Characteristic spatiotemporal features of large-scale functional network architecture in posttraumatic stress disorder

Abstract

Understanding how trauma reshapes the brain’s large-scale functional architecture requires a framework that integrates both spatial and temporal dimensions. In this study, we applied a multimodal analytic framework that combines functional connectivity gradient mapping with energy landscape analysis. This approach allowed us to characterize cortical reconfiguration in trauma-exposed controls (TEC) and individuals with posttraumatic stress disorder (PTSD). Gradient-based analyses revealed widespread spatial displacements along the first two functional dimensions – particularly within the visual, salience and default mode networks – for both TEC and PTSD participants. PTSD participants further displayed amplified disruptions along a higher-order cognitive gradient, indicating a more pronounced breakdown of large-scale functional organization. We also introduced a novel gradient-based distance metric to more precisely quantify topological alterations within the embedding space. State-based modeling of these network dynamics revealed that trauma exposure reshapes the brain’s coactivation landscape. While both trauma-exposed controls and PTSD participants deviated from normative dynamic profiles, individuals with PTSD exhibited less frequent but more stable anomalous states. This stability may be indicative of a shift away from flexible adaptation towards pathological consolidation. Our findings delineate how trauma reorganizes brain function across complementary facets of spatial embedding and dynamic coordination, highlighting the utility of integrative frameworks for advancing system-level models of trauma-related disorders. Thus, this work offers a conceptual foothold for future investigations into the functional architecture of PTSD.

Introduction

Understanding how the human brain organizes its function across space and time remains a central question in systems neuroscience [1, 2]. Early investigations emphasized localized circuits and modularity, wherein distinct cognitive functions were attributed to segregated units [3]. The expansion of functional magnetic resonance imaging (fMRI), however, shifted this “localizationist” view toward a network-centric framework – where cognitive functions emerged from coordinated interactions across distributed regions [4–6]. Correspondingly, psychiatric neuroscience has evolved from utilizing focal lesion models to investigating network-based theories of dysregulated large-scale systems [7–10]. This transition has provided a foundation for contemporary network approaches to be applied to neuropsychiatric disorders such as depression [11], schizophrenia [12], autism [13], and bipolar disorder [14].

As a major psychiatric disorder triggered by clearly identifiable external events [15], conceptualizing the neurobiological underpinnings of posttraumatic stress disorder (PTSD) is of particular interest. Pioneering neurocircuitry research – epitomized by the widely recognized “fear circuitry” model – has highlighted the involvement of a core aberrant amygdala–hippocampus–prefrontal circuit in those with PTSD [16]. Extending beyond circuit-level interpretations, more recent neurobiological models have emphasized the presence of disrupted large-scale brain networks in PTSD. Drawing on Menon’s triple-network framework [17], Akiki [18] proposed that PTSD exhibits dysregulation across three core systems: the default mode network (DMN), salience network (SN), and central executive network (CEN). Beyond these canonical systems, recent studies have revealed that PTSD is associated with widespread disruptions across sensory and limbic networks [19, 20]. Resting-state fMRI analyses have demonstrated altered intrinsic connectivity within the visual and somatomotor systems, implicating the presence of heightened sensory reactivity and dysregulated embodiment [21, 22]. Disruptions in medial temporal lobe circuits, particularly hippocampal-prefrontal pathways, have been linked to intrusive memory phenomena and impaired contextualization [23, 24]. Collectively, trauma appears to perturb both high-order cognitive and low-level perceptual systems, supporting a systems-level account of global disorganization in PTSD.

Despite increasing recognition of large-scale dysfunction in PTSD, most investigations remain constrained both by a narrow network focus, typically centered on canonical systems, and by analytic metrics that inadequately capture the brain’s continuous and hierarchical functional organization. These limitations underscore the need to establish a more comprehensive model of global functional architecture [25]. One promising approach toward this goal is functional gradient analysis. This technique models the cortex as a continuous, low-dimensional space defined by principal axes of functional variation, reflecting core organizational dimensions of brain function [26–28]. Recent applications have demonstrated its sensitivity to altered cortical hierarchy and diminished functional differentiation in neuropsychiatric disorders such as schizophrenia [29, 30], depression [31], and autism spectrum disorder [32]; this highlights its potential utility for elucidating large-scale network dysfunction in PTSD. While gradient-based methods reveal the spatial organization of large-scale brain networks, they remain limited in capturing how these static configurations give rise to time-varying patterns of network interactions. To address this, recent studies have increasingly adopted dynamic modeling approaches such as dynamic functional connectivity (dFC) [33], hidden Markov models (HMM) [34], and coactivation pattern analysis (CAPs) [35] to investigate transient brain states. Among these, energy landscape analysis (ELA) offers a conceptually tractable framework by focusing on large-scale networks [36–38]. Through this, ELA identifies discrete, recurring coactivation states and characterizes the transition structure between them.

Building on these theoretical and methodological foundations, the present study aims to integrate spatial and temporal perspectives to characterize global network reconfiguration in PTSD. By combining functional gradient mapping and energy landscape analysis, we seek to examine how trauma-related disruptions affect the brain’s macroscale architecture – capturing both fundamental functional dimensions and representative organizational patterns (Fig. 1). We hypothesize that PTSD is associated with widespread disruptions of cortical hierarchy, extending beyond canonical control–salience–default mode networks to include sensory and limbic systems. At the organizational level, we expect a fragmentation and reconfiguration of large-scale coactivation patterns, reflecting destabilized transitions between functional states.

Fig. 1. Schematic overview of the analytic pipeline.

a The cortex was parcellated into 1000 regions using the Schaefer atlas. fMRI time series were extracted from each parcel, and mean time courses were computed for seven canonical networks (DMN, FPN, LIM, SAN, DAN, SMN, VIS). b Gradient analysis workflow. Parcel-wise FC matrices (Fisher-z) were converted to cosine-similarity affinity matrices and embedded with diffusion maps to derive cortical gradients. Gradient 1 delineates a transmodal hierarchy from unimodal sensory regions to integrative association cortices. Gradient 2 captures sensory segregation, distinguishing visual areas from auditory and somatomotor systems. Gradient 3 reflects a control–default mode axis, separating frontoparietal control regions from default mode network regions. c Energy landscape analysis workflow. Network time series were binarized to construct brain states based on coactivation patterns across time. Pairwise maximum entropy modeling (pMEM) was then used to estimate the energy landscape underlying recurrent state transitions. The resulting landscape topology characterized brain states by their appearance probability and stability. d Gradient-based network centroids were embedded into a 3D coordinate system, and a novel Earth Mover’s Distance (EMD) metric was applied to quantify within- and between-network displacements across groups. VIS visual network, SMN somatomotor network, DAN dorsal attention network, SAN ventral attention/salience network, LIM limbic network, FPN frontoparietal network, DMN default-mode network.

Materials and methods

Participants and MRI procedures

Comprehensive details on participant recruitment and eligibility, clinical assessments, MRI acquisition parameters, and preprocessing workflows are provided in the Supplementary Methods, with demographic information provided in Supplementary Table S1.

A naturalistic trauma-exposed cohort was established following a Category 5 typhoon that struck a coastal region of Hainan Province in July 2014. From this event, we recruited 70 typhoon-exposed participants from the most severely affected area and 32 healthy controls (HC) from a neighboring city. Typhoon-exposed individuals were initially screened using the PTSD Checklist–Civilian Version (PCL) [39], and those above the screening threshold underwent a structured clinical interview (SCID-IV) [40] to confirm DSM-IV diagnoses. This process identified 36 individuals with PTSD and 34 trauma-exposed controls (TEC). PTSD symptom severity was further assessed using the Clinician-Administered PTSD Scale (CAPS) [41].

Participants were excluded if they were younger than 18 or older than 65 years old, had major neurological or medical conditions, a history of head injury or loss of consciousness, or were left-handed. Additional exclusion criteria included current or lifetime psychiatric disorders other than depression or anxiety, psychotropic medication use, alcohol/substance abuse, and any MRI contraindications (e.g., pregnancy, claustrophobia, ferromagnetic implants). After prespecified exclusions and imaging-quality checks, the final sample comprised 27 individuals with PTSD, 33 TEC, and 30 HC. Exclusions were due to incomplete imaging data (PTSD, n = 3), cerebral infarction (PTSD, n = 1; HC, n = 2), denture-related artifacts (PTSD, n = 2), pregnancy (PTSD, n = 1), and excessive head motion (PTSD, n = 2; TEC, n = 1). This study was conducted in full accordance with the Declaration of Helsinki and approved by the Ethics Committees of the Second Xiangya Hospital of Central South University and Hainan General Hospital (Number 20140306). All participants provided written informed consent after receiving a full explanation of the study procedures.

MRI data were acquired using a 3.0 Tesla MRI scanner and initial preprocessing was conducted with fMRIPrep version 20.2.1 [42]. Subsequent preprocessing was performed using Nilearn [43], including the regression of nuisance signals (head-motion parameters, white-matter and cerebrospinal fluid signals, global signal, and anatomical/temporal CompCor components), scrubbing of volumes with Jenkinson framewise displacement > 2 mm, temporal band-pass filtering at 0.01–0.1 Hz, and surface-based spatial smoothing with a 6 mm full width at half maximum (FWHM) kernel. The resulting preprocessed and smoothed surface-based fMRI data were used for region-level analyses in the cortical space.

Construction of cortical functional gradients

Cortical functional connectivity gradients were derived using the BrainSpace toolbox [44] implemented in MATLAB R2022b. Resting-state time series were extracted for each participant from 1000 cortical parcels defined by the Schaefer atlas [45].

For each subject, a functional connectivity matrix was obtained by computing parcel-wise Pearson correlation coefficients and applying Fisher’s r-to-z transformation. To emphasize the most informative connections, the matrices were thresholded to retain the top 10% of values. Cosine similarity between the connectivity profiles of parcels was then calculated to generate an affinity matrix, which served as the input for the diffusion map embedding. This nonlinear manifold learning technique reduces high-dimensional connectivity structures into low-dimensional representations, capturing principal axes of cortical organization. Group-average functional connectivity matrices were generated separately for the HC, TEC, and PTSD groups, and group-level gradient templates were derived accordingly. Individual gradients were aligned to the corresponding group template using Procrustes rotation. As a sensitivity analysis, we additionally aligned all participants to a single fixed HC-derived template to verify that the resulting gradient structure was not driven by the group-template alignment strategy. The corresponding results are shown in Supplementary Figs. 3–4.

Quantifying functional network geometry in gradient space

To enhance the comparability of gradient dimensions across groups, each subject’s gradients were scaled by the explained variance of the corresponding reference gradients. These were derived from group-level gradient templates to which individual gradients had been aligned. To evaluate the distances between functional networks in gradient space, we adopted the Earth Mover’s Distance (EMD), also known as the Wasserstein distance. This measure quantifies the minimal effort required to transform one spatial distribution into another, making it well-suited for comparing structured point sets. While distances are commonly estimated by computing the Euclidean distance between network centroids, this approach reduces each network to a single representative point and ignores spatial characteristics such as shape, extent, and dispersion. In contrast, EMD incorporates the entire distribution of parcels within a network, providing a more sensitive and geometrically informative comparison across groups.

We computed the EMD in gradient space as follows:

For each functional network, its spatial distribution in each clinical individual was compared to that of the corresponding network in a group-level reference gradient template derived from healthy controls. Each network was represented as a set of parcels in gradient space. Let $N_{\text{ind}}=\{r_1^{\text{ind}},\dots ,r_n^{\text{ind}}\}$ and $N_{\text{ref}}=\{r_1^{\text{ref}},\dots ,r_n^{\text{ref}}\}$, where $r_i^{\text{ind}},r_j^{\text{ref}}\in R^d,d\in \left\{\mathrm{1,2,3}\right\}$ represent the gradient coordinates of the i-th and j-th parcels from the individual and reference network in d-dimensional gradient space, respectively. A cost matrix M $\in R^{n\times n}$ was constructed by computing the pairwise Euclidean distances between parcels:

$$M_{\text{ij}}=\mathit{||}{r}_i^{\text{ind}}-r_j^{\text{ref}}|{|}_2$$

The EMD between the two spatial distributions was then defined as the optimal transport cost that minimizes the total effort required to transform one network distribution into the other:

$$\text{EMD}\left(N_{\text{ind},}{N}_{\text{ref}}\right)=\mathrm{mi}{\mathrm{n}}_{\left(\mathrm{\gamma }\in \mathrm{\Gamma }\left({\mathrm{N}}_{\text{ind}},{\mathrm{N}}_{\text{ref}}\right)\right)}⟨\mathrm{\gamma },\mathrm{M}⟩$$

where $\text{EMD}\left(N_{\text{ind}},N_{\text{ref}}\right)$, defined as network displacement, quantifies the minimal transport cost between the distributions of network $N_{\text{ind}}$ and reference network $N_{\text{ref}}$; ${\mathrm{\Gamma }}_{N_{\text{ind}},N_{\text{ref}}}$ is the set of valid transport plans, and the Frobenius inner product $\mathrm{\langle }\mathrm{\gamma },M\mathrm{\rangle }$ denotes the total transport cost.

$$⟨\mathrm{\gamma },M⟩=\sum _{i,j}{\mathrm{\gamma }}_{\mathit{ij}}{M}_{\mathit{ij}}$$

Here, ${\mathrm{\gamma }}_{\mathit{ij}}$ represents the amount of mass transported from the i-th parcel in the individual network to the j-th parcel in the reference network.

Between-network distances were computed at the individual level by comparing the spatial distributions of each pair of functional networks. Given two networks $N_i$ and $N_j$, we defined a cost matrix M $\in R^{n\times m}$. Due to the unequal number of parcels across networks, a partial matching strategy was applied, matching only the total mass of the smaller network. The Earth Mover’s Distance (EMD) was defined as:

$$\text{EMD}\left({\mathrm{N}}_{\text{i}},{\mathrm{N}}_j\right)=\mathrm{mi}{\mathrm{n}}_{\left(\mathrm{\gamma }\in {\mathrm{\Gamma }}_{\left(N_{\text{i}},N_{\text{j}};\mathrm{\tau }\right)}\right)}⟨\mathrm{\gamma },\mathrm{M}⟩$$

where $\mathrm{\tau }=\min \left(n,m\right)$, and $\mathrm{\Gamma }\left(N_i,N_j;\mathrm{\tau }\right)$ denotes the set of admissible partial transport plans such that the total transported mass equals $\mathrm{\tau }$.

Both full and partial EMD computations were performed using the Python Optimal Transport library, with ot.emd2 [46] and ot.partial.partial_wasserstein [47, 48] applied, respectively.

In addition to EMD-based metrics, we computed commonly used geometric measures, following that of previous studies [49, 50], including network range, within-network dispersion, and between-network dispersion. Moreover, matched supervised classification analyses using Random Forest models (scikit-learn) [51] were performed to evaluate and compare the discriminative capacity of the EMD- and Euclidean-based between-network representations. Further methodological details and corresponding results are presented in the Supplementary Materials.

As a supplementary measure, pairwise Euclidean distances between parcels were also computed to capture fine-scale spatial relationships within and across networks.

Characterizing functional network dynamics in the temporal domain

The spatial organization of functional networks in gradient space provides a basis for understanding their temporal interaction patterns, which can be viewed as a dynamic counterpart of this intrinsic architecture. Following previous studies [37, 38, 52, 53], we applied the pairwise maximum entropy model (pMEM) via the ELAT toolbox to characterize this temporal dimension. Each individual’s time series were binarized using the mean signal of each network as the threshold, resulting in binary activation patterns ${\sigma }_i^t\in \{-1,+1\}$ across $N=7$ networks. At each time point $t$, the brain is in one of $2^N$ = 128 possible activation patterns $V^t={\left[{\mathrm{\sigma }}_1^t,\dots ,{\mathrm{\sigma }}_N^t\right]}^{\text{T}}$. These patterns are indexed as $V_k$, and the probability of observing pattern $V_k$ is defined by the Boltzmann distribution:

$$P\left(V_k\right)=e^{-E\left(V_k\right)}/\sum _{i=1}^{2^N}{e}^{-E\left(V_i\right)}$$

where the energy of pattern $V_k$ is given by:

$$E\left(V_k\right)=-\sum _{i=1}^N{h}_i{\sigma }_i\left(V_k\right)-\frac{1}{2}\sum _{i\ne j}{J}_{\mathit{ij}}{\sigma }_i\left(V_k\right){\sigma }_j\left(V_k\right)$$

Here, $h_i$ denotes the individual bias of network i (its intrinsic activation tendency when isolated), and $J_{\mathit{ij}}$ represents the pairwise interaction between networks $i$ and $j$. These parameters were optimized via maximum likelihood using gradient ascent to match the empirical mean activity $⟨{\sigma }_i⟩=\frac{1}{T}{\sum }_{t=1}^T{\sigma }_i^t$ and co-activation $⟨{\sigma }_i{\sigma }_j⟩=\frac{1}{T}{\sum }_{t=1}^T{\sigma }_i^t{\sigma }_j^t.$

Model-predicted expectations were then computed as:

$${⟨{\sigma }_i⟩}_{\text{model}}=\sum _{k=1}^{2^N}{\sigma }_i\left(V_k\right)P\left(V_k\right),{⟨{\sigma }_i{\sigma }_j⟩}_{\text{model}}=\sum _{k=1}^{2^N}{\sigma }_i\left(V_k\right){\sigma }_j\left(V_k\right)P\left(V_k\right)$$

Model fit was evaluated using the relative reduction in Kullback–Leibler divergence between first-order and second-order models: $r_D=\frac{D_1-D_2}{D_1}$, and the Pearson correlation between empirical and model-predicted probabilities. In addition, to assess whether these fit indices generalized beyond the fitted data, we conducted a repeated random hold-out validation analysis. Detailed procedures and results are reported in the Supplementary Materials.

Based on the fitted model, we constructed an energy landscape to characterize the brain’s state-space structure and dynamics. A local minimum was defined as a brain state whose energy is lower than that of all its adjacent states – i.e., states that differ from it by the activity of only one network. Starting from each of the $2^N$ activity patterns, we iteratively moved to the neighboring state with the lowest energy until convergence to a local minimum. The set of states that converged to the same minimum defined a basin of attraction, allowing the reduction of high-dimensional brain activity into a small number of dominant attractor states.

We quantified each basin’s frequency as the empirical occurrence of those patterns in the time series. To analyze transition dynamics, we first computed the basin transition probability, which quantifies the likelihood of switching between distinct attractor basins. For any two basins, this was defined as the number of transitions between them at adjacent time points divided by the total number of transitions originating from the source basin [54].

While basin transition probability captures how often transitions occur, it does not reflect the extent of transitions – that is, whether transitions occur between local minimums, or merely wander through the peripheral regions of basins. To further characterize transition efficiency, we computed the efficiency score between pairs of basins, defined as the ratio of direct transitions between basin minima to those occurring between their peripheral patterns. As introduced by Ezaki [53], this measure reflects the stability and directness of transitions within a dynamical configuration, rather than its mere frequency of occurrence.

Statistical analysis

Our statistical analyses employed frequentist and Bayesian frameworks implemented within MATLAB, R, and JASP. We first performed classical statistical analyses to evaluate group differences across spatial and dynamic metrics. One-way ANOVA was used to assess group differences in gradient values across the three groups (HC, TEC, PTSD) for each of the three gradient dimensions at 1000 cortical parcels. For distance-based metrics, classical Student’s independent t-tests were conducted to compare the network displacement of each functional network in the clinical groups (TEC, PTSD) relative to the corresponding network in the healthy control reference template. ANOVA was also used to evaluate group differences in between-network distances and pairwise parcel distances. For energy landscape metrics, independent t-tests were applied to compare state frequency, basin transition probability, and efficiency score between TEC and PTSD groups, with ANOVA additionally applied to the efficiency score. To control for multiple comparisons, p-values were corrected using the Benjamini–Hochberg false discovery rate (FDR) method with a significance threshold of α = 0.05. Corrections were applied separately within each analysis family: across parcel-wise comparisons for parcel-level analyses, across each gradient-space family (G1, G2, G3, and G123) for network-level gradient metrics, and across the corresponding state-frequency and transition-probability tests for ELA analyses. We additionally conducted a permutation-based max-statistic FWE correction (5000 permutations) as a robustness analysis (Supplementary Fig. 5).

To provide a more nuanced interpretation of group differences beyond conventional p-value thresholds, Bayesian independent-sample t-tests and ANOVA were conducted using JASP (version 0.16.3; https://jasp-stats.org/) under default settings. A Cauchy prior distribution with a scale parameter of 0.707 was used for effect size estimation in the t-tests, and a scale of 0.5 was used for fixed effects in the ANOVA [55]. Results were summarized using Bayes factors (BF10), quantifying the likelihood of the data under the alternative hypothesis (H1: group difference) relative to the null hypothesis (H0: no difference). Interpretive guidelines were followed (e.g., BF10 > 3 indicates moderate evidence; BF10 > 10 indicates strong evidence), and posterior distributions were reported with median effect sizes and 95% credible intervals (CIs). Bayesian analyses were treated as complementary evidence, particularly for lower-dimensional hypothesis sets.

To investigate spatiotemporal coupling, we examined whether network displacement and between-network distance were associated with the empirical frequency of brain states. These analyses aimed to assess how deviations in spatial organization relate to the temporal prominence of brain dynamics. Additionally, we tested whether these spatial metrics were correlated with PTSD symptom severity, as measured by total CAPS scores and its subscales. Based on Shapiro–Wilk tests for normality and dependent on the distributional properties of each variable pair, classical correlation (Pearson or Kendall’s τ) and Bayesian correlation analyses were used.

Results

Fundamental features along gradient dimensions

Considering the curse of dimensionality inherent in high-dimensional gradient combinations – a feature that increases complexity and limits interpretability, we chose to focus only on the first three gradients. Of note, these gradients collectively accounted for 55.0% ± 3.5% of the variance in functional connectivity. Moreover, these gradients defined consistent macroscale organizational axes across the HC, TEC, and PTSD groups, providing a shared framework and principled foundation for subsequent spatial analyses (Fig. 2a and Fig. 2b).

Fig. 2. Gradient embedding and distribution across the three groups.

a Cortical gradient patterns visualized across the 2D and 3D combinations of gradient dimensions in healthy controls (HC), trauma-exposed controls (TEC), and individuals with posttraumatic stress disorder (PTSD). Each dot represents a cortical parcel, colored by its corresponding functional network. b One-dimensional gradient distributions and surface maps for each group, highlighting the spatial embedding and network-level topography along each gradient dimension. The cortical network visualizations correspond to the gradient templates of each group and were generated using the netplotbrain package (https://www.netplotbrain.org/) [107].

In accordance with prior work [28, 56–58], Gradient 1 delineated a principal axis extending from unimodal regions, including primary visual and somatosensory/motor cortices, to transmodal association areas such as the medial prefrontal cortex, posterior cingulate cortex, and inferior parietal lobule. This pattern aligns with the canonical hierarchy of cortical organization, ranging from externally oriented sensory processing to internally oriented integrative functions [27, 59]. In contrast to the transmodal hierarchy captured by Gradient 1, Gradient 2 revealed a sensory differentiation axis, primarily contrasting visual areas with auditory and somatomotor territories, suggesting segregation within the unimodal cortex based on sensory modality. Beyond sensory-based distinctions, Gradient 3 described a control–default mode axis, separating frontoparietal regions involved in cognitive control (such as the lateral prefrontal and inferior parietal cortices) from default mode regions (including the medial prefrontal cortex and posterior cingulate cortex). This distribution suggests a functional divergence between networks supporting goal-directed cognitive operations [60] and those involved in self-referential or internally directed processes.

To investigate the distributional characteristics of the gradient dimensions, we assessed several fundamental features of the first three gradients (Supplementary Data 1). The explained variance did not differ significantly, confirming that the three axes captured comparable proportions of connectivity variance. We then examined the range of each gradient dimension. Significant group differences were found in Gradient 1 and Gradient 3. Post hoc tests showed that the range of Gradient 1 was significantly lower in HC compared to both TEC and PTSD, while the range of Gradient 3 was reduced in PTSD relative to HC and TEC. The network range showed minimal differences between groups, suggesting that the global effects did not originate from an expansion confined to a single network. We next assessed global dispersion and within-network dispersion. Neither measure showed significant group differences, indicating that the spatial configuration of parcels in cortical embedding remained largely uniform across groups.

As a supplementary analysis, we examined group differences in parcel-level gradient values to capture more fine-grained variation in local gradient distributions (Fig. 3a and Fig. 3b). In Gradient 1 and Gradient 2, most differences were observed between HC and the clinical groups (TEC and PTSD), whereas Gradient 3 differences were predominantly associated with PTSD relative to both HC and TEC. In Gradient 1, parcels within the visual network (VIS) showed positive shifts in TEC (90.1%) and PTSD (95.7%), while the salience/ventral attention network (SAN/VAN) exhibited negative shifts (TEC: 66.9%; PTSD: 73.6%). In Gradient 2, the default mode network (DMN) shifted negatively (TEC: 81.1%; PTSD: 76.9%), while SAN (TEC: 50.4%; PTSD: 69.4%) and the somatomotor network (SMN) (TEC: 54.1%; PTSD: 33.0%) shifted positively. In Gradient 3, although network-level shifts were less pronounced, group differences were observed across a broader set of networks, including SMN, the dorsal attention network (DAN), SAN, the frontoparietal network (FPN), and DMN. Notably, the limbic network (LIM) remained relatively stable across all three gradient dimensions.

Fig. 3. Quantification of gradient metrics.

a Parcel counts exhibiting significant group differences (FDR < 0.05) within each large-scale functional network across Gradients 1, 2, and 3. Compared to healthy controls, both trauma-exposed controls and PTSD participants exhibited shared displacements along Gradients 1 and 2, whereas the latter exhibited broader disruptions along Gradient 3, with wide-ranging deviations across distributed networks. b Spatial maps of significant between-group differences projected onto the cortical surface, showing t-values (HC vs TEC, HC vs PTSD, TEC vs PTSD) per gradient dimension. c Violin plots showing network-wise displacement along each gradient dimension and across the joint embedding space (Gradients 1 & 2 & 3). Displacement was calculated relative to the HC’s template. d Detailed comparisons of between-network distances across groups in gradient space. * p_fdr < 0.05, ** p_fdr < 0.01, *** p_fdr < 0.001. e Radar plots illustrating participant-wise between-network distances across adopted combinations of gradient dimensions. f Visualization of pairwise Euclidean distances among cortical parcels within the joint 3D gradient embedding, capturing fine-scale topological organization both within and between functional networks. Grid lines indicate large-scale network boundaries. The heatmap in the upper-right corner summarizes the number of significantly altered inter-parcel distances between pairs of networks.

Together, these findings indicate that trauma exposure is associated with shared deviations along Gradient 1 and Gradient 2, particularly affecting the positioning of unimodal and transmodal networks. In contrast, PTSD-specific differences may preferentially manifest within Gradient 3, reflecting a more complex disruption in the balance between control-related and internally oriented systems.

Quantifying functional architecture in gradient space

Distance metrics were interpreted in one-dimensional (i.e., Gradient 1, 2, and 3 individually) and three-dimensional (i.e., Gradients 1–3 jointly) gradient spaces. Results from the two-dimensional combinations (i.e., 1 & 2, 1 & 3, 2 & 3) are not discussed here due to the difficulty in drawing distinct interpretations from these intermediate representations.

Network displacement analyses revealed that networks in the PTSD group exhibited greater deviations from their corresponding reference distributions (Fig. 3c, Supplementary Data 2). Along Gradient 1, this pattern was particularly evident in VIS, DAN, and SAN, all showing significantly larger displacements in PTSD. In Gradient 2, elevated displacements were observed in SAN and DMN in the PTSD group. For Gradient 3, SMN, DAN, and FPN exhibited greater shifts in PTSD. When considering the full three-dimensional gradient space, consistent group differences re-emerged in VIS, DAN, and SAN, further underscoring the pronounced reorganization of these networks in PTSD.

Building on the observed network-level shifts, we assessed between-network distances to further characterize large-scale architectural changes in gradient space (Fig. 3d and Fig. 3e; Supplementary Data 3). In Gradient 1, significant differences were mainly found in distances between VIS or SAN and other networks. Gradient 2 exhibited altered distances primarily for network pairs involving DMN and SAN. Across the first two gradients, group differences were most evident between HC and the clinical groups. Gradient 3 revealed a more extensive reconfiguration whereby network pairs with significant differences encompassed all canonical networks. Unlike Gradients 1 and 2, these effects were driven by PTSD-related deviations, as distances between HC and TEC remained largely comparable. These findings are consistent with the parcel-level analyses and thus further underscore that trauma exposure – especially in PTSD – is associated with widespread alterations in functional embedding.

In the three-dimensional gradient space, group differences were confined to a smaller subset of network pairs, specifically SMN–DAN, SMN–SAN, SAN–FPN, SAN–DMN, and LIM–FPN. We additionally computed pairwise Euclidean distances between parcels in gradient space, and the resulting network-level summaries closely mirrored the main between-network distance findings (Fig. 3f). This suggests that while substantial reconfiguration occurs along individual axes, large-scale shifts in the joint embedding space are relatively restrained.

Temporal architecture of large-scale brain networks

Next, we employed energy landscape analysis (ELA) to capture the temporal organization and dynamic interactions of large-scale brain networks (Fig. 4a). The pairwise maximum entropy model (pMEM) showed excellent fit across all three groups (Accuracy > 96.9%, r > 0.989, p < 1 × 10⁻⁵). The energy landscapes of PTSD and TEC exhibited identical hierarchies, characterized by the same four stable brain activity patterns (Fig. 4b). Two of these patterns corresponded to the canonical states observed in HC (State 1 and State 4), whereas the remaining two patterns (State 2 and State 3) represented abnormal, non-canonical network activation patterns identified in TEC and PTSD. Notably, the canonical and abnormal states showed complementary activation configurations.

Fig. 4. Energy landscape analysis of functional brain states.

a Energy landscapes separately visualized for HC, TEC, and PTSD groups. Each basin represents a stable brain state, with its depth corresponding to relative energy. The spatial layout of patterns (x and y axes) is arbitrary and derived from a directed acyclic graph structure used for landscape construction. For improved visualization, a smooth surface was fitted over the discrete energy states, and outlier states with disproportionately high energy were rescaled to zero. b Each stable brain state is represented by a local minimum, visualized through its spatial distribution on the cortical surface and accompanying bar plots that depict dominant network coactivation patterns. Violin plots summarize the frequency of each state across clinical groups, given that the HC group was characterized only by States 1 and 4. c Group comparisons of basin transition probabilities across stable brain states. d Efficiency scores quantifying the degree of transition between state pairs; higher values indicate more stable and accessible dynamic interactions. TEC showed stronger transitions between States 1 and 4, while PTSD participants exhibited greater efficiency between States 2 and 3, suggesting stabilization of an aberrant interaction pattern.

As TEC and PTSD exhibited the same set of local minima, we compared the morphological and dynamic properties of their energy landscapes. We first examined state frequency to evaluate the relative prominence of each stable state. State 1 (t ₍₅₈₎ = 12.409, p_unc = 5.772 × 10⁻¹⁸, BF_10,U = 5.611×10¹⁴, 3.144, [CI: 2.484, 3.893]) and State 4 (t ₍₅₈₎ = 5.403, p_unc = 1.282 × 10⁻⁶, BF_10,U = 10640.105, 1.308, [CI: 0.732, 1.889]) showed increased prevalence in the PTSD group, whereas the appearances of State 2 (t ₍₅₈₎ = 15.078, p_unc = 9.991 × 10⁻²², BF_10,U = 2.462×10¹⁸, 3.821, [CI: 2.996, 4.692]) and State 3 (t ₍₅₈₎ = 5.011, p_unc = 5.394 × 10⁻⁶, BF_10,U = 2919.184, 1.205, [CI: 0.639, 1.777]) were notably diminished.

To probe the dynamics within the energy landscape, we evaluated basin transition probability and efficiency score. Of the 16 possible basin transitions, those ending in State 1 or 4 exhibited higher probabilities in PTSD, while those ending in State 2 or 3 were more likely in TEC – indicating a global shift in transition preferences between the two groups (Fig. 4c; Supplementary Data 4). However, the efficiency score between States 1 and 4 was significantly higher in the TEC group (F_(2,87) = 6.284, p = 0.003, partial η² = 0.126, BF₁₀ = 13.624) compared to both HC (t₍₈₇₎ = 3.364, p_bonf = 0.003, BF_10,U = 11.213, 0.681, [CI: 0.186, 1.196]) and PTSD (t₍₈₇₎ = 2.548, p_bonf = 0.038, BF_10,U = 2.074, 0.499, [CI: 0.014, 1.012]), reflecting greater extent of transition between these two canonical states. In contrast, for transitions between States 2 and 3, the PTSD group exhibited significantly higher efficiency than the TEC group (t ₍₅₈₎ = 2.261, p = 0.027, BF₁₀ = 2.153, 0.504, [CI: 0.018, 1.017]), suggesting more efficient switching between these abnormal states (Fig. 4d).

Interrelations among functional spatiotemporal reorganization and PTSD symptomatology

To investigate how functional spatiotemporal reorganization relates to clinical symptomatology, we examined the associations among gradient-based spatial metrics, energy landscape-derived temporal features, and PTSD severity scores (Fig. 5a). In both clinical groups, the extent of VIS displacement in gradient space was strongly associated with the occurrence of states characterized by VIS activation: as displacement increased, the “normal” State 1 occurred less frequently, whereas the “abnormal” State 3 emerged more often. This trend extended to VIS–LIM spacing, where reduced distance was linked to the heightened presence of State 3.

Fig. 5. Correlations among spatiotemporal metrics and their associations with PTSD symptom severity.

a Correlation analyses examining interrelations between gradient-derived spatial metrics and energy landscape-derived dynamic measures. Visual network reconfiguration was significantly associated with the occupancy of State 3, while frontoparietal alterations were linked to increased engagement in State 2. Furthermore, state preference was significantly related to gradient-based metrics, suggesting aligned patterns of static and dynamic reorganization. b Distributions and group comparisons of key gradient-derived metrics (e.g., VIS displacement, FPN–DMN distance) illustrate progressive disruptions across HC, TEC, and PTSD, highlighting how gradient-based reconfiguration serves as a functional scaffold for the development of anomalous brain states in PTSD and is further associated with symptom severity.

We previously hypothesized that PTSD-specific functional architecture may remain latent in Gradient 3. Notably, it was along this dimension that we observed a significant negative association between the extent of FPN displacement, and the frequency of abnormal brain states characterized by FPN activation. In greater detail, the spacing between FPN and DMN reliably predicted the occurrence frequency of every representative state within the PTSD group.

Given that VIS–LIM and FPN–DMN spacing may underlie the atypical patterns of functional organization observed in the clinical groups, we next examined their associations with PTSD symptom severity as indexed by CAPS scores. Associations between network spacing and symptom severity were modest. As illustrated in Fig. 5a, shorter VIS–LIM distances were significantly associated with greater PTSD symptom severity, while the association with FPN–DMN spacing was weaker. Full results, including subscale-level statistics, are reported in the Supplementary Data 5, providing a more comprehensive view of symptom-specific associations.

Discussion

By jointly considering spatial gradient topology and temporal state dynamics, we explored the impact of trauma on the functional architecture of large-scale brain networks. Spatially, TEC and PTSD groups demonstrated convergent shifts along Gradient 1 and Gradient 2, particularly affecting the sensory and default mode networks. In contrast, PTSD showed unique alterations along Gradient 3, suggesting disrupted segregation between control and default systems. Temporally, both clinical groups expressed two additional abnormal brain states, interpreted as variants of the normative state architecture in healthy controls. Notably, PTSD exhibited enhanced transition efficiency between these anomalous states, whereas TEC displayed increased dwell frequency. Spatiotemporal interactions further revealed strong associations between specific gradient displacements (e.g., VIS and FPN) and the prevalence of corresponding brain states, with modest links to PTSD symptom severity. Together, these findings suggest that trauma induces a shared global reorganization of functional architecture, while PTSD-specific disruptions are anchored in Gradient 3 and manifest as a more consolidated abnormal network configuration.

As an initial observation, broad alignment in spatiotemporal organization was observed between healthy controls and clinical groups. While various affective and psychotic disorders exhibit diagnosis-specific alterations within localized regions of cortical gradients, the overall gradient architecture shows overarching stability, with each dimension preserving its characteristic functional hierarchy and organizational pattern [28, 30, 32]. Analogously, this conservative gradient reconfiguration profile is observed across both clinical groups and further substantiated by their shared engagement of State 1 and State 4 in temporal dynamics. These two complementary patterns of large-scale network coactivation have been widely identified across ELA studies and are considered a paradigm of dynamic brain network organization in healthy individuals [37, 61–63]. Similar bipartite has been consistently observed across other diverse frameworks, including the dual intertwined rings architecture [64], metastate analyses [65] and time-resolved modeling [66]. As further support for the canonical network configuration, we found no significant differences in transition efficiency between these complementary patterns in HC and PTSD, indicating the functional robustness of this network coordination across distinct clinical conditions. Within this model, the elevated efficiency score observed in TEC may be reflective of a compensatory response to the precipitous decline in the frequency of canonical states [67–70], rather than representing any fundamental alterations present in the interplay among constituent networks.

Beyond the shared expression of canonical dynamic patterns, we next considered the common alterations induced by trauma, which manifested across both spatial and temporal dimensions, suggesting that trauma exposure alone may engender a systematic response of large-scale brain networks, independent of diagnostic status. Representative features of this trauma-related response include consistent spatial displacements across both clinical groups: a rightward shift of VIS along Gradient 1, a leftward shift of the DMN along Gradient 2, and a leftward shift of the FPN along Gradient 3. In the temporal domain, both PTSD and TEC groups exhibited a pair of anomalous, complementary coactivation states –State 2 and State 3 – characterized by atypical engagement of these displaced networks. Similar network-level changes have been widely observed across both PTSD and trauma-exposed populations, supporting the notion of shared alterations following trauma [19, 20, 71, 72]. In our analysis of gradient space properties, we did not observe substantial changes in within-network spatial features such as gradient range or dispersion. At both the parcel and network levels, the observed alterations in clinical groups were primarily driven by global shifts of functional networks along gradient axes, rather than by extreme deviations in individual regions. Interestingly, from the perspective of the energy landscape model, States 1 & 2 and States 3 & 4 can be conceptualized as adjacent variants of a core brain state, which, by definition, indicates that they most frequently transition into one another. Moreover, the basins corresponding to both State 1 and State 4 in HC, and to States 1 & 2 and 3 & 4 in TEC and PTSD jointly cover roughly half of the energy landscape. Together, these regularities suggest that States 2 and 3 constitute deteriorated variants of the canonical State 1 and State 4 configurations.

Given that the hierarchical architecture of functional networks may underlie their temporal organization patterns [66, 73–76], the observed alignment between the spatial displacements of VIS, DMN, and FPN along gradient axes and the corresponding changes exhibited in their dynamic expression invites further speculation. One parsimonious interpretation is that trauma exposure perturbs large-scale coordination between perceptual, control, and internally oriented systems. This in turn disrupts threat detection, contextual appraisal, and regulation processes that are central to fear-related circuitry [77–79]. Consistent with this interpretation, the jointly displaced VIS, DMN, and FPN may reflect altered large-scale salience and executive-control dynamics. Similar cross-network alterations have been reported in trauma-exposed cohorts, including military veterans and trauma-exposed civilians [80–82]. Specifically, we hypothesize that the spatial distance between networks may directly influence their temporal interactions: when a network deviates from its canonical gradient-defined position, it may partially overlap with the gradient-defined functional territory of another network, thereby assuming some of the interactions typically mediated by the displaced system. This phenomenon could be conceptualized as a spatial substitution hypothesis, warranting further investigation in future studies.

Following the above discussion on shared spatiotemporal features, we next examined distinct characteristics that differentiate PTSD from trauma exposure alone. Widespread dissociations between TEC and PTSD emerged along Gradient 3, with most functional networks showing divergent directional shifts. Gradient 3 captures the dissociation between introspective/sensory systems and control-related networks. This axis corresponds to what is often described as the “task–rest” or “internal–external” cognitive dimension, reflecting core processes such as attention reorientation, control signaling, and executive functioning. Therefore, dysfunctions within these higher-order cognitive systems may underlie the genesis of PTSD symptoms, marking them as a principal differentiator from TEC [69, 70, 78, 83]. Another distinguishing feature lies in the altered frequency and efficiency score of the abnormal model (State 2 and State 3) seen within the clinical groups. Mirroring the earlier compensation hypothesis for the canonical model, the increased occurrence of the abnormal states in TEC may reflect a mixed antagonistic state, potentially balancing adaptive and maladaptive processes. By contrast, although the abnormal states were less frequent in PTSD than in TEC, the abnormal model showed a higher efficiency score in PTSD, indicating that its transitions were more directly and efficiently organized. In this sense, TEC may reflect a more frequent but less consolidated abnormal model, whereas PTSD may reflect a less frequent but more stabilized abnormal dynamical configuration. This perspective may also help explain why certain group-level differences, such as those in state frequency, network displacement, and between-network distance, fail to predict PTSD symptom severity. We interpret this pattern as a process of pathological consolidation, whereby PTSD is characterized not by a simple increase in abnormal-state prevalence, but by a qualitative stabilization of abnormal dynamics [84–86].

Taken together, the emergence and stabilization of this anomalous model may reflect an imbalance between top-down control and bottom-up responsivity, consistent with prevailing neurobiological accounts of PTSD. Reduced engagement of control-related systems may indicate weakened regulatory signaling, whereas exaggerated bottom-up coupling may reflect heightened sensory and salience-driven processing streams [18, 87]. This bidirectional abnormality may be interpreted as a vicious circle, consistent with the proposed sensory pathology framework of PTSD [88, 89].

Several limitations of the present study warrant consideration. First, we adopted a network-level perspective to characterize trauma-induced functional alterations. This choice inevitably comes at the cost of overlooking fine-grained regional deviations that may also carry clinical significance. Second, our framework did not incorporate subcortical regions that play central roles in PTSD-related circuitry, including the amygdala and hippocampus together with medial prefrontal and cingulate regulatory pathways [79, 90–92]. Methodologically, recent work has demonstrated that functional gradients can be derived from subcortical–cortical connectivity and that gradient correspondence can be examined across isocortex and subcortical structures [56, 93, 94]. However, integrating cortical and subcortical gradients into a unified interpretive framework remains an active area of research with non-trivial methodological and interpretational challenges [27, 95]. Accordingly, we prioritized a cortex-centered pipeline to maintain a clear and interpretable framework for cortical gradients and state dynamics. This choice limits our ability to characterize subcortical–cortical involvement within the present framework. Third, the modest sample size and relative homogeneity of the cohort may limit the generalizability of our findings. Variations in PTSD subtypes [20], demographic characteristics such as sex [96], and individual differences in cumulative neurobiological burden may all contribute to heterogeneity in functional alterations. This cumulative burden may be indexed by trajectory-related markers such as MRI-derived brain-age measures [97–99]. Fourth, gradient estimation and energy-landscape state metrics can be sensitive to preprocessing choices, including physiological noise handling, global signal regression, motion control, and temporal filtering. We followed preprocessing choices consistent with prior applications of energy landscape analysis and resting-state dynamics, but alternative pipelines may yield quantitative differences in gradients and state statistics. Future studies should therefore assess robustness across preprocessing variants and explicitly evaluate the impact of physiological confounds and denoising strategies on both gradient and ELA estimates. Finally, our distribution-based network metrics, especially EMD, depend on how networks are represented in parcellated space. Atlas choice and parcellation granularity can influence the stability and interpretability of network-level distributions and their distances in gradient space. Future work should evaluate robustness across parcellations and consider higher-resolution atlases or vertex-wise implementations where feasible.

Building on these limitations, future work should prioritize both generalizability and mechanistic interpretability of the proposed spatiotemporal framework. Replication in larger, multi-site and more diverse cohorts, leveraging large-scale collaborative efforts such as ENIGMA [100, 101], will be essential for establishing reproducibility and resolving clinical heterogeneity. In addition, extending the present cortex-centered pipeline to incorporate subcortical–cortical organization, including amygdala–hippocampal interactions, may provide a more complete whole-brain account of PTSD-related network architecture. Finally, convergent network-level evidence from structural modalities, including T1-derived morphological similarity networks [102, 103] and diffusion MRI structural connectivity [104, 105], suggests that PTSD-related alterations extend beyond functional organization, thereby motivating a unified multimodal framework. Given this need for integration, mechanistic modeling frameworks such as network control theory [106] may provide a principled bridge between structural topology and functional state dynamics, thereby clarifying the structural basis that constrains or shapes the functional abnormalities observed in PTSD.

Author contributions

J.W. and Y.Z. conceived the study. J.W., Z.C., and W.L. developed the methodology and implemented the analysis pipelines. F.C., Y.Z. and G.L. coordinated participant recruitment and MRI acquisition. J.W. conducted formal analyses, produced visualization, and drafted the manuscript. L.J.H reviewed and redrafted sections of the manuscript. All authors revised and approved the final version. G.L. and Y.Z. supervised the project and secured funding.

Funding

This research was funded by the National Natural Science Foundation of China (No. 32571269), National Key R&D Program of China (Grant No. 2023YFC3341300), Natural Science Foundation of Jiangsu Province (No. BK20241881), and “333 Talent Project” of Jiangsu Province.

Data availability

The derived data that support the main findings of this study, including the functional connectivity (FC) matrices, are available from the corresponding author upon reasonable request, subject to applicable ethical and data-governance requirements.

Code availability

The analysis code used in this study, including preprocessing, statistical analysis, and figure-generation scripts, is available at GitHub: https://github.com/Parasol0609/PTSD_Spatiotemporal_Features.

Competing interests

The authors declare no competing interest.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04216-x.