Source-linked AI summary
Decoupling of brain function from structure reveals regional behavioral specialization in humans
Maria Giulia Preti, Dimitri Van De Ville
TL;DR
The extent to which brain function is coupled to its structural wiring remains difficult to quantify. The paper introduces a structural-decoupling index and identifies a macroscale gradient from stronger structure-function coupling in lower-level sensory regions to weaker coupling in higher-level cognitive regions.
Problem
How strongly brain function is bound by structural wiring remains only partially answered and difficult to quantify.
Method
The study introduces a structural-decoupling index to quantify coupling strength between functional signals and underlying brain structure.
Results
The index reveals a macroscale gradient from regions more strongly coupled with structure to regions significantly less coupled, ordered from lower-level sensory to higher-level cognitive behavior.
Takeaways & Limitations
Structure-function coupling varies spatially across the human brain in a behaviorally relevant ordering from lower-level sensory to higher-level cognitive domains.
Takeaways & Limitations
The framework does not account for noise, so ignoring it can positively bias decoupling estimates.
Abstract
from arXiv · showhide
The brain is an assembly of neuronal populations interconnected by structural pathways. Brain activity is expressed on and constrained by this substrate. Therefore, statistical dependencies between functional signals in directly connected areas can be expected higher. However, the degree to which brain function is bound by the underlying wiring diagram remains a complex question that has been only partially answered. Here, we introduce the structural-decoupling index to quantify the coupling strength between structure and function, and we reveal a macroscale gradient from brain regions more strongly coupled, to regions more strongly decoupled, than expected by realistic surrogate data. This gradient spans behavioral domains from lower-level sensory function to high-level cognitive ones and shows for the first time that the strength of structure-function coupling is spatially varying in line with evidence derived from other modalities, such as functional connectivity, gene expression, microstructural properties and temporal hierarchy.
Introduction
The study introduces a structural-decoupling index and a non-parametric test to quantify and assess how brain function depends on anatomical structure. In HCP data, structure-function coupling varies across the cortex, linking sensory and motor regions to lower-level functions and decoupled regions to more complex cognition.
- Related work: Prior work related structural and functional connectivity through correlations, causal modeling, graph approaches, simulations, controllability, and harmonic analysis.Harmonic components summarize global network organization and resemble functional resting-state networks.
- Approach: The structural-decoupling index quantifies regional coupling by splitting activity into weakly and strongly structure-coupled components with, on average, equal energy.The index is based on the energy ratio of these components and measures functional signal smoothness on the structural graph.
- Approach: A new non-parametric test assesses the structural-decoupling index against a strong null model preserving selected properties of functional activity–connectome interplay.The analysis uses data from the Human Connectome Project.
- Results: Sensory regions are more strongly coupled with structure, whereas higher-level cognitive regions, including parietal, temporal, and orbitofrontal areas, are more strongly decoupled.The reported sensory regions include visual, auditory, and somatomotor areas; cognitive examples include executive control networks, the amygdala, and language areas.
- Results: Ranking regions by structural-decoupling index reveals a reliable macroscale cortical organization from lower-level sensory and motor functions to more complex memory, reward, and emotion functions.Behavioral relevance was evaluated through meta-analysis associating topic terms with brain areas, and the organization was highly reliable in test-retest analysis.
Results
Brain activity is preferentially expressed through low-frequency structural-connectome harmonics, while empirical functional connectivity reflects interactions beyond structural wiring. The structural-decoupling index reveals a macroscale gradient linking stronger structural coupling to sensory and motor functions and greater decoupling to higher-level cognition.
- Harmonics of the Structural Connectome: Low-frequency harmonics capture global, slow brain patterns, whereas higher frequencies encode increasingly complex and localized patterns.Harmonic components provide a spectral representation of graph signals with increasing complexity.
- Brain Activity Couples with the Structural Connectome: Resting-state activity is preferentially expressed by lower-frequency structural-connectome components.Activity was represented as weighted harmonic combinations, with time-averaged squared weights forming its energy spectral density.
- Null Models of Brain Activity Informed by the Structure: SC-informed surrogate functional connectivity tracks structural connectivity more strongly than empirical functional connectivity, with r = 0.93 versus r = 0.46.The difference was significant with p<0.05 using a nonparametric test across surrogates, indicating additional interactions in empirical functional connectivity.
- Brain Activity Decomposed According to Structural Coupling: Empirical functional timecourses show significantly greater decoupling in orbitofrontal, temporal, and parietal high-level cognition regions and greater coupling in primary sensory and somatomotor areas.The sensory pattern spans auditory, visual, and pre-/post-central networks, while structurally coupled posterior medial and parietal areas characterize SC-informed surrogates.
- Structural Decoupling Reveals Behaviorally Relevant Gradient: The structural-decoupling gradient associates structurally coupled regions with sensory, motor, and multisensory functions, and structurally decoupled regions with reward, emotion, social cognition, memory, and cognitive control.The gradient also includes auditory processing, verbal/visual semantics, affective processing, and motor/eye movements.
Reliability across sessions
Results were highly reliable across two test-retest resting-state acquisitions, with strong spatial agreement in structural-decoupling patterns and similar behavioral characterization of the gradient.
- Reliability across sessions: 0.99 spatial correlation between structural-decoupling index patterns was observed across the two acquisitions for the surrogates.The main influence in the surrogate results was the shared structural connectivity across datasets.
- Reliability across sessions: The meta-analysis provided very similar behavioral characterization of the structural-decoupling gradient across the two acquisitions.This similarity was reported in Fig. 3a and Supplementary Fig. 5.
Discussion
The study introduces a structural-decoupling index that quantifies regional coupling between functional brain activity and structural connectivity using empirical signals and surrogate data. It reveals a macroscale gradient from strongly structure-coupled sensory-motor regions to more decoupled higher-cognitive areas, while noting noise-related limitations.
- Main findings: Observed brain activity preferentially uses lower graph-frequency components that fit connectome constraints, indicating smoothness on the connectome.The functional energy spectral density projected onto structural harmonics is dominated by low frequencies.
- Method: The structural-decoupling index quantifies function–structure coupling by comparing low-frequency, structure-coupled and high-frequency, structure-decoupled components across empirical and surrogate functional data.Surrogates randomize combinations of structurally informed components while preserving amplitudes, generating a null distribution for detecting significant coupling.
- Main findings: The index reveals a macroscale gradient opposing significantly structure-coupled sensory-motor regions and significantly less coupled higher-cognitive areas.Meta-analysis confirms that this ordering corresponds to a behavioral progression from lower- to higher-level cognitive functions.
- Interpretation: Higher coupling in sensory-motor areas may support fast, reliable responses, whereas episodic memory and self-referential thought are less predictable and more decoupled from structural connectivity.Visual areas partially differentiate from the sensory-motor network in empirical-versus-surrogate structural coupling, consistent with a distinct sensory modality pattern.
- Limitations: Ignoring a uniform spectral-domain noise contribution would positively bias decoupling estimates, potentially making artifact-prone orbitomedial prefrontal regions appear more decoupled than they are.The meta-analysis provides evidence that the index captures functionally meaningful patterns beyond poor signal-to-noise ratio.
- Implications: The framework provides a principled way to quantify regional function–structure coupling and enables future study of inter- and intra-regional variation, including disease-related alterations.Suggested applications include changes over time, experimental conditions, and neurological disease or disorder.
Methods · Overview
The methods apply graph signal processing to human brain imaging by combining structural connectivity with time-dependent functional signals. Graph Laplacian eigendecomposition supplies harmonic components for graph Fourier analysis and brain-aware signal operations.
- Overview: The approach uses the emerging framework of graph signal processing for human brain imaging.
- Overview: The normalized graph adjacency matrix A is defined by the structural connectome.
- Overview: Time-dependent graph signals come from functional data, with activation levels assigned to graph nodes.
- Overview: The graph Laplacian operator L = I −A is eigendecomposed to provide harmonic components.
- Overview: These harmonic components form the graph Fourier transform, representing graph signals as weighted linear combinations.
- Overview: Graph filtering and randomization are introduced as operations that account for brain anatomy.
Data
The study included 56 healthy HCP volunteers and used diffusion-weighted and resting-state fMRI data. Structural connectivity was estimated across 360 cortical regions using tractography-based fiber counts normalized by regional volumes and group-level matrix normalization.
- Participants and ethics: 56 healthy volunteers from the HCP were included in the study, with informed consent and institutional ethical approvals obtained.The study followed relevant guidelines and regulations.
- Imaging acquisition: Diffusion-weighted imaging used three shells of b=1000, 2000, 3000 s mm-2 with 90 directions plus six b=0 acquisitions.The diffusion acquisition used TR=5520ms, TE=89.5ms, flip angle=78◦, and FOV=208x180.
- Imaging acquisition: Two 15-minute resting-state fMRI sessions were acquired, with two additional sessions analyzed separately for reliability.Two of the 56 subjects were excluded from the second resting-state dataset because of incomplete acquisition.
- Structural connectome: Glasser’s multimodal cortical atlas was used to parcellate the cortex into N = 360 regions of interest across both hemispheres.The atlas was converted to volume and split into 180 left-hemisphere and 180 right-hemisphere areas.
- Structural connectome: Structural connectivity was measured as the number of fibers connecting two regions divided by the sum of their connected-region volumes.A group connectome was formed by averaging subjects’ structural matrices, then normalized as A = D−1/2AunnormD−1/2.
Resting-State Functional Data
Resting-state fMRI volumes were preprocessed by removing initial transients, detrending, regressing nuisance signals, band-pass filtering, and parcellating regional signals. The resulting z-scored signals were stored in an N × T matrix for functional-connectome and graph-spectral analyses.
- Preprocessing: The first 10 volumes were discarded to achieve steady-state magnetization, resulting in T = 1190 time points.
- Preprocessing: Voxel fMRI timecourses were detrended and nuisance variables were regressed out, including six head-motion parameters and average cerebrospinal-fluid and white-matter signals.
- Preprocessing: Preprocessed voxel time courses were band-pass filtered at [0.01 −0.15Hz] to improve signal-to-noise ratio for typical resting-state fluctuations.
- Regional signals: Glasser’s multimodal parcellation was resliced to fMRI resolution to compute regionally averaged fMRI signals, which were z-scored and stored in S = [st]t=1,...,T.
- Functional connectivity: Functional-connectome node strengths were assessed as the sum of absolute correlation values for each connection.
- Graph spectral analysis: Graph Fourier analysis represented signals using structural-connectome harmonics, with low-λk eigenmodes encoding signals smooth relative to the structural network.
Null Model Generation
The null models generated surrogate functional signals by sign-randomizing graph spectral coefficients, either using configuration-model harmonics that preserve degree or empirical structural-connectome harmonics. For each resting-state session, 19 surrogates were used to compute averaged surrogate functional connectivity.
- Null model construction: Surrogate functional signals were generated by sign-randomizing graph spectral coefficients, with models either ignoring or incorporating structural connectivity.The sign randomization used diagonal matrices containing random +1/−1 values.
- Null model construction: The structure-agnostic model used a configuration-model graph preserving the degree of A and its harmonics for surrogate-signal reconstruction.The configuration-model graph was denoted A′, with harmonics U′ used in reconstruction.
- Null model construction: The structure-informed model reconstructed surrogate signals using empirical structural-connectome harmonics.This model used a diagonal matrix with random +1/−1 values for sign randomization.
- Surrogate connectivity analysis: 19 surrogates were generated for each resting-state session, and surrogate functional connectivity was computed with Pearson correlations and averaged across surrogates and subjects.FC node strength was assessed as the sum of absolute correlation values for each connection.
Structural-Decoupling Index
The structural-decoupling index quantifies regional structure–function coupling as the ratio between less-coupled and structure-coupled functional components. These components are separated from graph spectral eigenmodes using a median-split spectral filter.
- Structural-Decoupling Index: Functional signals are decomposed into structure-coupled components represented by low-frequency graph eigenmodes and less-coupled components represented by higher-frequency eigenmodes.The decomposition uses the graph Fourier transform and ideal low-pass/high-pass spectral filtering.
- Structural-Decoupling Index: The spectrum is split at a median cut-off into two portions with equal energy based on average energy spectral density across time and subjects.This median-split approach avoids selecting a difficult-to-determine cut-off frequency C.
- Structural-Decoupling Index: The structural-decoupling index is the ratio between the norms of sD and sC across time for each region.The filtered signals are constructed from complementary low- and high-frequency eigenmode matrices.
- Structural-Decoupling Index: Significance is assessed against SC-informed surrogates at α = 1/(19 + 1) = 0.05, with group averages thresholded using binomial detections and correction across N = 360 regions.The individual null threshold uses the maximal surrogate excursion, while group-level inference applies multiple-comparison correction.
Data availability
The study’s supporting data are available through the Human Connectome Project platform, with subject identifiers, analysis code, reporting information, and source data provided through specified repositories and files.
- Supporting data are available on the Human Connectome Project platform at db.humanconnectome.org.
- Included-subject identifiers are reported alongside the StructuralDecouplingIndex code repository at github.com/gpreti/GSP StructuralDecouplingIndex.