Functional connectivity gradients depend on cortical sampling position and brain state
nvidia/Gemma-4-26B-A4B-NVFP4 · academic_accessible/eval 94%/6 min read/Jul 26, 2026
Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.
Where You Look Matters: How Sampling Depth Shapes the Brain's Connectivity Map
Researchers found that where you measure brain activity—deep in the cortex versus near the white matter boundary—changes the map of how brain regions connect. This difference isn't fixed. It changes when you watch a movie, causing some networks to become more similar and others to become more distinct.
In the field of connectomics (the study of the brain's neural connections), scientists aim to map functional-connectivity gradients. These gradients act as a coordinate system. They situate different brain regions along a hierarchy from primary sensory systems to complex association cortex. Currently, most analyses treat signals sampled from nearby cortical positions as interchangeable. They assume a measurement taken from the middle of the gray matter is functionally identical to one taken near the boundary where the gray matter meets the white matter.
However, this assumption ignores anatomical and hemodynamic (blood flow-related) transitions at that boundary. As researchers push toward higher resolutions, a key question remains: does the specific depth at which we sample the BOLD (Blood Oxygen Level Dependent) signal—the proxy for neural activity used in fMRI (functional magnetic resonance imaging)—constitute a meaningful dimension of brain organization? This study suggests the answer is yes. Furthermore, this "sampling dimension" is sensitive to the brain's cognitive state.
The hidden variable in cortical sampling
The status quo in functional connectivity analysis typically relies on a single representative surface for the cortex. While this simplifies computation, it overlooks the biological complexity of the gray-white matter interface. This boundary marks a transition in tissue composition, fiber geometry, and vascular contributions.
If researchers sample from the "midthickness" (the core layer of the cortex), they capture a different signal than if they sample from the "gray-white-boundary" (the superficial edge). Previous studies have hinted that these signals differ. However, they haven't explored whether these differences represent a fundamental geometric property of the brain's connectivity. If the choice of sampling depth merely introduces a static offset, it might be ignored. But if the relationship between these depths changes depending on whether a person is resting or engaged in a task, then sampling position becomes a critical variable. It can shape scientific conclusions about macroscale brain organization.
Decomposing the geometry of connectivity
To investigate this, the authors employed a dual-surface sampling approach using 7 T (high-field strength) fMRI data from 176 participants. Instead of a single measurement, they extracted BOLD time series from two distinct subject-specific surfaces: the midthickness and the gray-white boundary.
The methodology follows a rigorous pipeline to ensure the differences observed are not mere artifacts:
Signal Extraction: BOLD signals are sampled from 400 cortical parcels on both surfaces using volume-to-surface mapping.
Gradient Embedding: The authors use diffusion-map embedding to transform massive connectivity matrices into a low-dimensional "gradient space." This acts like a dimensionality reduction technique. It is similar to how PCA (Principal Component Analysis) collapses complex data into a few essential axes.
Alignment: Different mathematical fits can rotate or scale these axes. Therefore, the authors align every participant's gradient to a fixed, consensus reference using a Procrustes transformation.
Shapley Decomposition: To understand how the state change (moving from rest to movie-watching) occurs, the authors apply Shapley decomposition. This mathematical tool assigns the total change in distance between the two surfaces to each surface individually. It helps determine if one surface is moving toward the other or if both are moving in a coordinated dance.
As shown in [Figure 2a], this decomposition allows the researchers to distinguish between a scenario where only one surface changes and one where both undergo a coordinated reconfiguration.
Network-specific reconfiguration during movies
The study's most striking finding is that movie viewing does not affect the entire brain uniformly. Instead, it induces a "convergence" or "differentiation" of the two sampled surfaces depending on the functional network involved.
The authors report that, on average, the separation between the midthickness and white-boundary representations decreases during movie watching [Figure 1d]. Specifically, the Visual, Limbic, and Default networks show significant convergence. This means the connectivity profiles sampled at different depths become more similar when the brain processes a naturalistic stimulus [Figure 1c]. Conversely, the Salience/ventral-attention cortex shows differentiation. In this region, the two surfaces actually become more distinct.
Crucially, the authors demonstrate that this isn't just a byproduct of the embedding process. They show that the changes in the original, high-dimensional connectivity profiles track the changes seen in the simplified gradients .
Figure 3. FC-profile reconfiguration provides an embedding-independent basis for the gradient pattern. a, FC-profile distance metric. For every source parcel and condition, midthickness and white-boundary FC rows were separately z-scored across the other 399 cortical targets. Their root-mean-square distance quantified cross-depth profile separation, and the state effect was defined as MOVIE distance minus REST distance. Negative values indicate that the two depth-specific FC profiles become more similar during MOVIE (profile convergence); positive values indicate greater dissimilarity (profile differentiation). b, Target-resolved MOVIE-minus-REST profile-distance changes for the four focal source networks and all seven target networks. Blue denotes convergence and red differentiation. Asterisks denote motion-adjusted BH-FDR across the complete 49 source-target tests. c, Whole-profile source-network summaries with bootstrap 95% confidence intervals. Unlike panel b, these values were calculated directly from each source parcel's complete 399-target FC profile and were not obtained by averaging the seven network-level heat-map cells. d, Associations between subject-level whole-profile change and G1 separation change after residualizing both variables for the MOVIE-minus-REST motion difference. Positive partial correlations indicate that stronger profile convergence accompanies stronger G1 convergence, and stronger profile differentiation accompanies stronger G1 differentiation. FC-profile correspondence does not establish causality.
Furthermore, they found that these changes aren't driven by simple vascular differences. A comparison with a venous atlas showed that the spatial pattern of separation did not follow the known distribution of large veins [Figure S8].
Limits of the sampling dimension
While the findings are robust, the authors highlight several important caveats. First, the spatial resolution of current fMRI is still not high enough to treat these two surfaces as truly isolated anatomical layers. They are essentially adjacent samples from a partially overlapping hemodynamic field. Think of it like trying to measure the temperature of a single room using two thermometers placed only an inch apart. You are capturing a gradient, but you cannot claim you are measuring two entirely different climates.
Second, the reliability of these measurements varies. While the group-level spatial maps are highly reproducible, the authors report that individual rankings are much less stable. This refers to how one person compares to another based on their separation score, especially when comparing data across different scanner strengths [Figure 5d]. This suggests that while the "depth effect" is a real phenomenon for understanding groups, it is not yet a reliable biomarker for individual clinical diagnosis.
Finally, the semantic analysis was constrained. The researchers used WordNet to quantify "semantic change" (the shift in meaning between consecutive moments in a movie). However, this is a purely lexical measure. It does not capture the full complexity of human narrative, social cues, or emotional arcs that characterize actual movie watching.
The verdict: A new dimension for connectomics
Is cortical sampling position a critical factor for neuroimaging? The evidence suggests yes. The study establishes that where you sample the BOLD signal is an interpretable dimension of functional connectivity. It responds dynamically to brain states.
The findings are highly reproducible across different hardware. The resting-state organization observed at 7 T successfully generalized to 3 T. Even though the absolute magnitude of the effect changed, the spatial pattern remained [Figure 5c]. This confirms that the spatial pattern is a fundamental property of the brain's organization rather than a quirk of high-field magnets. For researchers, the takeaway is clear. Any conclusion drawn about the "hierarchy" or "geometry" of the brain must account for the depth from which the signal was pulled. We cannot treat the cortex as a single, flat sheet of data.
Figures from the paper
Figure 1. Movie viewing induces network-specific cortical depth-gradient reconfiguration. a, Matchedstate analysis of 176 participants with four REST and four MOVIE runs. BOLD signals were sampled from subject-specific midthickness and gray-white-boundary surfaces, reduced to Schaefer-400 cortical parcels, converted to duration-matched functional-connectivity matrices and embedded using a fixed common gradient reference. Separation, Sep, is the within-subject scale-normalized absolute G1 distance between the two sampled surfaces. Δ Sep is Sep(MOVIE) minus Sep(REST); negative values indicate cross-depth convergence and positive values differentiation. b, Across-participant mean parcelwise Δ Sep displayed on fsLR-32k very-inflated surfaces. Blue denotes convergence, red differentiation and gray the medial wall without a Schaefer assignment. c, Seven-network mean Δ Sep. Points are group means and horizontal intervals are nonparametric bootstrap 95% confidence intervals; the vertical line denotes no state change. Blue and red encode effect direction, not statistical significance. Network inference used a linear model adjusting for the MOVIE-minus-REST motion difference and BH-FDR across seven networks. Visual, Limbic and Default showed significant convergence, whereas Salience/ventral attention showed significant differentiation. d, Paired wholecortex separation for all 176 participants. Translucent points are participants, gray lines join the same participant and large circles are group means. Mean separation decreased from 0.175 in REST to 0.155 in MOVIE (paired dz=-0.40; motion-adjusted p=3.21 × 10 -7 ). e, IntegrationGain across four prespecified duration-matched run pairs for the four networks surviving the panel-c motion-adjusted FDR analysis. IntegrationGain is Sep(REST) minus Sep(MOVIE), the negative of Δ Sep; positive values therefore indicate convergence and negative values differentiation. Run-pair estimates assess directional consistency within the same participants and are descriptive, not four independent replications.Figure 2. Distinct surface contributions generate convergence and G1-specific differentiation. a, Counterfactual state space used to decompose the observed REST-to-MOVIE separation change. Each D denotes the normalized G1 gradient-space distance between the two surface representations joined by the double arrow: D00 compares REST midthickness with REST white boundary, D10 compares MOVIE midthickness with REST white boundary, D01 compares REST midthickness with MOVIE white boundary and D11 compares MOVIE midthickness with MOVIE white boundary. Thus, the double arrow denotes a distance comparison, not addition. Horizontal arrows change midthickness while holding the white boundary fixed; vertical arrows change the white boundary while holding midthickness fixed. D00 and D11 are observed REST and MOVIE distances, whereas D10 and D01 are counterfactual mixed-state distances. The two Shapley contributions sum exactly to Δ Sep. b, Midthickness Shapley contributions, φ mid (orange circles), and white-boundary Shapley contributions, φ white (blue squares), for the four focal networks; intervals are bootstrap 95% confidence intervals. Negative φ reduces G1 separation and therefore promotes convergence, whereas positive φ increases G1 separation and promotes differentiation; φ mid plus φ white equals Δ Sep. c, REST-to-MOVIE movements in aligned G1-G2 space; hollow symbols denote REST, filled symbols MOVIE, orange circles midthickness and blue squares the white boundary. Panel b quantifies contributions to G1 separation, whereas panel c visualizes centroid movement along the first and second functional-connectivity gradients; an arrow's direction along either axis does not directly indicate the sign of φ . d, Direct one-surface-at-a-time distance changes. Positive values mean that changing either surface alone increases its distance from the other surface's REST representation. Thus, final convergence reflects coordinated repositioning of both representations, not one approaching the other's original position. These effects are defined in aligned functional-gradient space and do not imply anatomical movement or causal influence.Figure 4. Shared movie timing and semantic transitions organize cross-depth commonality. a, Stimuluslocking test performed on mean absolute Yeo-7 intersubject functional-connectivity (ISFC) edges before gradient embedding. Diamonds are observed values and translucent points are 1,000 null values generated by independently circularly shifting each participant within each clip. Observed midthickness and white-boundary values were 0.0361 and 0.0308, compared with null means of 0.00106 and 0.00094; both one-sided p=0.001. b, Spatial correspondence across 400 cortical parcels. The x axis is normalized G1 midthickness-to-white-boundary separation from ordinary within-participant MOVIE FC; the y axis is the corresponding separation from stimulus-shared ISFC. Colors identify Yeo-7 networks, and the black line is the least-squares fit. Inference used 10,000 hemisphere-preserving spins (Pearson r=0.410, pspin=0.00050), supporting a shared movie-state topology rather than showing that ISFC explains MOVIE-minus-REST Δ Sep. c, Semantic-change coefficients for the global cortex, Control and Salience/ventral-attention networks. Positive coefficients indicate that larger WordNet semantic transitions, shifted by 5 s, predict greater common-mode relative to differential-mode synchrony after MotionEnergy level/change, semantic-density and clip controls. Gray violins show the complete 1,000sample empirical circular-shift null distributions, dark horizontal segments their 2.5th-97.5th percentiles, short black ticks their means and colored diamonds the observed coefficients. These null intervals are not confidence intervals around the observed coefficients. Displayed q values are BH-FDR across Global plus seven networks. Control and Salience/ventral attention each had q=0.004, whereas the global effect was not significant. d, Clip-specific coefficients for all 13 movie clips. Control was positive in 9/13 clips and survived clip-as-replicate FDR (q=0.026); Salience/ventral attention was positive in 10/13 clips but did not survive this analysis (q=0.096). e, Dynamic-commonality workflow and nuisance controls. WordNet change is an automated feature transition, not a human-annotated narrative boundary. These associations do not establish cortical-layer or causal mechanisms.Figure 5. REST depth-gradient organization generalizes from 7T to 3T, but magnitude does not. a,b, Across-participant mean normalized G1 midthickness-to-white-boundary separation maps for 3T REST1 and matched 7T REST in the 172-participant complete-case cohort. Both maps use the same sequential blue scale. To prevent a small number of extreme parcels from compressing the visible cortical pattern, the display range is fixed to the pooled 2nd and 98th percentiles across both maps (0.103-0.407); color-bar extensions denote values outside this range. This clipping affects visualization only and was not applied to any statistical analysis. Gray denotes the medial wall without a Schaefer assignment. c, Spatial correspondence across the same 400 cortical parcels. The black line is the leastsquares fit; inference used 10,000 hemisphere-preserving spins to account for cortical spatial autocorrelation (Pearson r=0.826, pspin=0.00030). d, Cross-field individual ranking after residualizing each field's separation estimate for its own motion. Each point is one participant; inference used 10,000 subject-label permutations (r=0.239, pperm=0.0031; bootstrap 95% CI 0.106-0.390), indicating modest rather than strong individual generalization. e, Seven-network means. Blue points are 7T, orange points 3T and gray lines join the same network, showing preserved network organization alongside higher 3T values. f, Paired whole-cortex separation for all 172 participants. Gray lines join the same participant, translucent points show individuals and large circles group means. Mean separation was 0.174 at 7T and 0.235 at 3T (mean difference=0.061, paired dz=0.935; motion-adjusted p=2.97 × 10 -25 ). This is same-participant cross-field validation of REST spatial organization, not independent-cohort replication, numerical equivalence across acquisition conditions or validation of the 7T MOVIE effect.