functions from the Python package MNE⁶⁶ (https://github.com/mne-tools/mne-python) and its extension MNE_connectivity (https://github.com/mne-tools/mne-connectivity).

Tractography and connectome generation

Probabilistic tractography and construction of SC matrices were performed in subject-specific anatomical space using minimally processed diffusion-weighted imaging (DWI) data. The HCP minimal processing pipeline included B0 intensity normalization and corrections for geometric distortions, eddy currents, motion artifacts, and gradient nonlinearities⁶⁷. Starting from the minimally processed data, we applied Multi-shell Multi-tissue Constrained Spherical Deconvolution⁶⁸ to estimate fiber orientation distributions (FOD). We then segmented the high-resolution T1-weighted image to enable anatomically constrained tractography (ACT)⁶⁹ using a Hybrid Surface-Volume Segmentation approach⁷⁰. Next, a whole-brain tractogram consisting of 10 million streamlines was generated using the IFOD2 fiber-tracking algorithm with a dynamic seeding strategy in white matter⁷¹. To reduce biases inherent to tractography and improve the biological plausibility of the reconstructed streamlines, we applied the SIFT2 algorithm⁷¹. SC matrices were thresholded to achieve a connectance level of 10% while preserving network connectedness through a Minimum Spanning Tree approach, implemented via the NetNeuroTools package (https://github.com/netneurolab/netneurotools).

Structure–function coupling

Regional structure–function coupling was quantified using a multilinear regression model based on the dominant low-frequency eigenmodes of the SC’Laplacian^6,14. To identify these eigenmodes, we first computed the normalized Laplacian of each subject’s SC matrix as:

$$L=I-D^{-\mathrm{1/2}}{AD}^{-\mathrm{1/2}}$$

where I is the identity matrix, D is the degree matrix, and A the structural connectivity matrix. We then diagonalized L and ordered the eigenvalue-eigenvector pairs in ascending order according to the eigenvalues.The Laplacian eigenvectors provide an orthogonal basis describing spatial modes of signal propagation constrained by the network topology; this bears the advantage of capturing functional interactions beyond those supported by direct anatomical projections. The connectivity profile of each node was modeled as a linear combination of the first N = 10 eigenmodes, with coefficients estimated via ordinary least squares (OLS) regression. The local structural–functional decoupling index was computed as 1-R². In practice, this decoupling index represents the proportion of variance in functional connectivity (FC) that cannot be explained by the first 10 eigenmodes of the SC’s Laplacian.

PET receptor atlases

We considered a recently gathered set of PET receptor atlases for 19 different receptors/transporters based on 28 PET studies⁴⁶. The curated set of atlases comprises receptor maps of different neurotransmitter systems, including: Glutamate (mGluR5, NMDA), GABA (GABA_a), Serotonin (5-HT_1A, 5-HT_2A, 5-HT_1B, 5-HT₆, 5-HT₄, 5-HTT), Acetylcholine (α4β2, VAChT, M₁), Noradrenaline (NAT), Histamine (H₃), Opioid (MU), Cannabinoid (CB), and Dopamine (D₁, DAT, D₂). For full details, we refer to the original paper that gathered all the atlases together⁴⁶. All PET atlases are provided in MNI standard space. Each atlas was parcellated using the Schaefer 200 Regions 7 Networks atlas by calculating the mean value within each parcel.

Allen Human Brain Atlas (AHBA) preprocessing

Regional microarray expression data were obtained from six post-mortem brains (two female, ages 24–57, mean ± SD = 42.5 ± 11.3) provided by the Allen Human Brain Atlas²⁴. Data were processed with the abagen toolbox⁷² using the Schaefer 200 regions 7 networks atlas defined on the fsaverage5 surface³⁹. Microarray probes were reannotated using updated gene information⁷³, and those not matched to a valid Entrez ID were discarded. Probes were then filtered based on their expression intensity relative to background noise⁷⁴, excluding those with intensity below background in more than 50% of samples across donors. When multiple probes indexed the same gene, the probe with the most consistent spatial pattern across donors was selected⁷⁵. Probe selection was performed using data aggregated across all donors. The coordinates of tissue samples were updated using non-linear registration via Advanced Normalization Tools (ANTs), and samples were assigned to brain regions by minimizing the Euclidean distance between the fsaverage5 coordinates of each sample and the nearest surface vertex. Samples exceeding two standard deviations above the donor-specific mean distance were excluded. To reduce potential misassignments, sample-to-region matching was constrained by hemisphere and by major structural division (i.e., cortex, subcortex/brainstem, cerebellum). For regions not assigned any tissue samples, a dense interpolated expression map was generated by mapping each surface vertex to its nearest sample within each donor. The vertex-level expression values were averaged within each parcel, weighted by distance, to estimate expression for missing regions⁷⁶. Samples not assigned to any region were discarded. Expression values were normalized independently for each donor using a robust sigmoid transformation and rescaled to the unit interval⁷⁷. Gene-wise normalization used the same procedure. Samples assigned to the same parcel were averaged separately for each donor and then across donors.

Partial Least Squares (PLS) - correlation analyses

To relate structural–functional decoupling with the curated sets of Neurosynth maps and PET receptor/transporter atlases, we employed Partial Least Squares (PLS) correlation analysis. PLS seeks to maximize the covariance between latent dimensions of paired datasets and is based on the Singular Value Decomposition (SVD) of their cross-covariance matrix⁷⁸. For the analysis involving Neurosynth data, the first matrix was constructed by combining parcellated maps corresponding to 124 terms from the Cognitive Atlas, resulting in a matrix of dimensions N ROIs × N terms (200 × 124). The second matrix consisted of subject-specific decoupling maps, with dimensions N ROIs × N subjects (200 × 84). To relate decoupling with PET receptor/transporter atlases, we combined 19 parcellated receptor/transporter atlases into a matrix of dimensions N ROIs × N atlases (200 × 19), while the second matrix was the same subject-specific decoupling matrix described above. Statistical significance of each PLS component was assessed via spin tests, and only components explaining covariance significantly greater than chance were retained for further analysis. Reliability of the weights was evaluated through 10,000 bootstrap iterations, generating 95% confidence intervals (CI). We considered weights reliable if their 95% CI did not include zero. To assess the spatial generalizability of the PLS results, we implemented a distance-dependent cross-validation following Hansen et al.⁴⁵. For each node, 75% of the nearest regions (Euclidean distance) were used for training and the remaining 25% for testing. Model performance was quantified as the Spearman’s correlation between predicted and observed brain scores in the test set. PLS correlation analysis was performed using the Python package PyLS (https://github.com/rmarkello/pyls).

Null models

We conducted all statistical analyses while controlling for spatial autocorrelation by applying state-of-the-art spin tests, implemented using functions from the Python toolbox NetNeuroTools (https://github.com/netneurolab/netneurotools).

Statistics and reproducibility

Analyses were performed on a cohort of N = 89 participants from the HCP for whom both DWI and MEG data were available. The sample size reflects full data availability for this multimodal subset. Randomization and blinding were not applicable to this observational cohort study. Spatial associations were assessed using correlation coefficients, and statistical significance was evaluated via spin-test permutations (N = 10,000) to account for spatial autocorrelation, with a significance threshold of p_spin < 0.05. The reproducibility of the primary decoupling measure was confirmed using the three separate MEG sessions available per subject. Session-specific maps showed high similarity, assessed via Lin’s Concordance Correlation Coefficient (CCC) with a minimum value of 0.86. Furthermore, the robustness of the decoupling measure was assessed across variations in parcellation atlas, different functional connectivity metrics (i.e., amplitude envelope correlation with and without symmetric orthogonalization), and an independent phase-based connectivity measure. Multivariate analyses (i.e., PLS correlation with Neurosynth terms and neurotransmitter maps) were validated via bootstrap resampling (N = 10,000) to assess feature reliability and distance-dependent cross-validation to evaluate spatial generalizability.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

M.F.: Conceptualization, formal analysis, methodology, software, writing (original draft). A.D.F.: Conceptualization, supervision, writing (review & editing). A.B.: Conceptualization, supervision, methodology, writing (review & editing). All authors contributed to editing the manuscript and approved its final version.

Peer review

Peer review information

Communications Biology thanks Arnaud Messé, Ruixiang Li and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Michel Thiebaut de Schotten and Benjamin Bessieres. [A peer review file is available].

Data availability

MEG and DWI data are publicly available on the Human Connectome Project database (https://db.humanconnectome.org/) upon acceptance of HCP open access data use terms.

Code availability

Functions for source-localization of MEG data are available in the FieldTrip toolbox (https://github.com/fieldtrip/fieldtrip). Preprocessing of the Allen Human Brain Atlas (AHBA) was conducted using the publicly available Abagen toolbox (https://github.com/rmarkello/abagen). The code for the eigenmode mapping of structure–function coupling was obtained from a public GitHub repository (https://github.com/yaqiany/scfccoupling-high). Functions to perform the PLS analysis are publicly available in the Pyls toolbox (https://github.com/rmarkello/pyls). The spin test code is available in the NetNeuroTools toolbox (https://github.com/netneurolab/netneurotools). Source data underlying the figures of this work are provided in Supplementary Data 1.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-025-09444-3.