Mapping the white-matter functional connectome: a personal perspective

Abstract

In contemporary neuroscience, mapping the human brain’s functional connectomes is essential to understanding its functional organization. Functional organizations in the brain gray matter have been the subject of previous research, but the functional information in white matter (WM), the other half of the brain, has been relatively underexplored. However, the dynamics of functional magnetic resonance imaging (fMRI) have been reliably identified in the brain WM. This review summarizes current knowledge about task-free (resting-state) fMRI neuroimaging analyses for the WM functional connectome. We present comparative findings of the WM functional connectome, including its mapping, physiological underpinnings, cognitive neuroscience relationships, and clinical applications. Furthermore, we explore the emerging consensus that WM functional networks have valid topological characteristics that can distinguish between individuals with brain diseases and healthy controls, predict general intelligence, and identify inter-subject variabilities. Lastly, we emphasize the need for further studies and the limitations, challenges, and future directions for the WM functional connectome. An overview of these developments could lead to new directions for cognitive neuroscience and clinical neuropsychiatry.

Introduction

The human cerebral cortices are not discrete mosaics. Rather, each cortical region is tightly coupled with many geographically distinct regions to form a macroscale connectome (Sporns et al., 2005; Thiebaut de Schotten and Forkel, 2022). Historically, the structural connectivity connectome has been characterized by sets of neural fibers (anatomical tractography) connecting paired cortical regions (Hagmann et al., 2008). More recent studies of brain similarity networks (for systematic reviews see Wang and He, 2024; Sebenius et al., 2025) have quantified the similarities in morphological measures (architectonic similarity) in gray matter (GM) regions across both features (J. Li et al., 2021b, 2023a, 2024b; Yang et al., 2021; Meng et al., 2022; and subjects (He et al., 2007). The results of these studies have added insight into the basis of structural architectures (Park and Friston, 2013). Moreover, each cortical region can be functionally described by its neural activity and/or activation (Logothetis et al., 2001). Electroencephalography and magnetoencephalography can directly measure the cortical activation or activity (Mushtaq et al., 2024), or functional magentic resonance imaging (fMRI) can indirectly measure these parameters using the hemodynamic blood oxygen level-dependent (BOLD) signal (Poldrack and Farah, 2015). In general, the term “functional connectivity connectome” refers to temporal correlation or coherence between the activation/activity of cortical regions (Achard and Bullmore, 2007). These functional connectomes provide a network perspective on functional repertoires and dynamics (Smith, 2012). The structural/similarity and functional connectomes quantify distinct but biologically related physiological processes. Nonetheless, precise degenerate (many-to-one) structure–function mapping is crucial for understanding the nature of brain connectomes (Suarez et al., 2020; Fotiadis et al., 2024). The functional connectivity certainly involves the brain white matter (WM), that is, the other half of the brain (Fields, 2004, 2013), rather than in the cerebral cortices.

Previously, BOLD–fMRI dynamics were not discernible in the brain WM (Grajauskas et al., 2019) because they were thought to be noisy and unreliable. Paradoxically, the brain WM contains abundant functional information (Fields, 2013). Both theoretical models and experimental evidence have emerged over two decades to support this. First, there is no concrete evidence (Gawryluk et al., 2014), either theoretical or physiological, that BOLD–fMRI activation or activity can ot be measured in WM. Second, neuroimaging evidence from fMRI and diffusion MRI confirms that neural activation in the WM can be identified (Gore et al., 2019). According to initial diffusion MRI studies, external visual stimuli cause a transient decrease in the apparent diffusion coefficient of water in the human brain visual cortex (Darquie et al., 2001; Le Bihan, 2003; Le Bihan et al., 2006), which is linked to neuronal activation. The well-known example of fMRI activation studies in WM, involving the corpus callosum and internal capsule, is supported by growing BOLD–fMRI evidence from the interhemispheric transfer task test, as well as the Poffenberger and Sperry paradigms (Tettamanti et al., 2002; D’Arcy et al., 2006; Gawryluk et al., 2011, 2014; Courtemanche et al., 2018).

In contrast to the above external tasks, task-free (resting-state) fMRI (rs-fMRI), in which participants are instructed to rest, has revolutionized the study of neural architecture (for a historical overview see Biswal and Uddin, 2025). Using this neuroimaging paradigm, we first described how the distribution of BOLD fluctuations in WM exhibited a specific distribution instead of a random distribution of noise at rest (Ji et al., 2017). Almost at the same time, Peer et al. (2017) independently identified resting-state functional networks that were composed of interacting long-distance WM. Furthermore, new insights into the pathogenesis of major psychiatric diseases were revealed by imaging disruption of WM activity (Ji et al., 2023, 2025). Finally, in addition to WM activity, we mapped the correlation relationship between BOLD–fMRI time series, and thus constructed the functional connectome in WM (J. Li et al., 2019). We proposed that WM intrinsic functional organization is compatible with GM functional organization. Therefore, an in vivo method of measuring the functional properties of the human WM at the macroscale level was provided by the rs-fMRI paradigm.

Current evidence supports the existence of reliable BOLD signals not only in the human WM, but also in the human spinal cord. The first spinal-fMRI was conducted in 1996 to map spinal activity during unilateral hand closing (Yoshizawa et al., 1996). Subsequent spinal fMRI investigations have confirmed descending modulation of spinal cord activity during motor and sensory tasks (Stroman et al., 1999; Backes et al., 2001; Madi et al., 2001). Moreover, patterns of neural activity (Wei et al., 2010; Vahdat et al., 2020) and connectivity (Kinany et al., 2020) have been identified across various structures of the neural axis, thereby elucidating the functional architecture of the spinal cord (for a review see Harrison et al., 2021).

In this review, we summarize the current state of knowledge regarding the regional interdependency between the macroscale functions of the human brain WM. We begin by collating studies that map the human WM functional properties. Next, we survey studies reporting molecular attributes and electrophysiological fundamentals of the WM functional connectome. The potential use of WM functional connectome characteristics as effective disease biomarkers is then explored, including studies showing changes in the WM functional connectome in a variety of neurological and psychiatric disorders. In addition, we suggest future areas of study of WM functional connectomes.

Brain WM functional connectome mapping

The goal of neuroscience known as “brain connectome mapping” is to identify and analyze the intricate connections and interactions within the brain’s neural networks (Bassett and Sporns, 2017; Axer and Amunts, 2022; Fotiadis et al., 2024; Wang and He, 2024). Advancements in multimodal and cross-modal neuroimaging have made this field of study more productive (Collin and Whitfield-Gabrieli, 2023; Wheeler and Quintana, 2025). The significance of brain connectome mapping helps us to better understand how the brain works (Pini et al., 2024), identify the biological causes of psychiatric and neurological disorders (Bazinet et al., 2023; Hansen and Misic, 2025), and suggests clinical practices related to diagnoses and treatments (van den Heuvel and Sporns, 2019; Vogel et al., 2023). Hereafter, we focus on the functional connectome, which enables the real-time observation of brain dynamics and connectivity patterns.

Definition of node(s) and edge(s)

The brain is regarded as a network of nodes and edges (Bullmore and Sporns, 2009; Lynn and Bassett, 2019). When discussing brain networks, nodes usually refer to parcellation units within the brain (e.g. regions, vertices, or voxels) (Fig. 1a), whereas edges represent functional connectivities between nodes (Sporns, 2013). The node(s) of WM functional connectomes generally depend on brain parcellations, which are delineations of spatial partitions of the brain (Eickhoff et al., 2018). There are numerous parcellations (Revell et al., 2022), parcels or regions of interest (ROIs), which can differ based on levels (individual versus group) (C. Li et al., 2025), demographics (aged vs. younger) (Han et al., 2018), and modalities (function vs. structure) (Glasser et al., 2016; Schaefer et al., 2018). However, picking a parcellation within which to map WM functional connectomes remains a challenge (Arslan et al., 2018; Turnbull et al., 2025).

Figure 1:

Brain white-matter functional connectome mapping. (a) Node(s) of the functional connectome were defined using the random parcellation of the multimodal white-matter atlas of Zhou et al. (2025). (b) Two primary types of edge(s) of functional connectomes were defined as correlation and causal inferences. (c) Three fundamental topological properties of the white-matter functional connectome were observed across local, meso, and global scales. The white-matter functional connectome hubs are primarily distributed in the posterior thalamic radiation, superior corona radiation, and superior longitudinal fasciculus. Four modules suggest a non-random spatial distribution. The small-world organization indicates a more optimized and efficient parallel information transfer system in the brain. Panel (c) adapted from J. Li et al. (2019), © Wiley 2019, with permission.

Regardless of task-based or task-free fMRI imaging (Finn et al., 2023; Biswal and Uddin, 2025), prior research constructed functional connectivity connectomes within the GM (Liao et al., 2011; J. Li et al., 2024a), which were temporal synchronizations of neural dynamics (e.g. using Pearson’s correlation analysis and Granger causality analysis) between the paired regions (Fig. 1b). The edge(s) of WM functional connectomes were measured by Pearson’s correlation analyses between interregional time series, much like it was for the GM functional network (for a systematically comparative analysis see Liu et al., 2025).

To construct WM functional connectomes, brain parcellations are typically based on either structure (e.g. brain shape and/or where they connect) or function (e.g. brain activity/connectivity) (Table 1). Prior studies have used structural parcellations to map the WM functional connectome using random parcellation (J. Li et al., 2019, 2020a, 2020b) and JHU ICBM-DTI-81 WM atlas (Ding et al., 2018). For functional-based node definition, two types of data-driven methods—K-means clustering (Peer et al., 2017) and independent component analysis (ICA) (Basile et al., 2022; Nozais et al., 2023)—were previously applied in the functioning of WM. The 12 symmetrical WM functional networks, for instance, were found to correlate significantly with canonical GM networks, and corresponded to combinations of WM tracts throughout the brain (Peer et al., 2017). Communication within GM resting-state networks may therefore be greatly aided by WM functional networks (Peer et al., 2017). High resolution in WM functional network investigations, however, is limited by the possibility of wide, over-generalized parcellations produced by data-driven WM functional systems. However, the data-driven WM functional systems may produce broad, over-generalized parcellations, which limit high resolution in WM functional network investigations.

Table 1:

List of the commonly used white matter atlases and their advantages.

White matter atlases	Advantages
JHU ICBM-DTI-81	Structural atlas of 48 WM bundles
Random parcellation	Defining the number of ROIs with approximately identical sizes
ICA components	Spatially independent WM components corresponding to resting-state networks
K-means clusters	Obtaining WM functional networks by identifying clusters of WM voxels with similar connectivity patterns to the rest of the voxels
339 ROIs	Providing a probabilistic resection map of WM with 339 MNI coordinates
MWMA	A WM atlas based on multimodal 7T MRI data, and exhibiting good test–retest reliability

Abbreviations: ICA, independent component analysis; MRI, magnetic resonance imaging; WM, white matter; MWMA, multimodal white matter atlas; ROI, region of interest.

Mapping the cortical and subcortical fibers with direct electrical stimulation (DES) presents a special opportunity to characterize the brain’s functional connections (Duffau, 2015). Another study using this technique involved probabilistic resection mapping of WM from patients who underwent awake surgery for gliomas. The study then co-registered 339 ROIs into the Montreal Neurological Institute space, providing a valuable tool for cognitive neurosciences and clinical applications (Sarubbo et al., 2015). To advance this field of research, we recently used a 7T multimodal MRI dataset from the Human Connectome Project to construct a multimodal WM atlas throughout the whole-brain WM (Zhou et al., 2025). At 200 ROIs, this multimodal WM atlas demonstrated good test–retest reliability (Dice similarity coefficient: 0.93). When compared to the aforementioned JHU atlas and DES map, it was better at displaying WM functional connections at each spatial distance (P_FDR < 0.05). Moreover, when comparing three brain diseased individuals to healthy controls, the multimodal WM atlas demonstrated better classification accuracy than the JHU atlas. Using this multimodal WM atlas, we developed a standardized atlas to better understand the WM functional information of the human brain. This atlas can be used in future clinical research.

Topological properties

As imaging technologies improved, researchers began to adopt more complex analytical techniques, including graph-based methods, used to study brain networks (Rubinov and Sporns, 2010; Bullmore and Sporns, 2012; Avena-Koenigsberger et al., 2017). Once a brain network is constructed, researchers can utilize graph theory metrics to analyze the properties and behaviors of these networks. There are three fundamental topological properties of brain connectomes: hubs, modularity (Seguin et al., 2022), and small-world organization (Bassett and Bullmore, 2006, 2017), across topological scales (Betzel and Bassett, 2017). This mathematical framework facilitates the exploration of complex interactions within the brain, helping to uncover patterns that can aid our understanding of both healthy and dysfunctional brain states (van den Heuvel and Sporns, 2019). Based on longitudinal rs-fMRI datasets, J. Li et al. (2019) used graph theory to show that the WM functional connectome had reliable small-world topology and non-random modularity (Fig. 1c). Additionally, there was a positive correlation between individualized intelligence and the topological property of the WM functional connectome. This study elucidated the functional information in WM and provided additional ways to identify the relationships between brain network topological properties and cognitive traits and clinical diseases.

Physiological basis of the brain WM functional connectome

A crucial step to advancing the field of WM functional connectomes will involve identifying the neurophysiological basis of WM functional networks (Gawryluk et al., 2014) (Fig. 2a). The human brain comprises a multiscale network with multiple levels of organization. Empirical studies of multiple linkages among the molecular (brain-wide gene expression), energy metabolic (fluorodeoxyglucose positron emission tomography, FDG-PET), and macroscale neuroimaging features of brain architecture thus imply that different scales of organization of the human brain are probably not independent of each other, but rather contain several essential multiscale interactions (van den Heuvel et al., 2019).

Figure 2:

Physiological basis and cognitive and clinical applications of the brain white-matter functional connectome. (a) The physiological basis of the brain white-matter functional connectome generally ranged from molecular [Allen Human Brain Atlas (AHBA)], metabolic (FDG–functional PET), to cognitive (Neurosynth) associations. (b) A typical cognitive application uses a predictive model of the brain’s white-matter functional connectome to predict general intelligence (see J. Li et al., 2020a). Three steps are involved: connectome-based predictive modeling (CPM), and internal and external validations. (c) For clinical applications, the brain white-matter connectomes could be used as biomarkers, classifying patients and controls, and predicting the patient’s severity. The disease spectrum involved major depressive disorder (MDD), schizophrenia (SCZ), bipolar disorder (BD), and obsessive–compulsive disorder (OCD). Panel (a) adapted from J. Li et al. (2023b), © Oxford University Press 2023 and J. Li et al. (2021c), © Springer Nature 2021, with permission. Panel (b) adapted from J. Li et al. (2020a), © Springer Nature 2020, with permission. Panel (c) adapted from J. Li et al. (2020b), © Springer Nature 2020, Fan et al. (2020), © Wiley 2020, and Ji et al. (2023), © Springer Nature 2023, with permission.

Molecular association

Multiscale data collations of various open resources are acquired across different individuals; these include the brain-wide cortical gene expression data from the Allen Human Brain Atlas (Hawrylycz et al., 2012) and the density profiles of neurotransmitter receptors and transporters from PET/single-photon emission computed tomography (SPECT) tracer images (Dukart et al., 2021; Hansen et al., 2022; Markello et al., 2022; Hansen and Misic, 2025). To investigate the genetic architecture of WM functional networks, J. Li et al. (2021c) identified intersubject functional variabilities in the WM in longitudinal studies of healthy subjects. The authors found that the functional localization pattern of intersubject variabilities across the WM was heterogeneous, with most variabilities observed in the heteromodal cortex. They next decoded the map of intersubject variabilities in the WM using the whole-brain maps of gene expression using Neurosynth (Rubin et al., 2017; Lombardo et al., 2020). The variabilities of heteromodal regions in gene expression profiles were connected to neuronal cells and linked to mental illnesses. In contrast, genes expressed in the unimodal regions were related to neurological disorders and were mostly expressed in glial cells. Using convergent transcriptomics and cellular markers, the results provided hints on the intersubject variabilities of WM functional networks.

Energy metabolic association

To further characterize the psychological basis of WM functional networks from multiple perspectives, subsequent studies also determined whether the WM functional networks reflect underlying energy metabolism differences (Jamadar et al., 2025). Guo et al. (2022) found a substantial correlation between averaged WM functional connectivity and regional FDG uptake using simultaneous fMRI/PET (Jamadar et al., 2020), suggesting that BOLD signals in WM were directly related to variations in metabolic demands.

According to recent studies on the human brain WM, blood oxygenation and glucose metabolism are related both temporally and spatially (J. Li et al., 2023b). The authors found that BOLD signals and dynamic FDG shared mutual information in the default-mode network (DMN), and visual and sensorimotor–auditory networks on a temporal scale. Regarding spatial distributions, the WM functional networks obtained from BOLD–fMRI data were accompanied by substantial correspondences of FDG functional networks from degree centrality to global gradients. Specifically, eigenmodes of the FDG graph and BOLD fluctuations in the WM DMN were liberally aligned, indicating that metabolic dynamics constrained neurodynamic freedom in the DMN. These findings showed a strong association between WM functional networks and brain energy metabolism.

Electrophysiological association

To determine whether WM functional networks reflected underlying electrophysiological synchronization, Huang et al. (2023) used intracranial stereotactic-electroencephalography (SEEG) and rs-fMRI BOLD data from patients with drug-resistant epilepsy. They found that WM functional networks derived from BOLD were associated with those derived from SEEG. The result was consistent for each subject across a wide range of frequency bands. These findings supported the idea that WM functional networks have an electrophysiological basis. According to other research reports, the WM exhibits functional activation at a resting state (Greene et al., 2020) or is activated under the movement decoding state (G. Li et al., 2021a), but with a much lower amplitude than in GM. Additionally, when the brain WM and GM processes are combined, movement decoding accuracy is greatly improved, in contrast to when GM signals are used alone (G. Li et al., 2021a).

Taken together, these investigations provide a potential neurophysiological basis of brain WM functional networks ranging from gene expression and glucose metabolism to electrophysiology. Multiscale neuroimaging data can be used in future research to further identify the physiological basis of brain WM functional networks, such as distinct PET tracers, or to compare the differences in the neurophysiological basis between WM and GM. This information could confirm the interpretations of such results and increase trust in WM functional network data.

Cognitive neuroscience applications of the brain WM functional connectome

WM tracts mediate the essential connectivity by which human behavior is organized, working in concert with brain GM to enable the extraordinary repertoire of human cognitive capacities. Recent studies have investigated how WM functional networks relate to cognitive traits (Filley and Fields, 2016). To determine whether the WM functional connectome can predict human general fluid intelligence (Fig. 2b), Li et al. included a longitudinal rs-fMRI BOLD dataset of normal volunteers (J. Li et al., 2020a). Then, they constructed a WM functional connectome-based predictive model (CPM) (Shen et al., 2017). The results showed that the constructed predictive model was reliable in internal and external validation analyses, with the high predictive power of the WM functional network existing mainly in the deep frontal WM and ventral frontoparietal tracts. Future research should further examine whether integrating WM and GM functional connectomes can enhance the predictive capacity for advancing cognitive neuroscience.

Clinical applications of the brain WM functional connectome

A number of studies have also identified dysfunctions of WM functional networks in various diseases, including schizophrenia (Jiang et al., 2019a; Fan et al., 2020), major depressive disorder (MDD) (Zhao et al., 2021; Chu et al., 2025; Yu et al., 2025), epilepsy (Jiang et al., 2019b), pontine strokes (Wang et al., 2019), autism spectrum disorder (Chen et al., 2021), and attention-deficit hyperactivity disorder (Bu et al., 2020) (Fig. 2c). Further identification of functional networks within the WM could provide new possibilities for research in cognitive neuroscience and clinical neuropsychiatry.

For the above brain disorder applications, the brain WM connectomes, along with topological properties, could be used as biomarkers to classify patients and controls, and to predict the disease severity. To clarify whether patients with MDD showed disrupted topological properties of WM connectomes, Li and colleagues detected ubiquitous small-worldness organization and local information-processing capacity. The results showed that the MDD individuals had less salience of small-worldness than did the controls, both in discovery and in replication samples (J. Li et al., 2020b), implying a shift toward randomization in MDD WM functional connectomes. Moreover, the topologically based predictive model could predict disease severity in MDD patients, and the constructed classification model based on the discovery sample could be generalized to discriminate MDD patients from healthy controls in the replication sample (accuracy of 76%, sensitivity of 74%, and specificity of 80%). Ji et al. (2019) studied the small-worldness of WM functional networks in patients with Parkinson’s disease vs. healthy controls, where patients showed increased patterns when compared to healthy controls. The features of the WM functional connectome can also be used to classify patients with Parkinson’s disease vs. healthy controls (accuracy of 73%). These studies showed that graphing theoretical approaches in WM functional networks provided powerful tools for identifying biomarkers in brain diseases. This approach could enhance our understanding of the underlying mechanisms and facilitate the future development of targeted therapies based on network characteristics.

Key factors and recommendations for future research

To improve our understanding of WM functional networks, we emphasize four factors in the following section.

Considering harmonics in WM functional networks

One of the main objectives of neuroscience is to understand the intricate spatial structure of brain activity (Pang et al., 2023; Vohryzek et al., 2025). Recently, the concept of functional harmonics has emerged as a powerful framework for representing large-scale brain dynamics (Atasoy et al., 2016; Glomb et al., 2021). Inspired by principles from spectral graph theory and harmonic analysis, functional harmonics provides a set of spatial basis functions derived from the eigenmodes of brain connectivity networks. Typically, the eigenvectors of Laplacian operator of a functional connectivity graph form an orthogonal set of basis functions, known as graph harmonics (Atasoy et al., 2018). These harmonics reflect smooth spatial modes over the network, analogous to classical Fourier modes in Euclidean space, capturing intrinsic patterns of co-activation throughout the brain. They can be used to reconstruct and understand brain activity in a low-dimensional manner (Xia et al., 2024).

By decomposing brain signals into harmonic components, researchers can identify fundamental building blocks of functional organization, revealing how dispersed brain regions coordinate to promote cognition and behavior. One of the main benefits of functional harmonics (eigenmodes) is that they overcome the drawbacks of region-based parcellation techniques, by providing a spatially continuous depiction of brain activity (Glomb et al., 2021). The spectral index of eigenmodes determines how smooth they are; low-frequency harmonics capture global, long-range patterns, whereas higher-frequency modes reflect more localized or modular dynamics. This facilitates a multiscale analysis of brain function, where different cognitive processes or brain states can be interpreted as combinations of harmonics at varying scales (Aqil et al., 2021). This approach combines concepts from physics, signal processing, and network neuroscience to provide new insights into both typical brain functioning and its alterations in neurological and psychiatric disorders. Therefore, the next step is to show how gradient organization and the existence of discrete functional parcels may be explained by harmonic modes of WM functional networks.

Considering neurodevelopmental patterns of WM functional networks across the lifespan

Brain WM significantly grows and undergoes considerable remodeling during development and adolescence (Fields, 2008; Bethlehem et al., 2022). The WM undergoes extensive structural changes that support cognition across development, young adulthood, and during aging (Nazeri et al., 2022; Schilling et al., 2023). Considering spatially and temporally coordinated patterns, Bagautdinova et al. (2023) delineated distinct patterns of spatial covariance across distributed WM locations that supported cognition. The development of brain networks has gained attention in recent years, whereas traditional studies have mainly focused on GM functional (Sun et al., 2025) or structural networks (J. Li et al., 2024b). Given the reliable topological properties in the WM, it is crucial to conduct comprehensive investigations into WM development and its integration within brain networks, to advance our understanding of neurodevelopmental processes and to identify potential developmental abnormalities in brain disorders.

Utilizing the ICA method, Huang et al. (2024). identified WM functional networks in the developing brain. The authors found differences in inter- or intra-WM functional networks between neonates and 8-year-old children, reflecting the functional differentiation and integration in brain development. Focusing on rs-fMRI in full-term and preterm neonates, a previous study found that functional connectivity within the WM network identified by K-means clustering increased with age, with preterm infants exhibiting lower connectivity than full-term infants (Wang et al., 2025). Both studies depicted the age-related changes of WM functional networks in the developing brain. However, these two studies defined WM functional networks based on a data-driven approach. As described previously, future studies may define a set of generalized WM functional networks to compare the results among several studies. In addition, the two studies focused on a specific stage of life, mainly on neonatal and early childhood. Future studies should aggregate multisite fMRI scans to investigate age-related changes of within- or between-WM functional networks across the lifespan, thereby producing reproducible and generalizable brain charts.

Considering the integration of GM and WM functional networks in predictive and classified models

A primary goal of neuroimaging research is to associate brain organization with individual phenotypes. Learning the associations can advance our understanding of brain–behavior relationships that map brain organization to phenotypic measures and clinical applications in brain disorders. Numerous research reports have modeled brain–behavior associations. However, these models tended to be explanatory analyses that often cannot generalize to novel individuals and have inadequate clinical utility (Rosenberg et al., 2018). Predictive modeling approaches that define and validate models with independent datasets provide a solution to this problem. Furthermore, several rules for applying predictive modeling to brain connectivity data have been made (Shen et al., 2017; Scheinost et al., 2019), and many methodological considerations for brain-based predictive modeling in brain disorders have been considered (Dhamala et al., 2023). As the field of neuroimaging and brain connectivity analyses advances (Rosenberg et al., 2016), several promising avenues for future research on predicting and classifying brain disorders using functional connectomes have emerged.

One key area involves the integration of GM and WM functional networks. Previous studies have shown that intelligence is caused by a variety of complex neural mechanisms engaging an interacting functional network of brain GM regions (Jiang et al., 2020) or WM regions (J. Li et al., 2020a). Studies have also demonstrated that brain disorders can be differentiated from healthy controls based on GM functional connectivities (Winter et al., 2024) and WM functional connectivities (J. Li et al., 2020b; Zhou et al., 2025). However, to the best of our knowledge, no study has investigated whether combining GM and WM functional networks can increase the predictive power in brain–behaviour associations or clinical outcomes and classification performances in distinguishing brain disorders. Future studies should therefore focus on refining computational methods to combine functional connectivity data from different brain tissues seamlessly. In addition, considering that there might be an expected delay in fMRI responses from the WM versus GM regions (J. Li et al., 2019), future studies should determine the potential of the hemodynamic response function in combining GM and WM functional connectivity analyses. Finally, most studies used Pearson’s correlations, yielding functional connectivity networks by default. However, numerous pairwise interaction statistics exist in the scientific literature (Liu et al., 2025), which provide various methods that are sensitive to different features of brain organization. When future researchers combine GM and WM functional networks to differentiate individuals or predict individual differences in behavior, they should consider their choice of the functional connectivity method, involving the neurophysiological mechanism they are targeting, as well as their original research objective.

Considering technological challenges in WM functional networks

Although brain WM–fMRI, as well as spinal–fMRI are providing a new window into the central nervous system (Kinany et al., 2023; Stroman et al., 2025), the technical challenges and constraints of spatiotemporal resolution are apparent, such as bulk motion, magnetic field strength, and smoothing kernel. Bulk motion was the largest contributor to the signal variance (Harita and Stroman, 2017). Motion-correction methods and the RETROICOR method have been used to model and remove physiological noise (Barry et al., 2014; Kong et al., 2014). In addition, signal intensity changes with tasks varied in magnetic field strength. However, the reproducibility of low-frequency connectivity from independent groups adds credibility to the existence of functional information in the spinal cord and brain WM (Barry et al., 2016; J. Li et al., 2019). Finally, the proposed anisotropic spatial filtering has greater sensitivity and specificity for detecting slender anisotropic activations, compared to that achieved with conventional Gaussian filters (Abramian et al., 2021). At present, there is no complete solution.

Conclusions

In this review, we have highlighted WM functional networks from the perspectives of physiological, cognitive, and clinical applications, and we have addressed four key considerations and recommendations for the future. The reliable topological properties of WM functional networks allow us to predict general intelligence, capture inter-subject variabilities, and distinguish between normal and diseased brains. The interplay between cross-modal neuroimaging from transcriptomics, electrophysiology, glucose metabolism, and fMRI has advanced our understanding of WM functional networks, enabling us to apply brain functional connectomes to advance precision medicine in brain disorders. Future directions for WM functional network studies include harmonics, heterogeneity across individuals, the contribution to clinical uses and technological challenges. Taken together, the advances presented in this review provide support for a more straightforward approach to mapping WM functional networks.

Author contributions

Jiao Li (Conceptualization, Funding acquisition, Investigation, Project administration, Suppervision, Writing – original draft, Writing – review & editing), Huafu Chen (Funding acquisition, Writing – review & editing), and Wei Liao (Conceptualization, Funding acquisition, Project administration, Writing – original draft, Writing – review & editing)

Conflict of interests

The authors declare no competing interests.