Source-linked AI summary
Task-Based Core-Periphery Organisation of Human Brain Dynamics
Danielle S. Bassett, Nicholas F. Wymbs, M. Puck Rombach, Mason A. Porter, Peter J. Mucha, Scott T. Grafton
TL;DR
Understanding how brain networks reorganize during motor learning remains limited by the complexity of changing interactions across regions. Using dynamic community detection and core-periphery analysis of brain activity during motor sequencing, the paper finds that a stable sensorimotor and visual core plus a flexible multimodal periphery predicts individual learning differences.
Problem
Prior work shows that motor learning changes interactions across brain regions, but the network-level organization of these changes remains to be characterized.
Method
The study applies dynamic community detection to multilayer temporal brain networks and identifies temporal core-periphery structure from changes in regional module allegiance.
Results
A relatively stable sensorimotor and visual temporal core and a flexible multimodal association periphery emerge, with their separation predicting individual differences in extended learning.
Takeaways & Limitations
Temporal core-periphery organization provides a compact account of functional-module reconfiguration during sequential, goal-directed behavior.
Takeaways & Limitations
Functional connectivity based on coherence cannot distinguish shared information transfer from common activation by another brain region or external stimulus.
Abstract
from arXiv · showhide
As a person learns a new skill, distinct synapses, brain regions, and circuits are engaged and change over time. In this paper, we develop methods to examine patterns of correlated activity across a large set of brain regions. Our goal is to identify properties that enable robust learning of a motor skill. We measure brain activity during motor sequencing and characterize network properties based on coherent activity between brain regions. Using recently developed algorithms to detect time-evolving communities, we find that the complex reconfiguration patterns of the brain's putative functional modules that control learning can be described parsimoniously by the combined presence of a relatively stiff temporal core that is composed primarily of sensorimotor and visual regions whose connectivity changes little in time and a flexible temporal periphery that is composed primarily of multimodal association regions whose connectivity changes frequently. The separation between temporal core and periphery changes over the course of training and, importantly, is a good predictor of individual differences in learning success. The core of dynamically stiff regions exhibits dense connectivity, which is consistent with notions of core-periphery organization established previously in social networks. Our results demonstrate that core-periphery organization provides an insightful way to understand how putative functional modules are linked. This, in turn, enables the prediction of fundamental human capacities, including the production of complex goal-directed behavior.
Author Summary · Introduction
The study examines how brain-wide functional communities reorganize during motor-skill learning. It identifies a relatively stable sensorimotor and visual core alongside a more flexible multimodal association periphery, with community evolution shaped by training depth.
- Introduction: New motor skills alter brain activity from neuronal firing in motor cortex to interactions between primary motor and premotor areas.These large-scale interaction changes can influence the amount of learning.
- Introduction: The study extracts functional networks from task-based fMRI by dividing motor-learning time series into approximately two-minute intervals.Subjects learned 10-element motor sequences similar to piano arpeggios.
- Introduction: Multilayer brain-network representations are used to identify communities of regions with similar BOLD time courses in successive 2–3 minute windows.These communities represent putative functional modules linking brain regions through coherent activity.
- Introduction: Dynamic community-detection tools and diagnostic measures quantify modularization, module-allegiance variability, and changes in brain dynamics during learning.
- Introduction: Sensorimotor and visual cortices form a relatively stiff temporal core, whereas multimodal association areas form a relatively flexible temporal periphery.Core module affiliations change little over a scanning session, while peripheral affiliations are more variable.
- Introduction: The temporal evolution of community structure is strongly modulated by training depth, measured by the total number of practiced trials.
Results
Learning alters dynamic community structure: modularity initially rises then decreases with practice, while regional flexibility defines a stiff sensorimotor core and flexible multimodal periphery. Greater temporal core–periphery separation predicts better subsequent motor learning, and temporal core regions correspond to densely connected static core nodes.
- Results: After an initial increase from 50 to 200 practiced trials, multilayer modularity decreased as practice increased, indicating less pronounced community structure with learning.Community structure changed with the number of practiced trials independently of when practice occurred during the six weeks.
- Regional Variation in Flexibility: Regional flexibility ranged from approximately 0.04 to approximately 0.14, indicating that regions changed module allegiance across 4%–14% of trial blocks on average.The flexibility distribution was non-Gaussian, with mostly high-flexibility regions and a left-heavy tail.
- Defining the Temporal Core and Temporal Periphery: The temporal core comprised 19 predominantly primary sensorimotor regions, whereas the more flexible periphery comprised 25 predominantly multimodal regions.Most motor-related core regions were left-lateralized, consistent with participants’ right-handed motor-sequence performance.
- Geometrical Core-Periphery Organization: Regions with low temporal flexibility tended to be strongly connected core nodes in static network layers, and this relationship was reliable across training conditions and individual subjects.Geometrical core scores also remained relatively consistent across scanning sessions and sequence blocks.
Discussion
The study identifies temporal core-periphery organization as a predictive component of brain dynamics supporting sequential, goal-directed behavior. This framework complements geometrical core-periphery structure and parsimoniously characterizes dynamic reconfiguration of functional modules during learning.
- Network Predictions of Future Learning: Temporal core-periphery organization predicts individual differences in sequential motor learning, with poor learners showing weaker separation between core and periphery.The study summarizes this relationship in Fig. 7 and reports that temporal network properties can predict extended motor learning from early training.
- Network Predictions of Future Learning: The findings are not explained by regional size or task block design, and core-periphery measures predict learning better than regional signal power, mean connectivity strength, or general-linear-model estimates.These supporting analyses are reported in Text S1.
- Temporal Core-Periphery Organization: The temporal core comprises regions with fewer-than-expected changes in module allegiance, reflecting consistent task-based mesoscale functional connectivity over time.The method defines temporal-core membership relative to a dynamic-network null model.
- Core-Periphery Organization of Human Brain Structure and Function: Temporal and geometrical cores are complementary, and regions in the temporal core are also likely to occur in the geometrical core.Geometrical cores capture coherent activity, whereas temporal cores capture persistence of task-related organization throughout an experiment.
- Modular Versus Core-Periphery Organization: Temporal core-periphery organization parsimoniously describes the dynamic reconfiguration of putative functional modules and highlights different roles of brain regions in information processing.Community structure and core-periphery organization can coexist, with modules representing functional groupings and core-periphery structure characterizing role differences.
- Limitations: The conclusions are limited by large-scale parcellation and coherence-based functional connectivity, motivating finer-resolution studies and alternative association estimates or models.The study cautions that coherence does not by itself distinguish possible drivers of strong inter-regional associations.
Materials and Methods
The study examined motor-sequence learning in 20 participants through repeated behavioral training and four fMRI sessions. Brain activity was preprocessed, parcellated into 112 regions, and represented as window-specific functional-connectivity matrices, while learning was quantified from behavioral improvement.
- Participants and procedure: The final sample comprised 20 participants who completed at least 30 behavioral training sessions, three fMRI test sessions, and one pre-training scan.Two participants were excluded for noncompletion or excessive head motion; included participants had normal or corrected vision and no neurological or psychiatric history.
- Behavioral task: Participants practiced six visually presented 10-element sequences with their right hand in a discrete sequence-production task across three exposure levels.The same task and sequences were used during scanner testing, with unlimited response time and instructions to respond quickly while maintaining accuracy.
- Behavioral analysis: Learning was estimated from movement-time trajectories during early home-training sessions, separating the initial speed, asymptotic speed, and exponential rate of improvement.A negative learning parameter indicated decreasing movement time, interpreted as learning-related improvement.
- fMRI acquisition and preprocessing: Functional images were acquired with BOLD-sensitive echo-planar imaging on a 3.0 T Siemens Trio and processed by realignment, coregistration, MNI normalization, and 8 mm spatial smoothing.The acquisition used 37 slices, a 2000 ms TR, 30 ms TE, and 3 mm slice thickness with a 0.5 mm gap.
- Network construction: The brain was parcellated into 112 cortical and subcortical regions, and each block-specific dataset was converted into an N × N matrix of pairwise functional-connectivity weights.The parcellation used the structural Harvard-Oxford atlas, with N = 112 regions in the full-brain atlas.
Figure and Table Legends
The figures define temporal and geometrical brain-network core–periphery organization, track its changes during motor learning, and examine its relationship with community structure and learning. The accompanying results indicate that regional flexibility roles and geometrical core organization remain robust across training conditions and practice duration.
- Network analysis and experimental design: The figures relate temporal and geometrical core-periphery structure to brain-region flexibility, community structure, learning, and the experimental motor-sequence task.They also document temporal-network construction from fMRI coherence, dynamic-network diagnostics, trial structure, and the experiment timeline.
- Temporal core-periphery organization: Regional roles in the temporal core, bulk, and periphery remain robust across training levels, with similar flexibility across sequence-training conditions.This robustness is quantified using flexibility variability across regions and is illustrated for extensively, moderately, and minimally trained sequences.
- Temporal core-periphery organization: Regional flexibility roles are conserved across both training intensity and duration, with core regions remaining less flexible than bulk and periphery regions.The comparison spans minimally, moderately, and extensively trained sequences across scanning sessions 1–4.
- Geometrical core-periphery organization: Geometrical core-periphery organization remains consistent over 42 days of practice, across sequence types, training intensity, and training duration.The figure series tracks geometrical core scores across four scanning sessions and minimally, moderately, and extensively trained sequences.
A B extensive training moderate training minimal training
The study compares temporal core–periphery organization with community structure in functional brain networks during early learning. Communities identified by multilayer modularity significantly overlap with temporal core, bulk, and periphery partitions across training conditions.
- Community detection: A representative community partition was constructed after 100 single-layer modularity optimizations of the mean-coherence matrix.The representative partition was derived from the set of optimization results and included a community shown in Fig. 4.
- Methodological qualification: The observed core–periphery and community-structure relationship is established for networks encoded by mean matrices, whose averaged correlation-based forms may not capture individual-network topology or geometry.This motivates testing the relationship beyond mean-matrix representations.
- Partition comparison: The temporal core, bulk, and periphery define a functional brain-network partition whose similarity to algorithmic communities is quantified using the Rand coefficient z-score.For each participant, block, and optimization, the study compares the core–periphery partition with multilayer-modularity communities.
- Community structure: Multilayer-modularity communities significantly overlap with temporal core, bulk, and peripheral regions during early learning across extensively, moderately, and minimally trained sequences.The similarity was assessed across four scanning sessions over approximately six weeks.
Methodological Considerations · Experimental Factors · Effect of Region Size
The authors examined whether brain-region size could drive the observed core-periphery organization, finding no significant correlation between region size and flexibility.
- Effect of Region Size: Effect of Region Size: No significant correlation was observed between brain-region size and flexibility, arguing against region size as a driver of the functional core-periphery organization.Although region size affects hard-wired connectivity estimates in structural connectomes, this study evaluated functional connectomes.
Effect of Block Design
Block design can induce correlations among brain regions sharing trial-related on–off activity within a time window. The analysis uses much longer timescales—one window every 40–60 TRs—to examine learning independently of block-design effects.
- Block-design effects: Shared trial-related on–off activity can make brain regions appear highly correlated within a single multilayer time window.For example, motor cortex and supplementary motor area may be active during trials but quiet during inter-trial intervals, producing correlated activity with other regions showing the same pattern.
- Timescale: Long-timescale community-allegiance dynamics occur at 0.0083–0.012 Hz, roughly an order of magnitude below the associated block-design frequencies.These dynamics therefore probe connectivity changes over substantially longer timescales than activity correlations within a single time window.
- Learning analysis: Using one time window every 40–60 TRs enables separate examination of early and extended learning while reducing block-design confounding.At these longer timescales, the analysis can probe both learning phases independently of block-design effects.
Specificity of Dynamic Network Organization as a Predictor of Learning
Dynamic flexibility predicted learning better than simpler measures of brain activity, connectivity, or GLM parameter estimates. Mean pairwise coherence, mean power, and EXT-trial parameter estimates from the first scanning session showed no significant predictive relationship with later learning.
- Flexibility provided greater predictive power for learning than activity power, mean connectivity strength, or general linear model parameter estimates.
- −0.003 (p = 0.987) for mean pairwise coherence and −0.218 (p = 0.354) for mean regional activity power showed no predictive relationship with subsequent learning.Both measures were obtained during the first scanning session, before approximately 10 home training sessions.
- r = −0.10 (p = 0.65) showed no significant correlation between mean EXT-trial GLM parameter estimates in scanning session 1 and subsequent EXT-sequence learning.The estimates summarized BOLD-signal changes across brain regions.
Subject State-Dependence of Dynamic Network Organization
Temporal and geometrical core-periphery organization remained consistent across all four scanning sessions, while learning-related network changes persisted when analyses excluded the first session. These changes were not explained by altered task-related BOLD activation or its correlation with core-periphery structure.
- Session consistency: Core-periphery organization was observed consistently across all four scanning sessions, with stable anatomical identities for temporal and geometrical core and periphery nodes.Figures 2 and 3 documented session-consistent node identities for temporal and geometrical core-periphery organization.
- State-dependence control: Excluding the potentially higher-anxiety first session, scans 2–4 still showed decreased maximum modularity, more communities, increased flexibility, and reduced geometrical-core-score variance with learning.These analyses addressed whether longitudinal changes reflected task-related organization rather than early-session mental or physiological state.
- Activation control: Task-related fMRI BOLD activation magnitude in core, bulk, and peripheral regions did not change significantly across scanning sessions.A repeated-measures ANOVA analyzed training-depth-averaged GLM parameter estimates across network designations and sessions.
- Activation control: Mean GLM parameter estimates were not significantly correlated with learning-related changes in core-periphery structure for temporal core or bulk nodes.Reported correlations were r = 0.20, p = 0.52 for temporal-core nodes and r = −0.05, p = 0.86 for temporal-bulk nodes.
Temporal Core-Periphery Organization and Task-Related Activations
The temporal core tends to comprise regions with strong task-related activations, while whole-brain functional-connectivity analysis captures learning-related changes beyond traditional GLM analyses.
- Temporal-core regions tend to show strong task-related activations, as illustrated in Fig. 6.
- The whole-brain functional-connectivity approach remains sensitive to learning-related changes that traditional GLM analysis may not identify.
- The analysis explicitly examines the relationship between dynamic community structure and task-related activations.
Dynamic Community Detection · Effect of Structural Resolution Parameter · Effect of Temporal Resolution Parameter
The analysis tests how structural and temporal resolution choices affect multilayer community detection and the resulting temporal core–periphery organization. Results support a genuine mesoscale core–periphery structure across structural resolutions and assess its robustness across temporal coupling strengths.
- Dynamic Community Detection: The study evaluates structural resolution γ and temporal resolution ω because both parameters shape multilayer modularity optimization and dynamic community assignments.The main analysis used γ = 1 and ω = 1, while the supplementary analyses examine alternative values.
- Effect of Structural Resolution Parameter: The standard structural-resolution choice γ = 1 subtracts the optimization null model P from the adjacency tensor A in the modularity objective.Lower γ accesses smaller communities, whereas higher γ accesses larger community scales.
- Effect of Structural Resolution Parameter: At γ = 1, the bulk contains about 65 regions, while the temporal core and periphery contain about 20 and 25 regions, respectively.The structural-resolution analysis varied γ across broad and fine ranges to characterize these changes.
- Effect of Structural Resolution Parameter: Across γ ≈ 0.8–2.5, partitions span approximately 1–112 communities, while the examined core–periphery structure occurs at a mesoscale of approximately 3–20 communities.These partitions correspond to mean community sizes of approximately 6–37 brain regions, supporting the core–periphery structure as a genuine mesoscale feature of coherent brain dynamics.
- Effect of Temporal Resolution Parameter: Increasing ω from 0.1 to 2 decreases the number of detected communities, consistent with stronger inter-layer coupling allowing less variation in assignments across time.The analysis used increments of Δω = 0.1; smaller ω permits greater temporal variation in community assignments.
- Effect of Temporal Resolution Parameter: Temporal core, bulk, and periphery membership was tested across ω values using the same 95% confidence-interval criterion applied at ω = 1.Core and periphery regions were defined as those below and above the nodal null-model confidence interval, respectively.