Abstract The brain’s intrinsic organization into resting-state networks has long been suggested to be fundamental for the offline support of mental processes. Extensive task-based evidence support the relevance of the crosstalk between network segregation, supporting systems specialization, and network integration, allowing to flexibly implement complex behavior. However, only scarce evidence focusing on few behavioral measures directly link changes in these network properties at rest with interindividual differences in behavior. In this work, using a comprehensive set of behavioral measures together with resting-state functional magnetic resonance imaging from the human connectome project, we assessed whether connectivity patterns predictive of behavior reflected segregation or integration based on Groupe d’Imagerie Fonctionnelle Network Atlas, a 33 resting-state-networks atlas with cognitive characterization. We found that connectivity relevant for behavior organizes into three main latent dimensions, summarizing Cognition, Positive Affect and Negative Affect. Crucially, we found that connectivity predictive of Cognition, but not Affect, was associated with global network segregation and reduced network integration. We further reveal differential resting-state-network involvements, with Cognition associated with the segregation of higher-level resting-state networks, and the integration of lower-level, visual networks. All in all, the present results suggest that cognition may rest upon a segregated, modular intrinsic brain architecture.
Abstract I investigate a graph neural field model implemented on high-resolution multimodal individual human connectomes. The model, based on Wilson-Cowan and wave-diffusion equations, captures the harmonic power spectrum of functional magnetic resonance imaging and the temporal power spectrum of magnetoencephalography (MEG) over a wide range of scales. Additionally, the model displays properties of neuronal activity thought to be relevant for healthy brain function, such as proximity to instability and long-range temporal correlations (LRTCs), without being explicitly designed or optimized to achieve them. Finally, I find that model LRTCs originate in specific temporal frequency bands, display distinct patterns of spatial localization on the cortical surface, and are nontrivially linked to structural connectivity, with particular contributions of long-range white-matter fibers. Together, these findings extend the scope of graph neural fields as an effective framework for model-based investigations of multimodal neuroimaging data.
Abstract Dynamic functional connectivity (FC) analyses of resting-state functional magnetic resonance imaging (fMRI) commonly apply k-means clustering to time-resolved whole-brain connectivity matrices to identify recurring “brain states” presumed to reflect shared neural dynamics. However, whether this approach reliably identifies group-level states in relatively small and heterogeneous clinical cohorts remains unclear. Here, we systematically evaluated k-means clustering in temporal lobe epilepsy (TLE) subgroups (bilateral, left, and right TLE) and healthy controls by incrementally increasing the number of clusters (k) from two to the number of participants in each group. Across all groups, clustering solutions converged toward participant-specific rather than shared group-level FC patterns. Identified FC states increasingly resembled participants’ static FC matrices, with higher k values approaching near one-to-one correspondence between states and participants. Elbow and silhouette analyses failed to identify a consistent optimal k and generally favored larger clustering solutions, consistent with increasing participant-specific structure. State occupancy analyses further showed prolonged residence within a single dominant state, providing limited evidence for shared dynamic transitions. These findings persisted under a leave-same-participant-out condition designed to reduce participant-specific influence. Together, these findings suggest that whole-brain k-means clustering in small clinical cohorts may primarily capture interindividual FC differences rather than reproducible group-level dynamic brain states.
Functional connectivity is often constructed to understand functional organizations in neural systems, where the brain regions and their pairwise interactions are viewed as nodes and edges, respectively. In practice, functional connectivity is commonly estimated via the correlation of pairs of brain regions. One limitation is that the correlation coefficient captures only the pairwise linear dependence relationship between pairs of nodes and may fail to capture complex higher-order relationships. Recently, a novel concept known as edge-centric functional connectivity (eFC) has been introduced to measure interactions between pairs of edges based on the cofluctuation of two nodal time series, offering a new perspective for understanding brain networks. Nevertheless, eFC considers the absolute levels of edge time series, that is, their mean values, in estimation. If the parameter of interest is the covariation between a pair of edges, incorporating mean values of edge time series can introduce bias or deviation, resulting in a skewed estimation. In this manuscript, we propose an alternative approach to estimate the unbiased covariation between pairs of edges, termed centered edge functional connectivity (ceFC), with theoretical foundations. We demonstrate that the proposed estimator is consistent with a sufficient sample size or number of time frames. Additionally, we develop a multiple hypothesis testing framework with a controlled false discovery rate to evaluate the strength of the unbiased covariation among edges. Furthermore, we employ thresholding to obtain a thresholded estimator that has been shown to converge to the true ceFC matrix in high-dimensional settings in which the number of nodes is much larger than the number of samples or time frames. We validate the finite sample performance of the proposed methods via numerical studies and a data application using the Midnight Scan Club dataset.
Abstract We present a methodological framework for analyzing multifrequency dynamic functional connectivity (dFC) in electrophysiological recordings. The approach characterizes not only the magnitude of network reconfiguration over time but also whether these changes are spatially random or, instead, spatially organized in ways that drive a slower reconfiguration of modular structure. We define a generative null model of multiscale connectivity fluctuations that differ in their degree of spatiotemporal organization, and we describe dFC flows through the joint assessment of (a) instantaneous reconfiguration speed and (b) the extent and quality of ongoing modular reorganization. Different combinations of these features delineate distinct “flow styles,” ranging from more liquid to more frozen dynamics. As a case study, we apply this framework to stereo-electroencephalography recordings from epileptic patients. We identify transitions between dynamic “allegiance states,” whose flow styles closely mirror those of the null model. Seizure onset is associated with a pronounced slowing of dFC speed, while a specific postictal regime combines low speed with highly frozen allegiance and aligns most strongly with clinician-annotated aphasia. These pilot results suggest that temporal multiplex network analyses can reveal transient, frequency-specific network regimes linked to symptom expression and offers a generalizable tool for dissecting fast network dynamics in intracranial recordings.
Abstract Light influences human cognition and behavior, and neuroimaging studies show that brain activity is modulated by light intensity. However, how light affects temporal brain-state transitions and the control energy required for these transitions remains unclear. To investigate this, we applied a network control theory approach to fMRI data collected from 20 healthy participants who performed an auditory discrimination task under four light intensities. Despite similar task performance, higher light intensity increased the number of transitions between brain states. Increasing light intensity enhanced the occurrence of a visual network dominated brain state, while decreasing the occurrence of brain states characterized by suppressed default mode activity and elevated frontoparietal activity. Furthermore, light intensity affected transition probabilities among different brain states contributed by redistributions of energy demands with the dorso-posterior thalamus appearing to play a key role in mediating light-related effects. Regionally, higher light intensity was associated with a trend toward reduced control energy in the visual network; frontal, cingulate, and insular cortices; and caudate and task-related regions, while showing a trend toward increased control energy in the somatomotor network and temporal pole. These findings suggest that high-intensity light may enhance neural efficiency and flexibility by redistributing control energy demands across brain regions.
Abstract Integrating neuroimaging data enhances our understanding of the brain. Structural magnetic resonance imaging (sMRI) offers high-resolution anatomical detail, while functional MRI (fMRI) captures dynamic neural activity. Combining these modalities can reveal significant structure–function relationships in the brain. However, existing approaches typically link sMRI to only a single fMRI network, overlooking the spatial complexity of multiple networks and thereby missing distributed structure–function relationships. To address this limitation, we present parallel multilink group joint ICA (pmg-jICA), a data-driven framework that fuses gray matter images from sMRI with multiple intrinsic fMRI networks within a single model. pmg-jICA captures cross-network structure–function coupling, preserves subject-specific variability, and enables robust group-level statistical analyses. To demonstrate the approach, we applied pmg-jICA to an Alzheimer’s disease (AD) dataset, recovering linked structural and functional components for 53 brain networks. Notably, patients with AD exhibited alterations in subcortical, cognitive control and visual regions. Importantly, the subject loadings enabled the computation of functional network connectivity, revealing additional alterations in subcortical, visual, and cognitive control systems. Overall, our results demonstrate that pmg-jICA overcomes key limitations of existing multimodal fusion techniques, yielding deeper insights into structure–function disruptions in AD and potentially offering a flexible framework for studying other neurological and psychiatric disorders.
Understanding how the brain gives rise to social cognition has been a key goal of neuroimaging research. Both changes in regional activation as well as functional connectivity have been implicated as potential mechanisms underlying social cognition, but the two have rarely been examined concurrently. Moreover, because the neural processes underlying social cognition are dynamic, developing approaches to capture dynamic changes in regional activity and functional connectivity are critical. Here, we describe a novel analysis approach that captures both regional activity and dynamic functional connectivity simultaneously during a naturalistic, socially focused movie-watching task. We found that both regional activation and functional connectivity were uniquely related to awkwardness, a judgment associated with social faux pas detection and theory of mind. Regional activation within sensorimotor networks was positively associated with awkwardness, whereas activation in the default network was negatively associated. Models including functional connectivity accounted for unique variance beyond models with activity alone. Specifically, dynamic functional connectivity between networks, primarily the frontoparietal control network, was positively associated with awkwardness. Together, these findings suggest that both dynamic regional brain activity and functional connectivity each uniquely contribute to complex and dynamic social judgments.
Abstract Understanding how functional relationships relate to the brain’s structural architecture remains a central challenge in network neuroscience. Many Laplacian-based approaches describe function–structure coupling at the node level, which can make it difficult to identify the specific anatomical pathways that support observed functional relationships. This work introduces a constrained Laplacian formulation that incorporates externally specified pairwise functional relationships and yields a nodal field whose graph-gradient representation produces edge-level quantities describing how the structural network accommodates these relationships. The method is implemented using modified nodal analysis, enabling efficient computation on large connectomes. Given an observed pattern of functional associations and a structural connectivity graph, the proposed framework estimates which structural edges are most consistent with supporting the imposed pattern. The framework is demonstrated in multiple settings, including a single-subject example, a controlled diffusion phantom, an in silico function–structure simulation, and analyses of Human Connectome Project data at both group level (207 subjects) and test–retest conditions (three subjects). Across these applications, the method produces an edge-level representation of function–structure coupling, enabling pathway-specific analysis of brain connectivity.
Abstract Autism spectrum disorder is a pervasive developmental disorder with heterogeneous symptomatology. Currently, subjective evaluations of behavioral symptoms determine diagnosis and treatment, as the neurological interactions that underpin autism remain largely unknown. This study investigates the relationship between effective connectivity (EC) among resting-state networks and autism symptomatology in a large dataset. EC is estimated with spectral dynamic causal modeling for functional magnetic resonance imaging and then related to Autism Diagnostic Observation Schedule scores through parametric empirical Bayes and statistical correlation analyses. Furthermore, EC is used innovatively in machine learning for individual-level prediction. Group-level analyses reveal multiple connections associated with symptom domains, including the newly identified effective connection from the lateral visual network to the posterior default mode network, which is consistently negatively correlated with scores for restricted and repetitive behaviors. Despite the identified group-level associations between EC and symptom domains, EC shows limited generalizability in symptom prediction and heterogeneity in feature importance at the individual level. These findings suggest the presence of subgroups within the spectrum, where informative connectivity patterns vary between individuals.
Abstract Cognition arises from complex, distributed neural processes, which are often studied using fMRI. Most analyses focus on regional activation or (Pearson-based) functional connectivity. However, these measures provide a limited characterization of brain activity. Methods that provide a richer description, including directed and nonlinear relationships, are needed. Information-theoretic measures can be used to quantify such relationships, providing a characterization of information storage and transfer. Here, we propose an approach for estimating statistical measures of information processing: active information storage (AIS), transfer entropy (TE), and net synergy from task-based fMRI. AIS measures information maintained within a region, TE captures directed information transfer, and net synergy contrasts higher-order synergistic to redundant interactions. Crucially, to enable this framework we utilized a recently developed approach for calculating information-theoretic measures: the cross mutual information. This approach combines resting-state and task data to address the challenges of limited sample size, nonstationarity and context in task-based fMRI. We applied this framework to the working memory (N-back) task from the Human Connectome Project (470 participants). Results show that AIS increases in frontoparietal regions with working memory load, TE reveals enhanced directed information transfer across control pathways, and net synergy indicates a global shift to redundancy. This work establishes a novel methodology for quantifying information processing in task-based fMRI.
Constructing structural similarity networks from T1-weighted MRI offers a powerful means to characterize brain organization. Two prominent methods for constructing such networks, Morphometric Similarity Network (MSN) and Morphometric INverse Divergence (MIND), have been proposed. However, a systematic evaluation of the test-retest reliability and age sensitivity of both MIND and MSN is still lacking. The present study comprehensively assessed these properties to inform the reliability and validity of both approaches. Test-retest reliability was evaluated by the intraclass correlation coefficient (ICC) using two public datasets containing repeated MRI scans. Age sensitivity was examined by conducting edge-wise comparisons between younger and older age groups, as well as by training machine learning models to predict individual age, using two public lifespan datasets. Additionally, several practical variants of MIND and MSN were explored by constructing networks with different morphological feature sets. Results demonstrated that MSN exhibited higher test-retest reliability, whereas MIND showed greater age sensitivity when both methods employed the same five features. Both methods revealed distinct spatial patterns that differentiate older from younger adults. Notably, the choice of feature sets substantially influenced reliability and age sensitivity. These findings offer empirical guidance for methodological selection and highlight the importance of feature optimization in future studies.
Abstract To advance motor function assessments, there is a growing need for mechanistic approaches that offer deeper physiological insight. Here, we present a fully end-to-end framework that integrates network science, information theory, and machine-learning to generate targeted biomarkers from large-scale motion data. Showcasing our approach, we perform a comprehensive spatio-spectral decomposition of muscle activations into functionally diverse muscle networks. Then, by incorporating rigorous feature selection and our newly developed clustering algorithm, we identify motor features optimally associated with a chosen clinical measure and cluster participants in a targeted, clinically meaningful way across scales. Framework applications illustrate the mechanistic insights provided into the underlying physiological constructs of any clinical measure, uncovering data-driven population clusters of ageing and poststroke motor impairment chronicity and recovery. This adaptable framework bridges the underutilized large-scale motion data of clinical labs to the assessment tools they currently rely upon, offering in-depth characterizations of individual motor (dis)abilities, representing a powerful new assessment methodology. Future work should aim to establish concrete links between these biomarkers and underlying neurophysiological processes and test data capture protocols to optimize clinical relevance.
External brain stimulation is a promising tool for investigating and altering cognitive processes, with potential clinical applications to the restoration of dysfunctional neural dynamics. In line with experimental observations, we study how the effects of stimulation crucially depend on the ongoing dynamics of the brain, at the local level of the stimulated region but also of global coordinated brain activity. Specifically, we use connectome-based whole-brain computational modeling to explore how the effects of single-pulse stimulation to different regions strongly depend on both the phase of regional oscillatory activity and on the transiently occurring network of functional connectivity at the time of the applied stimulation. Importantly, we show that stimulation has not only state-dependent effects but can also induce global state switching. Lastly, predicting the effect of stimulation by using machine learning shows that functional network-aware measures (i.e., knowledge of either a discrete state of functional connectivity or of a detailed functional connectivity matrix) can increase the performance by up to 40%. Our results suggest that a fine characterization of intrinsic functional connectivity dynamics is essential for improving the reliability of exogenous stimulation.
Epilepsy remains a significant medical challenge, particularly in drug-resistant cases where surgical intervention may be the only viable treatment option. Identifying the epileptogenic zone, the brain region responsible for seizure initiation, is a critical step in surgical planning. Combining dynamical system models, machine learning, and the neuroimaging data of epileptic patients in the so-called Bayesian Virtual Epileptic Patient (VEP) framework has previously been shown to be a promising approach for identifying the epileptogenic zone. However, previous studies employed coupled neural mass models to describe the whole-brain seizure dynamics and, hence, could only provide a highly coarse spatial estimate of the epileptogenic zone. In this study, we propose an extension of the Bayesian VEP to a neural field model, which can improve the spatial resolution by several orders. Performing model inversion using neural field models is a challenging task as the parameter space is very high dimensional, and it becomes computationally expensive to compute gradients. We demonstrate that by using pseudospectral methods and spherical harmonic transforms, it is feasible to perform model inversion on a neural field extension. We found that the high-resolution Bayesian VEP not only improves the spatial resolution but also significantly reduces the number of false positives.
Healthy aging is marked by changes in both cognitive performance and the organization of brain networks. Declines in cognition have been linked to reductions in system segregation (SS), as older adults typically exhibit less segregated functional networks than younger adults. While lower segregation has been associated with diminished cognitive abilities, it remains unclear how individual variability in SS contributes to cognitive outcomes across the lifespan. Here, we examine relationships between SS and three cognitive domains (semantic, executive, episodic memory) using resting-state fMRI data from 179 younger (18-29 years) and 117 older adults (60-89 years). SS was measured globally and for specific networks using Schaefer's 7-network parcellation. Our findings confirmed a global age-related reduction in SS, particularly impacting the somatomotor, ventral attention, and frontoparietal networks. This reduction in global SS mediated negative effects of age group on semantic and executive performance. When examining younger and older groups separately, we found that higher SS was associated with better semantic performance in both groups, while observing a similar positive association with executive performance only in older adults, suggesting that executive function becomes increasingly dependent on preserved network architecture with age. Maintaining SS may therefore be critical for supporting healthy cognitive aging.
People with multiple sclerosis (MS) often present with cognitive deficits that cannot fully be attributed to focal brain alterations. Whole-brain network changes show stronger relations, but MS network insights have mostly focused on either structural or functional (single-layer) networks, while recent work has shown the importance of multilayer frontoparietal network integration for cognition. Here, we explored the cognitive relevance of multilayer integration of the frontoparietal network in relapsing-remitting MS (n = 780) using diffusion and resting-state fMRI. Cognitive relations were first assessed for nodal multilayer eigenvector centrality, averaged over frontoparietal network nodes as a measure of integration, and post hoc for mean eccentricity for both single layer and multilayers. Higher multilayer frontoparietal network centrality was associated with worse Symbol Digit Modalities Test (SDMT) performance (β = -.117, p = .005). Mean eccentricity of single-layer diffusion (β = -.123, p < .001) and multilayer networks (β = .085, p = .018) were associated with SDMT performance. However, results could not be replicated using a different anatomical parcellation. This study showed that cognition in MS is related to multilayer network parameters. Nevertheless, correlations were weak and atlas specific, suggesting that a binary structure-function multilayer network approach is not particularly relevant as a correlate of cognition in MS.
Network models are a key tool in human neuroscience, and translation into animal models is essential for interrogating mechanistic drivers of network organization. Using magnetic resonance imaging (MRI), we present the first in vivo network representation of the individual rat brain, a key animal model in neuroscience. We measured magnetization transfer ratio (MTR) at each of 53 distinct cortical areas and estimated a cortical similarity network for each scan across two independent cohorts. We characterized normative network development in rats scanned repeatedly between postnatal days 20 (weanling) and 290 (mid-adulthood; N = 47) and then contrasted these findings with a cohort exposed to early life stress (repeated maternal separation [RMS]; N = 40). The normative rat cortical similarity network exhibited biologically meaningful organization, consistent with cytoarchitectonic and tract-tracing data, and displayed complex topological features. Developmental analyses revealed increasing interregional similarity during early postnatal and adolescent periods, followed by divergence in mid-adulthood, particularly within fronto-hippocampal systems. RMS disrupted these trajectories, especially between frontal and parahippocampal regions that were also most dynamic during development and aging. These findings introduce a new network-based methodology for studying cortical organization in a model organism, providing a translational framework to understand how environmental risk factors alter brain network development.
Metastability of BOLD fMRI signals is a commonly used proxy of brain dynamics in behavioral and clinical studies. To date, little has been done to assess the confidence with which we can use estimates of metastability as reliable biomarkers of individual brain state. We analyze whole-brain and network-specific metastability for a highly sampled individual brain (84 sessions taken over 18 months) and quantify the within-subject reliability for the metrics as a function of the amount of data used, which we find to be comparable to that seen for static functional connectivity. As considerable variability is observed across networks in the required amount of data, we combine the networks' metrics in one novel feature vector that exhibits an order of magnitude improvement in reliability. We then test reproducibility by analyzing the Midnight Scan Club dataset (10 subjects imaged over 10 consecutive days). Finally, we examine the susceptibility to change of the proposed metastability measure in another dataset examining brain dynamics under the effect of psilocybin. We conclude that the networks' metastability feature vector exhibits strong within-subject reliability that renders it a promising candidate for the study of individual-specific biomarkers of brain dynamics and potential targets for precision neuromodulation.
The cerebellum, with its distinctive architecture and extensive cortical connections, has long been recognized for its highly structured interconnectivity with the cortex and has been proposed as part of a larger circuit that shapes brain network dynamics. Here, we evaluate dynamic network reconfigurations in resting-state fMRI connectivity pre- and post-noninvasive inhibitory repetitive transcranial magnetic stimulation targeting the right Crus I of the cerebellum. Using dynamic community detection to evaluate the stimulation's effect on modular network structures, we characterize the network properties by which cerebellar stimulation spreads through the cortex. We find that: (a) the flexibility, or the likelihood of network nodes to change module allegiances, increased post stimulation; (b) the dynamic patterns by which module allegiances emerged and evolved were highly individual and did not follow a single functional prototype; and (c) the cerebellar nodes had connectivity properties of integrators for distinct network modules. These results are consistent with the idea that cerebellum is pivotal in modulating distributed cortical activity by restructuring the integration and segregation of neural networks. This integrative capacity of the cerebellum may underlie its proposed role in coordinating neural systems, including those supporting higher cognitive function.
